From 80311efcf92cfcacccce16a39c529851bd2811d0 Mon Sep 17 00:00:00 2001 From: OnlyJundong Date: Mon, 6 Apr 2026 01:02:50 -0400 Subject: [PATCH 1/3] =?UTF-8?q?Add=20non-record=20submission:=20Extended?= =?UTF-8?q?=20Compute=20Scaling=20Analysis=20(20K=20steps,=20~6h,=204?= =?UTF-8?q?=C3=97A100=20MIG)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../README.md | 163 ++ .../run_scaling_20k.sh | 36 + .../submission.json | 20 + .../train_gpt.py | 1938 ++++++++++++++ .../train_step20k_seed1337.txt | 2241 ++++++++++++++++ .../train_step20k_seed2024.txt | 2242 +++++++++++++++++ .../train_step20k_seed42.txt | 2241 ++++++++++++++++ 7 files changed, 8881 insertions(+) create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/README.md create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/run_scaling_20k.sh create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/submission.json create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_gpt.py create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed1337.txt create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed2024.txt create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed42.txt diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/README.md b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/README.md new file mode 100644 index 0000000000..5ea583d0bd --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/README.md @@ -0,0 +1,163 @@ +# Extended Compute Scaling Analysis (20K Steps, ~6 Hours) + +**val_bpb: 1.0960** (20K steps, 3-seed mean, std 0.0003) | **~15.05 MB** | 4×A100 MIG (Unlimited Compute Track) + +## Summary + +This submission is a **non-record submission**. It studies how ([PR #549](https://github.com/openai/parameter-golf/pull/549) by @abaybektursun) scales under extended compute, removing the 10-minute wall-clock constraint. The same architecture and code are trained for 20K steps (~6 hours) on 4×A100 MIG instances (approximately 10× slower per step than 8×H100 SXM). + +Key findings: +- **20K steps achieves 1.0960 BPB post-TTT** (3-seed mean) +- **Artifact size balloons mid-training** (peaking at ~17.2MB around step 10K–15K) but **recovers to ~15.05MB** after warmdown completes — warmdown smooths weight entropy and restores compressibility +- **TTT gains scale with base model quality**: TTT provides -0.006 BPB on the 20K model + +## Results + +### 20K steps, ~6 hours (4×A100 MIG, 3-seed comparison) + +| Seed | step_avg | steps | Pre-TTT bpb | **Post-TTT bpb** | TTT gain | Artifact | +|------|----------|-------|-------------|-----------------|----------|----------| +| 1337 | 828.7ms | 20,000 | 1.1018 | **1.0957** | -0.0061 | 15,077,933 | +| 42 | 828.8ms | 20,000 | 1.1020 | **1.0962** | -0.0058 | 15,137,145 | +| 2024 | 839.8ms | 20,000 | 1.1017 | **1.0962** | -0.0055 | 14,942,394 | +| **Mean** | **832.4ms** | **20,000** | **1.1018** | **1.0960 (std 0.0003)** | **-0.0058** | **15,052,491** | + +## Scaling Analysis: BPB & Artifact Size vs Training Steps + +Artifact size is computed as int6+LZMA compressed model + code bytes. Every 2,000 steps. + +### Seed 2024 (2,000-step intervals) + +| Steps | Pre-TTT val_bpb | artifact_bytes | Under 16MB? | +|------:|--------:|---------------:|:-----------:| +| 0 | 4.1037 | 4,577,366 | Yes | +| 2,000 | 1.2651 | 13,800,942 | Yes | +| 4,000 | 1.2286 | 16,959,534 | **No** | +| 6,000 | 1.2122 | 17,243,366 | **No** | +| 8,000 | 1.2046 | 17,248,774 | **No** | +| 10,000 | 1.2007 | 17,246,738 | **No** | +| 12,000 | 1.1994 | 17,231,058 | **No** | +| 14,000 | 1.1835 | 16,929,622 | **No** | +| 16,000 | 1.1672 | 16,321,958 | **No** | +| 18,000 | 1.1429 | 15,534,274 | Yes | +| 20,000 | 1.1110* | 14,942,394 | Yes | + +*Step 20K artifact reflects the final model with full warmdown applied. Intermediate checkpoints without warmdown exceed 16MB. + +**Note:** Intermediate checkpoints do not benefit from warmdown. The final model has full warmdown applied, resulting in ~15.05MB — well under the 16MB limit. The artifact size peaks mid-training when weights are high-entropy, then drops as warmdown smooths them. + +### BPB vs Steps (ASCII plot) + +Power-law decay with two distinct phases: rapid early learning, then warmdown-driven final drop. + +``` +BPB +4.10 |* + | + | +2.50 | + | +1.26 | * +1.23 | * +1.22 | * +1.20 | * +1.19 | * * +1.18 | * +1.10 | * + +---------+--------+-> steps (K) + 0 10 20 + + ||| + (rapid) (sharp drop) +``` + +### Artifact Size vs Steps (ASCII plot) + +Non-monotonic: grows rapidly to a peak at ~15K steps, then shrinks back below budget during warmdown. + +``` +MB +17.2 | * * * * * * +16.0 |-------------------- 16MB limit +15.1 | * +14.1 | +13.1 | * + 4.6 |* + +---------+--------+-> steps (K) + 0 10 20 + + |<-fits->||| +``` + +Intermediate checkpoints between steps ~10K–17.5K exceed the 16MB budget and cannot be submitted as-is. Only the final model (with warmdown complete) fits. This means **early stopping is not viable** for this architecture without a separate warmdown phase. + +### Observations + +1. **BPB vs steps follows a power-law decay** with diminishing returns. The biggest gains are in the first 7,500 steps, with warmdown driving the final sharp drop. +2. **Artifact size is non-monotonic**: it grows rapidly from 4.6MB (init) to ~17.2MB (step 10K–15K peak), then shrinks back to ~15.1MB during warmdown. Intermediate checkpoints without warmdown exceed 16MB. +3. **TTT gain scales with compute**: At 20K steps, -0.006 BPB. The base model benefits from test-time adaptation. + +## Architecture + +Identical to [PR #549](https://github.com/openai/parameter-golf/pull/549) (LeakyReLU² + Legal TTT + Parallel Muon): + +| Component | Setting | +|-----------|---------| +| Layers | 11 (512d, 8H, 4KV) | +| MLP | 3× with LeakyReLU(0.5)² | +| BigramHash | 1536 | +| XSA | Last 4 layers | +| RoPE | Partial (16/64 dims) | +| LN Scale | 1/√(layer+1) | +| VE128 | Layers 9-10 | +| Weight avg | EMA(0.997) + Tight SWA(every 50) | +| Quantization | GPTQ-lite int6 + lzma | +| Optimizer | Parameter Banking + Parallel Muon | + +### Hyperparameter scaling for extended training + +LR schedule parameters scaled proportionally to maintain the same warmup/warmdown ratios: + +| Parameter | This work (20K steps) | +|-----------|-----------| +| ITERATIONS | 20,000 | +| WARMDOWN_ITERS | 7,800 (39.0%) | +| MUON_MOMENTUM_WARMUP_STEPS | 3,340 (16.7%) | +| MAX_WALLCLOCK_SECONDS | 0 (unlimited) | + +## Run Command + +```bash +RUN_ID=run_recordsota_nonrecord_seed1337 \ +NUM_LAYERS=11 BIGRAM_VOCAB_SIZE=1536 XSA_LAST_N=4 \ +EMA_ENABLED=1 EMA_DECAY=0.997 SWA_ENABLED=1 SWA_EVERY=50 \ +ROPE_DIMS=16 LN_SCALE=1 LATE_QAT=1 LATE_QAT_THRESHOLD=0.15 \ +VE_ENABLED=1 VE_DIM=128 VE_LAYERS=9,10 \ +TTT_ENABLED=1 TTT_LR=0.002 TTT_EPOCHS=3 TTT_CHUNK_TOKENS=32768 \ +TTT_FREEZE_BLOCKS=0 TTT_MOMENTUM=0.9 TTT_BATCH_SEQS=32 TTT_GRAD_CLIP=1.0 \ +MUON_WD=0.04 ADAM_WD=0.04 \ +MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 \ +MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 \ +MUON_MOMENTUM_WARMUP_STEPS=3340 WARMDOWN_ITERS=7800 \ +ITERATIONS=20000 MAX_WALLCLOCK_SECONDS=0 EVAL_STRIDE=64 \ +VAL_LOSS_EVERY=2000 \ +SEED=1337 \ +CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --standalone --nproc_per_node=4 train_gpt.py +``` + +## Hardware + +- 4×NVIDIA A100 MIG instances +- Seeds 1337 & 42: ~829ms/step | Seed 2024: ~840ms/step (vs 83ms/step on 8×H100 SXM — approximately 10× slower) +- grad_accum_steps=2 (to match 8-GPU effective batch size of 786,432 tokens) +- ~4.6–4.7h training + ~34min TTT eval = ~6 hours total per seed + +## Credits + +This submission uses the full architecture and code from the record-track PR #549 with no ML changes — only extended compute and proportionally scaled LR schedules. + +- **Base submission [PR #549](https://github.com/openai/parameter-golf/pull/549)**: @abaybektursun — LeakyReLU² + Legal TTT + Parallel Muon +- **LeakyReLU² activation**: [PR #493](https://github.com/openai/parameter-golf/pull/493) by @parinzee, [PR #518](https://github.com/openai/parameter-golf/pull/518) by @sofiabod +- **Optimizer (Parameter Banking + Parallel Muon)**: [PR #399](https://github.com/openai/parameter-golf/pull/399) by @abaybektursun +- **TTT recipe**: [PR #461](https://github.com/openai/parameter-golf/pull/461) by @Christopher-Lee-McClendon +- **Base model**: [PR #414](https://github.com/openai/parameter-golf/pull/414) by @signalrush diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/run_scaling_20k.sh b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/run_scaling_20k.sh new file mode 100644 index 0000000000..343ff9a6ce --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/run_scaling_20k.sh @@ -0,0 +1,36 @@ +#!/bin/bash +set -e + +cd "$(dirname "$0")" + +# 20k scaling run — intermediate val_bpb logged every 2K steps for scaling curve +# Warmdown and momentum warmup scaled proportionally (~39% and ~16.7%) +# 20k * 0.39 = 7800, 20k * 0.167 = 3340 + +echo "==========================================" +echo "Scaling run: 20k steps, seed=2024" +echo "Started at: $(date)" +echo "==========================================" + +env \ + RUN_ID=train_step20k_seed2024 \ + NUM_LAYERS=11 BIGRAM_VOCAB_SIZE=1536 XSA_LAST_N=4 \ + EMA_ENABLED=1 EMA_DECAY=0.997 SWA_ENABLED=1 SWA_EVERY=50 \ + ROPE_DIMS=16 LN_SCALE=1 LATE_QAT=1 LATE_QAT_THRESHOLD=0.15 \ + VE_ENABLED=1 VE_DIM=128 VE_LAYERS=9,10 \ + TTT_ENABLED=1 TTT_LR=0.002 TTT_EPOCHS=3 TTT_CHUNK_TOKENS=32768 \ + TTT_FREEZE_BLOCKS=0 TTT_MOMENTUM=0.9 TTT_BATCH_SEQS=32 TTT_GRAD_CLIP=1.0 \ + MUON_WD=0.04 ADAM_WD=0.04 \ + MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 \ + MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 \ + MUON_MOMENTUM_WARMUP_STEPS=3340 WARMDOWN_ITERS=7800 \ + ITERATIONS=20000 MAX_WALLCLOCK_SECONDS=0 EVAL_STRIDE=64 \ + VAL_LOSS_EVERY=2000 \ + SEED=2024 \ + CUDA_VISIBLE_DEVICES=0,1,2,3 \ + torchrun --standalone --nproc_per_node=4 train_gpt.py \ + 2>&1 | tee logs/train_step20k_seed2024.txt + +echo "==========================================" +echo "20k run finished at: $(date)" +echo "==========================================" diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/submission.json b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/submission.json new file mode 100644 index 0000000000..80070e8831 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/submission.json @@ -0,0 +1,20 @@ +{ + "name": "Extended Compute Scaling Analysis (20K Steps, ~6 Hours)", + "val_bpb": 1.0960, + "val_bpb_std": 0.0003, + "bytes_total": 15052491, + "seeds": [1337, 42, 2024], + "seed_results": { + "1337": { "val_bpb": 1.0957, "pre_ttt_bpb": 1.1018, "ttt_gain": -0.0061, "artifact_bytes": 15077933, "step_avg_ms": 828.7 }, + "42": { "val_bpb": 1.0962, "pre_ttt_bpb": 1.1020, "ttt_gain": -0.0058, "artifact_bytes": 15137145, "step_avg_ms": 828.8 }, + "2024": { "val_bpb": 1.0962, "pre_ttt_bpb": 1.1017, "ttt_gain": -0.0055, "artifact_bytes": 14942394, "step_avg_ms": 839.8 } + }, + "blurb": "Scaling analysis of the PR #549 SOTA stack (LeakyReLU² + TTT + Parallel Muon) under extended compute on 4×A100 MIG. 20K steps (~6 hours training) achieves 1.0960 BPB post-TTT (3-seed mean, std 0.0003) with ~15.05MB artifact. Includes compute scaling curve showing warmdown-driven compression recovery and artifact size non-monotonicity.", + "author": "Jundong Hu", + "github_id": "OnlyJundong", + "date": "2026-04-06", + "track": "non_record_16mb", + "hardware": "4×A100 MIG", + "steps": 20000, + "step_avg_ms_mean": 832.4 +} diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_gpt.py b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_gpt.py new file mode 100644 index 0000000000..248e83626b --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_gpt.py @@ -0,0 +1,1938 @@ +from __future__ import annotations +import copy +import glob +import io +import lzma +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +# Fix for MIG/multi-GPU: ensure each process only sees its assigned GPU +# Must be set before importing torch +if "LOCAL_RANK" in os.environ and "CUDA_VISIBLE_DEVICES" in os.environ: + visible_devices = os.environ["CUDA_VISIBLE_DEVICES"].split(",") + local_rank = int(os.environ["LOCAL_RANK"]) + if local_rank < len(visible_devices): + os.environ["CUDA_VISIBLE_DEVICES"] = visible_devices[local_rank] + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True +except ImportError: + _HAS_FA3 = False +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 2)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +# --- Data loading --- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + # Estimate artifact size (int6+lzma) at each validation step + try: + sd_snap = {k: v.detach().cpu() for k, v in base_model.state_dict().items() + if "mtp_heads" not in k} + unbanked_snap = _unbank_state_dict(sd_snap, args.num_layers) + q_snap, m_snap = mixed_quantize_int6(unbanked_snap, {"mlp", "attn"}) + snap_buf = io.BytesIO() + torch.save({"w": q_snap, "m": m_snap}, snap_buf) + snap_compressed = len(lzma.compress(snap_buf.getvalue(), preset=6)) + snap_code_bytes = len(code.encode("utf-8")) + snap_total = snap_compressed + snap_code_bytes + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_bytes:{snap_total} model_int6_lzma:{snap_compressed} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + except Exception as e: + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_size:error({e}) " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed1337.txt b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed1337.txt new file mode 100644 index 0000000000..6fa39e2dc6 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed1337.txt @@ -0,0 +1,2241 @@ +logs/run_recordsota_nonrecord_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9304 val_bpb:4.1046 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9317 train_time:850ms step_avg:849.64ms +step:2/20000 train_loss:8.6832 train_time:1394ms step_avg:696.84ms +step:3/20000 train_loss:7.6849 train_time:2245ms step_avg:748.41ms +step:4/20000 train_loss:7.1825 train_time:3097ms step_avg:774.24ms +step:5/20000 train_loss:7.1342 train_time:3949ms step_avg:789.77ms +step:6/20000 train_loss:7.1129 train_time:4801ms step_avg:800.14ms +step:7/20000 train_loss:7.1092 train_time:5653ms step_avg:807.50ms +step:8/20000 train_loss:6.9540 train_time:6505ms step_avg:813.11ms +step:9/20000 train_loss:6.6136 train_time:7357ms step_avg:817.44ms +step:10/20000 train_loss:6.2387 train_time:8218ms step_avg:821.76ms +step:500/20000 train_loss:2.3635 train_time:414424ms step_avg:828.85ms +step:1000/20000 train_loss:2.2672 train_time:828685ms step_avg:828.68ms +step:1500/20000 train_loss:2.1838 train_time:1243209ms step_avg:828.81ms +step:2000/20000 train_loss:2.0379 train_time:1657693ms step_avg:828.85ms +step:2500/20000 train_loss:2.1580 train_time:2071960ms step_avg:828.78ms +step:3000/20000 train_loss:2.1211 train_time:2486377ms step_avg:828.79ms +step:3500/20000 train_loss:2.1881 train_time:2900521ms step_avg:828.72ms +step:4000/20000 train_loss:2.1446 train_time:3314659ms step_avg:828.66ms +step:4000/20000 val_loss:2.0755 val_bpb:1.2292 train_time:3315001ms step_avg:828.75ms +step:4500/20000 train_loss:2.1098 train_time:3728932ms step_avg:828.65ms +step:5000/20000 train_loss:2.0727 train_time:4143237ms step_avg:828.65ms +step:5500/20000 train_loss:2.1133 train_time:4557500ms step_avg:828.64ms +step:6000/20000 train_loss:2.0121 train_time:4971750ms step_avg:828.63ms +step:6500/20000 train_loss:2.2772 train_time:5385925ms step_avg:828.60ms +step:7000/20000 train_loss:1.9077 train_time:5800366ms step_avg:828.62ms +step:7500/20000 train_loss:2.0332 train_time:6214436ms step_avg:828.59ms +step:8000/20000 train_loss:1.9856 train_time:6628761ms step_avg:828.60ms +step:8000/20000 val_loss:2.0333 val_bpb:1.2043 train_time:6629111ms step_avg:828.64ms +step:8500/20000 train_loss:1.9662 train_time:7043229ms step_avg:828.62ms +step:9000/20000 train_loss:2.0891 train_time:7457441ms step_avg:828.60ms +step:9500/20000 train_loss:2.2474 train_time:7872123ms step_avg:828.64ms +step:10000/20000 train_loss:2.0618 train_time:8286352ms step_avg:828.64ms +step:10500/20000 train_loss:2.0948 train_time:8700592ms step_avg:828.63ms +step:11000/20000 train_loss:1.9865 train_time:9115076ms step_avg:828.64ms +step:11500/20000 train_loss:1.9466 train_time:9529240ms step_avg:828.63ms +step:12000/20000 train_loss:1.9609 train_time:9943485ms step_avg:828.62ms +step:12000/20000 val_loss:2.0241 val_bpb:1.1988 train_time:9943828ms step_avg:828.65ms +step:12500/20000 train_loss:1.9691 train_time:10357780ms step_avg:828.62ms +step:13000/20000 train_loss:1.9781 train_time:10772024ms step_avg:828.62ms +step:13500/20000 train_loss:2.0928 train_time:11186188ms step_avg:828.61ms +step:14000/20000 train_loss:1.8907 train_time:11600289ms step_avg:828.59ms +step:14500/20000 train_loss:2.0651 train_time:12014544ms step_avg:828.59ms +step:15000/20000 train_loss:2.0078 train_time:12428690ms step_avg:828.58ms +step:15500/20000 train_loss:1.9744 train_time:12842898ms step_avg:828.57ms +step:16000/20000 train_loss:2.0443 train_time:13257276ms step_avg:828.58ms +step:16000/20000 val_loss:1.9697 val_bpb:1.1666 train_time:13257627ms step_avg:828.60ms +step:16500/20000 train_loss:2.0235 train_time:13671525ms step_avg:828.58ms +step:17000/20000 train_loss:2.0201 train_time:14085842ms step_avg:828.58ms +step:17500/20000 train_loss:1.9946 train_time:14500144ms step_avg:828.58ms +step:18000/20000 train_loss:1.8680 train_time:14914297ms step_avg:828.57ms +swa:start step:18450 +step:18500/20000 train_loss:1.8339 train_time:15328815ms step_avg:828.58ms +late_qat:enabled step:18831 scale:0.1499 +step:19000/20000 train_loss:1.8034 train_time:15744113ms step_avg:828.64ms +step:19500/20000 train_loss:1.9822 train_time:16159354ms step_avg:828.68ms +step:20000/20000 train_loss:1.8749 train_time:16574509ms step_avg:828.73ms +step:20000/20000 val_loss:1.8748 val_bpb:1.1104 train_time:16574910ms step_avg:828.75ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8738 val_bpb:1.1098 eval_time:19704ms +Serialized model: 106027446 bytes +Code size: 90361 bytes +Serialized model int6+lzma: 14987572 bytes +Total submission size int6+lzma: 15077933 bytes +final_int6_roundtrip val_loss:1.8998 val_bpb:1.1252 eval_time:24388ms +final_int6_roundtrip_exact val_loss:1.89979453 val_bpb:1.12516533 +final_int6_sliding_window val_loss:1.8604 val_bpb:1.1018 stride:64 eval_time:664624ms +final_int6_sliding_window_exact val_loss:1.86041005 val_bpb:1.10184254 +final_int8_zlib_roundtrip_exact val_loss:1.86041005 val_bpb:1.10184254 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.148962 time=1.1s + ttt_chunk [11/1893] bpb=1.106649 time=11.9s + ttt_chunk [21/1893] bpb=1.109183 time=22.7s + ttt_chunk [31/1893] bpb=1.112465 time=33.4s + ttt_chunk [41/1893] bpb=1.101158 time=44.2s + ttt_chunk [51/1893] bpb=1.097352 time=55.0s + ttt_chunk [61/1893] bpb=1.102163 time=65.7s + ttt_chunk [71/1893] bpb=1.099272 time=76.5s + ttt_chunk [81/1893] bpb=1.099287 time=87.3s + ttt_chunk [91/1893] bpb=1.097889 time=98.1s + ttt_chunk [101/1893] bpb=1.100512 time=108.9s + ttt_chunk [111/1893] bpb=1.101752 time=119.7s + ttt_chunk [121/1893] bpb=1.098599 time=130.4s + ttt_chunk [131/1893] bpb=1.098481 time=141.2s + ttt_chunk [141/1893] bpb=1.098188 time=152.0s + ttt_chunk [151/1893] bpb=1.101303 time=162.8s + ttt_chunk [161/1893] bpb=1.102934 time=173.6s + ttt_chunk [171/1893] bpb=1.103556 time=184.3s + ttt_chunk [181/1893] bpb=1.103556 time=195.1s + ttt_chunk [191/1893] bpb=1.106780 time=205.9s + ttt_chunk [201/1893] bpb=1.107089 time=216.7s + ttt_chunk [211/1893] bpb=1.104621 time=227.4s + ttt_chunk [221/1893] bpb=1.106682 time=238.2s + ttt_chunk [231/1893] bpb=1.106153 time=249.0s + ttt_chunk [241/1893] bpb=1.105951 time=259.8s + ttt_chunk [251/1893] bpb=1.104500 time=270.5s + ttt_chunk [261/1893] bpb=1.102854 time=281.3s + ttt_chunk [271/1893] bpb=1.101507 time=292.1s + ttt_chunk [281/1893] bpb=1.103788 time=302.9s + ttt_chunk [291/1893] bpb=1.104513 time=313.6s + ttt_chunk [301/1893] bpb=1.105062 time=324.4s + ttt_chunk [311/1893] bpb=1.106585 time=335.2s + ttt_chunk [321/1893] bpb=1.108036 time=346.0s + ttt_chunk [331/1893] bpb=1.107902 time=356.7s + ttt_chunk [341/1893] bpb=1.108209 time=367.5s + ttt_chunk [351/1893] bpb=1.109486 time=378.3s + ttt_chunk [361/1893] bpb=1.110674 time=389.1s + ttt_chunk [371/1893] bpb=1.110192 time=399.9s + ttt_chunk [381/1893] bpb=1.110144 time=410.6s + ttt_chunk [391/1893] bpb=1.109733 time=421.4s + ttt_chunk [401/1893] bpb=1.108356 time=432.2s + ttt_chunk [411/1893] bpb=1.107227 time=442.9s + ttt_chunk [421/1893] bpb=1.106719 time=453.7s + ttt_chunk [431/1893] bpb=1.107374 time=464.5s + ttt_chunk [441/1893] bpb=1.107195 time=475.3s + ttt_chunk [451/1893] bpb=1.106903 time=486.1s + ttt_chunk [461/1893] bpb=1.106178 time=496.8s + ttt_chunk [471/1893] bpb=1.105772 time=507.6s + ttt_chunk [481/1893] bpb=1.105518 time=518.4s + ttt_chunk [491/1893] bpb=1.105211 time=529.2s + ttt_chunk [501/1893] bpb=1.104631 time=539.9s + ttt_chunk [511/1893] bpb=1.104022 time=550.7s + ttt_chunk [521/1893] bpb=1.102977 time=561.5s + ttt_chunk [531/1893] bpb=1.103083 time=572.2s + ttt_chunk [541/1893] bpb=1.103047 time=583.0s + ttt_chunk [551/1893] bpb=1.101880 time=593.8s + ttt_chunk [561/1893] bpb=1.102420 time=604.6s + ttt_chunk [571/1893] bpb=1.101740 time=615.3s + ttt_chunk [581/1893] bpb=1.101206 time=626.1s + ttt_chunk [591/1893] bpb=1.100569 time=636.9s + ttt_chunk [601/1893] bpb=1.101233 time=647.7s + ttt_chunk [611/1893] bpb=1.100826 time=658.5s + ttt_chunk [621/1893] bpb=1.100733 time=669.2s + ttt_chunk [631/1893] bpb=1.101122 time=680.0s + ttt_chunk [641/1893] bpb=1.100907 time=690.8s + ttt_chunk [651/1893] bpb=1.100916 time=701.6s + ttt_chunk [661/1893] bpb=1.100822 time=712.3s + ttt_chunk [671/1893] bpb=1.100437 time=723.1s + ttt_chunk [681/1893] bpb=1.100704 time=733.9s + ttt_chunk [691/1893] bpb=1.101402 time=744.7s + ttt_chunk [701/1893] bpb=1.100649 time=755.4s + ttt_chunk [711/1893] bpb=1.101214 time=766.2s + ttt_chunk [721/1893] bpb=1.100855 time=777.0s + ttt_chunk [731/1893] bpb=1.101286 time=787.8s + ttt_chunk [741/1893] bpb=1.101246 time=798.6s + ttt_chunk [751/1893] bpb=1.100906 time=809.3s + ttt_chunk [761/1893] bpb=1.100754 time=820.1s + ttt_chunk [771/1893] bpb=1.100535 time=830.9s + ttt_chunk [781/1893] bpb=1.101052 time=841.6s + ttt_chunk [791/1893] bpb=1.100729 time=852.4s + ttt_chunk [801/1893] bpb=1.100817 time=863.2s + ttt_chunk [811/1893] bpb=1.100349 time=874.0s + ttt_chunk [821/1893] bpb=1.100161 time=884.7s + ttt_chunk [831/1893] bpb=1.099731 time=895.5s + ttt_chunk [841/1893] bpb=1.099163 time=906.3s + ttt_chunk [851/1893] bpb=1.099062 time=917.1s + ttt_chunk [861/1893] bpb=1.099192 time=927.9s + ttt_chunk [871/1893] bpb=1.099234 time=938.6s + ttt_chunk [881/1893] bpb=1.099269 time=949.4s + ttt_chunk [891/1893] bpb=1.099098 time=960.2s + ttt_chunk [901/1893] bpb=1.099091 time=971.0s + ttt_chunk [911/1893] bpb=1.099081 time=981.7s + ttt_chunk [921/1893] bpb=1.099457 time=992.5s + ttt_chunk [931/1893] bpb=1.099302 time=1003.3s + ttt_chunk [941/1893] bpb=1.099197 time=1014.1s + ttt_chunk [951/1893] bpb=1.099211 time=1024.8s + ttt_chunk [961/1893] bpb=1.098954 time=1035.6s + ttt_chunk [971/1893] bpb=1.099679 time=1046.4s + ttt_chunk [981/1893] bpb=1.099805 time=1057.1s + ttt_chunk [991/1893] bpb=1.099714 time=1067.9s + ttt_chunk [1001/1893] bpb=1.099874 time=1078.7s + ttt_chunk [1011/1893] bpb=1.100135 time=1089.5s + ttt_chunk [1021/1893] bpb=1.100290 time=1100.2s + ttt_chunk [1031/1893] bpb=1.100826 time=1111.0s + ttt_chunk [1041/1893] bpb=1.100503 time=1121.8s + ttt_chunk [1051/1893] bpb=1.100230 time=1132.6s + ttt_chunk [1061/1893] bpb=1.100500 time=1143.3s + ttt_chunk [1071/1893] bpb=1.100973 time=1154.1s + ttt_chunk [1081/1893] bpb=1.100979 time=1164.9s + ttt_chunk [1091/1893] bpb=1.101374 time=1175.7s + ttt_chunk [1101/1893] bpb=1.101517 time=1186.4s + ttt_chunk [1111/1893] bpb=1.101291 time=1197.2s + ttt_chunk [1121/1893] bpb=1.101240 time=1208.0s + ttt_chunk [1131/1893] bpb=1.101147 time=1218.8s + ttt_chunk [1141/1893] bpb=1.101001 time=1229.5s + ttt_chunk [1151/1893] bpb=1.101048 time=1240.3s + ttt_chunk [1161/1893] bpb=1.100357 time=1251.1s + ttt_chunk [1171/1893] bpb=1.100923 time=1261.9s + ttt_chunk [1181/1893] bpb=1.100425 time=1272.6s + ttt_chunk [1191/1893] bpb=1.100132 time=1283.4s + ttt_chunk [1201/1893] bpb=1.100718 time=1294.2s + ttt_chunk [1211/1893] bpb=1.100113 time=1304.9s + ttt_chunk [1221/1893] bpb=1.099789 time=1315.7s + ttt_chunk [1231/1893] bpb=1.099716 time=1326.5s + ttt_chunk [1241/1893] bpb=1.099538 time=1337.3s + ttt_chunk [1251/1893] bpb=1.099298 time=1348.0s + ttt_chunk [1261/1893] bpb=1.099230 time=1358.8s + ttt_chunk [1271/1893] bpb=1.099030 time=1369.6s + ttt_chunk [1281/1893] bpb=1.098825 time=1380.4s + ttt_chunk [1291/1893] bpb=1.098638 time=1391.1s + ttt_chunk [1301/1893] bpb=1.098263 time=1401.9s + ttt_chunk [1311/1893] bpb=1.097961 time=1412.7s + ttt_chunk [1321/1893] bpb=1.097795 time=1423.4s + ttt_chunk [1331/1893] bpb=1.097718 time=1434.2s + ttt_chunk [1341/1893] bpb=1.097600 time=1445.0s + ttt_chunk [1351/1893] bpb=1.097546 time=1455.8s + ttt_chunk [1361/1893] bpb=1.097752 time=1466.5s + ttt_chunk [1371/1893] bpb=1.097602 time=1477.3s + ttt_chunk [1381/1893] bpb=1.097542 time=1488.1s + ttt_chunk [1391/1893] bpb=1.097027 time=1498.8s + ttt_chunk [1401/1893] bpb=1.097052 time=1509.6s + ttt_chunk [1411/1893] bpb=1.097076 time=1520.4s + ttt_chunk [1421/1893] bpb=1.097336 time=1531.2s + ttt_chunk [1431/1893] bpb=1.097182 time=1541.9s + ttt_chunk [1441/1893] bpb=1.097860 time=1552.7s + ttt_chunk [1451/1893] bpb=1.097967 time=1563.5s + ttt_chunk [1461/1893] bpb=1.097711 time=1574.2s + ttt_chunk [1471/1893] bpb=1.098624 time=1585.0s + ttt_chunk [1481/1893] bpb=1.098435 time=1595.8s + ttt_chunk [1491/1893] bpb=1.098445 time=1606.5s + ttt_chunk [1501/1893] bpb=1.098608 time=1617.3s + ttt_chunk [1511/1893] bpb=1.098687 time=1628.1s + ttt_chunk [1521/1893] bpb=1.098715 time=1638.9s + ttt_chunk [1531/1893] bpb=1.098533 time=1649.6s + ttt_chunk [1541/1893] bpb=1.098501 time=1660.4s + ttt_chunk [1551/1893] bpb=1.098854 time=1671.2s + ttt_chunk [1561/1893] bpb=1.098993 time=1681.9s + ttt_chunk [1571/1893] bpb=1.099120 time=1692.7s + ttt_chunk [1581/1893] bpb=1.099236 time=1703.5s + ttt_chunk [1591/1893] bpb=1.099189 time=1714.3s + ttt_chunk [1601/1893] bpb=1.099373 time=1725.0s + ttt_chunk [1611/1893] bpb=1.099446 time=1735.8s + ttt_chunk [1621/1893] bpb=1.099315 time=1746.6s + ttt_chunk [1631/1893] bpb=1.099486 time=1757.4s + ttt_chunk [1641/1893] bpb=1.099365 time=1768.1s + ttt_chunk [1651/1893] bpb=1.099269 time=1778.9s + ttt_chunk [1661/1893] bpb=1.099155 time=1789.7s + ttt_chunk [1671/1893] bpb=1.099519 time=1800.4s + ttt_chunk [1681/1893] bpb=1.099767 time=1811.2s + ttt_chunk [1691/1893] bpb=1.099741 time=1822.0s + ttt_chunk [1701/1893] bpb=1.099731 time=1832.8s + ttt_chunk [1711/1893] bpb=1.099572 time=1843.5s + ttt_chunk [1721/1893] bpb=1.099426 time=1854.3s + ttt_chunk [1731/1893] bpb=1.099378 time=1865.1s + ttt_chunk [1741/1893] bpb=1.099166 time=1875.8s + ttt_chunk [1751/1893] bpb=1.099024 time=1886.6s + ttt_chunk [1761/1893] bpb=1.099109 time=1897.4s + ttt_chunk [1771/1893] bpb=1.099028 time=1908.2s + ttt_chunk [1781/1893] bpb=1.098987 time=1918.9s + ttt_chunk [1791/1893] bpb=1.098601 time=1929.7s + ttt_chunk [1801/1893] bpb=1.098603 time=1940.5s + ttt_chunk [1811/1893] bpb=1.098432 time=1951.3s + ttt_chunk [1821/1893] bpb=1.098469 time=1962.0s + ttt_chunk [1831/1893] bpb=1.098074 time=1972.8s + ttt_chunk [1841/1893] bpb=1.098044 time=1983.6s + ttt_chunk [1851/1893] bpb=1.097814 time=1994.3s + ttt_chunk [1861/1893] bpb=1.097345 time=2005.1s + ttt_chunk [1871/1893] bpb=1.097193 time=2015.9s + ttt_chunk [1881/1893] bpb=1.096820 time=2026.7s + ttt_chunk [1891/1893] bpb=1.096650 time=2037.4s + ttt_chunk [1893/1893] bpb=1.096669 time=2039.2s +ttt_sliding:done val_loss=1.850081 val_bpb=1.095725 elapsed=2039.2s +legal_ttt val_loss:1.8501 val_bpb:1.0957 eval_time:2039891ms +legal_ttt_exact val_loss:1.85008136 val_bpb:1.09572529 +tch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +# --- Data loading --- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.10.14 (main, May 6 2024, 19:42:50) [GCC 11.2.0] +Running PyTorch 2.10.0+cu128 +Fri Mar 27 04:07:26 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 550.127.08 Driver Version: 550.127.08 CUDA Version: 12.4 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA A100-SXM4-80GB On | 00000000:27:00.0 Off | On | +| N/A 38C P0 99W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA A100-SXM4-80GB On | 00000000:2A:00.0 Off | On | +| N/A 33C P0 89W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA A100-SXM4-80GB On | 00000000:51:00.0 Off | On | +| N/A 34C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA A100-SXM4-80GB On | 00000000:9E:00.0 Off | On | +| N/A 38C P0 99W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA A100-SXM4-80GB On | 00000000:A4:00.0 Off | On | +| N/A 33C P0 96W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA A100-SXM4-80GB On | 00000000:CA:00.0 Off | On | +| N/A 37C P0 97W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| MIG devices: | ++------------------+----------------------------------+-----------+-----------------------+ +| GPU GI CI MIG | Memory-Usage | Vol| Shared | +| ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG | +| | | ECC| | +|==================+==================================+===========+=======================| +| 0 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 2 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 3 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 2 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 1 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 3537273 C /opt/conda/bin/python3 0MiB | +| 1 N/A N/A 3537274 C /opt/conda/bin/python3 0MiB | +| 2 N/A N/A 3537275 C /opt/conda/bin/python3 0MiB | +| 3 N/A N/A 3537276 C /opt/conda/bin/python3 0MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9304 val_bpb:4.1046 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9317 train_time:850ms step_avg:849.64ms +step:2/20000 train_loss:8.6832 train_time:1394ms step_avg:696.84ms +step:3/20000 train_loss:7.6849 train_time:2245ms step_avg:748.41ms +step:4/20000 train_loss:7.1825 train_time:3097ms step_avg:774.24ms +step:5/20000 train_loss:7.1342 train_time:3949ms step_avg:789.77ms +step:6/20000 train_loss:7.1129 train_time:4801ms step_avg:800.14ms +step:7/20000 train_loss:7.1092 train_time:5653ms step_avg:807.50ms +step:8/20000 train_loss:6.9540 train_time:6505ms step_avg:813.11ms +step:9/20000 train_loss:6.6136 train_time:7357ms step_avg:817.44ms +step:10/20000 train_loss:6.2387 train_time:8218ms step_avg:821.76ms +step:500/20000 train_loss:2.3635 train_time:414424ms step_avg:828.85ms +step:1000/20000 train_loss:2.2672 train_time:828685ms step_avg:828.68ms +step:1500/20000 train_loss:2.1838 train_time:1243209ms step_avg:828.81ms +step:2000/20000 train_loss:2.0379 train_time:1657693ms step_avg:828.85ms +step:2500/20000 train_loss:2.1580 train_time:2071960ms step_avg:828.78ms +step:3000/20000 train_loss:2.1211 train_time:2486377ms step_avg:828.79ms +step:3500/20000 train_loss:2.1881 train_time:2900521ms step_avg:828.72ms +step:4000/20000 train_loss:2.1446 train_time:3314659ms step_avg:828.66ms +step:4000/20000 val_loss:2.0755 val_bpb:1.2292 train_time:3315001ms step_avg:828.75ms +step:4500/20000 train_loss:2.1098 train_time:3728932ms step_avg:828.65ms +step:5000/20000 train_loss:2.0727 train_time:4143237ms step_avg:828.65ms +step:5500/20000 train_loss:2.1133 train_time:4557500ms step_avg:828.64ms +step:6000/20000 train_loss:2.0121 train_time:4971750ms step_avg:828.63ms +step:6500/20000 train_loss:2.2772 train_time:5385925ms step_avg:828.60ms +step:7000/20000 train_loss:1.9077 train_time:5800366ms step_avg:828.62ms +step:7500/20000 train_loss:2.0332 train_time:6214436ms step_avg:828.59ms +step:8000/20000 train_loss:1.9856 train_time:6628761ms step_avg:828.60ms +step:8000/20000 val_loss:2.0333 val_bpb:1.2043 train_time:6629111ms step_avg:828.64ms +step:8500/20000 train_loss:1.9662 train_time:7043229ms step_avg:828.62ms +step:9000/20000 train_loss:2.0891 train_time:7457441ms step_avg:828.60ms +step:9500/20000 train_loss:2.2474 train_time:7872123ms step_avg:828.64ms +step:10000/20000 train_loss:2.0618 train_time:8286352ms step_avg:828.64ms +step:10500/20000 train_loss:2.0948 train_time:8700592ms step_avg:828.63ms +step:11000/20000 train_loss:1.9865 train_time:9115076ms step_avg:828.64ms +step:11500/20000 train_loss:1.9466 train_time:9529240ms step_avg:828.63ms +step:12000/20000 train_loss:1.9609 train_time:9943485ms step_avg:828.62ms +step:12000/20000 val_loss:2.0241 val_bpb:1.1988 train_time:9943828ms step_avg:828.65ms +step:12500/20000 train_loss:1.9691 train_time:10357780ms step_avg:828.62ms +step:13000/20000 train_loss:1.9781 train_time:10772024ms step_avg:828.62ms +step:13500/20000 train_loss:2.0928 train_time:11186188ms step_avg:828.61ms +step:14000/20000 train_loss:1.8907 train_time:11600289ms step_avg:828.59ms +step:14500/20000 train_loss:2.0651 train_time:12014544ms step_avg:828.59ms +step:15000/20000 train_loss:2.0078 train_time:12428690ms step_avg:828.58ms +step:15500/20000 train_loss:1.9744 train_time:12842898ms step_avg:828.57ms +step:16000/20000 train_loss:2.0443 train_time:13257276ms step_avg:828.58ms +step:16000/20000 val_loss:1.9697 val_bpb:1.1666 train_time:13257627ms step_avg:828.60ms +step:16500/20000 train_loss:2.0235 train_time:13671525ms step_avg:828.58ms +step:17000/20000 train_loss:2.0201 train_time:14085842ms step_avg:828.58ms +step:17500/20000 train_loss:1.9946 train_time:14500144ms step_avg:828.58ms +step:18000/20000 train_loss:1.8680 train_time:14914297ms step_avg:828.57ms +swa:start step:18450 +step:18500/20000 train_loss:1.8339 train_time:15328815ms step_avg:828.58ms +late_qat:enabled step:18831 scale:0.1499 +step:19000/20000 train_loss:1.8034 train_time:15744113ms step_avg:828.64ms +step:19500/20000 train_loss:1.9822 train_time:16159354ms step_avg:828.68ms +step:20000/20000 train_loss:1.8749 train_time:16574509ms step_avg:828.73ms +step:20000/20000 val_loss:1.8748 val_bpb:1.1104 train_time:16574910ms step_avg:828.75ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8738 val_bpb:1.1098 eval_time:19704ms +Serialized model: 106027446 bytes +Code size: 90361 bytes +Serialized model int6+lzma: 14987572 bytes +Total submission size int6+lzma: 15077933 bytes +final_int6_roundtrip val_loss:1.8998 val_bpb:1.1252 eval_time:24388ms +final_int6_roundtrip_exact val_loss:1.89979453 val_bpb:1.12516533 +final_int6_sliding_window val_loss:1.8604 val_bpb:1.1018 stride:64 eval_time:664624ms +final_int6_sliding_window_exact val_loss:1.86041005 val_bpb:1.10184254 +final_int8_zlib_roundtrip_exact val_loss:1.86041005 val_bpb:1.10184254 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.148962 time=1.1s + ttt_chunk [11/1893] bpb=1.106649 time=11.9s + ttt_chunk [21/1893] bpb=1.109183 time=22.7s + ttt_chunk [31/1893] bpb=1.112465 time=33.4s + ttt_chunk [41/1893] bpb=1.101158 time=44.2s + ttt_chunk [51/1893] bpb=1.097352 time=55.0s + ttt_chunk [61/1893] bpb=1.102163 time=65.7s + ttt_chunk [71/1893] bpb=1.099272 time=76.5s + ttt_chunk [81/1893] bpb=1.099287 time=87.3s + ttt_chunk [91/1893] bpb=1.097889 time=98.1s + ttt_chunk [101/1893] bpb=1.100512 time=108.9s + ttt_chunk [111/1893] bpb=1.101752 time=119.7s + ttt_chunk [121/1893] bpb=1.098599 time=130.4s + ttt_chunk [131/1893] bpb=1.098481 time=141.2s + ttt_chunk [141/1893] bpb=1.098188 time=152.0s + ttt_chunk [151/1893] bpb=1.101303 time=162.8s + ttt_chunk [161/1893] bpb=1.102934 time=173.6s + ttt_chunk [171/1893] bpb=1.103556 time=184.3s + ttt_chunk [181/1893] bpb=1.103556 time=195.1s + ttt_chunk [191/1893] bpb=1.106780 time=205.9s + ttt_chunk [201/1893] bpb=1.107089 time=216.7s + ttt_chunk [211/1893] bpb=1.104621 time=227.4s + ttt_chunk [221/1893] bpb=1.106682 time=238.2s + ttt_chunk [231/1893] bpb=1.106153 time=249.0s + ttt_chunk [241/1893] bpb=1.105951 time=259.8s + ttt_chunk [251/1893] bpb=1.104500 time=270.5s + ttt_chunk [261/1893] bpb=1.102854 time=281.3s + ttt_chunk [271/1893] bpb=1.101507 time=292.1s + ttt_chunk [281/1893] bpb=1.103788 time=302.9s + ttt_chunk [291/1893] bpb=1.104513 time=313.6s + ttt_chunk [301/1893] bpb=1.105062 time=324.4s + ttt_chunk [311/1893] bpb=1.106585 time=335.2s + ttt_chunk [321/1893] bpb=1.108036 time=346.0s + ttt_chunk [331/1893] bpb=1.107902 time=356.7s + ttt_chunk [341/1893] bpb=1.108209 time=367.5s + ttt_chunk [351/1893] bpb=1.109486 time=378.3s + ttt_chunk [361/1893] bpb=1.110674 time=389.1s + ttt_chunk [371/1893] bpb=1.110192 time=399.9s + ttt_chunk [381/1893] bpb=1.110144 time=410.6s + ttt_chunk [391/1893] bpb=1.109733 time=421.4s + ttt_chunk [401/1893] bpb=1.108356 time=432.2s + ttt_chunk [411/1893] bpb=1.107227 time=442.9s + ttt_chunk [421/1893] bpb=1.106719 time=453.7s + ttt_chunk [431/1893] bpb=1.107374 time=464.5s + ttt_chunk [441/1893] bpb=1.107195 time=475.3s + ttt_chunk [451/1893] bpb=1.106903 time=486.1s + ttt_chunk [461/1893] bpb=1.106178 time=496.8s + ttt_chunk [471/1893] bpb=1.105772 time=507.6s + ttt_chunk [481/1893] bpb=1.105518 time=518.4s + ttt_chunk [491/1893] bpb=1.105211 time=529.2s + ttt_chunk [501/1893] bpb=1.104631 time=539.9s + ttt_chunk [511/1893] bpb=1.104022 time=550.7s + ttt_chunk [521/1893] bpb=1.102977 time=561.5s + ttt_chunk [531/1893] bpb=1.103083 time=572.2s + ttt_chunk [541/1893] bpb=1.103047 time=583.0s + ttt_chunk [551/1893] bpb=1.101880 time=593.8s + ttt_chunk [561/1893] bpb=1.102420 time=604.6s + ttt_chunk [571/1893] bpb=1.101740 time=615.3s + ttt_chunk [581/1893] bpb=1.101206 time=626.1s + ttt_chunk [591/1893] bpb=1.100569 time=636.9s + ttt_chunk [601/1893] bpb=1.101233 time=647.7s + ttt_chunk [611/1893] bpb=1.100826 time=658.5s + ttt_chunk [621/1893] bpb=1.100733 time=669.2s + ttt_chunk [631/1893] bpb=1.101122 time=680.0s + ttt_chunk [641/1893] bpb=1.100907 time=690.8s + ttt_chunk [651/1893] bpb=1.100916 time=701.6s + ttt_chunk [661/1893] bpb=1.100822 time=712.3s + ttt_chunk [671/1893] bpb=1.100437 time=723.1s + ttt_chunk [681/1893] bpb=1.100704 time=733.9s + ttt_chunk [691/1893] bpb=1.101402 time=744.7s + ttt_chunk [701/1893] bpb=1.100649 time=755.4s + ttt_chunk [711/1893] bpb=1.101214 time=766.2s + ttt_chunk [721/1893] bpb=1.100855 time=777.0s + ttt_chunk [731/1893] bpb=1.101286 time=787.8s + ttt_chunk [741/1893] bpb=1.101246 time=798.6s + ttt_chunk [751/1893] bpb=1.100906 time=809.3s + ttt_chunk [761/1893] bpb=1.100754 time=820.1s + ttt_chunk [771/1893] bpb=1.100535 time=830.9s + ttt_chunk [781/1893] bpb=1.101052 time=841.6s + ttt_chunk [791/1893] bpb=1.100729 time=852.4s + ttt_chunk [801/1893] bpb=1.100817 time=863.2s + ttt_chunk [811/1893] bpb=1.100349 time=874.0s + ttt_chunk [821/1893] bpb=1.100161 time=884.7s + ttt_chunk [831/1893] bpb=1.099731 time=895.5s + ttt_chunk [841/1893] bpb=1.099163 time=906.3s + ttt_chunk [851/1893] bpb=1.099062 time=917.1s + ttt_chunk [861/1893] bpb=1.099192 time=927.9s + ttt_chunk [871/1893] bpb=1.099234 time=938.6s + ttt_chunk [881/1893] bpb=1.099269 time=949.4s + ttt_chunk [891/1893] bpb=1.099098 time=960.2s + ttt_chunk [901/1893] bpb=1.099091 time=971.0s + ttt_chunk [911/1893] bpb=1.099081 time=981.7s + ttt_chunk [921/1893] bpb=1.099457 time=992.5s + ttt_chunk [931/1893] bpb=1.099302 time=1003.3s + ttt_chunk [941/1893] bpb=1.099197 time=1014.1s + ttt_chunk [951/1893] bpb=1.099211 time=1024.8s + ttt_chunk [961/1893] bpb=1.098954 time=1035.6s + ttt_chunk [971/1893] bpb=1.099679 time=1046.4s + ttt_chunk [981/1893] bpb=1.099805 time=1057.1s + ttt_chunk [991/1893] bpb=1.099714 time=1067.9s + ttt_chunk [1001/1893] bpb=1.099874 time=1078.7s + ttt_chunk [1011/1893] bpb=1.100135 time=1089.5s + ttt_chunk [1021/1893] bpb=1.100290 time=1100.2s + ttt_chunk [1031/1893] bpb=1.100826 time=1111.0s + ttt_chunk [1041/1893] bpb=1.100503 time=1121.8s + ttt_chunk [1051/1893] bpb=1.100230 time=1132.6s + ttt_chunk [1061/1893] bpb=1.100500 time=1143.3s + ttt_chunk [1071/1893] bpb=1.100973 time=1154.1s + ttt_chunk [1081/1893] bpb=1.100979 time=1164.9s + ttt_chunk [1091/1893] bpb=1.101374 time=1175.7s + ttt_chunk [1101/1893] bpb=1.101517 time=1186.4s + ttt_chunk [1111/1893] bpb=1.101291 time=1197.2s + ttt_chunk [1121/1893] bpb=1.101240 time=1208.0s + ttt_chunk [1131/1893] bpb=1.101147 time=1218.8s + ttt_chunk [1141/1893] bpb=1.101001 time=1229.5s + ttt_chunk [1151/1893] bpb=1.101048 time=1240.3s + ttt_chunk [1161/1893] bpb=1.100357 time=1251.1s + ttt_chunk [1171/1893] bpb=1.100923 time=1261.9s + ttt_chunk [1181/1893] bpb=1.100425 time=1272.6s + ttt_chunk [1191/1893] bpb=1.100132 time=1283.4s + ttt_chunk [1201/1893] bpb=1.100718 time=1294.2s + ttt_chunk [1211/1893] bpb=1.100113 time=1304.9s + ttt_chunk [1221/1893] bpb=1.099789 time=1315.7s + ttt_chunk [1231/1893] bpb=1.099716 time=1326.5s + ttt_chunk [1241/1893] bpb=1.099538 time=1337.3s + ttt_chunk [1251/1893] bpb=1.099298 time=1348.0s + ttt_chunk [1261/1893] bpb=1.099230 time=1358.8s + ttt_chunk [1271/1893] bpb=1.099030 time=1369.6s + ttt_chunk [1281/1893] bpb=1.098825 time=1380.4s + ttt_chunk [1291/1893] bpb=1.098638 time=1391.1s + ttt_chunk [1301/1893] bpb=1.098263 time=1401.9s + ttt_chunk [1311/1893] bpb=1.097961 time=1412.7s + ttt_chunk [1321/1893] bpb=1.097795 time=1423.4s + ttt_chunk [1331/1893] bpb=1.097718 time=1434.2s + ttt_chunk [1341/1893] bpb=1.097600 time=1445.0s + ttt_chunk [1351/1893] bpb=1.097546 time=1455.8s + ttt_chunk [1361/1893] bpb=1.097752 time=1466.5s + ttt_chunk [1371/1893] bpb=1.097602 time=1477.3s + ttt_chunk [1381/1893] bpb=1.097542 time=1488.1s + ttt_chunk [1391/1893] bpb=1.097027 time=1498.8s + ttt_chunk [1401/1893] bpb=1.097052 time=1509.6s + ttt_chunk [1411/1893] bpb=1.097076 time=1520.4s + ttt_chunk [1421/1893] bpb=1.097336 time=1531.2s + ttt_chunk [1431/1893] bpb=1.097182 time=1541.9s + ttt_chunk [1441/1893] bpb=1.097860 time=1552.7s + ttt_chunk [1451/1893] bpb=1.097967 time=1563.5s + ttt_chunk [1461/1893] bpb=1.097711 time=1574.2s + ttt_chunk [1471/1893] bpb=1.098624 time=1585.0s + ttt_chunk [1481/1893] bpb=1.098435 time=1595.8s + ttt_chunk [1491/1893] bpb=1.098445 time=1606.5s + ttt_chunk [1501/1893] bpb=1.098608 time=1617.3s + ttt_chunk [1511/1893] bpb=1.098687 time=1628.1s + ttt_chunk [1521/1893] bpb=1.098715 time=1638.9s + ttt_chunk [1531/1893] bpb=1.098533 time=1649.6s + ttt_chunk [1541/1893] bpb=1.098501 time=1660.4s + ttt_chunk [1551/1893] bpb=1.098854 time=1671.2s + ttt_chunk [1561/1893] bpb=1.098993 time=1681.9s + ttt_chunk [1571/1893] bpb=1.099120 time=1692.7s + ttt_chunk [1581/1893] bpb=1.099236 time=1703.5s + ttt_chunk [1591/1893] bpb=1.099189 time=1714.3s + ttt_chunk [1601/1893] bpb=1.099373 time=1725.0s + ttt_chunk [1611/1893] bpb=1.099446 time=1735.8s + ttt_chunk [1621/1893] bpb=1.099315 time=1746.6s + ttt_chunk [1631/1893] bpb=1.099486 time=1757.4s + ttt_chunk [1641/1893] bpb=1.099365 time=1768.1s + ttt_chunk [1651/1893] bpb=1.099269 time=1778.9s + ttt_chunk [1661/1893] bpb=1.099155 time=1789.7s + ttt_chunk [1671/1893] bpb=1.099519 time=1800.4s + ttt_chunk [1681/1893] bpb=1.099767 time=1811.2s + ttt_chunk [1691/1893] bpb=1.099741 time=1822.0s + ttt_chunk [1701/1893] bpb=1.099731 time=1832.8s + ttt_chunk [1711/1893] bpb=1.099572 time=1843.5s + ttt_chunk [1721/1893] bpb=1.099426 time=1854.3s + ttt_chunk [1731/1893] bpb=1.099378 time=1865.1s + ttt_chunk [1741/1893] bpb=1.099166 time=1875.8s + ttt_chunk [1751/1893] bpb=1.099024 time=1886.6s + ttt_chunk [1761/1893] bpb=1.099109 time=1897.4s + ttt_chunk [1771/1893] bpb=1.099028 time=1908.2s + ttt_chunk [1781/1893] bpb=1.098987 time=1918.9s + ttt_chunk [1791/1893] bpb=1.098601 time=1929.7s + ttt_chunk [1801/1893] bpb=1.098603 time=1940.5s + ttt_chunk [1811/1893] bpb=1.098432 time=1951.3s + ttt_chunk [1821/1893] bpb=1.098469 time=1962.0s + ttt_chunk [1831/1893] bpb=1.098074 time=1972.8s + ttt_chunk [1841/1893] bpb=1.098044 time=1983.6s + ttt_chunk [1851/1893] bpb=1.097814 time=1994.3s + ttt_chunk [1861/1893] bpb=1.097345 time=2005.1s + ttt_chunk [1871/1893] bpb=1.097193 time=2015.9s + ttt_chunk [1881/1893] bpb=1.096820 time=2026.7s + ttt_chunk [1891/1893] bpb=1.096650 time=2037.4s + ttt_chunk [1893/1893] bpb=1.096669 time=2039.2s +ttt_sliding:done val_loss=1.850081 val_bpb=1.095725 elapsed=2039.2s +legal_ttt val_loss:1.8501 val_bpb:1.0957 eval_time:2039891ms +legal_ttt_exact val_loss:1.85008136 val_bpb:1.09572529 diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed2024.txt b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed2024.txt new file mode 100644 index 0000000000..ffda9fec99 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed2024.txt @@ -0,0 +1,2242 @@ +logs/train_step20k_seed2024.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:2024 +[rank1]:W0405 22:02:50.695000 884 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +[rank0]:W0405 22:02:50.822000 883 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +[rank2]:W0405 22:02:50.835000 885 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +[rank3]:W0405 22:02:51.189000 886 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9290 val_bpb:4.1037 artifact_bytes:4577366 model_int6_lzma:4485864 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:848ms step_avg:847.83ms +step:2/20000 train_loss:8.6930 train_time:1396ms step_avg:698.20ms +step:3/20000 train_loss:7.6769 train_time:2253ms step_avg:751.10ms +step:4/20000 train_loss:7.0844 train_time:3110ms step_avg:777.41ms +step:5/20000 train_loss:7.0620 train_time:3967ms step_avg:793.32ms +step:6/20000 train_loss:7.0071 train_time:4823ms step_avg:803.82ms +step:7/20000 train_loss:7.0103 train_time:5679ms step_avg:811.35ms +step:8/20000 train_loss:6.8079 train_time:6536ms step_avg:817.02ms +step:9/20000 train_loss:6.5751 train_time:7393ms step_avg:821.46ms +step:10/20000 train_loss:6.2326 train_time:8250ms step_avg:824.97ms +step:500/20000 train_loss:2.3519 train_time:423343ms step_avg:846.69ms +step:1000/20000 train_loss:2.2637 train_time:849745ms step_avg:849.75ms +step:1500/20000 train_loss:2.1797 train_time:1275018ms step_avg:850.01ms +step:2000/20000 train_loss:2.0328 train_time:1701329ms step_avg:850.66ms +step:2000/20000 val_loss:2.1360 val_bpb:1.2651 artifact_bytes:13800942 model_int6_lzma:13709440 train_time:1701663ms step_avg:850.83ms +step:2500/20000 train_loss:2.1597 train_time:2125765ms step_avg:850.31ms +step:3000/20000 train_loss:2.1205 train_time:2550537ms step_avg:850.18ms +step:3500/20000 train_loss:2.1840 train_time:2978216ms step_avg:850.92ms +step:4000/20000 train_loss:2.1437 train_time:3407355ms step_avg:851.84ms +step:4000/20000 val_loss:2.0744 val_bpb:1.2286 artifact_bytes:16959534 model_int6_lzma:16868032 train_time:3407699ms step_avg:851.92ms +step:4500/20000 train_loss:2.1109 train_time:3832530ms step_avg:851.67ms +step:5000/20000 train_loss:2.0706 train_time:4257966ms step_avg:851.59ms +step:5500/20000 train_loss:2.1133 train_time:4682702ms step_avg:851.40ms +step:6000/20000 train_loss:2.0121 train_time:5106901ms step_avg:851.15ms +step:6000/20000 val_loss:2.0467 val_bpb:1.2122 artifact_bytes:17243366 model_int6_lzma:17151864 train_time:5107244ms step_avg:851.21ms +step:6500/20000 train_loss:2.2763 train_time:5528450ms step_avg:850.53ms +step:7000/20000 train_loss:1.9070 train_time:5955474ms step_avg:850.78ms +step:7500/20000 train_loss:2.0288 train_time:6378084ms step_avg:850.41ms +step:8000/20000 train_loss:1.9864 train_time:6803482ms step_avg:850.44ms +step:8000/20000 val_loss:2.0339 val_bpb:1.2046 artifact_bytes:17248774 model_int6_lzma:17157272 train_time:6803837ms step_avg:850.48ms +step:8500/20000 train_loss:1.9686 train_time:7228939ms step_avg:850.46ms +step:9000/20000 train_loss:2.0894 train_time:7655159ms step_avg:850.57ms +step:9500/20000 train_loss:2.2479 train_time:8080823ms step_avg:850.61ms +step:10000/20000 train_loss:2.0598 train_time:8505899ms step_avg:850.59ms +step:10000/20000 val_loss:2.0273 val_bpb:1.2007 artifact_bytes:17246738 model_int6_lzma:17155236 train_time:8506253ms step_avg:850.63ms +step:10500/20000 train_loss:2.0975 train_time:8922740ms step_avg:849.78ms +step:11000/20000 train_loss:1.9898 train_time:9336990ms step_avg:848.82ms +step:11500/20000 train_loss:1.9468 train_time:9751240ms step_avg:847.93ms +step:12000/20000 train_loss:1.9626 train_time:10165487ms step_avg:847.12ms +step:12000/20000 val_loss:2.0252 val_bpb:1.1994 artifact_bytes:17231058 model_int6_lzma:17139556 train_time:10165832ms step_avg:847.15ms +step:12500/20000 train_loss:1.9657 train_time:10579788ms step_avg:846.38ms +step:13000/20000 train_loss:1.9791 train_time:10994081ms step_avg:845.70ms +step:13500/20000 train_loss:2.0937 train_time:11408330ms step_avg:845.06ms +step:14000/20000 train_loss:1.8916 train_time:11822591ms step_avg:844.47ms +step:14000/20000 val_loss:1.9983 val_bpb:1.1835 artifact_bytes:16929622 model_int6_lzma:16838120 train_time:11822934ms step_avg:844.50ms +step:14500/20000 train_loss:2.0665 train_time:12236977ms step_avg:843.93ms +step:15000/20000 train_loss:2.0097 train_time:12651249ms step_avg:843.42ms +step:15500/20000 train_loss:1.9744 train_time:13065485ms step_avg:842.93ms +step:16000/20000 train_loss:2.0459 train_time:13479771ms step_avg:842.49ms +step:16000/20000 val_loss:1.9707 val_bpb:1.1672 artifact_bytes:16321958 model_int6_lzma:16230456 train_time:13480122ms step_avg:842.51ms +step:16500/20000 train_loss:2.0228 train_time:13894217ms step_avg:842.07ms +step:17000/20000 train_loss:2.0202 train_time:14308482ms step_avg:841.68ms +step:17500/20000 train_loss:1.9969 train_time:14722729ms step_avg:841.30ms +step:18000/20000 train_loss:1.8682 train_time:15136963ms step_avg:840.94ms +step:18000/20000 val_loss:1.9297 val_bpb:1.1429 artifact_bytes:15534274 model_int6_lzma:15442772 train_time:15137307ms step_avg:840.96ms +swa:start step:18450 +step:18500/20000 train_loss:1.8317 train_time:15551945ms step_avg:840.65ms +late_qat:enabled step:18831 scale:0.1499 +step:19000/20000 train_loss:1.8044 train_time:15966815ms step_avg:840.36ms +step:19500/20000 train_loss:1.9819 train_time:16381777ms step_avg:840.09ms +step:20000/20000 train_loss:1.8757 train_time:16796717ms step_avg:839.84ms +step:20000/20000 val_loss:1.8759 val_bpb:1.1110 artifact_bytes:14942394 model_int6_lzma:14850892 train_time:16797108ms step_avg:839.86ms +peak memory allocated: 22092 MiB reserved: 22360 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8748 val_bpb:1.1104 eval_time:19929ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14892980 bytes +Total submission size int6+lzma: 14984482 bytes +final_int6_roundtrip val_loss:1.8998 val_bpb:1.1252 eval_time:37710ms +final_int6_roundtrip_exact val_loss:1.89978161 val_bpb:1.12515768 +final_int6_sliding_window val_loss:1.8602 val_bpb:1.1017 stride:64 eval_time:671665ms +final_int6_sliding_window_exact val_loss:1.86015002 val_bpb:1.10168853 +final_int8_zlib_roundtrip_exact val_loss:1.86015002 val_bpb:1.10168853 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.151935 time=1.1s + ttt_chunk [11/1893] bpb=1.105007 time=11.9s + ttt_chunk [21/1893] bpb=1.109129 time=22.7s + ttt_chunk [31/1893] bpb=1.112554 time=33.5s + ttt_chunk [41/1893] bpb=1.101507 time=44.2s + ttt_chunk [51/1893] bpb=1.097495 time=55.0s + ttt_chunk [61/1893] bpb=1.102100 time=65.8s + ttt_chunk [71/1893] bpb=1.098881 time=76.6s + ttt_chunk [81/1893] bpb=1.098894 time=87.4s + ttt_chunk [91/1893] bpb=1.097827 time=98.1s + ttt_chunk [101/1893] bpb=1.100631 time=108.9s + ttt_chunk [111/1893] bpb=1.101779 time=119.7s + ttt_chunk [121/1893] bpb=1.098637 time=130.5s + ttt_chunk [131/1893] bpb=1.098493 time=141.3s + ttt_chunk [141/1893] bpb=1.098221 time=152.1s + ttt_chunk [151/1893] bpb=1.101361 time=162.8s + ttt_chunk [161/1893] bpb=1.103076 time=173.6s + ttt_chunk [171/1893] bpb=1.103687 time=184.4s + ttt_chunk [181/1893] bpb=1.103813 time=195.2s + ttt_chunk [191/1893] bpb=1.106990 time=206.0s + ttt_chunk [201/1893] bpb=1.107242 time=216.8s + ttt_chunk [211/1893] bpb=1.104827 time=227.5s + ttt_chunk [221/1893] bpb=1.106837 time=238.3s + ttt_chunk [231/1893] bpb=1.106295 time=249.1s + ttt_chunk [241/1893] bpb=1.106172 time=259.9s + ttt_chunk [251/1893] bpb=1.104685 time=270.7s + ttt_chunk [261/1893] bpb=1.103090 time=281.5s + ttt_chunk [271/1893] bpb=1.101773 time=292.2s + ttt_chunk [281/1893] bpb=1.104028 time=303.0s + ttt_chunk [291/1893] bpb=1.104736 time=313.8s + ttt_chunk [301/1893] bpb=1.105353 time=324.6s + ttt_chunk [311/1893] bpb=1.106877 time=335.4s + ttt_chunk [321/1893] bpb=1.108352 time=346.1s + ttt_chunk [331/1893] bpb=1.108167 time=356.9s + ttt_chunk [341/1893] bpb=1.108495 time=367.7s + ttt_chunk [351/1893] bpb=1.109748 time=378.5s + ttt_chunk [361/1893] bpb=1.110945 time=389.3s + ttt_chunk [371/1893] bpb=1.110528 time=400.1s + ttt_chunk [381/1893] bpb=1.110477 time=410.8s + ttt_chunk [391/1893] bpb=1.110107 time=421.6s + ttt_chunk [401/1893] bpb=1.108760 time=432.4s + ttt_chunk [411/1893] bpb=1.107682 time=443.2s + ttt_chunk [421/1893] bpb=1.107120 time=454.0s + ttt_chunk [431/1893] bpb=1.107848 time=464.7s + ttt_chunk [441/1893] bpb=1.107673 time=475.5s + ttt_chunk [451/1893] bpb=1.107376 time=486.3s + ttt_chunk [461/1893] bpb=1.106651 time=497.1s + ttt_chunk [471/1893] bpb=1.106241 time=507.9s + ttt_chunk [481/1893] bpb=1.105941 time=518.7s + ttt_chunk [491/1893] bpb=1.105641 time=529.5s + ttt_chunk [501/1893] bpb=1.105093 time=540.2s + ttt_chunk [511/1893] bpb=1.104505 time=551.0s + ttt_chunk [521/1893] bpb=1.103482 time=561.8s + ttt_chunk [531/1893] bpb=1.103621 time=572.6s + ttt_chunk [541/1893] bpb=1.103601 time=583.3s + ttt_chunk [551/1893] bpb=1.102447 time=594.1s + ttt_chunk [561/1893] bpb=1.102973 time=604.9s + ttt_chunk [571/1893] bpb=1.102272 time=615.7s + ttt_chunk [581/1893] bpb=1.101751 time=626.5s + ttt_chunk [591/1893] bpb=1.101087 time=637.3s + ttt_chunk [601/1893] bpb=1.101777 time=648.0s + ttt_chunk [611/1893] bpb=1.101353 time=658.8s + ttt_chunk [621/1893] bpb=1.101270 time=669.6s + ttt_chunk [631/1893] bpb=1.101666 time=680.4s + ttt_chunk [641/1893] bpb=1.101459 time=691.2s + ttt_chunk [651/1893] bpb=1.101467 time=702.0s + ttt_chunk [661/1893] bpb=1.101370 time=712.7s + ttt_chunk [671/1893] bpb=1.100993 time=723.5s + ttt_chunk [681/1893] bpb=1.101258 time=734.3s + ttt_chunk [691/1893] bpb=1.101954 time=745.1s + ttt_chunk [701/1893] bpb=1.101187 time=755.9s + ttt_chunk [711/1893] bpb=1.101778 time=766.7s + ttt_chunk [721/1893] bpb=1.101410 time=777.5s + ttt_chunk [731/1893] bpb=1.101858 time=788.2s + ttt_chunk [741/1893] bpb=1.101812 time=799.0s + ttt_chunk [751/1893] bpb=1.101471 time=809.8s + ttt_chunk [761/1893] bpb=1.101283 time=820.6s + ttt_chunk [771/1893] bpb=1.101051 time=831.4s + ttt_chunk [781/1893] bpb=1.101576 time=842.2s + ttt_chunk [791/1893] bpb=1.101268 time=853.0s + ttt_chunk [801/1893] bpb=1.101383 time=863.8s + ttt_chunk [811/1893] bpb=1.100873 time=874.5s + ttt_chunk [821/1893] bpb=1.100672 time=885.3s + ttt_chunk [831/1893] bpb=1.100249 time=896.1s + ttt_chunk [841/1893] bpb=1.099666 time=906.9s + ttt_chunk [851/1893] bpb=1.099564 time=917.7s + ttt_chunk [861/1893] bpb=1.099674 time=928.5s + ttt_chunk [871/1893] bpb=1.099722 time=939.2s + ttt_chunk [881/1893] bpb=1.099752 time=950.0s + ttt_chunk [891/1893] bpb=1.099572 time=960.8s + ttt_chunk [901/1893] bpb=1.099562 time=971.6s + ttt_chunk [911/1893] bpb=1.099561 time=982.4s + ttt_chunk [921/1893] bpb=1.099952 time=993.2s + ttt_chunk [931/1893] bpb=1.099787 time=1004.0s + ttt_chunk [941/1893] bpb=1.099706 time=1014.7s + ttt_chunk [951/1893] bpb=1.099700 time=1025.5s + ttt_chunk [961/1893] bpb=1.099438 time=1036.3s + ttt_chunk [971/1893] bpb=1.100166 time=1047.1s + ttt_chunk [981/1893] bpb=1.100310 time=1057.9s + ttt_chunk [991/1893] bpb=1.100222 time=1068.7s + ttt_chunk [1001/1893] bpb=1.100375 time=1079.5s + ttt_chunk [1011/1893] bpb=1.100636 time=1090.2s + ttt_chunk [1021/1893] bpb=1.100805 time=1101.0s + ttt_chunk [1031/1893] bpb=1.101354 time=1111.8s + ttt_chunk [1041/1893] bpb=1.101015 time=1122.6s + ttt_chunk [1051/1893] bpb=1.100751 time=1133.4s + ttt_chunk [1061/1893] bpb=1.101009 time=1144.2s + ttt_chunk [1071/1893] bpb=1.101477 time=1155.0s + ttt_chunk [1081/1893] bpb=1.101501 time=1165.7s + ttt_chunk [1091/1893] bpb=1.101917 time=1176.5s + ttt_chunk [1101/1893] bpb=1.102051 time=1187.3s + ttt_chunk [1111/1893] bpb=1.101826 time=1198.1s + ttt_chunk [1121/1893] bpb=1.101755 time=1208.9s + ttt_chunk [1131/1893] bpb=1.101655 time=1219.7s + ttt_chunk [1141/1893] bpb=1.101510 time=1230.5s + ttt_chunk [1151/1893] bpb=1.101547 time=1241.3s + ttt_chunk [1161/1893] bpb=1.100839 time=1252.0s + ttt_chunk [1171/1893] bpb=1.101410 time=1262.8s + ttt_chunk [1181/1893] bpb=1.100911 time=1273.6s + ttt_chunk [1191/1893] bpb=1.100623 time=1284.4s + ttt_chunk [1201/1893] bpb=1.101185 time=1295.2s + ttt_chunk [1211/1893] bpb=1.100594 time=1306.0s + ttt_chunk [1221/1893] bpb=1.100288 time=1316.7s + ttt_chunk [1231/1893] bpb=1.100210 time=1327.5s + ttt_chunk [1241/1893] bpb=1.100019 time=1338.3s + ttt_chunk [1251/1893] bpb=1.099770 time=1349.1s + ttt_chunk [1261/1893] bpb=1.099713 time=1359.9s + ttt_chunk [1271/1893] bpb=1.099513 time=1370.7s + ttt_chunk [1281/1893] bpb=1.099319 time=1381.5s + ttt_chunk [1291/1893] bpb=1.099143 time=1392.2s + ttt_chunk [1301/1893] bpb=1.098779 time=1403.0s + ttt_chunk [1311/1893] bpb=1.098463 time=1413.8s + ttt_chunk [1321/1893] bpb=1.098280 time=1424.6s + ttt_chunk [1331/1893] bpb=1.098197 time=1435.4s + ttt_chunk [1341/1893] bpb=1.098087 time=1446.2s + ttt_chunk [1351/1893] bpb=1.098035 time=1456.9s + ttt_chunk [1361/1893] bpb=1.098243 time=1467.7s + ttt_chunk [1371/1893] bpb=1.098120 time=1478.5s + ttt_chunk [1381/1893] bpb=1.098055 time=1489.3s + ttt_chunk [1391/1893] bpb=1.097532 time=1500.1s + ttt_chunk [1401/1893] bpb=1.097552 time=1510.9s + ttt_chunk [1411/1893] bpb=1.097576 time=1521.7s + ttt_chunk [1421/1893] bpb=1.097828 time=1532.4s + ttt_chunk [1431/1893] bpb=1.097672 time=1543.2s + ttt_chunk [1441/1893] bpb=1.098339 time=1554.0s + ttt_chunk [1451/1893] bpb=1.098463 time=1564.8s + ttt_chunk [1461/1893] bpb=1.098206 time=1575.6s + ttt_chunk [1471/1893] bpb=1.099115 time=1586.4s + ttt_chunk [1481/1893] bpb=1.098922 time=1597.2s + ttt_chunk [1491/1893] bpb=1.098927 time=1608.0s + ttt_chunk [1501/1893] bpb=1.099079 time=1618.8s + ttt_chunk [1511/1893] bpb=1.099159 time=1629.6s + ttt_chunk [1521/1893] bpb=1.099190 time=1640.4s + ttt_chunk [1531/1893] bpb=1.099007 time=1651.1s + ttt_chunk [1541/1893] bpb=1.098972 time=1661.9s + ttt_chunk [1551/1893] bpb=1.099301 time=1672.7s + ttt_chunk [1561/1893] bpb=1.099433 time=1683.5s + ttt_chunk [1571/1893] bpb=1.099557 time=1694.3s + ttt_chunk [1581/1893] bpb=1.099678 time=1705.1s + ttt_chunk [1591/1893] bpb=1.099632 time=1715.9s + ttt_chunk [1601/1893] bpb=1.099808 time=1726.7s + ttt_chunk [1611/1893] bpb=1.099882 time=1737.5s + ttt_chunk [1621/1893] bpb=1.099745 time=1748.3s + ttt_chunk [1631/1893] bpb=1.099914 time=1759.1s + ttt_chunk [1641/1893] bpb=1.099799 time=1769.9s + ttt_chunk [1651/1893] bpb=1.099707 time=1780.7s + ttt_chunk [1661/1893] bpb=1.099601 time=1791.5s + ttt_chunk [1671/1893] bpb=1.099960 time=1802.3s + ttt_chunk [1681/1893] bpb=1.100212 time=1813.0s + ttt_chunk [1691/1893] bpb=1.100183 time=1823.8s + ttt_chunk [1701/1893] bpb=1.100186 time=1834.6s + ttt_chunk [1711/1893] bpb=1.100023 time=1845.4s + ttt_chunk [1721/1893] bpb=1.099882 time=1856.2s + ttt_chunk [1731/1893] bpb=1.099831 time=1867.0s + ttt_chunk [1741/1893] bpb=1.099617 time=1877.8s + ttt_chunk [1751/1893] bpb=1.099467 time=1888.6s + ttt_chunk [1761/1893] bpb=1.099542 time=1899.4s + ttt_chunk [1771/1893] bpb=1.099472 time=1910.2s + ttt_chunk [1781/1893] bpb=1.099430 time=1921.0s + ttt_chunk [1791/1893] bpb=1.099046 time=1931.7s + ttt_chunk [1801/1893] bpb=1.099051 time=1942.5s + ttt_chunk [1811/1893] bpb=1.098878 time=1953.3s + ttt_chunk [1821/1893] bpb=1.098909 time=1964.1s + ttt_chunk [1831/1893] bpb=1.098502 time=1974.9s + ttt_chunk [1841/1893] bpb=1.098474 time=1985.7s + ttt_chunk [1851/1893] bpb=1.098256 time=1996.5s + ttt_chunk [1861/1893] bpb=1.097788 time=2007.3s + ttt_chunk [1871/1893] bpb=1.097634 time=2018.1s + ttt_chunk [1881/1893] bpb=1.097253 time=2028.8s + ttt_chunk [1891/1893] bpb=1.097097 time=2039.6s + ttt_chunk [1893/1893] bpb=1.097114 time=2041.4s +ttt_sliding:done val_loss=1.850946 val_bpb=1.096237 elapsed=2041.4s +legal_ttt val_loss:1.8509 val_bpb:1.0962 eval_time:2042089ms +legal_ttt_exact val_loss:1.85094589 val_bpb:1.09623731 += torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +# --- Data loading --- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + # Estimate artifact size (int6+lzma) at each validation step + try: + sd_snap = {k: v.detach().cpu() for k, v in base_model.state_dict().items() + if "mtp_heads" not in k} + unbanked_snap = _unbank_state_dict(sd_snap, args.num_layers) + q_snap, m_snap = mixed_quantize_int6(unbanked_snap, {"mlp", "attn"}) + snap_buf = io.BytesIO() + torch.save({"w": q_snap, "m": m_snap}, snap_buf) + snap_compressed = len(lzma.compress(snap_buf.getvalue(), preset=6)) + snap_code_bytes = len(code.encode("utf-8")) + snap_total = snap_compressed + snap_code_bytes + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_bytes:{snap_total} model_int6_lzma:{snap_compressed} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + except Exception as e: + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_size:error({e}) " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.10.14 (main, May 6 2024, 19:42:50) [GCC 11.2.0] +Running PyTorch 2.10.0+cu128 +Sun Apr 5 22:02:22 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 550.127.08 Driver Version: 550.127.08 CUDA Version: 12.4 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA A100-SXM4-80GB On | 00000000:27:00.0 Off | On | +| N/A 36C P0 100W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA A100-SXM4-80GB On | 00000000:2A:00.0 Off | On | +| N/A 33C P0 99W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA A100-SXM4-80GB On | 00000000:51:00.0 Off | On | +| N/A 33C P0 97W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA A100-SXM4-80GB On | 00000000:57:00.0 Off | On | +| N/A 37C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA A100-SXM4-80GB On | 00000000:9E:00.0 Off | On | +| N/A 37C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA A100-SXM4-80GB On | 00000000:A4:00.0 Off | On | +| N/A 33C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA A100-SXM4-80GB On | 00000000:CA:00.0 Off | On | +| N/A 38C P0 102W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| MIG devices: | ++------------------+----------------------------------+-----------+-----------------------+ +| GPU GI CI MIG | Memory-Usage | Vol| Shared | +| ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG | +| | | ECC| | +|==================+==================================+===========+=======================| +| 0 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 2 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 3 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 1 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 2 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 6 2 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 883 C /opt/conda/bin/python3 0MiB | +| 1 N/A N/A 884 C /opt/conda/bin/python3 0MiB | +| 2 N/A N/A 885 C /opt/conda/bin/python3 0MiB | +| 3 N/A N/A 886 C /opt/conda/bin/python3 0MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:2024 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9290 val_bpb:4.1037 artifact_bytes:4577366 model_int6_lzma:4485864 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:848ms step_avg:847.83ms +step:2/20000 train_loss:8.6930 train_time:1396ms step_avg:698.20ms +step:3/20000 train_loss:7.6769 train_time:2253ms step_avg:751.10ms +step:4/20000 train_loss:7.0844 train_time:3110ms step_avg:777.41ms +step:5/20000 train_loss:7.0620 train_time:3967ms step_avg:793.32ms +step:6/20000 train_loss:7.0071 train_time:4823ms step_avg:803.82ms +step:7/20000 train_loss:7.0103 train_time:5679ms step_avg:811.35ms +step:8/20000 train_loss:6.8079 train_time:6536ms step_avg:817.02ms +step:9/20000 train_loss:6.5751 train_time:7393ms step_avg:821.46ms +step:10/20000 train_loss:6.2326 train_time:8250ms step_avg:824.97ms +step:500/20000 train_loss:2.3519 train_time:423343ms step_avg:846.69ms +step:1000/20000 train_loss:2.2637 train_time:849745ms step_avg:849.75ms +step:1500/20000 train_loss:2.1797 train_time:1275018ms step_avg:850.01ms +step:2000/20000 train_loss:2.0328 train_time:1701329ms step_avg:850.66ms +step:2000/20000 val_loss:2.1360 val_bpb:1.2651 artifact_bytes:13800942 model_int6_lzma:13709440 train_time:1701663ms step_avg:850.83ms +step:2500/20000 train_loss:2.1597 train_time:2125765ms step_avg:850.31ms +step:3000/20000 train_loss:2.1205 train_time:2550537ms step_avg:850.18ms +step:3500/20000 train_loss:2.1840 train_time:2978216ms step_avg:850.92ms +step:4000/20000 train_loss:2.1437 train_time:3407355ms step_avg:851.84ms +step:4000/20000 val_loss:2.0744 val_bpb:1.2286 artifact_bytes:16959534 model_int6_lzma:16868032 train_time:3407699ms step_avg:851.92ms +step:4500/20000 train_loss:2.1109 train_time:3832530ms step_avg:851.67ms +step:5000/20000 train_loss:2.0706 train_time:4257966ms step_avg:851.59ms +step:5500/20000 train_loss:2.1133 train_time:4682702ms step_avg:851.40ms +step:6000/20000 train_loss:2.0121 train_time:5106901ms step_avg:851.15ms +step:6000/20000 val_loss:2.0467 val_bpb:1.2122 artifact_bytes:17243366 model_int6_lzma:17151864 train_time:5107244ms step_avg:851.21ms +step:6500/20000 train_loss:2.2763 train_time:5528450ms step_avg:850.53ms +step:7000/20000 train_loss:1.9070 train_time:5955474ms step_avg:850.78ms +step:7500/20000 train_loss:2.0288 train_time:6378084ms step_avg:850.41ms +step:8000/20000 train_loss:1.9864 train_time:6803482ms step_avg:850.44ms +step:8000/20000 val_loss:2.0339 val_bpb:1.2046 artifact_bytes:17248774 model_int6_lzma:17157272 train_time:6803837ms step_avg:850.48ms +step:8500/20000 train_loss:1.9686 train_time:7228939ms step_avg:850.46ms +step:9000/20000 train_loss:2.0894 train_time:7655159ms step_avg:850.57ms +step:9500/20000 train_loss:2.2479 train_time:8080823ms step_avg:850.61ms +step:10000/20000 train_loss:2.0598 train_time:8505899ms step_avg:850.59ms +step:10000/20000 val_loss:2.0273 val_bpb:1.2007 artifact_bytes:17246738 model_int6_lzma:17155236 train_time:8506253ms step_avg:850.63ms +step:10500/20000 train_loss:2.0975 train_time:8922740ms step_avg:849.78ms +step:11000/20000 train_loss:1.9898 train_time:9336990ms step_avg:848.82ms +step:11500/20000 train_loss:1.9468 train_time:9751240ms step_avg:847.93ms +step:12000/20000 train_loss:1.9626 train_time:10165487ms step_avg:847.12ms +step:12000/20000 val_loss:2.0252 val_bpb:1.1994 artifact_bytes:17231058 model_int6_lzma:17139556 train_time:10165832ms step_avg:847.15ms +step:12500/20000 train_loss:1.9657 train_time:10579788ms step_avg:846.38ms +step:13000/20000 train_loss:1.9791 train_time:10994081ms step_avg:845.70ms +step:13500/20000 train_loss:2.0937 train_time:11408330ms step_avg:845.06ms +step:14000/20000 train_loss:1.8916 train_time:11822591ms step_avg:844.47ms +step:14000/20000 val_loss:1.9983 val_bpb:1.1835 artifact_bytes:16929622 model_int6_lzma:16838120 train_time:11822934ms step_avg:844.50ms +step:14500/20000 train_loss:2.0665 train_time:12236977ms step_avg:843.93ms +step:15000/20000 train_loss:2.0097 train_time:12651249ms step_avg:843.42ms +step:15500/20000 train_loss:1.9744 train_time:13065485ms step_avg:842.93ms +step:16000/20000 train_loss:2.0459 train_time:13479771ms step_avg:842.49ms +step:16000/20000 val_loss:1.9707 val_bpb:1.1672 artifact_bytes:16321958 model_int6_lzma:16230456 train_time:13480122ms step_avg:842.51ms +step:16500/20000 train_loss:2.0228 train_time:13894217ms step_avg:842.07ms +step:17000/20000 train_loss:2.0202 train_time:14308482ms step_avg:841.68ms +step:17500/20000 train_loss:1.9969 train_time:14722729ms step_avg:841.30ms +step:18000/20000 train_loss:1.8682 train_time:15136963ms step_avg:840.94ms +step:18000/20000 val_loss:1.9297 val_bpb:1.1429 artifact_bytes:15534274 model_int6_lzma:15442772 train_time:15137307ms step_avg:840.96ms +swa:start step:18450 +step:18500/20000 train_loss:1.8317 train_time:15551945ms step_avg:840.65ms +late_qat:enabled step:18831 scale:0.1499 +step:19000/20000 train_loss:1.8044 train_time:15966815ms step_avg:840.36ms +step:19500/20000 train_loss:1.9819 train_time:16381777ms step_avg:840.09ms +step:20000/20000 train_loss:1.8757 train_time:16796717ms step_avg:839.84ms +step:20000/20000 val_loss:1.8759 val_bpb:1.1110 artifact_bytes:14942394 model_int6_lzma:14850892 train_time:16797108ms step_avg:839.86ms +peak memory allocated: 22092 MiB reserved: 22360 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8748 val_bpb:1.1104 eval_time:19929ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14892980 bytes +Total submission size int6+lzma: 14984482 bytes +final_int6_roundtrip val_loss:1.8998 val_bpb:1.1252 eval_time:37710ms +final_int6_roundtrip_exact val_loss:1.89978161 val_bpb:1.12515768 +final_int6_sliding_window val_loss:1.8602 val_bpb:1.1017 stride:64 eval_time:671665ms +final_int6_sliding_window_exact val_loss:1.86015002 val_bpb:1.10168853 +final_int8_zlib_roundtrip_exact val_loss:1.86015002 val_bpb:1.10168853 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.151935 time=1.1s + ttt_chunk [11/1893] bpb=1.105007 time=11.9s + ttt_chunk [21/1893] bpb=1.109129 time=22.7s + ttt_chunk [31/1893] bpb=1.112554 time=33.5s + ttt_chunk [41/1893] bpb=1.101507 time=44.2s + ttt_chunk [51/1893] bpb=1.097495 time=55.0s + ttt_chunk [61/1893] bpb=1.102100 time=65.8s + ttt_chunk [71/1893] bpb=1.098881 time=76.6s + ttt_chunk [81/1893] bpb=1.098894 time=87.4s + ttt_chunk [91/1893] bpb=1.097827 time=98.1s + ttt_chunk [101/1893] bpb=1.100631 time=108.9s + ttt_chunk [111/1893] bpb=1.101779 time=119.7s + ttt_chunk [121/1893] bpb=1.098637 time=130.5s + ttt_chunk [131/1893] bpb=1.098493 time=141.3s + ttt_chunk [141/1893] bpb=1.098221 time=152.1s + ttt_chunk [151/1893] bpb=1.101361 time=162.8s + ttt_chunk [161/1893] bpb=1.103076 time=173.6s + ttt_chunk [171/1893] bpb=1.103687 time=184.4s + ttt_chunk [181/1893] bpb=1.103813 time=195.2s + ttt_chunk [191/1893] bpb=1.106990 time=206.0s + ttt_chunk [201/1893] bpb=1.107242 time=216.8s + ttt_chunk [211/1893] bpb=1.104827 time=227.5s + ttt_chunk [221/1893] bpb=1.106837 time=238.3s + ttt_chunk [231/1893] bpb=1.106295 time=249.1s + ttt_chunk [241/1893] bpb=1.106172 time=259.9s + ttt_chunk [251/1893] bpb=1.104685 time=270.7s + ttt_chunk [261/1893] bpb=1.103090 time=281.5s + ttt_chunk [271/1893] bpb=1.101773 time=292.2s + ttt_chunk [281/1893] bpb=1.104028 time=303.0s + ttt_chunk [291/1893] bpb=1.104736 time=313.8s + ttt_chunk [301/1893] bpb=1.105353 time=324.6s + ttt_chunk [311/1893] bpb=1.106877 time=335.4s + ttt_chunk [321/1893] bpb=1.108352 time=346.1s + ttt_chunk [331/1893] bpb=1.108167 time=356.9s + ttt_chunk [341/1893] bpb=1.108495 time=367.7s + ttt_chunk [351/1893] bpb=1.109748 time=378.5s + ttt_chunk [361/1893] bpb=1.110945 time=389.3s + ttt_chunk [371/1893] bpb=1.110528 time=400.1s + ttt_chunk [381/1893] bpb=1.110477 time=410.8s + ttt_chunk [391/1893] bpb=1.110107 time=421.6s + ttt_chunk [401/1893] bpb=1.108760 time=432.4s + ttt_chunk [411/1893] bpb=1.107682 time=443.2s + ttt_chunk [421/1893] bpb=1.107120 time=454.0s + ttt_chunk [431/1893] bpb=1.107848 time=464.7s + ttt_chunk [441/1893] bpb=1.107673 time=475.5s + ttt_chunk [451/1893] bpb=1.107376 time=486.3s + ttt_chunk [461/1893] bpb=1.106651 time=497.1s + ttt_chunk [471/1893] bpb=1.106241 time=507.9s + ttt_chunk [481/1893] bpb=1.105941 time=518.7s + ttt_chunk [491/1893] bpb=1.105641 time=529.5s + ttt_chunk [501/1893] bpb=1.105093 time=540.2s + ttt_chunk [511/1893] bpb=1.104505 time=551.0s + ttt_chunk [521/1893] bpb=1.103482 time=561.8s + ttt_chunk [531/1893] bpb=1.103621 time=572.6s + ttt_chunk [541/1893] bpb=1.103601 time=583.3s + ttt_chunk [551/1893] bpb=1.102447 time=594.1s + ttt_chunk [561/1893] bpb=1.102973 time=604.9s + ttt_chunk [571/1893] bpb=1.102272 time=615.7s + ttt_chunk [581/1893] bpb=1.101751 time=626.5s + ttt_chunk [591/1893] bpb=1.101087 time=637.3s + ttt_chunk [601/1893] bpb=1.101777 time=648.0s + ttt_chunk [611/1893] bpb=1.101353 time=658.8s + ttt_chunk [621/1893] bpb=1.101270 time=669.6s + ttt_chunk [631/1893] bpb=1.101666 time=680.4s + ttt_chunk [641/1893] bpb=1.101459 time=691.2s + ttt_chunk [651/1893] bpb=1.101467 time=702.0s + ttt_chunk [661/1893] bpb=1.101370 time=712.7s + ttt_chunk [671/1893] bpb=1.100993 time=723.5s + ttt_chunk [681/1893] bpb=1.101258 time=734.3s + ttt_chunk [691/1893] bpb=1.101954 time=745.1s + ttt_chunk [701/1893] bpb=1.101187 time=755.9s + ttt_chunk [711/1893] bpb=1.101778 time=766.7s + ttt_chunk [721/1893] bpb=1.101410 time=777.5s + ttt_chunk [731/1893] bpb=1.101858 time=788.2s + ttt_chunk [741/1893] bpb=1.101812 time=799.0s + ttt_chunk [751/1893] bpb=1.101471 time=809.8s + ttt_chunk [761/1893] bpb=1.101283 time=820.6s + ttt_chunk [771/1893] bpb=1.101051 time=831.4s + ttt_chunk [781/1893] bpb=1.101576 time=842.2s + ttt_chunk [791/1893] bpb=1.101268 time=853.0s + ttt_chunk [801/1893] bpb=1.101383 time=863.8s + ttt_chunk [811/1893] bpb=1.100873 time=874.5s + ttt_chunk [821/1893] bpb=1.100672 time=885.3s + ttt_chunk [831/1893] bpb=1.100249 time=896.1s + ttt_chunk [841/1893] bpb=1.099666 time=906.9s + ttt_chunk [851/1893] bpb=1.099564 time=917.7s + ttt_chunk [861/1893] bpb=1.099674 time=928.5s + ttt_chunk [871/1893] bpb=1.099722 time=939.2s + ttt_chunk [881/1893] bpb=1.099752 time=950.0s + ttt_chunk [891/1893] bpb=1.099572 time=960.8s + ttt_chunk [901/1893] bpb=1.099562 time=971.6s + ttt_chunk [911/1893] bpb=1.099561 time=982.4s + ttt_chunk [921/1893] bpb=1.099952 time=993.2s + ttt_chunk [931/1893] bpb=1.099787 time=1004.0s + ttt_chunk [941/1893] bpb=1.099706 time=1014.7s + ttt_chunk [951/1893] bpb=1.099700 time=1025.5s + ttt_chunk [961/1893] bpb=1.099438 time=1036.3s + ttt_chunk [971/1893] bpb=1.100166 time=1047.1s + ttt_chunk [981/1893] bpb=1.100310 time=1057.9s + ttt_chunk [991/1893] bpb=1.100222 time=1068.7s + ttt_chunk [1001/1893] bpb=1.100375 time=1079.5s + ttt_chunk [1011/1893] bpb=1.100636 time=1090.2s + ttt_chunk [1021/1893] bpb=1.100805 time=1101.0s + ttt_chunk [1031/1893] bpb=1.101354 time=1111.8s + ttt_chunk [1041/1893] bpb=1.101015 time=1122.6s + ttt_chunk [1051/1893] bpb=1.100751 time=1133.4s + ttt_chunk [1061/1893] bpb=1.101009 time=1144.2s + ttt_chunk [1071/1893] bpb=1.101477 time=1155.0s + ttt_chunk [1081/1893] bpb=1.101501 time=1165.7s + ttt_chunk [1091/1893] bpb=1.101917 time=1176.5s + ttt_chunk [1101/1893] bpb=1.102051 time=1187.3s + ttt_chunk [1111/1893] bpb=1.101826 time=1198.1s + ttt_chunk [1121/1893] bpb=1.101755 time=1208.9s + ttt_chunk [1131/1893] bpb=1.101655 time=1219.7s + ttt_chunk [1141/1893] bpb=1.101510 time=1230.5s + ttt_chunk [1151/1893] bpb=1.101547 time=1241.3s + ttt_chunk [1161/1893] bpb=1.100839 time=1252.0s + ttt_chunk [1171/1893] bpb=1.101410 time=1262.8s + ttt_chunk [1181/1893] bpb=1.100911 time=1273.6s + ttt_chunk [1191/1893] bpb=1.100623 time=1284.4s + ttt_chunk [1201/1893] bpb=1.101185 time=1295.2s + ttt_chunk [1211/1893] bpb=1.100594 time=1306.0s + ttt_chunk [1221/1893] bpb=1.100288 time=1316.7s + ttt_chunk [1231/1893] bpb=1.100210 time=1327.5s + ttt_chunk [1241/1893] bpb=1.100019 time=1338.3s + ttt_chunk [1251/1893] bpb=1.099770 time=1349.1s + ttt_chunk [1261/1893] bpb=1.099713 time=1359.9s + ttt_chunk [1271/1893] bpb=1.099513 time=1370.7s + ttt_chunk [1281/1893] bpb=1.099319 time=1381.5s + ttt_chunk [1291/1893] bpb=1.099143 time=1392.2s + ttt_chunk [1301/1893] bpb=1.098779 time=1403.0s + ttt_chunk [1311/1893] bpb=1.098463 time=1413.8s + ttt_chunk [1321/1893] bpb=1.098280 time=1424.6s + ttt_chunk [1331/1893] bpb=1.098197 time=1435.4s + ttt_chunk [1341/1893] bpb=1.098087 time=1446.2s + ttt_chunk [1351/1893] bpb=1.098035 time=1456.9s + ttt_chunk [1361/1893] bpb=1.098243 time=1467.7s + ttt_chunk [1371/1893] bpb=1.098120 time=1478.5s + ttt_chunk [1381/1893] bpb=1.098055 time=1489.3s + ttt_chunk [1391/1893] bpb=1.097532 time=1500.1s + ttt_chunk [1401/1893] bpb=1.097552 time=1510.9s + ttt_chunk [1411/1893] bpb=1.097576 time=1521.7s + ttt_chunk [1421/1893] bpb=1.097828 time=1532.4s + ttt_chunk [1431/1893] bpb=1.097672 time=1543.2s + ttt_chunk [1441/1893] bpb=1.098339 time=1554.0s + ttt_chunk [1451/1893] bpb=1.098463 time=1564.8s + ttt_chunk [1461/1893] bpb=1.098206 time=1575.6s + ttt_chunk [1471/1893] bpb=1.099115 time=1586.4s + ttt_chunk [1481/1893] bpb=1.098922 time=1597.2s + ttt_chunk [1491/1893] bpb=1.098927 time=1608.0s + ttt_chunk [1501/1893] bpb=1.099079 time=1618.8s + ttt_chunk [1511/1893] bpb=1.099159 time=1629.6s + ttt_chunk [1521/1893] bpb=1.099190 time=1640.4s + ttt_chunk [1531/1893] bpb=1.099007 time=1651.1s + ttt_chunk [1541/1893] bpb=1.098972 time=1661.9s + ttt_chunk [1551/1893] bpb=1.099301 time=1672.7s + ttt_chunk [1561/1893] bpb=1.099433 time=1683.5s + ttt_chunk [1571/1893] bpb=1.099557 time=1694.3s + ttt_chunk [1581/1893] bpb=1.099678 time=1705.1s + ttt_chunk [1591/1893] bpb=1.099632 time=1715.9s + ttt_chunk [1601/1893] bpb=1.099808 time=1726.7s + ttt_chunk [1611/1893] bpb=1.099882 time=1737.5s + ttt_chunk [1621/1893] bpb=1.099745 time=1748.3s + ttt_chunk [1631/1893] bpb=1.099914 time=1759.1s + ttt_chunk [1641/1893] bpb=1.099799 time=1769.9s + ttt_chunk [1651/1893] bpb=1.099707 time=1780.7s + ttt_chunk [1661/1893] bpb=1.099601 time=1791.5s + ttt_chunk [1671/1893] bpb=1.099960 time=1802.3s + ttt_chunk [1681/1893] bpb=1.100212 time=1813.0s + ttt_chunk [1691/1893] bpb=1.100183 time=1823.8s + ttt_chunk [1701/1893] bpb=1.100186 time=1834.6s + ttt_chunk [1711/1893] bpb=1.100023 time=1845.4s + ttt_chunk [1721/1893] bpb=1.099882 time=1856.2s + ttt_chunk [1731/1893] bpb=1.099831 time=1867.0s + ttt_chunk [1741/1893] bpb=1.099617 time=1877.8s + ttt_chunk [1751/1893] bpb=1.099467 time=1888.6s + ttt_chunk [1761/1893] bpb=1.099542 time=1899.4s + ttt_chunk [1771/1893] bpb=1.099472 time=1910.2s + ttt_chunk [1781/1893] bpb=1.099430 time=1921.0s + ttt_chunk [1791/1893] bpb=1.099046 time=1931.7s + ttt_chunk [1801/1893] bpb=1.099051 time=1942.5s + ttt_chunk [1811/1893] bpb=1.098878 time=1953.3s + ttt_chunk [1821/1893] bpb=1.098909 time=1964.1s + ttt_chunk [1831/1893] bpb=1.098502 time=1974.9s + ttt_chunk [1841/1893] bpb=1.098474 time=1985.7s + ttt_chunk [1851/1893] bpb=1.098256 time=1996.5s + ttt_chunk [1861/1893] bpb=1.097788 time=2007.3s + ttt_chunk [1871/1893] bpb=1.097634 time=2018.1s + ttt_chunk [1881/1893] bpb=1.097253 time=2028.8s + ttt_chunk [1891/1893] bpb=1.097097 time=2039.6s + ttt_chunk [1893/1893] bpb=1.097114 time=2041.4s +ttt_sliding:done val_loss=1.850946 val_bpb=1.096237 elapsed=2041.4s +legal_ttt val_loss:1.8509 val_bpb:1.0962 eval_time:2042089ms +legal_ttt_exact val_loss:1.85094589 val_bpb:1.09623731 diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed42.txt b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed42.txt new file mode 100644 index 0000000000..3c69276b30 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_step20k_seed42.txt @@ -0,0 +1,2241 @@ +logs/run_recordsota_nonrecord_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9293 val_bpb:4.1039 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9305 train_time:851ms step_avg:850.55ms +step:2/20000 train_loss:8.6548 train_time:1395ms step_avg:697.55ms +step:3/20000 train_loss:7.6769 train_time:2247ms step_avg:749.06ms +step:4/20000 train_loss:7.1975 train_time:3099ms step_avg:774.78ms +step:5/20000 train_loss:7.1772 train_time:3951ms step_avg:790.29ms +step:6/20000 train_loss:7.1087 train_time:4812ms step_avg:801.97ms +step:7/20000 train_loss:7.1576 train_time:5664ms step_avg:809.18ms +step:8/20000 train_loss:6.9941 train_time:6516ms step_avg:814.55ms +step:9/20000 train_loss:6.6553 train_time:7368ms step_avg:818.72ms +step:10/20000 train_loss:6.2949 train_time:8221ms step_avg:822.10ms +step:500/20000 train_loss:2.3555 train_time:414338ms step_avg:828.68ms +step:1000/20000 train_loss:2.2641 train_time:828509ms step_avg:828.51ms +step:1500/20000 train_loss:2.1807 train_time:1243119ms step_avg:828.75ms +step:2000/20000 train_loss:2.0310 train_time:1657855ms step_avg:828.93ms +step:2500/20000 train_loss:2.1602 train_time:2071904ms step_avg:828.76ms +step:3000/20000 train_loss:2.1249 train_time:2486661ms step_avg:828.89ms +step:3500/20000 train_loss:2.1866 train_time:2901085ms step_avg:828.88ms +step:4000/20000 train_loss:2.1438 train_time:3315094ms step_avg:828.77ms +step:4000/20000 val_loss:2.0753 val_bpb:1.2291 train_time:3315444ms step_avg:828.86ms +step:4500/20000 train_loss:2.1118 train_time:3729527ms step_avg:828.78ms +step:5000/20000 train_loss:2.0683 train_time:4143751ms step_avg:828.75ms +step:5500/20000 train_loss:2.1134 train_time:4558181ms step_avg:828.76ms +step:6000/20000 train_loss:2.0092 train_time:4972372ms step_avg:828.73ms +step:6500/20000 train_loss:2.2744 train_time:5386385ms step_avg:828.67ms +step:7000/20000 train_loss:1.9078 train_time:5800678ms step_avg:828.67ms +step:7500/20000 train_loss:2.0298 train_time:6214909ms step_avg:828.65ms +step:8000/20000 train_loss:1.9822 train_time:6629167ms step_avg:828.65ms +step:8000/20000 val_loss:2.0339 val_bpb:1.2046 train_time:6629524ms step_avg:828.69ms +step:8500/20000 train_loss:1.9647 train_time:7043393ms step_avg:828.63ms +step:9000/20000 train_loss:2.0911 train_time:7457618ms step_avg:828.62ms +step:9500/20000 train_loss:2.2465 train_time:7872112ms step_avg:828.64ms +step:10000/20000 train_loss:2.0596 train_time:8286348ms step_avg:828.63ms +step:10500/20000 train_loss:2.0969 train_time:8700627ms step_avg:828.63ms +step:11000/20000 train_loss:1.9872 train_time:9114964ms step_avg:828.63ms +step:11500/20000 train_loss:1.9445 train_time:9529091ms step_avg:828.62ms +step:12000/20000 train_loss:1.9632 train_time:9943448ms step_avg:828.62ms +step:12000/20000 val_loss:2.0250 val_bpb:1.1993 train_time:9943791ms step_avg:828.65ms +step:12500/20000 train_loss:1.9705 train_time:10357662ms step_avg:828.61ms +step:13000/20000 train_loss:1.9775 train_time:10771664ms step_avg:828.59ms +step:13500/20000 train_loss:2.0946 train_time:11185822ms step_avg:828.58ms +step:14000/20000 train_loss:1.8907 train_time:11599963ms step_avg:828.57ms +step:14500/20000 train_loss:2.0646 train_time:12013997ms step_avg:828.55ms +step:15000/20000 train_loss:2.0088 train_time:12428064ms step_avg:828.54ms +step:15500/20000 train_loss:1.9760 train_time:12842082ms step_avg:828.52ms +step:16000/20000 train_loss:2.0462 train_time:13256277ms step_avg:828.52ms +step:16000/20000 val_loss:1.9703 val_bpb:1.1669 train_time:13256627ms step_avg:828.54ms +step:16500/20000 train_loss:2.0226 train_time:13670137ms step_avg:828.49ms +step:17000/20000 train_loss:2.0200 train_time:14084149ms step_avg:828.48ms +step:17500/20000 train_loss:1.9948 train_time:14498193ms step_avg:828.47ms +step:18000/20000 train_loss:1.8682 train_time:14912122ms step_avg:828.45ms +swa:start step:18450 +step:18500/20000 train_loss:1.8336 train_time:15326272ms step_avg:828.45ms +late_qat:enabled step:18831 scale:0.1499 +step:19000/20000 train_loss:1.8044 train_time:15740755ms step_avg:828.46ms +step:19500/20000 train_loss:1.9852 train_time:16155668ms step_avg:828.50ms +step:20000/20000 train_loss:1.8751 train_time:16569987ms step_avg:828.50ms +step:20000/20000 val_loss:1.8757 val_bpb:1.1109 train_time:16570369ms step_avg:828.52ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8746 val_bpb:1.1103 eval_time:19693ms +Serialized model: 106027446 bytes +Code size: 90361 bytes +Serialized model int6+lzma: 15046784 bytes +Total submission size int6+lzma: 15137145 bytes +final_int6_roundtrip val_loss:1.9002 val_bpb:1.1254 eval_time:24369ms +final_int6_roundtrip_exact val_loss:1.90023350 val_bpb:1.12542531 +final_int6_sliding_window val_loss:1.8607 val_bpb:1.1020 stride:64 eval_time:664078ms +final_int6_sliding_window_exact val_loss:1.86067988 val_bpb:1.10200235 +final_int8_zlib_roundtrip_exact val_loss:1.86067988 val_bpb:1.10200235 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.149221 time=1.1s + ttt_chunk [11/1893] bpb=1.105427 time=11.9s + ttt_chunk [21/1893] bpb=1.108997 time=22.7s + ttt_chunk [31/1893] bpb=1.112744 time=33.4s + ttt_chunk [41/1893] bpb=1.101277 time=44.2s + ttt_chunk [51/1893] bpb=1.097692 time=55.0s + ttt_chunk [61/1893] bpb=1.102242 time=65.7s + ttt_chunk [71/1893] bpb=1.099299 time=76.5s + ttt_chunk [81/1893] bpb=1.099418 time=87.3s + ttt_chunk [91/1893] bpb=1.098337 time=98.0s + ttt_chunk [101/1893] bpb=1.101032 time=108.8s + ttt_chunk [111/1893] bpb=1.102383 time=119.6s + ttt_chunk [121/1893] bpb=1.099256 time=130.4s + ttt_chunk [131/1893] bpb=1.098961 time=141.1s + ttt_chunk [141/1893] bpb=1.098637 time=151.9s + ttt_chunk [151/1893] bpb=1.101853 time=162.7s + ttt_chunk [161/1893] bpb=1.103510 time=173.4s + ttt_chunk [171/1893] bpb=1.104085 time=184.2s + ttt_chunk [181/1893] bpb=1.104153 time=195.0s + ttt_chunk [191/1893] bpb=1.107342 time=205.7s + ttt_chunk [201/1893] bpb=1.107569 time=216.5s + ttt_chunk [211/1893] bpb=1.105080 time=227.3s + ttt_chunk [221/1893] bpb=1.107111 time=238.1s + ttt_chunk [231/1893] bpb=1.106532 time=248.8s + ttt_chunk [241/1893] bpb=1.106343 time=259.6s + ttt_chunk [251/1893] bpb=1.104851 time=270.4s + ttt_chunk [261/1893] bpb=1.103214 time=281.1s + ttt_chunk [271/1893] bpb=1.101864 time=291.9s + ttt_chunk [281/1893] bpb=1.104208 time=302.7s + ttt_chunk [291/1893] bpb=1.104921 time=313.4s + ttt_chunk [301/1893] bpb=1.105511 time=324.2s + ttt_chunk [311/1893] bpb=1.107095 time=335.0s + ttt_chunk [321/1893] bpb=1.108590 time=345.7s + ttt_chunk [331/1893] bpb=1.108461 time=356.5s + ttt_chunk [341/1893] bpb=1.108724 time=367.3s + ttt_chunk [351/1893] bpb=1.110014 time=378.0s + ttt_chunk [361/1893] bpb=1.111235 time=388.8s + ttt_chunk [371/1893] bpb=1.110810 time=399.6s + ttt_chunk [381/1893] bpb=1.110775 time=410.3s + ttt_chunk [391/1893] bpb=1.110407 time=421.1s + ttt_chunk [401/1893] bpb=1.109049 time=431.9s + ttt_chunk [411/1893] bpb=1.107960 time=442.6s + ttt_chunk [421/1893] bpb=1.107430 time=453.4s + ttt_chunk [431/1893] bpb=1.108117 time=464.2s + ttt_chunk [441/1893] bpb=1.107977 time=474.9s + ttt_chunk [451/1893] bpb=1.107702 time=485.7s + ttt_chunk [461/1893] bpb=1.106981 time=496.5s + ttt_chunk [471/1893] bpb=1.106584 time=507.2s + ttt_chunk [481/1893] bpb=1.106271 time=518.0s + ttt_chunk [491/1893] bpb=1.105997 time=528.8s + ttt_chunk [501/1893] bpb=1.105439 time=539.6s + ttt_chunk [511/1893] bpb=1.104831 time=550.3s + ttt_chunk [521/1893] bpb=1.103810 time=561.2s + ttt_chunk [531/1893] bpb=1.103933 time=571.9s + ttt_chunk [541/1893] bpb=1.103845 time=582.7s + ttt_chunk [551/1893] bpb=1.102632 time=593.5s + ttt_chunk [561/1893] bpb=1.103166 time=604.2s + ttt_chunk [571/1893] bpb=1.102478 time=615.0s + ttt_chunk [581/1893] bpb=1.101930 time=625.8s + ttt_chunk [591/1893] bpb=1.101262 time=636.5s + ttt_chunk [601/1893] bpb=1.101952 time=647.3s + ttt_chunk [611/1893] bpb=1.101538 time=658.1s + ttt_chunk [621/1893] bpb=1.101453 time=668.9s + ttt_chunk [631/1893] bpb=1.101855 time=679.6s + ttt_chunk [641/1893] bpb=1.101660 time=690.4s + ttt_chunk [651/1893] bpb=1.101678 time=701.2s + ttt_chunk [661/1893] bpb=1.101577 time=711.9s + ttt_chunk [671/1893] bpb=1.101226 time=722.7s + ttt_chunk [681/1893] bpb=1.101501 time=733.5s + ttt_chunk [691/1893] bpb=1.102163 time=744.2s + ttt_chunk [701/1893] bpb=1.101383 time=755.0s + ttt_chunk [711/1893] bpb=1.101968 time=765.8s + ttt_chunk [721/1893] bpb=1.101606 time=776.6s + ttt_chunk [731/1893] bpb=1.102036 time=787.3s + ttt_chunk [741/1893] bpb=1.101977 time=798.1s + ttt_chunk [751/1893] bpb=1.101623 time=808.9s + ttt_chunk [761/1893] bpb=1.101455 time=819.6s + ttt_chunk [771/1893] bpb=1.101210 time=830.4s + ttt_chunk [781/1893] bpb=1.101740 time=841.2s + ttt_chunk [791/1893] bpb=1.101420 time=851.9s + ttt_chunk [801/1893] bpb=1.101479 time=862.7s + ttt_chunk [811/1893] bpb=1.100996 time=873.5s + ttt_chunk [821/1893] bpb=1.100793 time=884.3s + ttt_chunk [831/1893] bpb=1.100359 time=895.0s + ttt_chunk [841/1893] bpb=1.099773 time=905.8s + ttt_chunk [851/1893] bpb=1.099667 time=916.6s + ttt_chunk [861/1893] bpb=1.099807 time=927.3s + ttt_chunk [871/1893] bpb=1.099860 time=938.1s + ttt_chunk [881/1893] bpb=1.099862 time=948.9s + ttt_chunk [891/1893] bpb=1.099699 time=959.6s + ttt_chunk [901/1893] bpb=1.099687 time=970.4s + ttt_chunk [911/1893] bpb=1.099686 time=981.2s + ttt_chunk [921/1893] bpb=1.100066 time=991.9s + ttt_chunk [931/1893] bpb=1.099886 time=1002.7s + ttt_chunk [941/1893] bpb=1.099777 time=1013.5s + ttt_chunk [951/1893] bpb=1.099765 time=1024.2s + ttt_chunk [961/1893] bpb=1.099516 time=1035.0s + ttt_chunk [971/1893] bpb=1.100237 time=1045.8s + ttt_chunk [981/1893] bpb=1.100402 time=1056.5s + ttt_chunk [991/1893] bpb=1.100302 time=1067.3s + ttt_chunk [1001/1893] bpb=1.100447 time=1078.1s + ttt_chunk [1011/1893] bpb=1.100704 time=1088.9s + ttt_chunk [1021/1893] bpb=1.100855 time=1099.6s + ttt_chunk [1031/1893] bpb=1.101407 time=1110.4s + ttt_chunk [1041/1893] bpb=1.101071 time=1121.2s + ttt_chunk [1051/1893] bpb=1.100782 time=1131.9s + ttt_chunk [1061/1893] bpb=1.101022 time=1142.7s + ttt_chunk [1071/1893] bpb=1.101493 time=1153.4s + ttt_chunk [1081/1893] bpb=1.101509 time=1164.2s + ttt_chunk [1091/1893] bpb=1.101914 time=1175.0s + ttt_chunk [1101/1893] bpb=1.102047 time=1185.8s + ttt_chunk [1111/1893] bpb=1.101822 time=1196.5s + ttt_chunk [1121/1893] bpb=1.101747 time=1207.3s + ttt_chunk [1131/1893] bpb=1.101647 time=1218.1s + ttt_chunk [1141/1893] bpb=1.101496 time=1228.8s + ttt_chunk [1151/1893] bpb=1.101538 time=1239.6s + ttt_chunk [1161/1893] bpb=1.100850 time=1250.4s + ttt_chunk [1171/1893] bpb=1.101414 time=1261.1s + ttt_chunk [1181/1893] bpb=1.100908 time=1271.9s + ttt_chunk [1191/1893] bpb=1.100618 time=1282.7s + ttt_chunk [1201/1893] bpb=1.101200 time=1293.4s + ttt_chunk [1211/1893] bpb=1.100606 time=1304.2s + ttt_chunk [1221/1893] bpb=1.100309 time=1315.0s + ttt_chunk [1231/1893] bpb=1.100236 time=1325.8s + ttt_chunk [1241/1893] bpb=1.100026 time=1336.5s + ttt_chunk [1251/1893] bpb=1.099790 time=1347.3s + ttt_chunk [1261/1893] bpb=1.099730 time=1358.1s + ttt_chunk [1271/1893] bpb=1.099522 time=1368.8s + ttt_chunk [1281/1893] bpb=1.099310 time=1379.6s + ttt_chunk [1291/1893] bpb=1.099133 time=1390.3s + ttt_chunk [1301/1893] bpb=1.098747 time=1401.1s + ttt_chunk [1311/1893] bpb=1.098432 time=1411.9s + ttt_chunk [1321/1893] bpb=1.098267 time=1422.6s + ttt_chunk [1331/1893] bpb=1.098192 time=1433.4s + ttt_chunk [1341/1893] bpb=1.098083 time=1444.2s + ttt_chunk [1351/1893] bpb=1.098029 time=1455.0s + ttt_chunk [1361/1893] bpb=1.098231 time=1465.7s + ttt_chunk [1371/1893] bpb=1.098089 time=1476.5s + ttt_chunk [1381/1893] bpb=1.098019 time=1487.3s + ttt_chunk [1391/1893] bpb=1.097495 time=1498.0s + ttt_chunk [1401/1893] bpb=1.097514 time=1508.8s + ttt_chunk [1411/1893] bpb=1.097548 time=1519.6s + ttt_chunk [1421/1893] bpb=1.097802 time=1530.3s + ttt_chunk [1431/1893] bpb=1.097657 time=1541.1s + ttt_chunk [1441/1893] bpb=1.098321 time=1551.9s + ttt_chunk [1451/1893] bpb=1.098447 time=1562.6s + ttt_chunk [1461/1893] bpb=1.098192 time=1573.4s + ttt_chunk [1471/1893] bpb=1.099100 time=1584.2s + ttt_chunk [1481/1893] bpb=1.098893 time=1594.9s + ttt_chunk [1491/1893] bpb=1.098882 time=1605.7s + ttt_chunk [1501/1893] bpb=1.099047 time=1616.5s + ttt_chunk [1511/1893] bpb=1.099129 time=1627.2s + ttt_chunk [1521/1893] bpb=1.099154 time=1638.0s + ttt_chunk [1531/1893] bpb=1.098978 time=1648.8s + ttt_chunk [1541/1893] bpb=1.098944 time=1659.6s + ttt_chunk [1551/1893] bpb=1.099287 time=1670.3s + ttt_chunk [1561/1893] bpb=1.099428 time=1681.1s + ttt_chunk [1571/1893] bpb=1.099561 time=1691.9s + ttt_chunk [1581/1893] bpb=1.099680 time=1702.6s + ttt_chunk [1591/1893] bpb=1.099632 time=1713.4s + ttt_chunk [1601/1893] bpb=1.099814 time=1724.2s + ttt_chunk [1611/1893] bpb=1.099890 time=1734.9s + ttt_chunk [1621/1893] bpb=1.099746 time=1745.7s + ttt_chunk [1631/1893] bpb=1.099923 time=1756.5s + ttt_chunk [1641/1893] bpb=1.099803 time=1767.2s + ttt_chunk [1651/1893] bpb=1.099721 time=1778.0s + ttt_chunk [1661/1893] bpb=1.099604 time=1788.8s + ttt_chunk [1671/1893] bpb=1.099950 time=1799.6s + ttt_chunk [1681/1893] bpb=1.100198 time=1810.3s + ttt_chunk [1691/1893] bpb=1.100167 time=1821.1s + ttt_chunk [1701/1893] bpb=1.100167 time=1831.9s + ttt_chunk [1711/1893] bpb=1.100005 time=1842.6s + ttt_chunk [1721/1893] bpb=1.099862 time=1853.4s + ttt_chunk [1731/1893] bpb=1.099816 time=1864.2s + ttt_chunk [1741/1893] bpb=1.099595 time=1874.9s + ttt_chunk [1751/1893] bpb=1.099450 time=1885.7s + ttt_chunk [1761/1893] bpb=1.099527 time=1896.5s + ttt_chunk [1771/1893] bpb=1.099457 time=1907.2s + ttt_chunk [1781/1893] bpb=1.099414 time=1918.0s + ttt_chunk [1791/1893] bpb=1.099023 time=1928.8s + ttt_chunk [1801/1893] bpb=1.099031 time=1939.5s + ttt_chunk [1811/1893] bpb=1.098867 time=1950.3s + ttt_chunk [1821/1893] bpb=1.098905 time=1961.1s + ttt_chunk [1831/1893] bpb=1.098518 time=1971.9s + ttt_chunk [1841/1893] bpb=1.098497 time=1982.6s + ttt_chunk [1851/1893] bpb=1.098282 time=1993.4s + ttt_chunk [1861/1893] bpb=1.097807 time=2004.2s + ttt_chunk [1871/1893] bpb=1.097656 time=2014.9s + ttt_chunk [1881/1893] bpb=1.097279 time=2025.7s + ttt_chunk [1891/1893] bpb=1.097109 time=2036.5s + ttt_chunk [1893/1893] bpb=1.097126 time=2038.2s +ttt_sliding:done val_loss=1.850831 val_bpb=1.096169 elapsed=2038.3s +legal_ttt val_loss:1.8508 val_bpb:1.0962 eval_time:2038958ms +legal_ttt_exact val_loss:1.85083110 val_bpb:1.09616933 +batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +# --- Data loading --- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.10.14 (main, May 6 2024, 19:42:50) [GCC 11.2.0] +Running PyTorch 2.10.0+cu128 +Fri Mar 27 09:32:41 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 550.127.08 Driver Version: 550.127.08 CUDA Version: 12.4 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA A100-SXM4-80GB On | 00000000:27:00.0 Off | On | +| N/A 43C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA A100-SXM4-80GB On | 00000000:2A:00.0 Off | On | +| N/A 36C P0 91W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA A100-SXM4-80GB On | 00000000:51:00.0 Off | On | +| N/A 36C P0 103W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA A100-SXM4-80GB On | 00000000:9E:00.0 Off | On | +| N/A 40C P0 100W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA A100-SXM4-80GB On | 00000000:A4:00.0 Off | On | +| N/A 32C P0 95W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA A100-SXM4-80GB On | 00000000:CA:00.0 Off | On | +| N/A 37C P0 97W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| MIG devices: | ++------------------+----------------------------------+-----------+-----------------------+ +| GPU GI CI MIG | Memory-Usage | Vol| Shared | +| ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG | +| | | ECC| | +|==================+==================================+===========+=======================| +| 0 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 2 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 3 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 2 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 1 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 3538291 C /opt/conda/bin/python3 0MiB | +| 1 N/A N/A 3538292 C /opt/conda/bin/python3 0MiB | +| 2 N/A N/A 3538293 C /opt/conda/bin/python3 0MiB | +| 3 N/A N/A 3538294 C /opt/conda/bin/python3 0MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9293 val_bpb:4.1039 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9305 train_time:851ms step_avg:850.55ms +step:2/20000 train_loss:8.6548 train_time:1395ms step_avg:697.55ms +step:3/20000 train_loss:7.6769 train_time:2247ms step_avg:749.06ms +step:4/20000 train_loss:7.1975 train_time:3099ms step_avg:774.78ms +step:5/20000 train_loss:7.1772 train_time:3951ms step_avg:790.29ms +step:6/20000 train_loss:7.1087 train_time:4812ms step_avg:801.97ms +step:7/20000 train_loss:7.1576 train_time:5664ms step_avg:809.18ms +step:8/20000 train_loss:6.9941 train_time:6516ms step_avg:814.55ms +step:9/20000 train_loss:6.6553 train_time:7368ms step_avg:818.72ms +step:10/20000 train_loss:6.2949 train_time:8221ms step_avg:822.10ms +step:500/20000 train_loss:2.3555 train_time:414338ms step_avg:828.68ms +step:1000/20000 train_loss:2.2641 train_time:828509ms step_avg:828.51ms +step:1500/20000 train_loss:2.1807 train_time:1243119ms step_avg:828.75ms +step:2000/20000 train_loss:2.0310 train_time:1657855ms step_avg:828.93ms +step:2500/20000 train_loss:2.1602 train_time:2071904ms step_avg:828.76ms +step:3000/20000 train_loss:2.1249 train_time:2486661ms step_avg:828.89ms +step:3500/20000 train_loss:2.1866 train_time:2901085ms step_avg:828.88ms +step:4000/20000 train_loss:2.1438 train_time:3315094ms step_avg:828.77ms +step:4000/20000 val_loss:2.0753 val_bpb:1.2291 train_time:3315444ms step_avg:828.86ms +step:4500/20000 train_loss:2.1118 train_time:3729527ms step_avg:828.78ms +step:5000/20000 train_loss:2.0683 train_time:4143751ms step_avg:828.75ms +step:5500/20000 train_loss:2.1134 train_time:4558181ms step_avg:828.76ms +step:6000/20000 train_loss:2.0092 train_time:4972372ms step_avg:828.73ms +step:6500/20000 train_loss:2.2744 train_time:5386385ms step_avg:828.67ms +step:7000/20000 train_loss:1.9078 train_time:5800678ms step_avg:828.67ms +step:7500/20000 train_loss:2.0298 train_time:6214909ms step_avg:828.65ms +step:8000/20000 train_loss:1.9822 train_time:6629167ms step_avg:828.65ms +step:8000/20000 val_loss:2.0339 val_bpb:1.2046 train_time:6629524ms step_avg:828.69ms +step:8500/20000 train_loss:1.9647 train_time:7043393ms step_avg:828.63ms +step:9000/20000 train_loss:2.0911 train_time:7457618ms step_avg:828.62ms +step:9500/20000 train_loss:2.2465 train_time:7872112ms step_avg:828.64ms +step:10000/20000 train_loss:2.0596 train_time:8286348ms step_avg:828.63ms +step:10500/20000 train_loss:2.0969 train_time:8700627ms step_avg:828.63ms +step:11000/20000 train_loss:1.9872 train_time:9114964ms step_avg:828.63ms +step:11500/20000 train_loss:1.9445 train_time:9529091ms step_avg:828.62ms +step:12000/20000 train_loss:1.9632 train_time:9943448ms step_avg:828.62ms +step:12000/20000 val_loss:2.0250 val_bpb:1.1993 train_time:9943791ms step_avg:828.65ms +step:12500/20000 train_loss:1.9705 train_time:10357662ms step_avg:828.61ms +step:13000/20000 train_loss:1.9775 train_time:10771664ms step_avg:828.59ms +step:13500/20000 train_loss:2.0946 train_time:11185822ms step_avg:828.58ms +step:14000/20000 train_loss:1.8907 train_time:11599963ms step_avg:828.57ms +step:14500/20000 train_loss:2.0646 train_time:12013997ms step_avg:828.55ms +step:15000/20000 train_loss:2.0088 train_time:12428064ms step_avg:828.54ms +step:15500/20000 train_loss:1.9760 train_time:12842082ms step_avg:828.52ms +step:16000/20000 train_loss:2.0462 train_time:13256277ms step_avg:828.52ms +step:16000/20000 val_loss:1.9703 val_bpb:1.1669 train_time:13256627ms step_avg:828.54ms +step:16500/20000 train_loss:2.0226 train_time:13670137ms step_avg:828.49ms +step:17000/20000 train_loss:2.0200 train_time:14084149ms step_avg:828.48ms +step:17500/20000 train_loss:1.9948 train_time:14498193ms step_avg:828.47ms +step:18000/20000 train_loss:1.8682 train_time:14912122ms step_avg:828.45ms +swa:start step:18450 +step:18500/20000 train_loss:1.8336 train_time:15326272ms step_avg:828.45ms +late_qat:enabled step:18831 scale:0.1499 +step:19000/20000 train_loss:1.8044 train_time:15740755ms step_avg:828.46ms +step:19500/20000 train_loss:1.9852 train_time:16155668ms step_avg:828.50ms +step:20000/20000 train_loss:1.8751 train_time:16569987ms step_avg:828.50ms +step:20000/20000 val_loss:1.8757 val_bpb:1.1109 train_time:16570369ms step_avg:828.52ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8746 val_bpb:1.1103 eval_time:19693ms +Serialized model: 106027446 bytes +Code size: 90361 bytes +Serialized model int6+lzma: 15046784 bytes +Total submission size int6+lzma: 15137145 bytes +final_int6_roundtrip val_loss:1.9002 val_bpb:1.1254 eval_time:24369ms +final_int6_roundtrip_exact val_loss:1.90023350 val_bpb:1.12542531 +final_int6_sliding_window val_loss:1.8607 val_bpb:1.1020 stride:64 eval_time:664078ms +final_int6_sliding_window_exact val_loss:1.86067988 val_bpb:1.10200235 +final_int8_zlib_roundtrip_exact val_loss:1.86067988 val_bpb:1.10200235 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.149221 time=1.1s + ttt_chunk [11/1893] bpb=1.105427 time=11.9s + ttt_chunk [21/1893] bpb=1.108997 time=22.7s + ttt_chunk [31/1893] bpb=1.112744 time=33.4s + ttt_chunk [41/1893] bpb=1.101277 time=44.2s + ttt_chunk [51/1893] bpb=1.097692 time=55.0s + ttt_chunk [61/1893] bpb=1.102242 time=65.7s + ttt_chunk [71/1893] bpb=1.099299 time=76.5s + ttt_chunk [81/1893] bpb=1.099418 time=87.3s + ttt_chunk [91/1893] bpb=1.098337 time=98.0s + ttt_chunk [101/1893] bpb=1.101032 time=108.8s + ttt_chunk [111/1893] bpb=1.102383 time=119.6s + ttt_chunk [121/1893] bpb=1.099256 time=130.4s + ttt_chunk [131/1893] bpb=1.098961 time=141.1s + ttt_chunk [141/1893] bpb=1.098637 time=151.9s + ttt_chunk [151/1893] bpb=1.101853 time=162.7s + ttt_chunk [161/1893] bpb=1.103510 time=173.4s + ttt_chunk [171/1893] bpb=1.104085 time=184.2s + ttt_chunk [181/1893] bpb=1.104153 time=195.0s + ttt_chunk [191/1893] bpb=1.107342 time=205.7s + ttt_chunk [201/1893] bpb=1.107569 time=216.5s + ttt_chunk [211/1893] bpb=1.105080 time=227.3s + ttt_chunk [221/1893] bpb=1.107111 time=238.1s + ttt_chunk [231/1893] bpb=1.106532 time=248.8s + ttt_chunk [241/1893] bpb=1.106343 time=259.6s + ttt_chunk [251/1893] bpb=1.104851 time=270.4s + ttt_chunk [261/1893] bpb=1.103214 time=281.1s + ttt_chunk [271/1893] bpb=1.101864 time=291.9s + ttt_chunk [281/1893] bpb=1.104208 time=302.7s + ttt_chunk [291/1893] bpb=1.104921 time=313.4s + ttt_chunk [301/1893] bpb=1.105511 time=324.2s + ttt_chunk [311/1893] bpb=1.107095 time=335.0s + ttt_chunk [321/1893] bpb=1.108590 time=345.7s + ttt_chunk [331/1893] bpb=1.108461 time=356.5s + ttt_chunk [341/1893] bpb=1.108724 time=367.3s + ttt_chunk [351/1893] bpb=1.110014 time=378.0s + ttt_chunk [361/1893] bpb=1.111235 time=388.8s + ttt_chunk [371/1893] bpb=1.110810 time=399.6s + ttt_chunk [381/1893] bpb=1.110775 time=410.3s + ttt_chunk [391/1893] bpb=1.110407 time=421.1s + ttt_chunk [401/1893] bpb=1.109049 time=431.9s + ttt_chunk [411/1893] bpb=1.107960 time=442.6s + ttt_chunk [421/1893] bpb=1.107430 time=453.4s + ttt_chunk [431/1893] bpb=1.108117 time=464.2s + ttt_chunk [441/1893] bpb=1.107977 time=474.9s + ttt_chunk [451/1893] bpb=1.107702 time=485.7s + ttt_chunk [461/1893] bpb=1.106981 time=496.5s + ttt_chunk [471/1893] bpb=1.106584 time=507.2s + ttt_chunk [481/1893] bpb=1.106271 time=518.0s + ttt_chunk [491/1893] bpb=1.105997 time=528.8s + ttt_chunk [501/1893] bpb=1.105439 time=539.6s + ttt_chunk [511/1893] bpb=1.104831 time=550.3s + ttt_chunk [521/1893] bpb=1.103810 time=561.2s + ttt_chunk [531/1893] bpb=1.103933 time=571.9s + ttt_chunk [541/1893] bpb=1.103845 time=582.7s + ttt_chunk [551/1893] bpb=1.102632 time=593.5s + ttt_chunk [561/1893] bpb=1.103166 time=604.2s + ttt_chunk [571/1893] bpb=1.102478 time=615.0s + ttt_chunk [581/1893] bpb=1.101930 time=625.8s + ttt_chunk [591/1893] bpb=1.101262 time=636.5s + ttt_chunk [601/1893] bpb=1.101952 time=647.3s + ttt_chunk [611/1893] bpb=1.101538 time=658.1s + ttt_chunk [621/1893] bpb=1.101453 time=668.9s + ttt_chunk [631/1893] bpb=1.101855 time=679.6s + ttt_chunk [641/1893] bpb=1.101660 time=690.4s + ttt_chunk [651/1893] bpb=1.101678 time=701.2s + ttt_chunk [661/1893] bpb=1.101577 time=711.9s + ttt_chunk [671/1893] bpb=1.101226 time=722.7s + ttt_chunk [681/1893] bpb=1.101501 time=733.5s + ttt_chunk [691/1893] bpb=1.102163 time=744.2s + ttt_chunk [701/1893] bpb=1.101383 time=755.0s + ttt_chunk [711/1893] bpb=1.101968 time=765.8s + ttt_chunk [721/1893] bpb=1.101606 time=776.6s + ttt_chunk [731/1893] bpb=1.102036 time=787.3s + ttt_chunk [741/1893] bpb=1.101977 time=798.1s + ttt_chunk [751/1893] bpb=1.101623 time=808.9s + ttt_chunk [761/1893] bpb=1.101455 time=819.6s + ttt_chunk [771/1893] bpb=1.101210 time=830.4s + ttt_chunk [781/1893] bpb=1.101740 time=841.2s + ttt_chunk [791/1893] bpb=1.101420 time=851.9s + ttt_chunk [801/1893] bpb=1.101479 time=862.7s + ttt_chunk [811/1893] bpb=1.100996 time=873.5s + ttt_chunk [821/1893] bpb=1.100793 time=884.3s + ttt_chunk [831/1893] bpb=1.100359 time=895.0s + ttt_chunk [841/1893] bpb=1.099773 time=905.8s + ttt_chunk [851/1893] bpb=1.099667 time=916.6s + ttt_chunk [861/1893] bpb=1.099807 time=927.3s + ttt_chunk [871/1893] bpb=1.099860 time=938.1s + ttt_chunk [881/1893] bpb=1.099862 time=948.9s + ttt_chunk [891/1893] bpb=1.099699 time=959.6s + ttt_chunk [901/1893] bpb=1.099687 time=970.4s + ttt_chunk [911/1893] bpb=1.099686 time=981.2s + ttt_chunk [921/1893] bpb=1.100066 time=991.9s + ttt_chunk [931/1893] bpb=1.099886 time=1002.7s + ttt_chunk [941/1893] bpb=1.099777 time=1013.5s + ttt_chunk [951/1893] bpb=1.099765 time=1024.2s + ttt_chunk [961/1893] bpb=1.099516 time=1035.0s + ttt_chunk [971/1893] bpb=1.100237 time=1045.8s + ttt_chunk [981/1893] bpb=1.100402 time=1056.5s + ttt_chunk [991/1893] bpb=1.100302 time=1067.3s + ttt_chunk [1001/1893] bpb=1.100447 time=1078.1s + ttt_chunk [1011/1893] bpb=1.100704 time=1088.9s + ttt_chunk [1021/1893] bpb=1.100855 time=1099.6s + ttt_chunk [1031/1893] bpb=1.101407 time=1110.4s + ttt_chunk [1041/1893] bpb=1.101071 time=1121.2s + ttt_chunk [1051/1893] bpb=1.100782 time=1131.9s + ttt_chunk [1061/1893] bpb=1.101022 time=1142.7s + ttt_chunk [1071/1893] bpb=1.101493 time=1153.4s + ttt_chunk [1081/1893] bpb=1.101509 time=1164.2s + ttt_chunk [1091/1893] bpb=1.101914 time=1175.0s + ttt_chunk [1101/1893] bpb=1.102047 time=1185.8s + ttt_chunk [1111/1893] bpb=1.101822 time=1196.5s + ttt_chunk [1121/1893] bpb=1.101747 time=1207.3s + ttt_chunk [1131/1893] bpb=1.101647 time=1218.1s + ttt_chunk [1141/1893] bpb=1.101496 time=1228.8s + ttt_chunk [1151/1893] bpb=1.101538 time=1239.6s + ttt_chunk [1161/1893] bpb=1.100850 time=1250.4s + ttt_chunk [1171/1893] bpb=1.101414 time=1261.1s + ttt_chunk [1181/1893] bpb=1.100908 time=1271.9s + ttt_chunk [1191/1893] bpb=1.100618 time=1282.7s + ttt_chunk [1201/1893] bpb=1.101200 time=1293.4s + ttt_chunk [1211/1893] bpb=1.100606 time=1304.2s + ttt_chunk [1221/1893] bpb=1.100309 time=1315.0s + ttt_chunk [1231/1893] bpb=1.100236 time=1325.8s + ttt_chunk [1241/1893] bpb=1.100026 time=1336.5s + ttt_chunk [1251/1893] bpb=1.099790 time=1347.3s + ttt_chunk [1261/1893] bpb=1.099730 time=1358.1s + ttt_chunk [1271/1893] bpb=1.099522 time=1368.8s + ttt_chunk [1281/1893] bpb=1.099310 time=1379.6s + ttt_chunk [1291/1893] bpb=1.099133 time=1390.3s + ttt_chunk [1301/1893] bpb=1.098747 time=1401.1s + ttt_chunk [1311/1893] bpb=1.098432 time=1411.9s + ttt_chunk [1321/1893] bpb=1.098267 time=1422.6s + ttt_chunk [1331/1893] bpb=1.098192 time=1433.4s + ttt_chunk [1341/1893] bpb=1.098083 time=1444.2s + ttt_chunk [1351/1893] bpb=1.098029 time=1455.0s + ttt_chunk [1361/1893] bpb=1.098231 time=1465.7s + ttt_chunk [1371/1893] bpb=1.098089 time=1476.5s + ttt_chunk [1381/1893] bpb=1.098019 time=1487.3s + ttt_chunk [1391/1893] bpb=1.097495 time=1498.0s + ttt_chunk [1401/1893] bpb=1.097514 time=1508.8s + ttt_chunk [1411/1893] bpb=1.097548 time=1519.6s + ttt_chunk [1421/1893] bpb=1.097802 time=1530.3s + ttt_chunk [1431/1893] bpb=1.097657 time=1541.1s + ttt_chunk [1441/1893] bpb=1.098321 time=1551.9s + ttt_chunk [1451/1893] bpb=1.098447 time=1562.6s + ttt_chunk [1461/1893] bpb=1.098192 time=1573.4s + ttt_chunk [1471/1893] bpb=1.099100 time=1584.2s + ttt_chunk [1481/1893] bpb=1.098893 time=1594.9s + ttt_chunk [1491/1893] bpb=1.098882 time=1605.7s + ttt_chunk [1501/1893] bpb=1.099047 time=1616.5s + ttt_chunk [1511/1893] bpb=1.099129 time=1627.2s + ttt_chunk [1521/1893] bpb=1.099154 time=1638.0s + ttt_chunk [1531/1893] bpb=1.098978 time=1648.8s + ttt_chunk [1541/1893] bpb=1.098944 time=1659.6s + ttt_chunk [1551/1893] bpb=1.099287 time=1670.3s + ttt_chunk [1561/1893] bpb=1.099428 time=1681.1s + ttt_chunk [1571/1893] bpb=1.099561 time=1691.9s + ttt_chunk [1581/1893] bpb=1.099680 time=1702.6s + ttt_chunk [1591/1893] bpb=1.099632 time=1713.4s + ttt_chunk [1601/1893] bpb=1.099814 time=1724.2s + ttt_chunk [1611/1893] bpb=1.099890 time=1734.9s + ttt_chunk [1621/1893] bpb=1.099746 time=1745.7s + ttt_chunk [1631/1893] bpb=1.099923 time=1756.5s + ttt_chunk [1641/1893] bpb=1.099803 time=1767.2s + ttt_chunk [1651/1893] bpb=1.099721 time=1778.0s + ttt_chunk [1661/1893] bpb=1.099604 time=1788.8s + ttt_chunk [1671/1893] bpb=1.099950 time=1799.6s + ttt_chunk [1681/1893] bpb=1.100198 time=1810.3s + ttt_chunk [1691/1893] bpb=1.100167 time=1821.1s + ttt_chunk [1701/1893] bpb=1.100167 time=1831.9s + ttt_chunk [1711/1893] bpb=1.100005 time=1842.6s + ttt_chunk [1721/1893] bpb=1.099862 time=1853.4s + ttt_chunk [1731/1893] bpb=1.099816 time=1864.2s + ttt_chunk [1741/1893] bpb=1.099595 time=1874.9s + ttt_chunk [1751/1893] bpb=1.099450 time=1885.7s + ttt_chunk [1761/1893] bpb=1.099527 time=1896.5s + ttt_chunk [1771/1893] bpb=1.099457 time=1907.2s + ttt_chunk [1781/1893] bpb=1.099414 time=1918.0s + ttt_chunk [1791/1893] bpb=1.099023 time=1928.8s + ttt_chunk [1801/1893] bpb=1.099031 time=1939.5s + ttt_chunk [1811/1893] bpb=1.098867 time=1950.3s + ttt_chunk [1821/1893] bpb=1.098905 time=1961.1s + ttt_chunk [1831/1893] bpb=1.098518 time=1971.9s + ttt_chunk [1841/1893] bpb=1.098497 time=1982.6s + ttt_chunk [1851/1893] bpb=1.098282 time=1993.4s + ttt_chunk [1861/1893] bpb=1.097807 time=2004.2s + ttt_chunk [1871/1893] bpb=1.097656 time=2014.9s + ttt_chunk [1881/1893] bpb=1.097279 time=2025.7s + ttt_chunk [1891/1893] bpb=1.097109 time=2036.5s + ttt_chunk [1893/1893] bpb=1.097126 time=2038.2s +ttt_sliding:done val_loss=1.850831 val_bpb=1.096169 elapsed=2038.3s +legal_ttt val_loss:1.8508 val_bpb:1.0962 eval_time:2038958ms +legal_ttt_exact val_loss:1.85083110 val_bpb:1.09616933 From 0e44679960b8a7e16429bcfcf01a2a065a235427 Mon Sep 17 00:00:00 2001 From: OnlyJundong Date: Mon, 6 Apr 2026 14:49:36 -0400 Subject: [PATCH 2/3] =?UTF-8?q?Add=20non-record=20submission:=20Extended?= =?UTF-8?q?=20Compute=20Scaling=20Analysis=20(50K=20steps,=20~12h,=204?= =?UTF-8?q?=C3=97A100=20MIG)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../README.md | 188 ++ .../run_scaling_50k.sh | 36 + .../submission.json | 20 + .../train_gpt.py | 1938 ++++++++++++++ .../train_step50k_seed1337.txt | 2260 ++++++++++++++++ .../train_step50k_seed2024.txt | 2264 +++++++++++++++++ .../train_step50k_seed42.txt | 2252 ++++++++++++++++ 7 files changed, 8958 insertions(+) create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/run_scaling_50k.sh create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/submission.json create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_gpt.py create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed1337.txt create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed2024.txt create mode 100644 records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed42.txt diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md new file mode 100644 index 0000000000..1b45425ab5 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md @@ -0,0 +1,188 @@ +# Extended Compute Scaling Analysis (50K Steps, ~12 Hours) + +**val_bpb: 1.0858** (50K steps, 3-seed mean, std 0.0005) | **~14.30 MB** | 4×A100 MIG (Unlimited Compute Track) + +## Summary + +This submission is a **non-record submission**. It extends the [20K scaling run](../2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/) to 50K steps (~12 hours), using the same architecture and code from ([PR #549](https://github.com/openai/parameter-golf/pull/549) by @abaybektursun). Training runs on 4×A100 MIG instances (approximately 10× slower per step than 8×H100 SXM). + +Key findings: +- **50K steps achieves 1.0858 BPB post-TTT** (3-seed mean, std 0.0005) — improvement of **−0.0102 BPB** over the 20K run (1.0960) +- **Artifact shrinks further to ~14.30MB** as warmdown smooths weight entropy more aggressively at longer horizons +- **TTT gains scale with compute**: TTT provides −0.0089 BPB at 50K steps vs −0.0058 at 20K steps (+53% more TTT benefit) +- **Artifact size plateau at 10K–30K steps** (~17.2MB) confirms the non-monotonic pattern; warmdown recovery is steeper and more complete at 50K than at 20K + +## Results + +### 50K steps, ~12 hours (4×A100 MIG, 3-seed comparison) + +| Seed | step_avg | steps | Pre-TTT bpb | **Post-TTT bpb** | TTT gain | Artifact | +|------|----------|-------|-------------|-----------------|----------|----------| +| 1337 | 829.0ms | 50,000 | 1.0942 | **1.0853** | -0.0089 | 14,330,478 | +| 42 | 830.9ms | 50,000 | 1.0945 | **1.0857** | -0.0088 | 14,206,210 | +| 2024 | 828.7ms | 50,000 | 1.0953 | **1.0865** | -0.0088 | 14,353,974 | +| **Mean** | **829.5ms** | **50,000** | **1.0947** | **1.0858 (std 0.0005)** | **-0.0088** | **14,296,887** | + +### Comparison: 20K vs 50K + +| Metric | 20K steps (~6h) | 50K steps (~12h) | Δ | +|--------|----------------|-----------------|---| +| Post-TTT BPB | 1.0960 | 1.0858 | **−0.0102** | +| Artifact size | ~15.05 MB | ~14.30 MB | −0.75 MB | +| TTT gain | −0.0058 | −0.0088 | −0.0030 (53% more) | + +## Scaling Analysis: BPB & Artifact Size vs Training Steps + +Artifact size is computed as int6+LZMA compressed model + code bytes. Every 2,500 steps. + +### Seed 2024 (2,500-step intervals) + +| Steps | Pre-TTT val_bpb | artifact_bytes | Under 16MB? | +|------:|--------:|---------------:|:-----------:| +| 0 | 4.1037 | 4,577,366 | Yes | +| 2,500 | 1.2552 | 13,069,870 | Yes | +| 5,000 | 1.2282 | 13,892,598 | Yes | +| 7,500 | 1.2185 | 15,496,510 | Yes | +| 10,000 | 1.2027 | 17,215,490 | **No** | +| 12,500 | 1.1996 | 17,232,630 | **No** | +| 15,000 | 1.1945 | 17,237,242 | **No** | +| 17,500 | 1.1923 | 17,219,546 | **No** | +| 20,000 | 1.1912 | 17,213,066 | **No** | +| 22,500 | 1.1903 | 17,206,306 | **No** | +| 25,000 | 1.1904 | 17,201,238 | **No** | +| 27,500 | 1.1883 | 17,193,574 | **No** | +| 30,000 | 1.1865 | 17,190,822 | **No** | +| 32,500 | 1.1813 | 17,052,022 | **No** | +| 35,000 | 1.1742 | 16,786,866 | **No** | +| 37,500 | 1.1677 | 16,553,402 | **No** | +| 40,000 | 1.1592 | 16,067,974 | **No** | +| 42,500 | 1.1493 | 15,598,778 | Yes | +| 45,000 | 1.1350 | 15,025,010 | Yes | +| 47,500 | 1.1178 | 14,743,726 | Yes | +| 50,000 | 1.0953* | 14,353,974 | Yes | + +*Step 50K artifact reflects the final model with full warmdown applied. + +**Note:** The artifact peaks at ~17.24MB between steps 10K–30K — a much wider over-budget window than at 20K training. Warmdown fully recovers compressibility, landing at ~14.30MB, ~1.7MB below the 16MB limit. Early stopping without a dedicated warmdown phase remains non-viable for this architecture. + +### BPB vs Steps (ASCII plot) + +Power-law decay with two distinct phases: rapid early learning (0–10K), plateau (10K–30K), then warmdown-driven final drop. + +``` +BPB +4.10 |* + | + | +2.50 | + | +1.26 | * +1.23 | * +1.22 | * +1.20 | * * * * * * * * * +1.18 | * +1.16 | * +1.14 | * +1.12 | * +1.10 | * + +---------+--------+---------+-------> steps (K) + 0 10 20 30 50 + + ||<--- plateau --->|| + (rapid) (BPB ≈1.19–1.20) (sharp drop) +``` + +### Artifact Size vs Steps (ASCII plot) + +Non-monotonic: grows to peak at ~10K–30K steps (~17.2MB), then shrinks rapidly during warmdown. + +``` +MB +17.2 | * * * * * * * * * * * * +16.0 |-------------------------------------- 16MB limit +15.5 | +14.9 | * * +14.4 | * +13.9 | * +13.1 | * + 4.6 |* + +---------+--------+---------+-------> steps (K) + 0 10 20 30 50 + + ||<-------over budget------->|| +``` + +Intermediate checkpoints between steps ~10K–42K exceed the 16MB budget. Only the final model (with warmdown complete) fits at ~14.30MB. + +### Observations + +1. **BPB vs steps: plateau then sharp warmdown drop.** Unlike the 20K run (where warmdown is visible but brief), the 50K warmdown spans 19,500 steps and drives a dramatic BPB drop from 1.1904 (step 25K) to 1.0953 (step 50K) — a drop of 0.095 BPB in the final half of training. +2. **Artifact size plateau is wider at 50K.** The over-budget window spans steps 10K–42K (32K steps vs ~10K steps at 20K scale). Warmdown at larger scale is both longer and more effective — final artifact is 14.30MB vs 15.05MB at 20K. +3. **TTT scales with compute.** At 50K steps, TTT provides −0.0088 BPB (up from −0.0058 at 20K steps). A stronger base model enables more adaptation at test time. +4. **Diminishing returns on BPB.** Doubling compute from 20K→50K (steps) yields −0.0102 BPB improvement. Power-law decay holds but with diminishing marginal returns as expected. + +## Architecture + +Identical to [PR #549](https://github.com/openai/parameter-golf/pull/549) (LeakyReLU² + Legal TTT + Parallel Muon): + +| Component | Setting | +|-----------|---------| +| Layers | 11 (512d, 8H, 4KV) | +| MLP | 3× with LeakyReLU(0.5)² | +| BigramHash | 1536 | +| XSA | Last 4 layers | +| RoPE | Partial (16/64 dims) | +| LN Scale | 1/√(layer+1) | +| VE128 | Layers 9-10 | +| Weight avg | EMA(0.997) + Tight SWA(every 50) | +| Quantization | GPTQ-lite int6 + lzma | +| Optimizer | Parameter Banking + Parallel Muon | + +### Hyperparameter scaling for extended training + +LR schedule parameters scaled proportionally to maintain the same warmup/warmdown ratios: + +| Parameter | 20K steps | This work (50K steps) | +|-----------|-----------|-----------| +| ITERATIONS | 20,000 | 50,000 | +| WARMDOWN_ITERS | 7,800 (39.0%) | 19,500 (39.0%) | +| MUON_MOMENTUM_WARMUP_STEPS | 3,340 (16.7%) | 8,350 (16.7%) | +| VAL_LOSS_EVERY | 2,000 | 2,500 | +| MAX_WALLCLOCK_SECONDS | 0 (unlimited) | 0 (unlimited) | + +## Run Command + +```bash +RUN_ID=train_step50k_seed1337 \ +NUM_LAYERS=11 BIGRAM_VOCAB_SIZE=1536 XSA_LAST_N=4 \ +EMA_ENABLED=1 EMA_DECAY=0.997 SWA_ENABLED=1 SWA_EVERY=50 \ +ROPE_DIMS=16 LN_SCALE=1 LATE_QAT=1 LATE_QAT_THRESHOLD=0.15 \ +VE_ENABLED=1 VE_DIM=128 VE_LAYERS=9,10 \ +TTT_ENABLED=1 TTT_LR=0.002 TTT_EPOCHS=3 TTT_CHUNK_TOKENS=32768 \ +TTT_FREEZE_BLOCKS=0 TTT_MOMENTUM=0.9 TTT_BATCH_SEQS=32 TTT_GRAD_CLIP=1.0 \ +MUON_WD=0.04 ADAM_WD=0.04 \ +MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 \ +MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 \ +MUON_MOMENTUM_WARMUP_STEPS=8350 WARMDOWN_ITERS=19500 \ +ITERATIONS=50000 MAX_WALLCLOCK_SECONDS=0 EVAL_STRIDE=64 \ +VAL_LOSS_EVERY=2500 \ +SEED=1337 \ +CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --standalone --nproc_per_node=4 train_gpt.py +``` + +## Hardware + +- 4×NVIDIA A100 MIG instances +- Seeds 1337 & 2024: ~829ms/step | Seed 42: ~831ms/step (vs 83ms/step on 8×H100 SXM — approximately 10× slower) +- grad_accum_steps=2 (to match 8-GPU effective batch size of 786,432 tokens) +- ~11.5h training + ~34min TTT eval = ~12 hours total per seed + +## Credits + +This submission uses the full architecture and code from the record-track PR #549 with no ML changes — only extended compute and proportionally scaled LR schedules. + +- **Base submission [PR #549](https://github.com/openai/parameter-golf/pull/549)**: @abaybektursun — LeakyReLU² + Legal TTT + Parallel Muon +- **LeakyReLU² activation**: [PR #493](https://github.com/openai/parameter-golf/pull/493) by @parinzee, [PR #518](https://github.com/openai/parameter-golf/pull/518) by @sofiabod +- **Optimizer (Parameter Banking + Parallel Muon)**: [PR #399](https://github.com/openai/parameter-golf/pull/399) by @abaybektursun +- **TTT recipe**: [PR #461](https://github.com/openai/parameter-golf/pull/461) by @Christopher-Lee-McClendon +- **Base model**: [PR #414](https://github.com/openai/parameter-golf/pull/414) by @signalrush diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/run_scaling_50k.sh b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/run_scaling_50k.sh new file mode 100644 index 0000000000..24533c568c --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/run_scaling_50k.sh @@ -0,0 +1,36 @@ +#!/bin/bash +set -e + +cd /projects/jundhu-v2/jundhu_projects/parameter_golf/parameter_golf_exp + +# 50K scaling run — intermediate val_bpb logged every 2.5K steps for scaling curve +# Warmdown and momentum warmup scaled proportionally (~39% and ~16.7%) +# 50K * 0.39 = 19500, 50K * 0.167 = 8350 + +echo "==========================================" +echo "Scaling run: 50K steps, seed=2024" +echo "Started at: $(date)" +echo "==========================================" + +env \ + RUN_ID=train_step50k_seed2024 \ + NUM_LAYERS=11 BIGRAM_VOCAB_SIZE=1536 XSA_LAST_N=4 \ + EMA_ENABLED=1 EMA_DECAY=0.997 SWA_ENABLED=1 SWA_EVERY=50 \ + ROPE_DIMS=16 LN_SCALE=1 LATE_QAT=1 LATE_QAT_THRESHOLD=0.15 \ + VE_ENABLED=1 VE_DIM=128 VE_LAYERS=9,10 \ + TTT_ENABLED=1 TTT_LR=0.002 TTT_EPOCHS=3 TTT_CHUNK_TOKENS=32768 \ + TTT_FREEZE_BLOCKS=0 TTT_MOMENTUM=0.9 TTT_BATCH_SEQS=32 TTT_GRAD_CLIP=1.0 \ + MUON_WD=0.04 ADAM_WD=0.04 \ + MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 \ + MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 \ + MUON_MOMENTUM_WARMUP_STEPS=8350 WARMDOWN_ITERS=19500 \ + ITERATIONS=50000 MAX_WALLCLOCK_SECONDS=0 EVAL_STRIDE=64 \ + VAL_LOSS_EVERY=2500 \ + SEED=2024 \ + CUDA_VISIBLE_DEVICES=0,1,2,3 \ + torchrun --standalone --nproc_per_node=4 train_gpt.py \ + 2>&1 | tee logs/train_step50k_seed2024.txt + +echo "==========================================" +echo "50K run finished at: $(date)" +echo "==========================================" diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/submission.json b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/submission.json new file mode 100644 index 0000000000..03a4a72836 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/submission.json @@ -0,0 +1,20 @@ +{ + "name": "Extended Compute Scaling Analysis (50K Steps, ~12 Hours)", + "val_bpb": 1.0858, + "val_bpb_std": 0.0005, + "bytes_total": 14296887, + "seeds": [1337, 42, 2024], + "seed_results": { + "1337": { "val_bpb": 1.0853, "pre_ttt_bpb": 1.0942, "ttt_gain": -0.0089, "artifact_bytes": 14330478, "step_avg_ms": 829.02 }, + "42": { "val_bpb": 1.0857, "pre_ttt_bpb": 1.0945, "ttt_gain": -0.0088, "artifact_bytes": 14206210, "step_avg_ms": 830.86 }, + "2024": { "val_bpb": 1.0865, "pre_ttt_bpb": 1.0953, "ttt_gain": -0.0088, "artifact_bytes": 14353974, "step_avg_ms": 828.72 } + }, + "blurb": "Scaling analysis of the PR #549 SOTA stack (LeakyReLU² + TTT + Parallel Muon) under extended compute on 4×A100 MIG. 50K steps (~12 hours training) achieves 1.0858 BPB post-TTT (3-seed mean, std 0.0005) with ~14.30MB artifact. Extends the 20K scaling curve showing continued power-law decay and TTT gains scaling with compute (-0.0089 BPB vs -0.0058 at 20K).", + "author": "Jundong Hu", + "github_id": "OnlyJundong", + "date": "2026-04-06", + "track": "non_record_16mb", + "hardware": "4×A100 MIG", + "steps": 50000, + "step_avg_ms_mean": 829.53 +} diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_gpt.py b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_gpt.py new file mode 100644 index 0000000000..248e83626b --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_gpt.py @@ -0,0 +1,1938 @@ +from __future__ import annotations +import copy +import glob +import io +import lzma +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +# Fix for MIG/multi-GPU: ensure each process only sees its assigned GPU +# Must be set before importing torch +if "LOCAL_RANK" in os.environ and "CUDA_VISIBLE_DEVICES" in os.environ: + visible_devices = os.environ["CUDA_VISIBLE_DEVICES"].split(",") + local_rank = int(os.environ["LOCAL_RANK"]) + if local_rank < len(visible_devices): + os.environ["CUDA_VISIBLE_DEVICES"] = visible_devices[local_rank] + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True +except ImportError: + _HAS_FA3 = False +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 2)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +# --- Data loading --- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + # Estimate artifact size (int6+lzma) at each validation step + try: + sd_snap = {k: v.detach().cpu() for k, v in base_model.state_dict().items() + if "mtp_heads" not in k} + unbanked_snap = _unbank_state_dict(sd_snap, args.num_layers) + q_snap, m_snap = mixed_quantize_int6(unbanked_snap, {"mlp", "attn"}) + snap_buf = io.BytesIO() + torch.save({"w": q_snap, "m": m_snap}, snap_buf) + snap_compressed = len(lzma.compress(snap_buf.getvalue(), preset=6)) + snap_code_bytes = len(code.encode("utf-8")) + snap_total = snap_compressed + snap_code_bytes + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_bytes:{snap_total} model_int6_lzma:{snap_compressed} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + except Exception as e: + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_size:error({e}) " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed1337.txt b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed1337.txt new file mode 100644 index 0000000000..1273bc65b0 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed1337.txt @@ -0,0 +1,2260 @@ +logs/scaling_50k_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:50000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/50000 val_loss:6.9304 val_bpb:4.1046 artifact_bytes:4577902 model_int6_lzma:4486400 train_time:0ms step_avg:0.02ms +step:1/50000 train_loss:6.9317 train_time:1023ms step_avg:1022.65ms +step:2/50000 train_loss:8.6832 train_time:1567ms step_avg:783.57ms +step:3/50000 train_loss:7.6849 train_time:2419ms step_avg:806.34ms +step:4/50000 train_loss:7.1826 train_time:3279ms step_avg:819.82ms +step:5/50000 train_loss:7.1343 train_time:4131ms step_avg:826.29ms +step:6/50000 train_loss:7.1132 train_time:4984ms step_avg:830.61ms +step:7/50000 train_loss:7.1096 train_time:5836ms step_avg:833.65ms +step:8/50000 train_loss:6.9542 train_time:6688ms step_avg:835.94ms +step:9/50000 train_loss:6.6144 train_time:7540ms step_avg:837.75ms +step:10/50000 train_loss:6.2392 train_time:8392ms step_avg:839.20ms +step:500/50000 train_loss:2.3566 train_time:414391ms step_avg:828.78ms +step:1000/50000 train_loss:2.2649 train_time:828718ms step_avg:828.72ms +step:1500/50000 train_loss:2.1835 train_time:1242997ms step_avg:828.66ms +step:2000/50000 train_loss:2.0346 train_time:1657066ms step_avg:828.53ms +step:2500/50000 train_loss:2.1555 train_time:2071255ms step_avg:828.50ms +step:2500/50000 val_loss:2.1197 val_bpb:1.2554 artifact_bytes:13062362 model_int6_lzma:12970860 train_time:2071607ms step_avg:828.64ms +step:3000/50000 train_loss:2.1230 train_time:2486329ms step_avg:828.78ms +step:3500/50000 train_loss:2.1888 train_time:2900436ms step_avg:828.70ms +step:4000/50000 train_loss:2.1608 train_time:3314653ms step_avg:828.66ms +step:4500/50000 train_loss:2.1304 train_time:3729010ms step_avg:828.67ms +step:5000/50000 train_loss:2.0927 train_time:4142940ms step_avg:828.59ms +step:5000/50000 val_loss:2.0751 val_bpb:1.2290 artifact_bytes:14137046 model_int6_lzma:14045544 train_time:4143282ms step_avg:828.66ms +step:5500/50000 train_loss:2.1367 train_time:4561300ms step_avg:829.33ms +step:6000/50000 train_loss:2.0319 train_time:4975437ms step_avg:829.24ms +step:6500/50000 train_loss:2.3011 train_time:5389560ms step_avg:829.16ms +step:7000/50000 train_loss:1.9272 train_time:5803732ms step_avg:829.10ms +step:7500/50000 train_loss:2.0523 train_time:6217640ms step_avg:829.02ms +step:7500/50000 val_loss:2.0601 val_bpb:1.2201 artifact_bytes:15465582 model_int6_lzma:15374080 train_time:6217998ms step_avg:829.07ms +step:8000/50000 train_loss:2.0043 train_time:6635416ms step_avg:829.43ms +step:8500/50000 train_loss:1.9835 train_time:7049580ms step_avg:829.36ms +step:9000/50000 train_loss:2.0981 train_time:7463731ms step_avg:829.30ms +step:9500/50000 train_loss:2.2500 train_time:7878137ms step_avg:829.28ms +step:10000/50000 train_loss:2.0623 train_time:8292109ms step_avg:829.21ms +step:10000/50000 val_loss:2.0298 val_bpb:1.2022 artifact_bytes:17185494 model_int6_lzma:17093992 train_time:8292463ms step_avg:829.25ms +step:10500/50000 train_loss:2.0985 train_time:8706332ms step_avg:829.17ms +step:11000/50000 train_loss:1.9900 train_time:9120685ms step_avg:829.15ms +step:11500/50000 train_loss:1.9463 train_time:9534878ms step_avg:829.12ms +step:12000/50000 train_loss:1.9647 train_time:9948914ms step_avg:829.08ms +step:12500/50000 train_loss:1.9733 train_time:10363165ms step_avg:829.05ms +step:12500/50000 val_loss:2.0239 val_bpb:1.1987 artifact_bytes:17195458 model_int6_lzma:17103956 train_time:10363511ms step_avg:829.08ms +step:13000/50000 train_loss:1.9893 train_time:10777330ms step_avg:829.03ms +step:13500/50000 train_loss:2.1058 train_time:11191724ms step_avg:829.02ms +step:14000/50000 train_loss:1.9129 train_time:11605951ms step_avg:829.00ms +step:14500/50000 train_loss:2.0925 train_time:12020167ms step_avg:828.98ms +step:15000/50000 train_loss:2.0414 train_time:12434266ms step_avg:828.95ms +step:15000/50000 val_loss:2.0153 val_bpb:1.1936 artifact_bytes:17238818 model_int6_lzma:17147316 train_time:12434621ms step_avg:828.97ms +step:15500/50000 train_loss:2.0143 train_time:12849006ms step_avg:828.97ms +step:16000/50000 train_loss:2.0965 train_time:13263279ms step_avg:828.95ms +step:16500/50000 train_loss:2.0796 train_time:13677134ms step_avg:828.92ms +step:17000/50000 train_loss:2.0793 train_time:14090858ms step_avg:828.87ms +step:17500/50000 train_loss:2.0703 train_time:14505088ms step_avg:828.86ms +step:17500/50000 val_loss:2.0123 val_bpb:1.1918 artifact_bytes:17178826 model_int6_lzma:17087324 train_time:14505424ms step_avg:828.88ms +step:18000/50000 train_loss:1.9518 train_time:14920145ms step_avg:828.90ms +step:18500/50000 train_loss:1.9218 train_time:15334455ms step_avg:828.89ms +step:19000/50000 train_loss:1.9064 train_time:15748579ms step_avg:828.87ms +step:19500/50000 train_loss:2.1016 train_time:16162691ms step_avg:828.86ms +step:20000/50000 train_loss:2.0155 train_time:16576715ms step_avg:828.84ms +step:20000/50000 val_loss:2.0102 val_bpb:1.1905 artifact_bytes:17182262 model_int6_lzma:17090760 train_time:16577058ms step_avg:828.85ms +step:20500/50000 train_loss:2.0779 train_time:16990982ms step_avg:828.83ms +step:21000/50000 train_loss:1.9839 train_time:17405163ms step_avg:828.82ms +step:21500/50000 train_loss:2.0096 train_time:17819403ms step_avg:828.81ms +step:22000/50000 train_loss:2.0285 train_time:18233536ms step_avg:828.80ms +step:22500/50000 train_loss:2.0467 train_time:18647640ms step_avg:828.78ms +step:22500/50000 val_loss:2.0092 val_bpb:1.1900 artifact_bytes:17168362 model_int6_lzma:17076860 train_time:18647994ms step_avg:828.80ms +step:23000/50000 train_loss:1.9741 train_time:19063091ms step_avg:828.83ms +step:23500/50000 train_loss:2.0698 train_time:19477367ms step_avg:828.82ms +step:24000/50000 train_loss:1.9651 train_time:19891565ms step_avg:828.82ms +step:24500/50000 train_loss:2.0126 train_time:20305658ms step_avg:828.80ms +step:25000/50000 train_loss:2.0209 train_time:20719923ms step_avg:828.80ms +step:25000/50000 val_loss:2.0089 val_bpb:1.1898 artifact_bytes:17167526 model_int6_lzma:17076024 train_time:20720282ms step_avg:828.81ms +step:25500/50000 train_loss:2.0290 train_time:21135316ms step_avg:828.84ms +step:26000/50000 train_loss:2.0744 train_time:21549311ms step_avg:828.82ms +step:26500/50000 train_loss:2.0755 train_time:21963268ms step_avg:828.80ms +step:27000/50000 train_loss:1.9917 train_time:22377137ms step_avg:828.78ms +step:27500/50000 train_loss:1.9870 train_time:22791091ms step_avg:828.77ms +step:27500/50000 val_loss:2.0054 val_bpb:1.1877 artifact_bytes:17158910 model_int6_lzma:17067408 train_time:22791427ms step_avg:828.78ms +step:28000/50000 train_loss:1.8265 train_time:23207952ms step_avg:828.86ms +step:28500/50000 train_loss:2.0216 train_time:23621876ms step_avg:828.84ms +step:29000/50000 train_loss:2.0337 train_time:24035834ms step_avg:828.82ms +step:29500/50000 train_loss:2.0410 train_time:24449558ms step_avg:828.80ms +step:30000/50000 train_loss:1.9772 train_time:24863967ms step_avg:828.80ms +step:30000/50000 val_loss:2.0026 val_bpb:1.1861 artifact_bytes:17148538 model_int6_lzma:17057036 train_time:24864321ms step_avg:828.81ms +step:30500/50000 train_loss:2.0227 train_time:25280953ms step_avg:828.88ms +step:31000/50000 train_loss:2.0236 train_time:25694982ms step_avg:828.87ms +step:31500/50000 train_loss:1.9577 train_time:26109139ms step_avg:828.86ms +step:32000/50000 train_loss:2.0406 train_time:26522737ms step_avg:828.84ms +step:32500/50000 train_loss:1.9898 train_time:26936766ms step_avg:828.82ms +step:32500/50000 val_loss:1.9930 val_bpb:1.1804 artifact_bytes:17016594 model_int6_lzma:16925092 train_time:26937120ms step_avg:828.83ms +step:33000/50000 train_loss:1.9735 train_time:27354051ms step_avg:828.91ms +step:33500/50000 train_loss:1.9539 train_time:27767883ms step_avg:828.89ms +step:34000/50000 train_loss:2.0578 train_time:28182056ms step_avg:828.88ms +step:34500/50000 train_loss:1.9867 train_time:28595937ms step_avg:828.87ms +step:35000/50000 train_loss:1.9751 train_time:29010008ms step_avg:828.86ms +step:35000/50000 val_loss:1.9822 val_bpb:1.1740 artifact_bytes:16753850 model_int6_lzma:16662348 train_time:29010353ms step_avg:828.87ms +step:35500/50000 train_loss:2.0141 train_time:29428063ms step_avg:828.96ms +step:36000/50000 train_loss:1.9957 train_time:29841900ms step_avg:828.94ms +step:36500/50000 train_loss:2.0234 train_time:30255712ms step_avg:828.92ms +step:37000/50000 train_loss:1.8832 train_time:30669416ms step_avg:828.90ms +step:37500/50000 train_loss:2.0084 train_time:31083261ms step_avg:828.89ms +step:37500/50000 val_loss:1.9702 val_bpb:1.1669 artifact_bytes:16422374 model_int6_lzma:16330872 train_time:31083616ms step_avg:828.90ms +step:38000/50000 train_loss:1.9965 train_time:31497684ms step_avg:828.89ms +step:38500/50000 train_loss:1.9659 train_time:31911447ms step_avg:828.87ms +step:39000/50000 train_loss:1.8706 train_time:32325400ms step_avg:828.86ms +step:39500/50000 train_loss:1.9308 train_time:32739167ms step_avg:828.84ms +step:40000/50000 train_loss:1.9201 train_time:33153087ms step_avg:828.83ms +step:40000/50000 val_loss:1.9556 val_bpb:1.1582 artifact_bytes:16034950 model_int6_lzma:15943448 train_time:33153441ms step_avg:828.84ms +step:40500/50000 train_loss:2.3030 train_time:33568094ms step_avg:828.84ms +step:41000/50000 train_loss:1.8509 train_time:33981981ms step_avg:828.83ms +step:41500/50000 train_loss:1.7751 train_time:34396141ms step_avg:828.82ms +step:42000/50000 train_loss:1.9640 train_time:34810273ms step_avg:828.82ms +step:42500/50000 train_loss:1.9338 train_time:35224133ms step_avg:828.80ms +step:42500/50000 val_loss:1.9395 val_bpb:1.1487 artifact_bytes:15698694 model_int6_lzma:15607192 train_time:35224478ms step_avg:828.81ms +step:43000/50000 train_loss:2.0258 train_time:35642593ms step_avg:828.90ms +step:43500/50000 train_loss:2.1260 train_time:36056389ms step_avg:828.88ms +step:44000/50000 train_loss:1.8069 train_time:36470177ms step_avg:828.87ms +step:44500/50000 train_loss:1.8500 train_time:36883956ms step_avg:828.85ms +step:45000/50000 train_loss:1.9481 train_time:37297763ms step_avg:828.84ms +step:45000/50000 val_loss:1.9147 val_bpb:1.1340 artifact_bytes:15144098 model_int6_lzma:15052596 train_time:37298113ms step_avg:828.85ms +step:45500/50000 train_loss:1.9780 train_time:37716132ms step_avg:828.93ms +step:46000/50000 train_loss:1.8996 train_time:38129711ms step_avg:828.91ms +swa:start step:46150 +step:46500/50000 train_loss:1.8117 train_time:38544019ms step_avg:828.90ms +step:47000/50000 train_loss:1.9247 train_time:38958530ms step_avg:828.90ms +late_qat:enabled step:47076 scale:0.1499 +step:47500/50000 train_loss:1.8909 train_time:39372954ms step_avg:828.90ms +step:47500/50000 val_loss:1.8854 val_bpb:1.1167 artifact_bytes:14726650 model_int6_lzma:14635148 train_time:39373341ms step_avg:828.91ms +step:48000/50000 train_loss:1.8819 train_time:39791819ms step_avg:829.00ms +step:48500/50000 train_loss:1.8370 train_time:40206398ms step_avg:829.00ms +step:49000/50000 train_loss:1.7870 train_time:40621323ms step_avg:829.01ms +step:49500/50000 train_loss:1.8224 train_time:41036021ms step_avg:829.01ms +step:50000/50000 train_loss:1.8391 train_time:41450533ms step_avg:829.01ms +step:50000/50000 val_loss:1.8475 val_bpb:1.0942 artifact_bytes:14330478 model_int6_lzma:14238976 train_time:41450929ms step_avg:829.02ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8469 val_bpb:1.0939 eval_time:23715ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14257144 bytes +Total submission size int6+lzma: 14348646 bytes +final_int6_roundtrip val_loss:1.8963 val_bpb:1.1231 eval_time:24360ms +final_int6_roundtrip_exact val_loss:1.89626397 val_bpb:1.12307434 +final_int6_sliding_window val_loss:1.8566 val_bpb:1.0996 stride:64 eval_time:663400ms +final_int6_sliding_window_exact val_loss:1.85662085 val_bpb:1.09959836 +final_int8_zlib_roundtrip_exact val_loss:1.85662085 val_bpb:1.09959836 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.143264 time=1.1s + ttt_chunk [11/1893] bpb=1.102386 time=11.9s + ttt_chunk [21/1893] bpb=1.104574 time=22.6s + ttt_chunk [31/1893] bpb=1.106762 time=33.4s + ttt_chunk [41/1893] bpb=1.094977 time=44.2s + ttt_chunk [51/1893] bpb=1.090986 time=54.9s + ttt_chunk [61/1893] bpb=1.095411 time=65.7s + ttt_chunk [71/1893] bpb=1.092186 time=76.5s + ttt_chunk [81/1893] bpb=1.092049 time=87.3s + ttt_chunk [91/1893] bpb=1.090636 time=98.0s + ttt_chunk [101/1893] bpb=1.093373 time=108.8s + ttt_chunk [111/1893] bpb=1.094380 time=119.6s + ttt_chunk [121/1893] bpb=1.090917 time=130.3s + ttt_chunk [131/1893] bpb=1.090511 time=141.1s + ttt_chunk [141/1893] bpb=1.090135 time=151.9s + ttt_chunk [151/1893] bpb=1.093283 time=162.6s + ttt_chunk [161/1893] bpb=1.094916 time=173.4s + ttt_chunk [171/1893] bpb=1.095509 time=184.2s + ttt_chunk [181/1893] bpb=1.095422 time=194.9s + ttt_chunk [191/1893] bpb=1.098575 time=205.7s + ttt_chunk [201/1893] bpb=1.098760 time=216.5s + ttt_chunk [211/1893] bpb=1.096088 time=227.3s + ttt_chunk [221/1893] bpb=1.098044 time=238.0s + ttt_chunk [231/1893] bpb=1.097377 time=248.8s + ttt_chunk [241/1893] bpb=1.097047 time=259.6s + ttt_chunk [251/1893] bpb=1.095568 time=270.4s + ttt_chunk [261/1893] bpb=1.093933 time=281.1s + ttt_chunk [271/1893] bpb=1.092596 time=291.9s + ttt_chunk [281/1893] bpb=1.094857 time=302.7s + ttt_chunk [291/1893] bpb=1.095580 time=313.4s + ttt_chunk [301/1893] bpb=1.096123 time=324.2s + ttt_chunk [311/1893] bpb=1.097615 time=335.0s + ttt_chunk [321/1893] bpb=1.099057 time=345.8s + ttt_chunk [331/1893] bpb=1.098819 time=356.5s + ttt_chunk [341/1893] bpb=1.099094 time=367.3s + ttt_chunk [351/1893] bpb=1.100367 time=378.1s + ttt_chunk [361/1893] bpb=1.101515 time=388.8s + ttt_chunk [371/1893] bpb=1.100995 time=399.6s + ttt_chunk [381/1893] bpb=1.100931 time=410.4s + ttt_chunk [391/1893] bpb=1.100561 time=421.2s + ttt_chunk [401/1893] bpb=1.099187 time=431.9s + ttt_chunk [411/1893] bpb=1.098041 time=442.7s + ttt_chunk [421/1893] bpb=1.097510 time=453.5s + ttt_chunk [431/1893] bpb=1.098206 time=464.3s + ttt_chunk [441/1893] bpb=1.098037 time=475.0s + ttt_chunk [451/1893] bpb=1.097717 time=485.8s + ttt_chunk [461/1893] bpb=1.097002 time=496.6s + ttt_chunk [471/1893] bpb=1.096618 time=507.3s + ttt_chunk [481/1893] bpb=1.096311 time=518.1s + ttt_chunk [491/1893] bpb=1.095995 time=528.9s + ttt_chunk [501/1893] bpb=1.095459 time=539.7s + ttt_chunk [511/1893] bpb=1.094842 time=550.4s + ttt_chunk [521/1893] bpb=1.093730 time=561.2s + ttt_chunk [531/1893] bpb=1.093812 time=572.0s + ttt_chunk [541/1893] bpb=1.093750 time=582.7s + ttt_chunk [551/1893] bpb=1.092578 time=593.5s + ttt_chunk [561/1893] bpb=1.093122 time=604.3s + ttt_chunk [571/1893] bpb=1.092406 time=615.1s + ttt_chunk [581/1893] bpb=1.091865 time=625.8s + ttt_chunk [591/1893] bpb=1.091172 time=636.6s + ttt_chunk [601/1893] bpb=1.091831 time=647.4s + ttt_chunk [611/1893] bpb=1.091385 time=658.2s + ttt_chunk [621/1893] bpb=1.091308 time=668.9s + ttt_chunk [631/1893] bpb=1.091699 time=679.7s + ttt_chunk [641/1893] bpb=1.091471 time=690.4s + ttt_chunk [651/1893] bpb=1.091489 time=701.2s + ttt_chunk [661/1893] bpb=1.091359 time=712.0s + ttt_chunk [671/1893] bpb=1.090971 time=722.8s + ttt_chunk [681/1893] bpb=1.091218 time=733.5s + ttt_chunk [691/1893] bpb=1.091894 time=744.3s + ttt_chunk [701/1893] bpb=1.091104 time=755.1s + ttt_chunk [711/1893] bpb=1.091702 time=765.8s + ttt_chunk [721/1893] bpb=1.091323 time=776.6s + ttt_chunk [731/1893] bpb=1.091750 time=787.4s + ttt_chunk [741/1893] bpb=1.091690 time=798.2s + ttt_chunk [751/1893] bpb=1.091341 time=808.9s + ttt_chunk [761/1893] bpb=1.091177 time=819.7s + ttt_chunk [771/1893] bpb=1.090948 time=830.5s + ttt_chunk [781/1893] bpb=1.091445 time=841.2s + ttt_chunk [791/1893] bpb=1.091118 time=852.0s + ttt_chunk [801/1893] bpb=1.091173 time=862.8s + ttt_chunk [811/1893] bpb=1.090683 time=873.6s + ttt_chunk [821/1893] bpb=1.090471 time=884.3s + ttt_chunk [831/1893] bpb=1.089997 time=895.1s + ttt_chunk [841/1893] bpb=1.089406 time=905.9s + ttt_chunk [851/1893] bpb=1.089279 time=916.7s + ttt_chunk [861/1893] bpb=1.089409 time=927.4s + ttt_chunk [871/1893] bpb=1.089437 time=938.2s + ttt_chunk [881/1893] bpb=1.089446 time=949.0s + ttt_chunk [891/1893] bpb=1.089263 time=959.7s + ttt_chunk [901/1893] bpb=1.089245 time=970.5s + ttt_chunk [911/1893] bpb=1.089239 time=981.3s + ttt_chunk [921/1893] bpb=1.089612 time=992.0s + ttt_chunk [931/1893] bpb=1.089451 time=1002.8s + ttt_chunk [941/1893] bpb=1.089356 time=1013.6s + ttt_chunk [951/1893] bpb=1.089337 time=1024.3s + ttt_chunk [961/1893] bpb=1.089082 time=1035.1s + ttt_chunk [971/1893] bpb=1.089790 time=1045.9s + ttt_chunk [981/1893] bpb=1.089929 time=1056.7s + ttt_chunk [991/1893] bpb=1.089809 time=1067.5s + ttt_chunk [1001/1893] bpb=1.089976 time=1078.2s + ttt_chunk [1011/1893] bpb=1.090238 time=1089.0s + ttt_chunk [1021/1893] bpb=1.090390 time=1099.8s + ttt_chunk [1031/1893] bpb=1.090917 time=1110.6s + ttt_chunk [1041/1893] bpb=1.090578 time=1121.3s + ttt_chunk [1051/1893] bpb=1.090273 time=1132.1s + ttt_chunk [1061/1893] bpb=1.090511 time=1142.9s + ttt_chunk [1071/1893] bpb=1.090954 time=1153.6s + ttt_chunk [1081/1893] bpb=1.090943 time=1164.4s + ttt_chunk [1091/1893] bpb=1.091356 time=1175.2s + ttt_chunk [1101/1893] bpb=1.091488 time=1186.0s + ttt_chunk [1111/1893] bpb=1.091260 time=1196.7s + ttt_chunk [1121/1893] bpb=1.091181 time=1207.5s + ttt_chunk [1131/1893] bpb=1.091072 time=1218.3s + ttt_chunk [1141/1893] bpb=1.090937 time=1229.0s + ttt_chunk [1151/1893] bpb=1.090970 time=1239.8s + ttt_chunk [1161/1893] bpb=1.090188 time=1250.6s + ttt_chunk [1171/1893] bpb=1.090766 time=1261.4s + ttt_chunk [1181/1893] bpb=1.090269 time=1272.1s + ttt_chunk [1191/1893] bpb=1.089988 time=1282.9s + ttt_chunk [1201/1893] bpb=1.090561 time=1293.7s + ttt_chunk [1211/1893] bpb=1.089962 time=1304.4s + ttt_chunk [1221/1893] bpb=1.089661 time=1315.2s + ttt_chunk [1231/1893] bpb=1.089599 time=1326.0s + ttt_chunk [1241/1893] bpb=1.089410 time=1336.8s + ttt_chunk [1251/1893] bpb=1.089171 time=1347.5s + ttt_chunk [1261/1893] bpb=1.089113 time=1358.3s + ttt_chunk [1271/1893] bpb=1.088914 time=1369.1s + ttt_chunk [1281/1893] bpb=1.088710 time=1379.9s + ttt_chunk [1291/1893] bpb=1.088523 time=1390.6s + ttt_chunk [1301/1893] bpb=1.088161 time=1401.4s + ttt_chunk [1311/1893] bpb=1.087859 time=1412.2s + ttt_chunk [1321/1893] bpb=1.087701 time=1423.0s + ttt_chunk [1331/1893] bpb=1.087611 time=1433.7s + ttt_chunk [1341/1893] bpb=1.087498 time=1444.5s + ttt_chunk [1351/1893] bpb=1.087448 time=1455.3s + ttt_chunk [1361/1893] bpb=1.087658 time=1466.0s + ttt_chunk [1371/1893] bpb=1.087513 time=1476.8s + ttt_chunk [1381/1893] bpb=1.087452 time=1487.6s + ttt_chunk [1391/1893] bpb=1.086938 time=1498.4s + ttt_chunk [1401/1893] bpb=1.086952 time=1509.2s + ttt_chunk [1411/1893] bpb=1.086986 time=1519.9s + ttt_chunk [1421/1893] bpb=1.087243 time=1530.7s + ttt_chunk [1431/1893] bpb=1.087094 time=1541.5s + ttt_chunk [1441/1893] bpb=1.087766 time=1552.2s + ttt_chunk [1451/1893] bpb=1.087875 time=1563.0s + ttt_chunk [1461/1893] bpb=1.087616 time=1573.8s + ttt_chunk [1471/1893] bpb=1.088508 time=1584.6s + ttt_chunk [1481/1893] bpb=1.088318 time=1595.3s + ttt_chunk [1491/1893] bpb=1.088295 time=1606.1s + ttt_chunk [1501/1893] bpb=1.088428 time=1616.9s + ttt_chunk [1511/1893] bpb=1.088500 time=1627.7s + ttt_chunk [1521/1893] bpb=1.088516 time=1638.4s + ttt_chunk [1531/1893] bpb=1.088335 time=1649.2s + ttt_chunk [1541/1893] bpb=1.088300 time=1660.0s + ttt_chunk [1551/1893] bpb=1.088644 time=1670.8s + ttt_chunk [1561/1893] bpb=1.088779 time=1681.5s + ttt_chunk [1571/1893] bpb=1.088893 time=1692.3s + ttt_chunk [1581/1893] bpb=1.088999 time=1703.1s + ttt_chunk [1591/1893] bpb=1.088948 time=1713.9s + ttt_chunk [1601/1893] bpb=1.089129 time=1724.6s + ttt_chunk [1611/1893] bpb=1.089211 time=1735.4s + ttt_chunk [1621/1893] bpb=1.089078 time=1746.2s + ttt_chunk [1631/1893] bpb=1.089247 time=1756.9s + ttt_chunk [1641/1893] bpb=1.089113 time=1767.7s + ttt_chunk [1651/1893] bpb=1.089025 time=1778.5s + ttt_chunk [1661/1893] bpb=1.088917 time=1789.3s + ttt_chunk [1671/1893] bpb=1.089273 time=1800.0s + ttt_chunk [1681/1893] bpb=1.089519 time=1810.8s + ttt_chunk [1691/1893] bpb=1.089483 time=1821.6s + ttt_chunk [1701/1893] bpb=1.089475 time=1832.3s + ttt_chunk [1711/1893] bpb=1.089308 time=1843.1s + ttt_chunk [1721/1893] bpb=1.089158 time=1853.9s + ttt_chunk [1731/1893] bpb=1.089108 time=1864.6s + ttt_chunk [1741/1893] bpb=1.088877 time=1875.4s + ttt_chunk [1751/1893] bpb=1.088730 time=1886.2s + ttt_chunk [1761/1893] bpb=1.088814 time=1897.0s + ttt_chunk [1771/1893] bpb=1.088740 time=1907.7s + ttt_chunk [1781/1893] bpb=1.088687 time=1918.5s + ttt_chunk [1791/1893] bpb=1.088300 time=1929.3s + ttt_chunk [1801/1893] bpb=1.088297 time=1940.0s + ttt_chunk [1811/1893] bpb=1.088139 time=1950.8s + ttt_chunk [1821/1893] bpb=1.088176 time=1961.6s + ttt_chunk [1831/1893] bpb=1.087789 time=1972.3s + ttt_chunk [1841/1893] bpb=1.087730 time=1983.1s + ttt_chunk [1851/1893] bpb=1.087510 time=1993.9s + ttt_chunk [1861/1893] bpb=1.087051 time=2004.6s + ttt_chunk [1871/1893] bpb=1.086903 time=2015.4s + ttt_chunk [1881/1893] bpb=1.086533 time=2026.2s + ttt_chunk [1891/1893] bpb=1.086383 time=2037.0s + ttt_chunk [1893/1893] bpb=1.086404 time=2038.7s +ttt_sliding:done val_loss=1.832524 val_bpb=1.085327 elapsed=2038.8s +legal_ttt val_loss:1.8325 val_bpb:1.0853 eval_time:2039427ms +legal_ttt_exact val_loss:1.83252444 val_bpb:1.08532707 +.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + # Estimate artifact size (int6+lzma) at each validation step + try: + sd_snap = {k: v.detach().cpu() for k, v in base_model.state_dict().items() + if "mtp_heads" not in k} + unbanked_snap = _unbank_state_dict(sd_snap, args.num_layers) + q_snap, m_snap = mixed_quantize_int6(unbanked_snap, {"mlp", "attn"}) + snap_buf = io.BytesIO() + torch.save({"w": q_snap, "m": m_snap}, snap_buf) + snap_compressed = len(lzma.compress(snap_buf.getvalue(), preset=6)) + snap_code_bytes = len(code.encode("utf-8")) + snap_total = snap_compressed + snap_code_bytes + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_bytes:{snap_total} model_int6_lzma:{snap_compressed} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + except Exception as e: + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_size:error({e}) " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.10.14 (main, May 6 2024, 19:42:50) [GCC 11.2.0] +Running PyTorch 2.10.0+cu128 +Fri Mar 27 15:38:32 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 550.127.08 Driver Version: 550.127.08 CUDA Version: 12.4 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA A100-SXM4-80GB On | 00000000:27:00.0 Off | On | +| N/A 37C P0 98W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA A100-SXM4-80GB On | 00000000:2A:00.0 Off | On | +| N/A 33C P0 91W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA A100-SXM4-80GB On | 00000000:51:00.0 Off | On | +| N/A 34C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA A100-SXM4-80GB On | 00000000:9E:00.0 Off | On | +| N/A 37C P0 99W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA A100-SXM4-80GB On | 00000000:A4:00.0 Off | On | +| N/A 32C P0 95W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA A100-SXM4-80GB On | 00000000:CA:00.0 Off | On | +| N/A 37C P0 97W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| MIG devices: | ++------------------+----------------------------------+-----------+-----------------------+ +| GPU GI CI MIG | Memory-Usage | Vol| Shared | +| ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG | +| | | ECC| | +|==================+==================================+===========+=======================| +| 0 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 2 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 3 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 2 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 1 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 3539864 C /opt/conda/bin/python3 0MiB | +| 1 N/A N/A 3539865 C /opt/conda/bin/python3 0MiB | +| 2 N/A N/A 3539866 C /opt/conda/bin/python3 0MiB | +| 3 N/A N/A 3539867 C /opt/conda/bin/python3 0MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:50000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/50000 val_loss:6.9304 val_bpb:4.1046 artifact_bytes:4577902 model_int6_lzma:4486400 train_time:0ms step_avg:0.02ms +step:1/50000 train_loss:6.9317 train_time:1023ms step_avg:1022.65ms +step:2/50000 train_loss:8.6832 train_time:1567ms step_avg:783.57ms +step:3/50000 train_loss:7.6849 train_time:2419ms step_avg:806.34ms +step:4/50000 train_loss:7.1826 train_time:3279ms step_avg:819.82ms +step:5/50000 train_loss:7.1343 train_time:4131ms step_avg:826.29ms +step:6/50000 train_loss:7.1132 train_time:4984ms step_avg:830.61ms +step:7/50000 train_loss:7.1096 train_time:5836ms step_avg:833.65ms +step:8/50000 train_loss:6.9542 train_time:6688ms step_avg:835.94ms +step:9/50000 train_loss:6.6144 train_time:7540ms step_avg:837.75ms +step:10/50000 train_loss:6.2392 train_time:8392ms step_avg:839.20ms +step:500/50000 train_loss:2.3566 train_time:414391ms step_avg:828.78ms +step:1000/50000 train_loss:2.2649 train_time:828718ms step_avg:828.72ms +step:1500/50000 train_loss:2.1835 train_time:1242997ms step_avg:828.66ms +step:2000/50000 train_loss:2.0346 train_time:1657066ms step_avg:828.53ms +step:2500/50000 train_loss:2.1555 train_time:2071255ms step_avg:828.50ms +step:2500/50000 val_loss:2.1197 val_bpb:1.2554 artifact_bytes:13062362 model_int6_lzma:12970860 train_time:2071607ms step_avg:828.64ms +step:3000/50000 train_loss:2.1230 train_time:2486329ms step_avg:828.78ms +step:3500/50000 train_loss:2.1888 train_time:2900436ms step_avg:828.70ms +step:4000/50000 train_loss:2.1608 train_time:3314653ms step_avg:828.66ms +step:4500/50000 train_loss:2.1304 train_time:3729010ms step_avg:828.67ms +step:5000/50000 train_loss:2.0927 train_time:4142940ms step_avg:828.59ms +step:5000/50000 val_loss:2.0751 val_bpb:1.2290 artifact_bytes:14137046 model_int6_lzma:14045544 train_time:4143282ms step_avg:828.66ms +step:5500/50000 train_loss:2.1367 train_time:4561300ms step_avg:829.33ms +step:6000/50000 train_loss:2.0319 train_time:4975437ms step_avg:829.24ms +step:6500/50000 train_loss:2.3011 train_time:5389560ms step_avg:829.16ms +step:7000/50000 train_loss:1.9272 train_time:5803732ms step_avg:829.10ms +step:7500/50000 train_loss:2.0523 train_time:6217640ms step_avg:829.02ms +step:7500/50000 val_loss:2.0601 val_bpb:1.2201 artifact_bytes:15465582 model_int6_lzma:15374080 train_time:6217998ms step_avg:829.07ms +step:8000/50000 train_loss:2.0043 train_time:6635416ms step_avg:829.43ms +step:8500/50000 train_loss:1.9835 train_time:7049580ms step_avg:829.36ms +step:9000/50000 train_loss:2.0981 train_time:7463731ms step_avg:829.30ms +step:9500/50000 train_loss:2.2500 train_time:7878137ms step_avg:829.28ms +step:10000/50000 train_loss:2.0623 train_time:8292109ms step_avg:829.21ms +step:10000/50000 val_loss:2.0298 val_bpb:1.2022 artifact_bytes:17185494 model_int6_lzma:17093992 train_time:8292463ms step_avg:829.25ms +step:10500/50000 train_loss:2.0985 train_time:8706332ms step_avg:829.17ms +step:11000/50000 train_loss:1.9900 train_time:9120685ms step_avg:829.15ms +step:11500/50000 train_loss:1.9463 train_time:9534878ms step_avg:829.12ms +step:12000/50000 train_loss:1.9647 train_time:9948914ms step_avg:829.08ms +step:12500/50000 train_loss:1.9733 train_time:10363165ms step_avg:829.05ms +step:12500/50000 val_loss:2.0239 val_bpb:1.1987 artifact_bytes:17195458 model_int6_lzma:17103956 train_time:10363511ms step_avg:829.08ms +step:13000/50000 train_loss:1.9893 train_time:10777330ms step_avg:829.03ms +step:13500/50000 train_loss:2.1058 train_time:11191724ms step_avg:829.02ms +step:14000/50000 train_loss:1.9129 train_time:11605951ms step_avg:829.00ms +step:14500/50000 train_loss:2.0925 train_time:12020167ms step_avg:828.98ms +step:15000/50000 train_loss:2.0414 train_time:12434266ms step_avg:828.95ms +step:15000/50000 val_loss:2.0153 val_bpb:1.1936 artifact_bytes:17238818 model_int6_lzma:17147316 train_time:12434621ms step_avg:828.97ms +step:15500/50000 train_loss:2.0143 train_time:12849006ms step_avg:828.97ms +step:16000/50000 train_loss:2.0965 train_time:13263279ms step_avg:828.95ms +step:16500/50000 train_loss:2.0796 train_time:13677134ms step_avg:828.92ms +step:17000/50000 train_loss:2.0793 train_time:14090858ms step_avg:828.87ms +step:17500/50000 train_loss:2.0703 train_time:14505088ms step_avg:828.86ms +step:17500/50000 val_loss:2.0123 val_bpb:1.1918 artifact_bytes:17178826 model_int6_lzma:17087324 train_time:14505424ms step_avg:828.88ms +step:18000/50000 train_loss:1.9518 train_time:14920145ms step_avg:828.90ms +step:18500/50000 train_loss:1.9218 train_time:15334455ms step_avg:828.89ms +step:19000/50000 train_loss:1.9064 train_time:15748579ms step_avg:828.87ms +step:19500/50000 train_loss:2.1016 train_time:16162691ms step_avg:828.86ms +step:20000/50000 train_loss:2.0155 train_time:16576715ms step_avg:828.84ms +step:20000/50000 val_loss:2.0102 val_bpb:1.1905 artifact_bytes:17182262 model_int6_lzma:17090760 train_time:16577058ms step_avg:828.85ms +step:20500/50000 train_loss:2.0779 train_time:16990982ms step_avg:828.83ms +step:21000/50000 train_loss:1.9839 train_time:17405163ms step_avg:828.82ms +step:21500/50000 train_loss:2.0096 train_time:17819403ms step_avg:828.81ms +step:22000/50000 train_loss:2.0285 train_time:18233536ms step_avg:828.80ms +step:22500/50000 train_loss:2.0467 train_time:18647640ms step_avg:828.78ms +step:22500/50000 val_loss:2.0092 val_bpb:1.1900 artifact_bytes:17168362 model_int6_lzma:17076860 train_time:18647994ms step_avg:828.80ms +step:23000/50000 train_loss:1.9741 train_time:19063091ms step_avg:828.83ms +step:23500/50000 train_loss:2.0698 train_time:19477367ms step_avg:828.82ms +step:24000/50000 train_loss:1.9651 train_time:19891565ms step_avg:828.82ms +step:24500/50000 train_loss:2.0126 train_time:20305658ms step_avg:828.80ms +step:25000/50000 train_loss:2.0209 train_time:20719923ms step_avg:828.80ms +step:25000/50000 val_loss:2.0089 val_bpb:1.1898 artifact_bytes:17167526 model_int6_lzma:17076024 train_time:20720282ms step_avg:828.81ms +step:25500/50000 train_loss:2.0290 train_time:21135316ms step_avg:828.84ms +step:26000/50000 train_loss:2.0744 train_time:21549311ms step_avg:828.82ms +step:26500/50000 train_loss:2.0755 train_time:21963268ms step_avg:828.80ms +step:27000/50000 train_loss:1.9917 train_time:22377137ms step_avg:828.78ms +step:27500/50000 train_loss:1.9870 train_time:22791091ms step_avg:828.77ms +step:27500/50000 val_loss:2.0054 val_bpb:1.1877 artifact_bytes:17158910 model_int6_lzma:17067408 train_time:22791427ms step_avg:828.78ms +step:28000/50000 train_loss:1.8265 train_time:23207952ms step_avg:828.86ms +step:28500/50000 train_loss:2.0216 train_time:23621876ms step_avg:828.84ms +step:29000/50000 train_loss:2.0337 train_time:24035834ms step_avg:828.82ms +step:29500/50000 train_loss:2.0410 train_time:24449558ms step_avg:828.80ms +step:30000/50000 train_loss:1.9772 train_time:24863967ms step_avg:828.80ms +step:30000/50000 val_loss:2.0026 val_bpb:1.1861 artifact_bytes:17148538 model_int6_lzma:17057036 train_time:24864321ms step_avg:828.81ms +step:30500/50000 train_loss:2.0227 train_time:25280953ms step_avg:828.88ms +step:31000/50000 train_loss:2.0236 train_time:25694982ms step_avg:828.87ms +step:31500/50000 train_loss:1.9577 train_time:26109139ms step_avg:828.86ms +step:32000/50000 train_loss:2.0406 train_time:26522737ms step_avg:828.84ms +step:32500/50000 train_loss:1.9898 train_time:26936766ms step_avg:828.82ms +step:32500/50000 val_loss:1.9930 val_bpb:1.1804 artifact_bytes:17016594 model_int6_lzma:16925092 train_time:26937120ms step_avg:828.83ms +step:33000/50000 train_loss:1.9735 train_time:27354051ms step_avg:828.91ms +step:33500/50000 train_loss:1.9539 train_time:27767883ms step_avg:828.89ms +step:34000/50000 train_loss:2.0578 train_time:28182056ms step_avg:828.88ms +step:34500/50000 train_loss:1.9867 train_time:28595937ms step_avg:828.87ms +step:35000/50000 train_loss:1.9751 train_time:29010008ms step_avg:828.86ms +step:35000/50000 val_loss:1.9822 val_bpb:1.1740 artifact_bytes:16753850 model_int6_lzma:16662348 train_time:29010353ms step_avg:828.87ms +step:35500/50000 train_loss:2.0141 train_time:29428063ms step_avg:828.96ms +step:36000/50000 train_loss:1.9957 train_time:29841900ms step_avg:828.94ms +step:36500/50000 train_loss:2.0234 train_time:30255712ms step_avg:828.92ms +step:37000/50000 train_loss:1.8832 train_time:30669416ms step_avg:828.90ms +step:37500/50000 train_loss:2.0084 train_time:31083261ms step_avg:828.89ms +step:37500/50000 val_loss:1.9702 val_bpb:1.1669 artifact_bytes:16422374 model_int6_lzma:16330872 train_time:31083616ms step_avg:828.90ms +step:38000/50000 train_loss:1.9965 train_time:31497684ms step_avg:828.89ms +step:38500/50000 train_loss:1.9659 train_time:31911447ms step_avg:828.87ms +step:39000/50000 train_loss:1.8706 train_time:32325400ms step_avg:828.86ms +step:39500/50000 train_loss:1.9308 train_time:32739167ms step_avg:828.84ms +step:40000/50000 train_loss:1.9201 train_time:33153087ms step_avg:828.83ms +step:40000/50000 val_loss:1.9556 val_bpb:1.1582 artifact_bytes:16034950 model_int6_lzma:15943448 train_time:33153441ms step_avg:828.84ms +step:40500/50000 train_loss:2.3030 train_time:33568094ms step_avg:828.84ms +step:41000/50000 train_loss:1.8509 train_time:33981981ms step_avg:828.83ms +step:41500/50000 train_loss:1.7751 train_time:34396141ms step_avg:828.82ms +step:42000/50000 train_loss:1.9640 train_time:34810273ms step_avg:828.82ms +step:42500/50000 train_loss:1.9338 train_time:35224133ms step_avg:828.80ms +step:42500/50000 val_loss:1.9395 val_bpb:1.1487 artifact_bytes:15698694 model_int6_lzma:15607192 train_time:35224478ms step_avg:828.81ms +step:43000/50000 train_loss:2.0258 train_time:35642593ms step_avg:828.90ms +step:43500/50000 train_loss:2.1260 train_time:36056389ms step_avg:828.88ms +step:44000/50000 train_loss:1.8069 train_time:36470177ms step_avg:828.87ms +step:44500/50000 train_loss:1.8500 train_time:36883956ms step_avg:828.85ms +step:45000/50000 train_loss:1.9481 train_time:37297763ms step_avg:828.84ms +step:45000/50000 val_loss:1.9147 val_bpb:1.1340 artifact_bytes:15144098 model_int6_lzma:15052596 train_time:37298113ms step_avg:828.85ms +step:45500/50000 train_loss:1.9780 train_time:37716132ms step_avg:828.93ms +step:46000/50000 train_loss:1.8996 train_time:38129711ms step_avg:828.91ms +swa:start step:46150 +step:46500/50000 train_loss:1.8117 train_time:38544019ms step_avg:828.90ms +step:47000/50000 train_loss:1.9247 train_time:38958530ms step_avg:828.90ms +late_qat:enabled step:47076 scale:0.1499 +step:47500/50000 train_loss:1.8909 train_time:39372954ms step_avg:828.90ms +step:47500/50000 val_loss:1.8854 val_bpb:1.1167 artifact_bytes:14726650 model_int6_lzma:14635148 train_time:39373341ms step_avg:828.91ms +step:48000/50000 train_loss:1.8819 train_time:39791819ms step_avg:829.00ms +step:48500/50000 train_loss:1.8370 train_time:40206398ms step_avg:829.00ms +step:49000/50000 train_loss:1.7870 train_time:40621323ms step_avg:829.01ms +step:49500/50000 train_loss:1.8224 train_time:41036021ms step_avg:829.01ms +step:50000/50000 train_loss:1.8391 train_time:41450533ms step_avg:829.01ms +step:50000/50000 val_loss:1.8475 val_bpb:1.0942 artifact_bytes:14330478 model_int6_lzma:14238976 train_time:41450929ms step_avg:829.02ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8469 val_bpb:1.0939 eval_time:23715ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14257144 bytes +Total submission size int6+lzma: 14348646 bytes +final_int6_roundtrip val_loss:1.8963 val_bpb:1.1231 eval_time:24360ms +final_int6_roundtrip_exact val_loss:1.89626397 val_bpb:1.12307434 +final_int6_sliding_window val_loss:1.8566 val_bpb:1.0996 stride:64 eval_time:663400ms +final_int6_sliding_window_exact val_loss:1.85662085 val_bpb:1.09959836 +final_int8_zlib_roundtrip_exact val_loss:1.85662085 val_bpb:1.09959836 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.143264 time=1.1s + ttt_chunk [11/1893] bpb=1.102386 time=11.9s + ttt_chunk [21/1893] bpb=1.104574 time=22.6s + ttt_chunk [31/1893] bpb=1.106762 time=33.4s + ttt_chunk [41/1893] bpb=1.094977 time=44.2s + ttt_chunk [51/1893] bpb=1.090986 time=54.9s + ttt_chunk [61/1893] bpb=1.095411 time=65.7s + ttt_chunk [71/1893] bpb=1.092186 time=76.5s + ttt_chunk [81/1893] bpb=1.092049 time=87.3s + ttt_chunk [91/1893] bpb=1.090636 time=98.0s + ttt_chunk [101/1893] bpb=1.093373 time=108.8s + ttt_chunk [111/1893] bpb=1.094380 time=119.6s + ttt_chunk [121/1893] bpb=1.090917 time=130.3s + ttt_chunk [131/1893] bpb=1.090511 time=141.1s + ttt_chunk [141/1893] bpb=1.090135 time=151.9s + ttt_chunk [151/1893] bpb=1.093283 time=162.6s + ttt_chunk [161/1893] bpb=1.094916 time=173.4s + ttt_chunk [171/1893] bpb=1.095509 time=184.2s + ttt_chunk [181/1893] bpb=1.095422 time=194.9s + ttt_chunk [191/1893] bpb=1.098575 time=205.7s + ttt_chunk [201/1893] bpb=1.098760 time=216.5s + ttt_chunk [211/1893] bpb=1.096088 time=227.3s + ttt_chunk [221/1893] bpb=1.098044 time=238.0s + ttt_chunk [231/1893] bpb=1.097377 time=248.8s + ttt_chunk [241/1893] bpb=1.097047 time=259.6s + ttt_chunk [251/1893] bpb=1.095568 time=270.4s + ttt_chunk [261/1893] bpb=1.093933 time=281.1s + ttt_chunk [271/1893] bpb=1.092596 time=291.9s + ttt_chunk [281/1893] bpb=1.094857 time=302.7s + ttt_chunk [291/1893] bpb=1.095580 time=313.4s + ttt_chunk [301/1893] bpb=1.096123 time=324.2s + ttt_chunk [311/1893] bpb=1.097615 time=335.0s + ttt_chunk [321/1893] bpb=1.099057 time=345.8s + ttt_chunk [331/1893] bpb=1.098819 time=356.5s + ttt_chunk [341/1893] bpb=1.099094 time=367.3s + ttt_chunk [351/1893] bpb=1.100367 time=378.1s + ttt_chunk [361/1893] bpb=1.101515 time=388.8s + ttt_chunk [371/1893] bpb=1.100995 time=399.6s + ttt_chunk [381/1893] bpb=1.100931 time=410.4s + ttt_chunk [391/1893] bpb=1.100561 time=421.2s + ttt_chunk [401/1893] bpb=1.099187 time=431.9s + ttt_chunk [411/1893] bpb=1.098041 time=442.7s + ttt_chunk [421/1893] bpb=1.097510 time=453.5s + ttt_chunk [431/1893] bpb=1.098206 time=464.3s + ttt_chunk [441/1893] bpb=1.098037 time=475.0s + ttt_chunk [451/1893] bpb=1.097717 time=485.8s + ttt_chunk [461/1893] bpb=1.097002 time=496.6s + ttt_chunk [471/1893] bpb=1.096618 time=507.3s + ttt_chunk [481/1893] bpb=1.096311 time=518.1s + ttt_chunk [491/1893] bpb=1.095995 time=528.9s + ttt_chunk [501/1893] bpb=1.095459 time=539.7s + ttt_chunk [511/1893] bpb=1.094842 time=550.4s + ttt_chunk [521/1893] bpb=1.093730 time=561.2s + ttt_chunk [531/1893] bpb=1.093812 time=572.0s + ttt_chunk [541/1893] bpb=1.093750 time=582.7s + ttt_chunk [551/1893] bpb=1.092578 time=593.5s + ttt_chunk [561/1893] bpb=1.093122 time=604.3s + ttt_chunk [571/1893] bpb=1.092406 time=615.1s + ttt_chunk [581/1893] bpb=1.091865 time=625.8s + ttt_chunk [591/1893] bpb=1.091172 time=636.6s + ttt_chunk [601/1893] bpb=1.091831 time=647.4s + ttt_chunk [611/1893] bpb=1.091385 time=658.2s + ttt_chunk [621/1893] bpb=1.091308 time=668.9s + ttt_chunk [631/1893] bpb=1.091699 time=679.7s + ttt_chunk [641/1893] bpb=1.091471 time=690.4s + ttt_chunk [651/1893] bpb=1.091489 time=701.2s + ttt_chunk [661/1893] bpb=1.091359 time=712.0s + ttt_chunk [671/1893] bpb=1.090971 time=722.8s + ttt_chunk [681/1893] bpb=1.091218 time=733.5s + ttt_chunk [691/1893] bpb=1.091894 time=744.3s + ttt_chunk [701/1893] bpb=1.091104 time=755.1s + ttt_chunk [711/1893] bpb=1.091702 time=765.8s + ttt_chunk [721/1893] bpb=1.091323 time=776.6s + ttt_chunk [731/1893] bpb=1.091750 time=787.4s + ttt_chunk [741/1893] bpb=1.091690 time=798.2s + ttt_chunk [751/1893] bpb=1.091341 time=808.9s + ttt_chunk [761/1893] bpb=1.091177 time=819.7s + ttt_chunk [771/1893] bpb=1.090948 time=830.5s + ttt_chunk [781/1893] bpb=1.091445 time=841.2s + ttt_chunk [791/1893] bpb=1.091118 time=852.0s + ttt_chunk [801/1893] bpb=1.091173 time=862.8s + ttt_chunk [811/1893] bpb=1.090683 time=873.6s + ttt_chunk [821/1893] bpb=1.090471 time=884.3s + ttt_chunk [831/1893] bpb=1.089997 time=895.1s + ttt_chunk [841/1893] bpb=1.089406 time=905.9s + ttt_chunk [851/1893] bpb=1.089279 time=916.7s + ttt_chunk [861/1893] bpb=1.089409 time=927.4s + ttt_chunk [871/1893] bpb=1.089437 time=938.2s + ttt_chunk [881/1893] bpb=1.089446 time=949.0s + ttt_chunk [891/1893] bpb=1.089263 time=959.7s + ttt_chunk [901/1893] bpb=1.089245 time=970.5s + ttt_chunk [911/1893] bpb=1.089239 time=981.3s + ttt_chunk [921/1893] bpb=1.089612 time=992.0s + ttt_chunk [931/1893] bpb=1.089451 time=1002.8s + ttt_chunk [941/1893] bpb=1.089356 time=1013.6s + ttt_chunk [951/1893] bpb=1.089337 time=1024.3s + ttt_chunk [961/1893] bpb=1.089082 time=1035.1s + ttt_chunk [971/1893] bpb=1.089790 time=1045.9s + ttt_chunk [981/1893] bpb=1.089929 time=1056.7s + ttt_chunk [991/1893] bpb=1.089809 time=1067.5s + ttt_chunk [1001/1893] bpb=1.089976 time=1078.2s + ttt_chunk [1011/1893] bpb=1.090238 time=1089.0s + ttt_chunk [1021/1893] bpb=1.090390 time=1099.8s + ttt_chunk [1031/1893] bpb=1.090917 time=1110.6s + ttt_chunk [1041/1893] bpb=1.090578 time=1121.3s + ttt_chunk [1051/1893] bpb=1.090273 time=1132.1s + ttt_chunk [1061/1893] bpb=1.090511 time=1142.9s + ttt_chunk [1071/1893] bpb=1.090954 time=1153.6s + ttt_chunk [1081/1893] bpb=1.090943 time=1164.4s + ttt_chunk [1091/1893] bpb=1.091356 time=1175.2s + ttt_chunk [1101/1893] bpb=1.091488 time=1186.0s + ttt_chunk [1111/1893] bpb=1.091260 time=1196.7s + ttt_chunk [1121/1893] bpb=1.091181 time=1207.5s + ttt_chunk [1131/1893] bpb=1.091072 time=1218.3s + ttt_chunk [1141/1893] bpb=1.090937 time=1229.0s + ttt_chunk [1151/1893] bpb=1.090970 time=1239.8s + ttt_chunk [1161/1893] bpb=1.090188 time=1250.6s + ttt_chunk [1171/1893] bpb=1.090766 time=1261.4s + ttt_chunk [1181/1893] bpb=1.090269 time=1272.1s + ttt_chunk [1191/1893] bpb=1.089988 time=1282.9s + ttt_chunk [1201/1893] bpb=1.090561 time=1293.7s + ttt_chunk [1211/1893] bpb=1.089962 time=1304.4s + ttt_chunk [1221/1893] bpb=1.089661 time=1315.2s + ttt_chunk [1231/1893] bpb=1.089599 time=1326.0s + ttt_chunk [1241/1893] bpb=1.089410 time=1336.8s + ttt_chunk [1251/1893] bpb=1.089171 time=1347.5s + ttt_chunk [1261/1893] bpb=1.089113 time=1358.3s + ttt_chunk [1271/1893] bpb=1.088914 time=1369.1s + ttt_chunk [1281/1893] bpb=1.088710 time=1379.9s + ttt_chunk [1291/1893] bpb=1.088523 time=1390.6s + ttt_chunk [1301/1893] bpb=1.088161 time=1401.4s + ttt_chunk [1311/1893] bpb=1.087859 time=1412.2s + ttt_chunk [1321/1893] bpb=1.087701 time=1423.0s + ttt_chunk [1331/1893] bpb=1.087611 time=1433.7s + ttt_chunk [1341/1893] bpb=1.087498 time=1444.5s + ttt_chunk [1351/1893] bpb=1.087448 time=1455.3s + ttt_chunk [1361/1893] bpb=1.087658 time=1466.0s + ttt_chunk [1371/1893] bpb=1.087513 time=1476.8s + ttt_chunk [1381/1893] bpb=1.087452 time=1487.6s + ttt_chunk [1391/1893] bpb=1.086938 time=1498.4s + ttt_chunk [1401/1893] bpb=1.086952 time=1509.2s + ttt_chunk [1411/1893] bpb=1.086986 time=1519.9s + ttt_chunk [1421/1893] bpb=1.087243 time=1530.7s + ttt_chunk [1431/1893] bpb=1.087094 time=1541.5s + ttt_chunk [1441/1893] bpb=1.087766 time=1552.2s + ttt_chunk [1451/1893] bpb=1.087875 time=1563.0s + ttt_chunk [1461/1893] bpb=1.087616 time=1573.8s + ttt_chunk [1471/1893] bpb=1.088508 time=1584.6s + ttt_chunk [1481/1893] bpb=1.088318 time=1595.3s + ttt_chunk [1491/1893] bpb=1.088295 time=1606.1s + ttt_chunk [1501/1893] bpb=1.088428 time=1616.9s + ttt_chunk [1511/1893] bpb=1.088500 time=1627.7s + ttt_chunk [1521/1893] bpb=1.088516 time=1638.4s + ttt_chunk [1531/1893] bpb=1.088335 time=1649.2s + ttt_chunk [1541/1893] bpb=1.088300 time=1660.0s + ttt_chunk [1551/1893] bpb=1.088644 time=1670.8s + ttt_chunk [1561/1893] bpb=1.088779 time=1681.5s + ttt_chunk [1571/1893] bpb=1.088893 time=1692.3s + ttt_chunk [1581/1893] bpb=1.088999 time=1703.1s + ttt_chunk [1591/1893] bpb=1.088948 time=1713.9s + ttt_chunk [1601/1893] bpb=1.089129 time=1724.6s + ttt_chunk [1611/1893] bpb=1.089211 time=1735.4s + ttt_chunk [1621/1893] bpb=1.089078 time=1746.2s + ttt_chunk [1631/1893] bpb=1.089247 time=1756.9s + ttt_chunk [1641/1893] bpb=1.089113 time=1767.7s + ttt_chunk [1651/1893] bpb=1.089025 time=1778.5s + ttt_chunk [1661/1893] bpb=1.088917 time=1789.3s + ttt_chunk [1671/1893] bpb=1.089273 time=1800.0s + ttt_chunk [1681/1893] bpb=1.089519 time=1810.8s + ttt_chunk [1691/1893] bpb=1.089483 time=1821.6s + ttt_chunk [1701/1893] bpb=1.089475 time=1832.3s + ttt_chunk [1711/1893] bpb=1.089308 time=1843.1s + ttt_chunk [1721/1893] bpb=1.089158 time=1853.9s + ttt_chunk [1731/1893] bpb=1.089108 time=1864.6s + ttt_chunk [1741/1893] bpb=1.088877 time=1875.4s + ttt_chunk [1751/1893] bpb=1.088730 time=1886.2s + ttt_chunk [1761/1893] bpb=1.088814 time=1897.0s + ttt_chunk [1771/1893] bpb=1.088740 time=1907.7s + ttt_chunk [1781/1893] bpb=1.088687 time=1918.5s + ttt_chunk [1791/1893] bpb=1.088300 time=1929.3s + ttt_chunk [1801/1893] bpb=1.088297 time=1940.0s + ttt_chunk [1811/1893] bpb=1.088139 time=1950.8s + ttt_chunk [1821/1893] bpb=1.088176 time=1961.6s + ttt_chunk [1831/1893] bpb=1.087789 time=1972.3s + ttt_chunk [1841/1893] bpb=1.087730 time=1983.1s + ttt_chunk [1851/1893] bpb=1.087510 time=1993.9s + ttt_chunk [1861/1893] bpb=1.087051 time=2004.6s + ttt_chunk [1871/1893] bpb=1.086903 time=2015.4s + ttt_chunk [1881/1893] bpb=1.086533 time=2026.2s + ttt_chunk [1891/1893] bpb=1.086383 time=2037.0s + ttt_chunk [1893/1893] bpb=1.086404 time=2038.7s +ttt_sliding:done val_loss=1.832524 val_bpb=1.085327 elapsed=2038.8s +legal_ttt val_loss:1.8325 val_bpb:1.0853 eval_time:2039427ms +legal_ttt_exact val_loss:1.83252444 val_bpb:1.08532707 diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed2024.txt b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed2024.txt new file mode 100644 index 0000000000..43da236768 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed2024.txt @@ -0,0 +1,2264 @@ +logs/train_step50k_seed2024.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:50000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:2024 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/50000 val_loss:6.9290 val_bpb:4.1037 artifact_bytes:4577366 model_int6_lzma:4485864 train_time:0ms step_avg:0.02ms +step:1/50000 train_loss:6.9301 train_time:1034ms step_avg:1033.70ms +step:2/50000 train_loss:8.6930 train_time:1582ms step_avg:790.83ms +step:3/50000 train_loss:7.6769 train_time:2438ms step_avg:812.61ms +step:4/50000 train_loss:7.0843 train_time:3295ms step_avg:823.64ms +step:5/50000 train_loss:7.0619 train_time:4160ms step_avg:832.05ms +step:6/50000 train_loss:7.0070 train_time:5017ms step_avg:836.17ms +step:7/50000 train_loss:7.0103 train_time:5874ms step_avg:839.17ms +step:8/50000 train_loss:6.8082 train_time:6731ms step_avg:841.36ms +step:9/50000 train_loss:6.5753 train_time:7588ms step_avg:843.06ms +step:10/50000 train_loss:6.2322 train_time:8444ms step_avg:844.43ms +step:500/50000 train_loss:2.3518 train_time:414482ms step_avg:828.96ms +step:1000/50000 train_loss:2.2638 train_time:828787ms step_avg:828.79ms +step:1500/50000 train_loss:2.1800 train_time:1243082ms step_avg:828.72ms +step:2000/50000 train_loss:2.0335 train_time:1657356ms step_avg:828.68ms +step:2500/50000 train_loss:2.1582 train_time:2071616ms step_avg:828.65ms +step:2500/50000 val_loss:2.1193 val_bpb:1.2552 artifact_bytes:13069870 model_int6_lzma:12978368 train_time:2071970ms step_avg:828.79ms +step:3000/50000 train_loss:2.1231 train_time:2486102ms step_avg:828.70ms +step:3500/50000 train_loss:2.1889 train_time:2900363ms step_avg:828.68ms +step:4000/50000 train_loss:2.1587 train_time:3314567ms step_avg:828.64ms +step:4500/50000 train_loss:2.1296 train_time:3728769ms step_avg:828.62ms +step:5000/50000 train_loss:2.0917 train_time:4142951ms step_avg:828.59ms +step:5000/50000 val_loss:2.0738 val_bpb:1.2282 artifact_bytes:13892598 model_int6_lzma:13801096 train_time:4143296ms step_avg:828.66ms +step:5500/50000 train_loss:2.1354 train_time:4558121ms step_avg:828.75ms +step:6000/50000 train_loss:2.0288 train_time:4972348ms step_avg:828.72ms +step:6500/50000 train_loss:2.2960 train_time:5386597ms step_avg:828.71ms +step:7000/50000 train_loss:1.9276 train_time:5800868ms step_avg:828.70ms +step:7500/50000 train_loss:2.0520 train_time:6215138ms step_avg:828.69ms +step:7500/50000 val_loss:2.0574 val_bpb:1.2185 artifact_bytes:15496510 model_int6_lzma:15405008 train_time:6215491ms step_avg:828.73ms +step:8000/50000 train_loss:2.0031 train_time:6629599ms step_avg:828.70ms +step:8500/50000 train_loss:1.9843 train_time:7043865ms step_avg:828.69ms +step:9000/50000 train_loss:2.0993 train_time:7458118ms step_avg:828.68ms +step:9500/50000 train_loss:2.2518 train_time:7872395ms step_avg:828.67ms +step:10000/50000 train_loss:2.0636 train_time:8286710ms step_avg:828.67ms +step:10000/50000 val_loss:2.0307 val_bpb:1.2027 artifact_bytes:17215490 model_int6_lzma:17123988 train_time:8287065ms step_avg:828.71ms +step:10500/50000 train_loss:2.0990 train_time:8701002ms step_avg:828.67ms +step:11000/50000 train_loss:1.9894 train_time:9115302ms step_avg:828.66ms +step:11500/50000 train_loss:1.9473 train_time:9529567ms step_avg:828.66ms +step:12000/50000 train_loss:1.9629 train_time:9943872ms step_avg:828.66ms +step:12500/50000 train_loss:1.9684 train_time:10358175ms step_avg:828.65ms +step:12500/50000 val_loss:2.0255 val_bpb:1.1996 artifact_bytes:17232630 model_int6_lzma:17141128 train_time:10358520ms step_avg:828.68ms +step:13000/50000 train_loss:1.9917 train_time:10772470ms step_avg:828.65ms +step:13500/50000 train_loss:2.1074 train_time:11186772ms step_avg:828.65ms +step:14000/50000 train_loss:1.9099 train_time:11601067ms step_avg:828.65ms +step:14500/50000 train_loss:2.0934 train_time:12015363ms step_avg:828.65ms +step:15000/50000 train_loss:2.0417 train_time:12429650ms step_avg:828.64ms +step:15000/50000 val_loss:2.0169 val_bpb:1.1945 artifact_bytes:17237242 model_int6_lzma:17145740 train_time:12430005ms step_avg:828.67ms +step:15500/50000 train_loss:2.0146 train_time:12844165ms step_avg:828.66ms +step:16000/50000 train_loss:2.0983 train_time:13258417ms step_avg:828.65ms +step:16500/50000 train_loss:2.0789 train_time:13672658ms step_avg:828.65ms +step:17000/50000 train_loss:2.0796 train_time:14086881ms step_avg:828.64ms +step:17500/50000 train_loss:2.0700 train_time:14501139ms step_avg:828.64ms +step:17500/50000 val_loss:2.0131 val_bpb:1.1923 artifact_bytes:17219546 model_int6_lzma:17128044 train_time:14501493ms step_avg:828.66ms +step:18000/50000 train_loss:1.9560 train_time:14915434ms step_avg:828.64ms +step:18500/50000 train_loss:1.9202 train_time:15329716ms step_avg:828.63ms +step:19000/50000 train_loss:1.9069 train_time:15743980ms step_avg:828.63ms +step:19500/50000 train_loss:2.1009 train_time:16158242ms step_avg:828.63ms +step:20000/50000 train_loss:2.0161 train_time:16572503ms step_avg:828.63ms +step:20000/50000 val_loss:2.0113 val_bpb:1.1912 artifact_bytes:17213066 model_int6_lzma:17121564 train_time:16572848ms step_avg:828.64ms +step:20500/50000 train_loss:2.0783 train_time:16986763ms step_avg:828.62ms +step:21000/50000 train_loss:1.9828 train_time:17401026ms step_avg:828.62ms +step:21500/50000 train_loss:2.0095 train_time:17815277ms step_avg:828.62ms +step:22000/50000 train_loss:2.0267 train_time:18229521ms step_avg:828.61ms +step:22500/50000 train_loss:2.0491 train_time:18643764ms step_avg:828.61ms +step:22500/50000 val_loss:2.0098 val_bpb:1.1903 artifact_bytes:17206306 model_int6_lzma:17114804 train_time:18644117ms step_avg:828.63ms +step:23000/50000 train_loss:1.9758 train_time:19058049ms step_avg:828.61ms +step:23500/50000 train_loss:2.0687 train_time:19472294ms step_avg:828.61ms +step:24000/50000 train_loss:1.9674 train_time:19886543ms step_avg:828.61ms +step:24500/50000 train_loss:2.0122 train_time:20300768ms step_avg:828.60ms +step:25000/50000 train_loss:2.0222 train_time:20714987ms step_avg:828.60ms +step:25000/50000 val_loss:2.0100 val_bpb:1.1904 artifact_bytes:17201238 model_int6_lzma:17109736 train_time:20715346ms step_avg:828.61ms +step:25500/50000 train_loss:2.0314 train_time:21129712ms step_avg:828.62ms +step:26000/50000 train_loss:2.0764 train_time:21543928ms step_avg:828.61ms +step:26500/50000 train_loss:2.0794 train_time:21958152ms step_avg:828.61ms +step:27000/50000 train_loss:1.9939 train_time:22372402ms step_avg:828.61ms +step:27500/50000 train_loss:1.9893 train_time:22786672ms step_avg:828.61ms +step:27500/50000 val_loss:2.0064 val_bpb:1.1883 artifact_bytes:17193574 model_int6_lzma:17102072 train_time:22787027ms step_avg:828.62ms +step:28000/50000 train_loss:1.8326 train_time:23201412ms step_avg:828.62ms +step:28500/50000 train_loss:2.0206 train_time:23615709ms step_avg:828.62ms +step:29000/50000 train_loss:2.0331 train_time:24029950ms step_avg:828.62ms +step:29500/50000 train_loss:2.0431 train_time:24444160ms step_avg:828.62ms +step:30000/50000 train_loss:1.9783 train_time:24858509ms step_avg:828.62ms +step:30000/50000 val_loss:2.0033 val_bpb:1.1865 artifact_bytes:17190822 model_int6_lzma:17099320 train_time:24858862ms step_avg:828.63ms +step:30500/50000 train_loss:2.0243 train_time:25272792ms step_avg:828.62ms +step:31000/50000 train_loss:2.0254 train_time:25687051ms step_avg:828.61ms +step:31500/50000 train_loss:1.9576 train_time:26101287ms step_avg:828.61ms +step:32000/50000 train_loss:2.0433 train_time:26515527ms step_avg:828.61ms +step:32500/50000 train_loss:1.9925 train_time:26929743ms step_avg:828.61ms +step:32500/50000 val_loss:1.9945 val_bpb:1.1813 artifact_bytes:17052022 model_int6_lzma:16960520 train_time:26930097ms step_avg:828.62ms +step:33000/50000 train_loss:1.9740 train_time:27344003ms step_avg:828.61ms +step:33500/50000 train_loss:1.9531 train_time:27758223ms step_avg:828.60ms +step:34000/50000 train_loss:2.0600 train_time:28172443ms step_avg:828.60ms +step:34500/50000 train_loss:1.9874 train_time:28586649ms step_avg:828.60ms +step:35000/50000 train_loss:1.9774 train_time:29000890ms step_avg:828.60ms +step:35000/50000 val_loss:1.9826 val_bpb:1.1742 artifact_bytes:16786866 model_int6_lzma:16695364 train_time:29001242ms step_avg:828.61ms +step:35500/50000 train_loss:2.0157 train_time:29415143ms step_avg:828.60ms +step:36000/50000 train_loss:1.9959 train_time:29829409ms step_avg:828.59ms +step:36500/50000 train_loss:2.0246 train_time:30243660ms step_avg:828.59ms +step:37000/50000 train_loss:1.8805 train_time:30657895ms step_avg:828.59ms +step:37500/50000 train_loss:2.0089 train_time:31072136ms step_avg:828.59ms +step:37500/50000 val_loss:1.9715 val_bpb:1.1677 artifact_bytes:16553402 model_int6_lzma:16461900 train_time:31072495ms step_avg:828.60ms +step:38000/50000 train_loss:1.9977 train_time:31486795ms step_avg:828.60ms +step:38500/50000 train_loss:1.9700 train_time:31901084ms step_avg:828.60ms +step:39000/50000 train_loss:1.8706 train_time:32315328ms step_avg:828.60ms +step:39500/50000 train_loss:1.9330 train_time:32729566ms step_avg:828.60ms +step:40000/50000 train_loss:1.9198 train_time:33143840ms step_avg:828.60ms +step:40000/50000 val_loss:1.9573 val_bpb:1.1592 artifact_bytes:16067974 model_int6_lzma:15976472 train_time:33144193ms step_avg:828.60ms +step:40500/50000 train_loss:2.3025 train_time:33559060ms step_avg:828.62ms +step:41000/50000 train_loss:1.8542 train_time:33973289ms step_avg:828.62ms +step:41500/50000 train_loss:1.7735 train_time:34387489ms step_avg:828.61ms +step:42000/50000 train_loss:1.9641 train_time:34801713ms step_avg:828.61ms +step:42500/50000 train_loss:1.9355 train_time:35215965ms step_avg:828.61ms +step:42500/50000 val_loss:1.9406 val_bpb:1.1493 artifact_bytes:15598778 model_int6_lzma:15507276 train_time:35216309ms step_avg:828.62ms +step:43000/50000 train_loss:2.0262 train_time:35630707ms step_avg:828.62ms +step:43500/50000 train_loss:2.1248 train_time:36044972ms step_avg:828.62ms +step:44000/50000 train_loss:1.8095 train_time:36459194ms step_avg:828.62ms +step:44500/50000 train_loss:1.8511 train_time:36873375ms step_avg:828.62ms +step:45000/50000 train_loss:1.9509 train_time:37287600ms step_avg:828.61ms +step:45000/50000 val_loss:1.9164 val_bpb:1.1350 artifact_bytes:15025010 model_int6_lzma:14933508 train_time:37287936ms step_avg:828.62ms +step:45500/50000 train_loss:1.9787 train_time:37701819ms step_avg:828.61ms +step:46000/50000 train_loss:1.9022 train_time:38116019ms step_avg:828.61ms +swa:start step:46150 +step:46500/50000 train_loss:1.8170 train_time:38530797ms step_avg:828.62ms +step:47000/50000 train_loss:1.9271 train_time:38945732ms step_avg:828.63ms +late_qat:enabled step:47076 scale:0.1499 +step:47500/50000 train_loss:1.8898 train_time:39360682ms step_avg:828.65ms +step:47500/50000 val_loss:1.8873 val_bpb:1.1178 artifact_bytes:14743726 model_int6_lzma:14652224 train_time:39361076ms step_avg:828.65ms +step:48000/50000 train_loss:1.8844 train_time:39775639ms step_avg:828.66ms +step:48500/50000 train_loss:1.8388 train_time:40190570ms step_avg:828.67ms +step:49000/50000 train_loss:1.7883 train_time:40605639ms step_avg:828.69ms +step:49500/50000 train_loss:1.8231 train_time:41020558ms step_avg:828.70ms +step:50000/50000 train_loss:1.8412 train_time:41435439ms step_avg:828.71ms +step:50000/50000 val_loss:1.8494 val_bpb:1.0953 artifact_bytes:14353974 model_int6_lzma:14262472 train_time:41435835ms step_avg:828.72ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8488 val_bpb:1.0949 eval_time:23122ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14279664 bytes +Total submission size int6+lzma: 14371166 bytes +final_int6_roundtrip val_loss:1.8970 val_bpb:1.1235 eval_time:23858ms +final_int6_roundtrip_exact val_loss:1.89698682 val_bpb:1.12350245 +final_int6_sliding_window val_loss:1.8571 val_bpb:1.0999 stride:64 eval_time:662955ms +final_int6_sliding_window_exact val_loss:1.85714441 val_bpb:1.09990844 +final_int8_zlib_roundtrip_exact val_loss:1.85714441 val_bpb:1.09990844 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.148757 time=1.2s + ttt_chunk [11/1893] bpb=1.102178 time=11.9s + ttt_chunk [21/1893] bpb=1.105358 time=22.7s + ttt_chunk [31/1893] bpb=1.108851 time=33.5s + ttt_chunk [41/1893] bpb=1.096845 time=44.3s + ttt_chunk [51/1893] bpb=1.092842 time=55.1s + ttt_chunk [61/1893] bpb=1.097112 time=65.9s + ttt_chunk [71/1893] bpb=1.093608 time=76.6s + ttt_chunk [81/1893] bpb=1.093503 time=87.4s + ttt_chunk [91/1893] bpb=1.092102 time=98.2s + ttt_chunk [101/1893] bpb=1.094825 time=109.0s + ttt_chunk [111/1893] bpb=1.095924 time=119.8s + ttt_chunk [121/1893] bpb=1.092419 time=130.6s + ttt_chunk [131/1893] bpb=1.091950 time=141.4s + ttt_chunk [141/1893] bpb=1.091511 time=152.2s + ttt_chunk [151/1893] bpb=1.094588 time=163.0s + ttt_chunk [161/1893] bpb=1.096045 time=173.7s + ttt_chunk [171/1893] bpb=1.096592 time=184.5s + ttt_chunk [181/1893] bpb=1.096579 time=195.3s + ttt_chunk [191/1893] bpb=1.099678 time=206.1s + ttt_chunk [201/1893] bpb=1.099857 time=216.9s + ttt_chunk [211/1893] bpb=1.097320 time=227.7s + ttt_chunk [221/1893] bpb=1.099322 time=238.5s + ttt_chunk [231/1893] bpb=1.098750 time=249.2s + ttt_chunk [241/1893] bpb=1.098538 time=260.0s + ttt_chunk [251/1893] bpb=1.096996 time=270.8s + ttt_chunk [261/1893] bpb=1.095315 time=281.6s + ttt_chunk [271/1893] bpb=1.093958 time=292.4s + ttt_chunk [281/1893] bpb=1.096219 time=303.2s + ttt_chunk [291/1893] bpb=1.096848 time=314.0s + ttt_chunk [301/1893] bpb=1.097386 time=324.7s + ttt_chunk [311/1893] bpb=1.098887 time=335.5s + ttt_chunk [321/1893] bpb=1.100300 time=346.3s + ttt_chunk [331/1893] bpb=1.100039 time=357.1s + ttt_chunk [341/1893] bpb=1.100387 time=367.9s + ttt_chunk [351/1893] bpb=1.101641 time=378.7s + ttt_chunk [361/1893] bpb=1.102780 time=389.5s + ttt_chunk [371/1893] bpb=1.102319 time=400.3s + ttt_chunk [381/1893] bpb=1.102277 time=411.0s + ttt_chunk [391/1893] bpb=1.101936 time=421.8s + ttt_chunk [401/1893] bpb=1.100543 time=432.6s + ttt_chunk [411/1893] bpb=1.099425 time=443.4s + ttt_chunk [421/1893] bpb=1.098865 time=454.2s + ttt_chunk [431/1893] bpb=1.099505 time=465.0s + ttt_chunk [441/1893] bpb=1.099299 time=475.8s + ttt_chunk [451/1893] bpb=1.099028 time=486.5s + ttt_chunk [461/1893] bpb=1.098360 time=497.3s + ttt_chunk [471/1893] bpb=1.097998 time=508.1s + ttt_chunk [481/1893] bpb=1.097693 time=518.9s + ttt_chunk [491/1893] bpb=1.097377 time=529.7s + ttt_chunk [501/1893] bpb=1.096823 time=540.5s + ttt_chunk [511/1893] bpb=1.096195 time=551.3s + ttt_chunk [521/1893] bpb=1.095062 time=562.1s + ttt_chunk [531/1893] bpb=1.095173 time=572.8s + ttt_chunk [541/1893] bpb=1.095103 time=583.6s + ttt_chunk [551/1893] bpb=1.093925 time=594.4s + ttt_chunk [561/1893] bpb=1.094427 time=605.2s + ttt_chunk [571/1893] bpb=1.093719 time=616.0s + ttt_chunk [581/1893] bpb=1.093184 time=626.8s + ttt_chunk [591/1893] bpb=1.092495 time=637.6s + ttt_chunk [601/1893] bpb=1.093161 time=648.3s + ttt_chunk [611/1893] bpb=1.092714 time=659.1s + ttt_chunk [621/1893] bpb=1.092635 time=669.9s + ttt_chunk [631/1893] bpb=1.093026 time=680.7s + ttt_chunk [641/1893] bpb=1.092802 time=691.5s + ttt_chunk [651/1893] bpb=1.092816 time=702.3s + ttt_chunk [661/1893] bpb=1.092694 time=713.1s + ttt_chunk [671/1893] bpb=1.092341 time=723.9s + ttt_chunk [681/1893] bpb=1.092594 time=734.7s + ttt_chunk [691/1893] bpb=1.093260 time=745.4s + ttt_chunk [701/1893] bpb=1.092485 time=756.2s + ttt_chunk [711/1893] bpb=1.093061 time=767.0s + ttt_chunk [721/1893] bpb=1.092678 time=777.8s + ttt_chunk [731/1893] bpb=1.093086 time=788.6s + ttt_chunk [741/1893] bpb=1.093053 time=799.4s + ttt_chunk [751/1893] bpb=1.092696 time=810.2s + ttt_chunk [761/1893] bpb=1.092495 time=820.9s + ttt_chunk [771/1893] bpb=1.092281 time=831.7s + ttt_chunk [781/1893] bpb=1.092808 time=842.5s + ttt_chunk [791/1893] bpb=1.092483 time=853.3s + ttt_chunk [801/1893] bpb=1.092537 time=864.1s + ttt_chunk [811/1893] bpb=1.092037 time=874.9s + ttt_chunk [821/1893] bpb=1.091822 time=885.7s + ttt_chunk [831/1893] bpb=1.091358 time=896.5s + ttt_chunk [841/1893] bpb=1.090731 time=907.3s + ttt_chunk [851/1893] bpb=1.090589 time=918.1s + ttt_chunk [861/1893] bpb=1.090713 time=928.9s + ttt_chunk [871/1893] bpb=1.090745 time=939.7s + ttt_chunk [881/1893] bpb=1.090745 time=950.5s + ttt_chunk [891/1893] bpb=1.090552 time=961.2s + ttt_chunk [901/1893] bpb=1.090540 time=972.0s + ttt_chunk [911/1893] bpb=1.090551 time=982.8s + ttt_chunk [921/1893] bpb=1.090909 time=993.6s + ttt_chunk [931/1893] bpb=1.090718 time=1004.4s + ttt_chunk [941/1893] bpb=1.090610 time=1015.2s + ttt_chunk [951/1893] bpb=1.090584 time=1026.0s + ttt_chunk [961/1893] bpb=1.090279 time=1036.8s + ttt_chunk [971/1893] bpb=1.090998 time=1047.5s + ttt_chunk [981/1893] bpb=1.091160 time=1058.3s + ttt_chunk [991/1893] bpb=1.091060 time=1069.1s + ttt_chunk [1001/1893] bpb=1.091203 time=1079.9s + ttt_chunk [1011/1893] bpb=1.091460 time=1090.7s + ttt_chunk [1021/1893] bpb=1.091612 time=1101.5s + ttt_chunk [1031/1893] bpb=1.092160 time=1112.3s + ttt_chunk [1041/1893] bpb=1.091833 time=1123.1s + ttt_chunk [1051/1893] bpb=1.091517 time=1133.9s + ttt_chunk [1061/1893] bpb=1.091746 time=1144.7s + ttt_chunk [1071/1893] bpb=1.092177 time=1155.4s + ttt_chunk [1081/1893] bpb=1.092185 time=1166.2s + ttt_chunk [1091/1893] bpb=1.092584 time=1177.0s + ttt_chunk [1101/1893] bpb=1.092718 time=1187.8s + ttt_chunk [1111/1893] bpb=1.092497 time=1198.6s + ttt_chunk [1121/1893] bpb=1.092419 time=1209.4s + ttt_chunk [1131/1893] bpb=1.092305 time=1220.2s + ttt_chunk [1141/1893] bpb=1.092149 time=1231.0s + ttt_chunk [1151/1893] bpb=1.092182 time=1241.8s + ttt_chunk [1161/1893] bpb=1.091393 time=1252.6s + ttt_chunk [1171/1893] bpb=1.091978 time=1263.4s + ttt_chunk [1181/1893] bpb=1.091471 time=1274.2s + ttt_chunk [1191/1893] bpb=1.091173 time=1284.9s + ttt_chunk [1201/1893] bpb=1.091725 time=1295.7s + ttt_chunk [1211/1893] bpb=1.091131 time=1306.5s + ttt_chunk [1221/1893] bpb=1.090838 time=1317.3s + ttt_chunk [1231/1893] bpb=1.090780 time=1328.1s + ttt_chunk [1241/1893] bpb=1.090593 time=1338.9s + ttt_chunk [1251/1893] bpb=1.090354 time=1349.7s + ttt_chunk [1261/1893] bpb=1.090289 time=1360.5s + ttt_chunk [1271/1893] bpb=1.090085 time=1371.2s + ttt_chunk [1281/1893] bpb=1.089887 time=1382.0s + ttt_chunk [1291/1893] bpb=1.089707 time=1392.8s + ttt_chunk [1301/1893] bpb=1.089346 time=1403.6s + ttt_chunk [1311/1893] bpb=1.089041 time=1414.4s + ttt_chunk [1321/1893] bpb=1.088865 time=1425.2s + ttt_chunk [1331/1893] bpb=1.088783 time=1436.0s + ttt_chunk [1341/1893] bpb=1.088666 time=1446.8s + ttt_chunk [1351/1893] bpb=1.088607 time=1457.6s + ttt_chunk [1361/1893] bpb=1.088806 time=1468.4s + ttt_chunk [1371/1893] bpb=1.088660 time=1479.1s + ttt_chunk [1381/1893] bpb=1.088583 time=1489.9s + ttt_chunk [1391/1893] bpb=1.088057 time=1500.7s + ttt_chunk [1401/1893] bpb=1.088066 time=1511.5s + ttt_chunk [1411/1893] bpb=1.088092 time=1522.3s + ttt_chunk [1421/1893] bpb=1.088333 time=1533.1s + ttt_chunk [1431/1893] bpb=1.088173 time=1543.9s + ttt_chunk [1441/1893] bpb=1.088817 time=1554.7s + ttt_chunk [1451/1893] bpb=1.088928 time=1565.4s + ttt_chunk [1461/1893] bpb=1.088670 time=1576.2s + ttt_chunk [1471/1893] bpb=1.089574 time=1587.0s + ttt_chunk [1481/1893] bpb=1.089388 time=1597.8s + ttt_chunk [1491/1893] bpb=1.089376 time=1608.6s + ttt_chunk [1501/1893] bpb=1.089522 time=1619.4s + ttt_chunk [1511/1893] bpb=1.089582 time=1630.2s + ttt_chunk [1521/1893] bpb=1.089593 time=1641.0s + ttt_chunk [1531/1893] bpb=1.089397 time=1651.8s + ttt_chunk [1541/1893] bpb=1.089357 time=1662.6s + ttt_chunk [1551/1893] bpb=1.089710 time=1673.3s + ttt_chunk [1561/1893] bpb=1.089842 time=1684.1s + ttt_chunk [1571/1893] bpb=1.089962 time=1694.9s + ttt_chunk [1581/1893] bpb=1.090055 time=1705.7s + ttt_chunk [1591/1893] bpb=1.090000 time=1716.5s + ttt_chunk [1601/1893] bpb=1.090185 time=1727.3s + ttt_chunk [1611/1893] bpb=1.090261 time=1738.1s + ttt_chunk [1621/1893] bpb=1.090129 time=1748.9s + ttt_chunk [1631/1893] bpb=1.090299 time=1759.6s + ttt_chunk [1641/1893] bpb=1.090181 time=1770.4s + ttt_chunk [1651/1893] bpb=1.090078 time=1781.2s + ttt_chunk [1661/1893] bpb=1.089971 time=1792.0s + ttt_chunk [1671/1893] bpb=1.090309 time=1802.8s + ttt_chunk [1681/1893] bpb=1.090559 time=1813.6s + ttt_chunk [1691/1893] bpb=1.090524 time=1824.4s + ttt_chunk [1701/1893] bpb=1.090516 time=1835.2s + ttt_chunk [1711/1893] bpb=1.090351 time=1846.0s + ttt_chunk [1721/1893] bpb=1.090207 time=1856.8s + ttt_chunk [1731/1893] bpb=1.090145 time=1867.5s + ttt_chunk [1741/1893] bpb=1.089909 time=1878.3s + ttt_chunk [1751/1893] bpb=1.089769 time=1889.1s + ttt_chunk [1761/1893] bpb=1.089845 time=1899.9s + ttt_chunk [1771/1893] bpb=1.089780 time=1910.7s + ttt_chunk [1781/1893] bpb=1.089728 time=1921.5s + ttt_chunk [1791/1893] bpb=1.089337 time=1932.3s + ttt_chunk [1801/1893] bpb=1.089329 time=1943.1s + ttt_chunk [1811/1893] bpb=1.089165 time=1953.9s + ttt_chunk [1821/1893] bpb=1.089201 time=1964.7s + ttt_chunk [1831/1893] bpb=1.088803 time=1975.5s + ttt_chunk [1841/1893] bpb=1.088748 time=1986.3s + ttt_chunk [1851/1893] bpb=1.088540 time=1997.0s + ttt_chunk [1861/1893] bpb=1.088077 time=2007.8s + ttt_chunk [1871/1893] bpb=1.087916 time=2018.6s + ttt_chunk [1881/1893] bpb=1.087550 time=2029.4s + ttt_chunk [1891/1893] bpb=1.087393 time=2040.2s + ttt_chunk [1893/1893] bpb=1.087411 time=2041.9s +ttt_sliding:done val_loss=1.834459 val_bpb=1.086473 elapsed=2042.0s +legal_ttt val_loss:1.8345 val_bpb:1.0865 eval_time:2042650ms +legal_ttt_exact val_loss:1.83445894 val_bpb:1.08647279 +.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + # Estimate artifact size (int6+lzma) at each validation step + try: + sd_snap = {k: v.detach().cpu() for k, v in base_model.state_dict().items() + if "mtp_heads" not in k} + unbanked_snap = _unbank_state_dict(sd_snap, args.num_layers) + q_snap, m_snap = mixed_quantize_int6(unbanked_snap, {"mlp", "attn"}) + snap_buf = io.BytesIO() + torch.save({"w": q_snap, "m": m_snap}, snap_buf) + snap_compressed = len(lzma.compress(snap_buf.getvalue(), preset=6)) + snap_code_bytes = len(code.encode("utf-8")) + snap_total = snap_compressed + snap_code_bytes + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_bytes:{snap_total} model_int6_lzma:{snap_compressed} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + except Exception as e: + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_size:error({e}) " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.10.14 (main, May 6 2024, 19:42:50) [GCC 11.2.0] +Running PyTorch 2.10.0+cu128 +Mon Apr 6 04:20:28 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 550.127.08 Driver Version: 550.127.08 CUDA Version: 12.4 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA A100-SXM4-80GB On | 00000000:27:00.0 Off | On | +| N/A 36C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA A100-SXM4-80GB On | 00000000:2A:00.0 Off | On | +| N/A 33C P0 96W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA A100-SXM4-80GB On | 00000000:51:00.0 Off | On | +| N/A 33C P0 100W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA A100-SXM4-80GB On | 00000000:57:00.0 Off | On | +| N/A 38C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA A100-SXM4-80GB On | 00000000:9E:00.0 Off | On | +| N/A 37C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA A100-SXM4-80GB On | 00000000:A4:00.0 Off | On | +| N/A 33C P0 101W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA A100-SXM4-80GB On | 00000000:CA:00.0 Off | On | +| N/A 38C P0 103W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| MIG devices: | ++------------------+----------------------------------+-----------+-----------------------+ +| GPU GI CI MIG | Memory-Usage | Vol| Shared | +| ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG | +| | | ECC| | +|==================+==================================+===========+=======================| +| 0 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 2 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 3 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 1 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 2 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 6 2 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 5474 C /opt/conda/bin/python3 0MiB | +| 1 N/A N/A 5475 C /opt/conda/bin/python3 0MiB | +| 2 N/A N/A 5476 C /opt/conda/bin/python3 0MiB | +| 3 N/A N/A 5477 C /opt/conda/bin/python3 0MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:50000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:2024 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/50000 val_loss:6.9290 val_bpb:4.1037 artifact_bytes:4577366 model_int6_lzma:4485864 train_time:0ms step_avg:0.02ms +step:1/50000 train_loss:6.9301 train_time:1034ms step_avg:1033.70ms +step:2/50000 train_loss:8.6930 train_time:1582ms step_avg:790.83ms +step:3/50000 train_loss:7.6769 train_time:2438ms step_avg:812.61ms +step:4/50000 train_loss:7.0843 train_time:3295ms step_avg:823.64ms +step:5/50000 train_loss:7.0619 train_time:4160ms step_avg:832.05ms +step:6/50000 train_loss:7.0070 train_time:5017ms step_avg:836.17ms +step:7/50000 train_loss:7.0103 train_time:5874ms step_avg:839.17ms +step:8/50000 train_loss:6.8082 train_time:6731ms step_avg:841.36ms +step:9/50000 train_loss:6.5753 train_time:7588ms step_avg:843.06ms +step:10/50000 train_loss:6.2322 train_time:8444ms step_avg:844.43ms +step:500/50000 train_loss:2.3518 train_time:414482ms step_avg:828.96ms +step:1000/50000 train_loss:2.2638 train_time:828787ms step_avg:828.79ms +step:1500/50000 train_loss:2.1800 train_time:1243082ms step_avg:828.72ms +step:2000/50000 train_loss:2.0335 train_time:1657356ms step_avg:828.68ms +step:2500/50000 train_loss:2.1582 train_time:2071616ms step_avg:828.65ms +step:2500/50000 val_loss:2.1193 val_bpb:1.2552 artifact_bytes:13069870 model_int6_lzma:12978368 train_time:2071970ms step_avg:828.79ms +step:3000/50000 train_loss:2.1231 train_time:2486102ms step_avg:828.70ms +step:3500/50000 train_loss:2.1889 train_time:2900363ms step_avg:828.68ms +step:4000/50000 train_loss:2.1587 train_time:3314567ms step_avg:828.64ms +step:4500/50000 train_loss:2.1296 train_time:3728769ms step_avg:828.62ms +step:5000/50000 train_loss:2.0917 train_time:4142951ms step_avg:828.59ms +step:5000/50000 val_loss:2.0738 val_bpb:1.2282 artifact_bytes:13892598 model_int6_lzma:13801096 train_time:4143296ms step_avg:828.66ms +step:5500/50000 train_loss:2.1354 train_time:4558121ms step_avg:828.75ms +step:6000/50000 train_loss:2.0288 train_time:4972348ms step_avg:828.72ms +step:6500/50000 train_loss:2.2960 train_time:5386597ms step_avg:828.71ms +step:7000/50000 train_loss:1.9276 train_time:5800868ms step_avg:828.70ms +step:7500/50000 train_loss:2.0520 train_time:6215138ms step_avg:828.69ms +step:7500/50000 val_loss:2.0574 val_bpb:1.2185 artifact_bytes:15496510 model_int6_lzma:15405008 train_time:6215491ms step_avg:828.73ms +step:8000/50000 train_loss:2.0031 train_time:6629599ms step_avg:828.70ms +step:8500/50000 train_loss:1.9843 train_time:7043865ms step_avg:828.69ms +step:9000/50000 train_loss:2.0993 train_time:7458118ms step_avg:828.68ms +step:9500/50000 train_loss:2.2518 train_time:7872395ms step_avg:828.67ms +step:10000/50000 train_loss:2.0636 train_time:8286710ms step_avg:828.67ms +step:10000/50000 val_loss:2.0307 val_bpb:1.2027 artifact_bytes:17215490 model_int6_lzma:17123988 train_time:8287065ms step_avg:828.71ms +step:10500/50000 train_loss:2.0990 train_time:8701002ms step_avg:828.67ms +step:11000/50000 train_loss:1.9894 train_time:9115302ms step_avg:828.66ms +step:11500/50000 train_loss:1.9473 train_time:9529567ms step_avg:828.66ms +step:12000/50000 train_loss:1.9629 train_time:9943872ms step_avg:828.66ms +step:12500/50000 train_loss:1.9684 train_time:10358175ms step_avg:828.65ms +step:12500/50000 val_loss:2.0255 val_bpb:1.1996 artifact_bytes:17232630 model_int6_lzma:17141128 train_time:10358520ms step_avg:828.68ms +step:13000/50000 train_loss:1.9917 train_time:10772470ms step_avg:828.65ms +step:13500/50000 train_loss:2.1074 train_time:11186772ms step_avg:828.65ms +step:14000/50000 train_loss:1.9099 train_time:11601067ms step_avg:828.65ms +step:14500/50000 train_loss:2.0934 train_time:12015363ms step_avg:828.65ms +step:15000/50000 train_loss:2.0417 train_time:12429650ms step_avg:828.64ms +step:15000/50000 val_loss:2.0169 val_bpb:1.1945 artifact_bytes:17237242 model_int6_lzma:17145740 train_time:12430005ms step_avg:828.67ms +step:15500/50000 train_loss:2.0146 train_time:12844165ms step_avg:828.66ms +step:16000/50000 train_loss:2.0983 train_time:13258417ms step_avg:828.65ms +step:16500/50000 train_loss:2.0789 train_time:13672658ms step_avg:828.65ms +step:17000/50000 train_loss:2.0796 train_time:14086881ms step_avg:828.64ms +step:17500/50000 train_loss:2.0700 train_time:14501139ms step_avg:828.64ms +step:17500/50000 val_loss:2.0131 val_bpb:1.1923 artifact_bytes:17219546 model_int6_lzma:17128044 train_time:14501493ms step_avg:828.66ms +step:18000/50000 train_loss:1.9560 train_time:14915434ms step_avg:828.64ms +step:18500/50000 train_loss:1.9202 train_time:15329716ms step_avg:828.63ms +step:19000/50000 train_loss:1.9069 train_time:15743980ms step_avg:828.63ms +step:19500/50000 train_loss:2.1009 train_time:16158242ms step_avg:828.63ms +step:20000/50000 train_loss:2.0161 train_time:16572503ms step_avg:828.63ms +step:20000/50000 val_loss:2.0113 val_bpb:1.1912 artifact_bytes:17213066 model_int6_lzma:17121564 train_time:16572848ms step_avg:828.64ms +step:20500/50000 train_loss:2.0783 train_time:16986763ms step_avg:828.62ms +step:21000/50000 train_loss:1.9828 train_time:17401026ms step_avg:828.62ms +step:21500/50000 train_loss:2.0095 train_time:17815277ms step_avg:828.62ms +step:22000/50000 train_loss:2.0267 train_time:18229521ms step_avg:828.61ms +step:22500/50000 train_loss:2.0491 train_time:18643764ms step_avg:828.61ms +step:22500/50000 val_loss:2.0098 val_bpb:1.1903 artifact_bytes:17206306 model_int6_lzma:17114804 train_time:18644117ms step_avg:828.63ms +step:23000/50000 train_loss:1.9758 train_time:19058049ms step_avg:828.61ms +step:23500/50000 train_loss:2.0687 train_time:19472294ms step_avg:828.61ms +step:24000/50000 train_loss:1.9674 train_time:19886543ms step_avg:828.61ms +step:24500/50000 train_loss:2.0122 train_time:20300768ms step_avg:828.60ms +step:25000/50000 train_loss:2.0222 train_time:20714987ms step_avg:828.60ms +step:25000/50000 val_loss:2.0100 val_bpb:1.1904 artifact_bytes:17201238 model_int6_lzma:17109736 train_time:20715346ms step_avg:828.61ms +step:25500/50000 train_loss:2.0314 train_time:21129712ms step_avg:828.62ms +step:26000/50000 train_loss:2.0764 train_time:21543928ms step_avg:828.61ms +step:26500/50000 train_loss:2.0794 train_time:21958152ms step_avg:828.61ms +step:27000/50000 train_loss:1.9939 train_time:22372402ms step_avg:828.61ms +step:27500/50000 train_loss:1.9893 train_time:22786672ms step_avg:828.61ms +step:27500/50000 val_loss:2.0064 val_bpb:1.1883 artifact_bytes:17193574 model_int6_lzma:17102072 train_time:22787027ms step_avg:828.62ms +step:28000/50000 train_loss:1.8326 train_time:23201412ms step_avg:828.62ms +step:28500/50000 train_loss:2.0206 train_time:23615709ms step_avg:828.62ms +step:29000/50000 train_loss:2.0331 train_time:24029950ms step_avg:828.62ms +step:29500/50000 train_loss:2.0431 train_time:24444160ms step_avg:828.62ms +step:30000/50000 train_loss:1.9783 train_time:24858509ms step_avg:828.62ms +step:30000/50000 val_loss:2.0033 val_bpb:1.1865 artifact_bytes:17190822 model_int6_lzma:17099320 train_time:24858862ms step_avg:828.63ms +step:30500/50000 train_loss:2.0243 train_time:25272792ms step_avg:828.62ms +step:31000/50000 train_loss:2.0254 train_time:25687051ms step_avg:828.61ms +step:31500/50000 train_loss:1.9576 train_time:26101287ms step_avg:828.61ms +step:32000/50000 train_loss:2.0433 train_time:26515527ms step_avg:828.61ms +step:32500/50000 train_loss:1.9925 train_time:26929743ms step_avg:828.61ms +step:32500/50000 val_loss:1.9945 val_bpb:1.1813 artifact_bytes:17052022 model_int6_lzma:16960520 train_time:26930097ms step_avg:828.62ms +step:33000/50000 train_loss:1.9740 train_time:27344003ms step_avg:828.61ms +step:33500/50000 train_loss:1.9531 train_time:27758223ms step_avg:828.60ms +step:34000/50000 train_loss:2.0600 train_time:28172443ms step_avg:828.60ms +step:34500/50000 train_loss:1.9874 train_time:28586649ms step_avg:828.60ms +step:35000/50000 train_loss:1.9774 train_time:29000890ms step_avg:828.60ms +step:35000/50000 val_loss:1.9826 val_bpb:1.1742 artifact_bytes:16786866 model_int6_lzma:16695364 train_time:29001242ms step_avg:828.61ms +step:35500/50000 train_loss:2.0157 train_time:29415143ms step_avg:828.60ms +step:36000/50000 train_loss:1.9959 train_time:29829409ms step_avg:828.59ms +step:36500/50000 train_loss:2.0246 train_time:30243660ms step_avg:828.59ms +step:37000/50000 train_loss:1.8805 train_time:30657895ms step_avg:828.59ms +step:37500/50000 train_loss:2.0089 train_time:31072136ms step_avg:828.59ms +step:37500/50000 val_loss:1.9715 val_bpb:1.1677 artifact_bytes:16553402 model_int6_lzma:16461900 train_time:31072495ms step_avg:828.60ms +step:38000/50000 train_loss:1.9977 train_time:31486795ms step_avg:828.60ms +step:38500/50000 train_loss:1.9700 train_time:31901084ms step_avg:828.60ms +step:39000/50000 train_loss:1.8706 train_time:32315328ms step_avg:828.60ms +step:39500/50000 train_loss:1.9330 train_time:32729566ms step_avg:828.60ms +step:40000/50000 train_loss:1.9198 train_time:33143840ms step_avg:828.60ms +step:40000/50000 val_loss:1.9573 val_bpb:1.1592 artifact_bytes:16067974 model_int6_lzma:15976472 train_time:33144193ms step_avg:828.60ms +step:40500/50000 train_loss:2.3025 train_time:33559060ms step_avg:828.62ms +step:41000/50000 train_loss:1.8542 train_time:33973289ms step_avg:828.62ms +step:41500/50000 train_loss:1.7735 train_time:34387489ms step_avg:828.61ms +step:42000/50000 train_loss:1.9641 train_time:34801713ms step_avg:828.61ms +step:42500/50000 train_loss:1.9355 train_time:35215965ms step_avg:828.61ms +step:42500/50000 val_loss:1.9406 val_bpb:1.1493 artifact_bytes:15598778 model_int6_lzma:15507276 train_time:35216309ms step_avg:828.62ms +step:43000/50000 train_loss:2.0262 train_time:35630707ms step_avg:828.62ms +step:43500/50000 train_loss:2.1248 train_time:36044972ms step_avg:828.62ms +step:44000/50000 train_loss:1.8095 train_time:36459194ms step_avg:828.62ms +step:44500/50000 train_loss:1.8511 train_time:36873375ms step_avg:828.62ms +step:45000/50000 train_loss:1.9509 train_time:37287600ms step_avg:828.61ms +step:45000/50000 val_loss:1.9164 val_bpb:1.1350 artifact_bytes:15025010 model_int6_lzma:14933508 train_time:37287936ms step_avg:828.62ms +step:45500/50000 train_loss:1.9787 train_time:37701819ms step_avg:828.61ms +step:46000/50000 train_loss:1.9022 train_time:38116019ms step_avg:828.61ms +swa:start step:46150 +step:46500/50000 train_loss:1.8170 train_time:38530797ms step_avg:828.62ms +step:47000/50000 train_loss:1.9271 train_time:38945732ms step_avg:828.63ms +late_qat:enabled step:47076 scale:0.1499 +step:47500/50000 train_loss:1.8898 train_time:39360682ms step_avg:828.65ms +step:47500/50000 val_loss:1.8873 val_bpb:1.1178 artifact_bytes:14743726 model_int6_lzma:14652224 train_time:39361076ms step_avg:828.65ms +step:48000/50000 train_loss:1.8844 train_time:39775639ms step_avg:828.66ms +step:48500/50000 train_loss:1.8388 train_time:40190570ms step_avg:828.67ms +step:49000/50000 train_loss:1.7883 train_time:40605639ms step_avg:828.69ms +step:49500/50000 train_loss:1.8231 train_time:41020558ms step_avg:828.70ms +step:50000/50000 train_loss:1.8412 train_time:41435439ms step_avg:828.71ms +step:50000/50000 val_loss:1.8494 val_bpb:1.0953 artifact_bytes:14353974 model_int6_lzma:14262472 train_time:41435835ms step_avg:828.72ms +peak memory allocated: 22092 MiB reserved: 22340 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8488 val_bpb:1.0949 eval_time:23122ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14279664 bytes +Total submission size int6+lzma: 14371166 bytes +final_int6_roundtrip val_loss:1.8970 val_bpb:1.1235 eval_time:23858ms +final_int6_roundtrip_exact val_loss:1.89698682 val_bpb:1.12350245 +final_int6_sliding_window val_loss:1.8571 val_bpb:1.0999 stride:64 eval_time:662955ms +final_int6_sliding_window_exact val_loss:1.85714441 val_bpb:1.09990844 +final_int8_zlib_roundtrip_exact val_loss:1.85714441 val_bpb:1.09990844 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.148757 time=1.2s + ttt_chunk [11/1893] bpb=1.102178 time=11.9s + ttt_chunk [21/1893] bpb=1.105358 time=22.7s + ttt_chunk [31/1893] bpb=1.108851 time=33.5s + ttt_chunk [41/1893] bpb=1.096845 time=44.3s + ttt_chunk [51/1893] bpb=1.092842 time=55.1s + ttt_chunk [61/1893] bpb=1.097112 time=65.9s + ttt_chunk [71/1893] bpb=1.093608 time=76.6s + ttt_chunk [81/1893] bpb=1.093503 time=87.4s + ttt_chunk [91/1893] bpb=1.092102 time=98.2s + ttt_chunk [101/1893] bpb=1.094825 time=109.0s + ttt_chunk [111/1893] bpb=1.095924 time=119.8s + ttt_chunk [121/1893] bpb=1.092419 time=130.6s + ttt_chunk [131/1893] bpb=1.091950 time=141.4s + ttt_chunk [141/1893] bpb=1.091511 time=152.2s + ttt_chunk [151/1893] bpb=1.094588 time=163.0s + ttt_chunk [161/1893] bpb=1.096045 time=173.7s + ttt_chunk [171/1893] bpb=1.096592 time=184.5s + ttt_chunk [181/1893] bpb=1.096579 time=195.3s + ttt_chunk [191/1893] bpb=1.099678 time=206.1s + ttt_chunk [201/1893] bpb=1.099857 time=216.9s + ttt_chunk [211/1893] bpb=1.097320 time=227.7s + ttt_chunk [221/1893] bpb=1.099322 time=238.5s + ttt_chunk [231/1893] bpb=1.098750 time=249.2s + ttt_chunk [241/1893] bpb=1.098538 time=260.0s + ttt_chunk [251/1893] bpb=1.096996 time=270.8s + ttt_chunk [261/1893] bpb=1.095315 time=281.6s + ttt_chunk [271/1893] bpb=1.093958 time=292.4s + ttt_chunk [281/1893] bpb=1.096219 time=303.2s + ttt_chunk [291/1893] bpb=1.096848 time=314.0s + ttt_chunk [301/1893] bpb=1.097386 time=324.7s + ttt_chunk [311/1893] bpb=1.098887 time=335.5s + ttt_chunk [321/1893] bpb=1.100300 time=346.3s + ttt_chunk [331/1893] bpb=1.100039 time=357.1s + ttt_chunk [341/1893] bpb=1.100387 time=367.9s + ttt_chunk [351/1893] bpb=1.101641 time=378.7s + ttt_chunk [361/1893] bpb=1.102780 time=389.5s + ttt_chunk [371/1893] bpb=1.102319 time=400.3s + ttt_chunk [381/1893] bpb=1.102277 time=411.0s + ttt_chunk [391/1893] bpb=1.101936 time=421.8s + ttt_chunk [401/1893] bpb=1.100543 time=432.6s + ttt_chunk [411/1893] bpb=1.099425 time=443.4s + ttt_chunk [421/1893] bpb=1.098865 time=454.2s + ttt_chunk [431/1893] bpb=1.099505 time=465.0s + ttt_chunk [441/1893] bpb=1.099299 time=475.8s + ttt_chunk [451/1893] bpb=1.099028 time=486.5s + ttt_chunk [461/1893] bpb=1.098360 time=497.3s + ttt_chunk [471/1893] bpb=1.097998 time=508.1s + ttt_chunk [481/1893] bpb=1.097693 time=518.9s + ttt_chunk [491/1893] bpb=1.097377 time=529.7s + ttt_chunk [501/1893] bpb=1.096823 time=540.5s + ttt_chunk [511/1893] bpb=1.096195 time=551.3s + ttt_chunk [521/1893] bpb=1.095062 time=562.1s + ttt_chunk [531/1893] bpb=1.095173 time=572.8s + ttt_chunk [541/1893] bpb=1.095103 time=583.6s + ttt_chunk [551/1893] bpb=1.093925 time=594.4s + ttt_chunk [561/1893] bpb=1.094427 time=605.2s + ttt_chunk [571/1893] bpb=1.093719 time=616.0s + ttt_chunk [581/1893] bpb=1.093184 time=626.8s + ttt_chunk [591/1893] bpb=1.092495 time=637.6s + ttt_chunk [601/1893] bpb=1.093161 time=648.3s + ttt_chunk [611/1893] bpb=1.092714 time=659.1s + ttt_chunk [621/1893] bpb=1.092635 time=669.9s + ttt_chunk [631/1893] bpb=1.093026 time=680.7s + ttt_chunk [641/1893] bpb=1.092802 time=691.5s + ttt_chunk [651/1893] bpb=1.092816 time=702.3s + ttt_chunk [661/1893] bpb=1.092694 time=713.1s + ttt_chunk [671/1893] bpb=1.092341 time=723.9s + ttt_chunk [681/1893] bpb=1.092594 time=734.7s + ttt_chunk [691/1893] bpb=1.093260 time=745.4s + ttt_chunk [701/1893] bpb=1.092485 time=756.2s + ttt_chunk [711/1893] bpb=1.093061 time=767.0s + ttt_chunk [721/1893] bpb=1.092678 time=777.8s + ttt_chunk [731/1893] bpb=1.093086 time=788.6s + ttt_chunk [741/1893] bpb=1.093053 time=799.4s + ttt_chunk [751/1893] bpb=1.092696 time=810.2s + ttt_chunk [761/1893] bpb=1.092495 time=820.9s + ttt_chunk [771/1893] bpb=1.092281 time=831.7s + ttt_chunk [781/1893] bpb=1.092808 time=842.5s + ttt_chunk [791/1893] bpb=1.092483 time=853.3s + ttt_chunk [801/1893] bpb=1.092537 time=864.1s + ttt_chunk [811/1893] bpb=1.092037 time=874.9s + ttt_chunk [821/1893] bpb=1.091822 time=885.7s + ttt_chunk [831/1893] bpb=1.091358 time=896.5s + ttt_chunk [841/1893] bpb=1.090731 time=907.3s + ttt_chunk [851/1893] bpb=1.090589 time=918.1s + ttt_chunk [861/1893] bpb=1.090713 time=928.9s + ttt_chunk [871/1893] bpb=1.090745 time=939.7s + ttt_chunk [881/1893] bpb=1.090745 time=950.5s + ttt_chunk [891/1893] bpb=1.090552 time=961.2s + ttt_chunk [901/1893] bpb=1.090540 time=972.0s + ttt_chunk [911/1893] bpb=1.090551 time=982.8s + ttt_chunk [921/1893] bpb=1.090909 time=993.6s + ttt_chunk [931/1893] bpb=1.090718 time=1004.4s + ttt_chunk [941/1893] bpb=1.090610 time=1015.2s + ttt_chunk [951/1893] bpb=1.090584 time=1026.0s + ttt_chunk [961/1893] bpb=1.090279 time=1036.8s + ttt_chunk [971/1893] bpb=1.090998 time=1047.5s + ttt_chunk [981/1893] bpb=1.091160 time=1058.3s + ttt_chunk [991/1893] bpb=1.091060 time=1069.1s + ttt_chunk [1001/1893] bpb=1.091203 time=1079.9s + ttt_chunk [1011/1893] bpb=1.091460 time=1090.7s + ttt_chunk [1021/1893] bpb=1.091612 time=1101.5s + ttt_chunk [1031/1893] bpb=1.092160 time=1112.3s + ttt_chunk [1041/1893] bpb=1.091833 time=1123.1s + ttt_chunk [1051/1893] bpb=1.091517 time=1133.9s + ttt_chunk [1061/1893] bpb=1.091746 time=1144.7s + ttt_chunk [1071/1893] bpb=1.092177 time=1155.4s + ttt_chunk [1081/1893] bpb=1.092185 time=1166.2s + ttt_chunk [1091/1893] bpb=1.092584 time=1177.0s + ttt_chunk [1101/1893] bpb=1.092718 time=1187.8s + ttt_chunk [1111/1893] bpb=1.092497 time=1198.6s + ttt_chunk [1121/1893] bpb=1.092419 time=1209.4s + ttt_chunk [1131/1893] bpb=1.092305 time=1220.2s + ttt_chunk [1141/1893] bpb=1.092149 time=1231.0s + ttt_chunk [1151/1893] bpb=1.092182 time=1241.8s + ttt_chunk [1161/1893] bpb=1.091393 time=1252.6s + ttt_chunk [1171/1893] bpb=1.091978 time=1263.4s + ttt_chunk [1181/1893] bpb=1.091471 time=1274.2s + ttt_chunk [1191/1893] bpb=1.091173 time=1284.9s + ttt_chunk [1201/1893] bpb=1.091725 time=1295.7s + ttt_chunk [1211/1893] bpb=1.091131 time=1306.5s + ttt_chunk [1221/1893] bpb=1.090838 time=1317.3s + ttt_chunk [1231/1893] bpb=1.090780 time=1328.1s + ttt_chunk [1241/1893] bpb=1.090593 time=1338.9s + ttt_chunk [1251/1893] bpb=1.090354 time=1349.7s + ttt_chunk [1261/1893] bpb=1.090289 time=1360.5s + ttt_chunk [1271/1893] bpb=1.090085 time=1371.2s + ttt_chunk [1281/1893] bpb=1.089887 time=1382.0s + ttt_chunk [1291/1893] bpb=1.089707 time=1392.8s + ttt_chunk [1301/1893] bpb=1.089346 time=1403.6s + ttt_chunk [1311/1893] bpb=1.089041 time=1414.4s + ttt_chunk [1321/1893] bpb=1.088865 time=1425.2s + ttt_chunk [1331/1893] bpb=1.088783 time=1436.0s + ttt_chunk [1341/1893] bpb=1.088666 time=1446.8s + ttt_chunk [1351/1893] bpb=1.088607 time=1457.6s + ttt_chunk [1361/1893] bpb=1.088806 time=1468.4s + ttt_chunk [1371/1893] bpb=1.088660 time=1479.1s + ttt_chunk [1381/1893] bpb=1.088583 time=1489.9s + ttt_chunk [1391/1893] bpb=1.088057 time=1500.7s + ttt_chunk [1401/1893] bpb=1.088066 time=1511.5s + ttt_chunk [1411/1893] bpb=1.088092 time=1522.3s + ttt_chunk [1421/1893] bpb=1.088333 time=1533.1s + ttt_chunk [1431/1893] bpb=1.088173 time=1543.9s + ttt_chunk [1441/1893] bpb=1.088817 time=1554.7s + ttt_chunk [1451/1893] bpb=1.088928 time=1565.4s + ttt_chunk [1461/1893] bpb=1.088670 time=1576.2s + ttt_chunk [1471/1893] bpb=1.089574 time=1587.0s + ttt_chunk [1481/1893] bpb=1.089388 time=1597.8s + ttt_chunk [1491/1893] bpb=1.089376 time=1608.6s + ttt_chunk [1501/1893] bpb=1.089522 time=1619.4s + ttt_chunk [1511/1893] bpb=1.089582 time=1630.2s + ttt_chunk [1521/1893] bpb=1.089593 time=1641.0s + ttt_chunk [1531/1893] bpb=1.089397 time=1651.8s + ttt_chunk [1541/1893] bpb=1.089357 time=1662.6s + ttt_chunk [1551/1893] bpb=1.089710 time=1673.3s + ttt_chunk [1561/1893] bpb=1.089842 time=1684.1s + ttt_chunk [1571/1893] bpb=1.089962 time=1694.9s + ttt_chunk [1581/1893] bpb=1.090055 time=1705.7s + ttt_chunk [1591/1893] bpb=1.090000 time=1716.5s + ttt_chunk [1601/1893] bpb=1.090185 time=1727.3s + ttt_chunk [1611/1893] bpb=1.090261 time=1738.1s + ttt_chunk [1621/1893] bpb=1.090129 time=1748.9s + ttt_chunk [1631/1893] bpb=1.090299 time=1759.6s + ttt_chunk [1641/1893] bpb=1.090181 time=1770.4s + ttt_chunk [1651/1893] bpb=1.090078 time=1781.2s + ttt_chunk [1661/1893] bpb=1.089971 time=1792.0s + ttt_chunk [1671/1893] bpb=1.090309 time=1802.8s + ttt_chunk [1681/1893] bpb=1.090559 time=1813.6s + ttt_chunk [1691/1893] bpb=1.090524 time=1824.4s + ttt_chunk [1701/1893] bpb=1.090516 time=1835.2s + ttt_chunk [1711/1893] bpb=1.090351 time=1846.0s + ttt_chunk [1721/1893] bpb=1.090207 time=1856.8s + ttt_chunk [1731/1893] bpb=1.090145 time=1867.5s + ttt_chunk [1741/1893] bpb=1.089909 time=1878.3s + ttt_chunk [1751/1893] bpb=1.089769 time=1889.1s + ttt_chunk [1761/1893] bpb=1.089845 time=1899.9s + ttt_chunk [1771/1893] bpb=1.089780 time=1910.7s + ttt_chunk [1781/1893] bpb=1.089728 time=1921.5s + ttt_chunk [1791/1893] bpb=1.089337 time=1932.3s + ttt_chunk [1801/1893] bpb=1.089329 time=1943.1s + ttt_chunk [1811/1893] bpb=1.089165 time=1953.9s + ttt_chunk [1821/1893] bpb=1.089201 time=1964.7s + ttt_chunk [1831/1893] bpb=1.088803 time=1975.5s + ttt_chunk [1841/1893] bpb=1.088748 time=1986.3s + ttt_chunk [1851/1893] bpb=1.088540 time=1997.0s + ttt_chunk [1861/1893] bpb=1.088077 time=2007.8s + ttt_chunk [1871/1893] bpb=1.087916 time=2018.6s + ttt_chunk [1881/1893] bpb=1.087550 time=2029.4s + ttt_chunk [1891/1893] bpb=1.087393 time=2040.2s + ttt_chunk [1893/1893] bpb=1.087411 time=2041.9s +ttt_sliding:done val_loss=1.834459 val_bpb=1.086473 elapsed=2042.0s +legal_ttt val_loss:1.8345 val_bpb:1.0865 eval_time:2042650ms +legal_ttt_exact val_loss:1.83445894 val_bpb:1.08647279 diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed42.txt b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed42.txt new file mode 100644 index 0000000000..cd0c9db772 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/train_step50k_seed42.txt @@ -0,0 +1,2252 @@ +logs/scaling_50k_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:50000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:42 +[rank1]:W0405 22:24:23.225000 574 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +[rank2]:W0405 22:24:23.318000 575 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +[rank3]:W0405 22:24:23.475000 576 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +[rank0]:W0405 22:24:23.631000 573 site-packages/torch/_inductor/utils.py:1679] [0/0] Not enough SMs to use max_autotune_gemm mode +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/50000 val_loss:6.9293 val_bpb:4.1039 artifact_bytes:4578194 model_int6_lzma:4486692 train_time:0ms step_avg:0.01ms +step:1/50000 train_loss:6.9305 train_time:915ms step_avg:914.86ms +step:2/50000 train_loss:8.6548 train_time:1459ms step_avg:729.33ms +step:3/50000 train_loss:7.6769 train_time:2309ms step_avg:769.55ms +step:4/50000 train_loss:7.1975 train_time:3160ms step_avg:789.92ms +step:5/50000 train_loss:7.1772 train_time:4010ms step_avg:802.07ms +step:6/50000 train_loss:7.1089 train_time:4861ms step_avg:810.12ms +step:7/50000 train_loss:7.1578 train_time:5711ms step_avg:815.90ms +step:8/50000 train_loss:6.9943 train_time:6562ms step_avg:820.22ms +step:9/50000 train_loss:6.6552 train_time:7412ms step_avg:823.56ms +step:10/50000 train_loss:6.2954 train_time:8263ms step_avg:826.26ms +step:500/50000 train_loss:2.3628 train_time:418023ms step_avg:836.05ms +step:1000/50000 train_loss:2.2661 train_time:839366ms step_avg:839.37ms +step:1500/50000 train_loss:2.1786 train_time:1263023ms step_avg:842.02ms +step:2000/50000 train_loss:2.0345 train_time:1689029ms step_avg:844.51ms +step:2500/50000 train_loss:2.1572 train_time:2114326ms step_avg:845.73ms +step:2500/50000 val_loss:2.1187 val_bpb:1.2548 artifact_bytes:13056334 model_int6_lzma:12964832 train_time:2114678ms step_avg:845.87ms +step:3000/50000 train_loss:2.1194 train_time:2539037ms step_avg:846.35ms +step:3500/50000 train_loss:2.1855 train_time:2963217ms step_avg:846.63ms +step:4000/50000 train_loss:2.1617 train_time:3387141ms step_avg:846.79ms +step:4500/50000 train_loss:2.1300 train_time:3811231ms step_avg:846.94ms +step:5000/50000 train_loss:2.0913 train_time:4234820ms step_avg:846.96ms +step:5000/50000 val_loss:2.0752 val_bpb:1.2291 artifact_bytes:13885946 model_int6_lzma:13794444 train_time:4235166ms step_avg:847.03ms +step:5500/50000 train_loss:2.1352 train_time:4660884ms step_avg:847.43ms +step:6000/50000 train_loss:2.0313 train_time:5085107ms step_avg:847.52ms +step:6500/50000 train_loss:2.2975 train_time:5508951ms step_avg:847.53ms +step:7000/50000 train_loss:1.9296 train_time:5933455ms step_avg:847.64ms +step:7500/50000 train_loss:2.0541 train_time:6354607ms step_avg:847.28ms +step:7500/50000 val_loss:2.0583 val_bpb:1.2190 artifact_bytes:15486174 model_int6_lzma:15394672 train_time:6354962ms step_avg:847.33ms +step:8000/50000 train_loss:2.0065 train_time:6778752ms step_avg:847.34ms +step:8500/50000 train_loss:1.9858 train_time:7202637ms step_avg:847.37ms +step:9000/50000 train_loss:2.0977 train_time:7626364ms step_avg:847.37ms +step:9500/50000 train_loss:2.2507 train_time:8052455ms step_avg:847.63ms +step:10000/50000 train_loss:2.0620 train_time:8478882ms step_avg:847.89ms +step:10000/50000 val_loss:2.0289 val_bpb:1.2016 artifact_bytes:17216690 model_int6_lzma:17125188 train_time:8479234ms step_avg:847.92ms +step:10500/50000 train_loss:2.0994 train_time:8895727ms step_avg:847.21ms +step:11000/50000 train_loss:1.9885 train_time:9308890ms step_avg:846.26ms +step:11500/50000 train_loss:1.9475 train_time:9722041ms step_avg:845.39ms +step:12000/50000 train_loss:1.9648 train_time:10135212ms step_avg:844.60ms +step:12500/50000 train_loss:1.9724 train_time:10548368ms step_avg:843.87ms +step:12500/50000 val_loss:2.0248 val_bpb:1.1992 artifact_bytes:17234558 model_int6_lzma:17143056 train_time:10548712ms step_avg:843.90ms +step:13000/50000 train_loss:1.9881 train_time:10961705ms step_avg:843.21ms +step:13500/50000 train_loss:2.1075 train_time:11374805ms step_avg:842.58ms +step:14000/50000 train_loss:1.9118 train_time:11787919ms step_avg:841.99ms +step:14500/50000 train_loss:2.0919 train_time:12201017ms step_avg:841.45ms +step:15000/50000 train_loss:2.0424 train_time:12614116ms step_avg:840.94ms +step:15000/50000 val_loss:2.0156 val_bpb:1.1937 artifact_bytes:17224954 model_int6_lzma:17133452 train_time:12614468ms step_avg:840.96ms +step:15500/50000 train_loss:2.0146 train_time:13027262ms step_avg:840.47ms +step:16000/50000 train_loss:2.0999 train_time:13440439ms step_avg:840.03ms +step:16500/50000 train_loss:2.0789 train_time:13853591ms step_avg:839.61ms +step:17000/50000 train_loss:2.0820 train_time:14266752ms step_avg:839.22ms +step:17500/50000 train_loss:2.0686 train_time:14679921ms step_avg:838.85ms +step:17500/50000 val_loss:2.0127 val_bpb:1.1920 artifact_bytes:17211154 model_int6_lzma:17119652 train_time:14680273ms step_avg:838.87ms +step:18000/50000 train_loss:1.9524 train_time:15093131ms step_avg:838.51ms +step:18500/50000 train_loss:1.9213 train_time:15506326ms step_avg:838.18ms +step:19000/50000 train_loss:1.9063 train_time:15919452ms step_avg:837.87ms +step:19500/50000 train_loss:2.0999 train_time:16332601ms step_avg:837.57ms +step:20000/50000 train_loss:2.0165 train_time:16745754ms step_avg:837.29ms +step:20000/50000 val_loss:2.0099 val_bpb:1.1904 artifact_bytes:17211686 model_int6_lzma:17120184 train_time:16746098ms step_avg:837.30ms +step:20500/50000 train_loss:2.0790 train_time:17158955ms step_avg:837.02ms +step:21000/50000 train_loss:1.9808 train_time:17572067ms step_avg:836.77ms +step:21500/50000 train_loss:2.0100 train_time:17985182ms step_avg:836.52ms +step:22000/50000 train_loss:2.0250 train_time:18398314ms step_avg:836.29ms +step:22500/50000 train_loss:2.0465 train_time:18811473ms step_avg:836.07ms +step:22500/50000 val_loss:2.0089 val_bpb:1.1898 artifact_bytes:17203554 model_int6_lzma:17112052 train_time:18811825ms step_avg:836.08ms +step:23000/50000 train_loss:1.9704 train_time:19224740ms step_avg:835.86ms +step:23500/50000 train_loss:2.0687 train_time:19637845ms step_avg:835.65ms +step:24000/50000 train_loss:1.9694 train_time:20050933ms step_avg:835.46ms +step:24500/50000 train_loss:2.0109 train_time:20464050ms step_avg:835.27ms +step:25000/50000 train_loss:2.0203 train_time:20877142ms step_avg:835.09ms +step:25000/50000 val_loss:2.0083 val_bpb:1.1894 artifact_bytes:17197970 model_int6_lzma:17106468 train_time:20877502ms step_avg:835.10ms +step:25500/50000 train_loss:2.0304 train_time:21290451ms step_avg:834.92ms +step:26000/50000 train_loss:2.0734 train_time:21703553ms step_avg:834.75ms +step:26500/50000 train_loss:2.0779 train_time:22116656ms step_avg:834.59ms +step:27000/50000 train_loss:1.9940 train_time:22529783ms step_avg:834.44ms +step:27500/50000 train_loss:1.9872 train_time:22942904ms step_avg:834.29ms +step:27500/50000 val_loss:2.0049 val_bpb:1.1874 artifact_bytes:17187270 model_int6_lzma:17095768 train_time:22943247ms step_avg:834.30ms +step:28000/50000 train_loss:1.8291 train_time:23356073ms step_avg:834.15ms +step:28500/50000 train_loss:2.0221 train_time:23769189ms step_avg:834.01ms +step:29000/50000 train_loss:2.0353 train_time:24182357ms step_avg:833.87ms +step:29500/50000 train_loss:2.0412 train_time:24595502ms step_avg:833.75ms +step:30000/50000 train_loss:1.9789 train_time:25008691ms step_avg:833.62ms +step:30000/50000 val_loss:2.0028 val_bpb:1.1861 artifact_bytes:17180426 model_int6_lzma:17088924 train_time:25009042ms step_avg:833.63ms +step:30500/50000 train_loss:2.0251 train_time:25422171ms step_avg:833.51ms +step:31000/50000 train_loss:2.0238 train_time:25835309ms step_avg:833.40ms +step:31500/50000 train_loss:1.9553 train_time:26248470ms step_avg:833.28ms +step:32000/50000 train_loss:2.0435 train_time:26661656ms step_avg:833.18ms +step:32500/50000 train_loss:1.9911 train_time:27074838ms step_avg:833.07ms +step:32500/50000 val_loss:1.9930 val_bpb:1.1803 artifact_bytes:17036114 model_int6_lzma:16944612 train_time:27075190ms step_avg:833.08ms +step:33000/50000 train_loss:1.9705 train_time:27488204ms step_avg:832.98ms +step:33500/50000 train_loss:1.9551 train_time:27901355ms step_avg:832.88ms +step:34000/50000 train_loss:2.0590 train_time:28314513ms step_avg:832.78ms +step:34500/50000 train_loss:1.9835 train_time:28727649ms step_avg:832.69ms +step:35000/50000 train_loss:1.9767 train_time:29140764ms step_avg:832.59ms +step:35000/50000 val_loss:1.9818 val_bpb:1.1737 artifact_bytes:16778662 model_int6_lzma:16687160 train_time:29141107ms step_avg:832.60ms +step:35500/50000 train_loss:2.0147 train_time:29553964ms step_avg:832.51ms +step:36000/50000 train_loss:1.9922 train_time:29967104ms step_avg:832.42ms +step:36500/50000 train_loss:2.0225 train_time:30380305ms step_avg:832.34ms +step:37000/50000 train_loss:1.8803 train_time:30793462ms step_avg:832.26ms +step:37500/50000 train_loss:2.0098 train_time:31206646ms step_avg:832.18ms +step:37500/50000 val_loss:1.9707 val_bpb:1.1671 artifact_bytes:16452362 model_int6_lzma:16360860 train_time:31206997ms step_avg:832.19ms +step:38000/50000 train_loss:2.0002 train_time:31619858ms step_avg:832.10ms +step:38500/50000 train_loss:1.9651 train_time:32033002ms step_avg:832.03ms +step:39000/50000 train_loss:1.8694 train_time:32446145ms step_avg:831.95ms +step:39500/50000 train_loss:1.9330 train_time:32859246ms step_avg:831.88ms +step:40000/50000 train_loss:1.9198 train_time:33272339ms step_avg:831.81ms +step:40000/50000 val_loss:1.9555 val_bpb:1.1582 artifact_bytes:16058426 model_int6_lzma:15966924 train_time:33272691ms step_avg:831.82ms +step:40500/50000 train_loss:2.3043 train_time:33685481ms step_avg:831.74ms +step:41000/50000 train_loss:1.8533 train_time:34098619ms step_avg:831.67ms +step:41500/50000 train_loss:1.7756 train_time:34511772ms step_avg:831.61ms +step:42000/50000 train_loss:1.9655 train_time:34924894ms step_avg:831.55ms +step:42500/50000 train_loss:1.9346 train_time:35338023ms step_avg:831.48ms +step:42500/50000 val_loss:1.9397 val_bpb:1.1488 artifact_bytes:15590046 model_int6_lzma:15498544 train_time:35338366ms step_avg:831.49ms +step:43000/50000 train_loss:2.0274 train_time:35751720ms step_avg:831.44ms +step:43500/50000 train_loss:2.1275 train_time:36164908ms step_avg:831.38ms +step:44000/50000 train_loss:1.8064 train_time:36578068ms step_avg:831.32ms +step:44500/50000 train_loss:1.8508 train_time:36991266ms step_avg:831.26ms +step:45000/50000 train_loss:1.9490 train_time:37404434ms step_avg:831.21ms +step:45000/50000 val_loss:1.9152 val_bpb:1.1343 artifact_bytes:15147478 model_int6_lzma:15055976 train_time:37404786ms step_avg:831.22ms +step:45500/50000 train_loss:1.9782 train_time:37817815ms step_avg:831.16ms +step:46000/50000 train_loss:1.8995 train_time:38231032ms step_avg:831.11ms +swa:start step:46150 +step:46500/50000 train_loss:1.8161 train_time:38644896ms step_avg:831.07ms +step:47000/50000 train_loss:1.9230 train_time:39058825ms step_avg:831.04ms +late_qat:enabled step:47076 scale:0.1499 +step:47500/50000 train_loss:1.8891 train_time:39472823ms step_avg:831.01ms +step:47500/50000 val_loss:1.8860 val_bpb:1.1170 artifact_bytes:14621598 model_int6_lzma:14530096 train_time:39473219ms step_avg:831.02ms +step:48000/50000 train_loss:1.8819 train_time:39886757ms step_avg:830.97ms +step:48500/50000 train_loss:1.8381 train_time:40300754ms step_avg:830.94ms +step:49000/50000 train_loss:1.7874 train_time:40714693ms step_avg:830.91ms +step:49500/50000 train_loss:1.8217 train_time:41128655ms step_avg:830.88ms +step:50000/50000 train_loss:1.8393 train_time:41542652ms step_avg:830.85ms +step:50000/50000 val_loss:1.8480 val_bpb:1.0945 artifact_bytes:14206210 model_int6_lzma:14114708 train_time:41543044ms step_avg:830.86ms +peak memory allocated: 22092 MiB reserved: 22360 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8474 val_bpb:1.0941 eval_time:19647ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14275888 bytes +Total submission size int6+lzma: 14367390 bytes +final_int6_roundtrip val_loss:1.8956 val_bpb:1.1227 eval_time:37404ms +final_int6_roundtrip_exact val_loss:1.89563091 val_bpb:1.12269940 +final_int6_sliding_window val_loss:1.8561 val_bpb:1.0993 stride:64 eval_time:671805ms +final_int6_sliding_window_exact val_loss:1.85611057 val_bpb:1.09929614 +final_int8_zlib_roundtrip_exact val_loss:1.85611057 val_bpb:1.09929614 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.148732 time=1.1s + ttt_chunk [11/1893] bpb=1.103345 time=11.9s + ttt_chunk [21/1893] bpb=1.106382 time=22.7s + ttt_chunk [31/1893] bpb=1.109180 time=33.4s + ttt_chunk [41/1893] bpb=1.096909 time=44.2s + ttt_chunk [51/1893] bpb=1.092659 time=54.9s + ttt_chunk [61/1893] bpb=1.097005 time=65.7s + ttt_chunk [71/1893] bpb=1.093883 time=76.5s + ttt_chunk [81/1893] bpb=1.093465 time=87.3s + ttt_chunk [91/1893] bpb=1.091886 time=98.0s + ttt_chunk [101/1893] bpb=1.094456 time=108.8s + ttt_chunk [111/1893] bpb=1.095400 time=119.5s + ttt_chunk [121/1893] bpb=1.091789 time=130.3s + ttt_chunk [131/1893] bpb=1.091323 time=141.1s + ttt_chunk [141/1893] bpb=1.090902 time=151.9s + ttt_chunk [151/1893] bpb=1.093933 time=162.6s + ttt_chunk [161/1893] bpb=1.095461 time=173.4s + ttt_chunk [171/1893] bpb=1.096016 time=184.2s + ttt_chunk [181/1893] bpb=1.096039 time=194.9s + ttt_chunk [191/1893] bpb=1.099175 time=205.7s + ttt_chunk [201/1893] bpb=1.099375 time=216.5s + ttt_chunk [211/1893] bpb=1.096758 time=227.2s + ttt_chunk [221/1893] bpb=1.098693 time=238.0s + ttt_chunk [231/1893] bpb=1.098051 time=248.8s + ttt_chunk [241/1893] bpb=1.097751 time=259.5s + ttt_chunk [251/1893] bpb=1.096160 time=270.3s + ttt_chunk [261/1893] bpb=1.094537 time=281.1s + ttt_chunk [271/1893] bpb=1.093177 time=291.8s + ttt_chunk [281/1893] bpb=1.095466 time=302.6s + ttt_chunk [291/1893] bpb=1.096133 time=313.4s + ttt_chunk [301/1893] bpb=1.096660 time=324.1s + ttt_chunk [311/1893] bpb=1.098199 time=334.9s + ttt_chunk [321/1893] bpb=1.099626 time=345.7s + ttt_chunk [331/1893] bpb=1.099448 time=356.4s + ttt_chunk [341/1893] bpb=1.099686 time=367.2s + ttt_chunk [351/1893] bpb=1.100967 time=378.0s + ttt_chunk [361/1893] bpb=1.102133 time=388.7s + ttt_chunk [371/1893] bpb=1.101643 time=399.5s + ttt_chunk [381/1893] bpb=1.101562 time=410.3s + ttt_chunk [391/1893] bpb=1.101139 time=421.0s + ttt_chunk [401/1893] bpb=1.099731 time=431.8s + ttt_chunk [411/1893] bpb=1.098607 time=442.6s + ttt_chunk [421/1893] bpb=1.098052 time=453.3s + ttt_chunk [431/1893] bpb=1.098693 time=464.1s + ttt_chunk [441/1893] bpb=1.098511 time=474.9s + ttt_chunk [451/1893] bpb=1.098270 time=485.6s + ttt_chunk [461/1893] bpb=1.097540 time=496.4s + ttt_chunk [471/1893] bpb=1.097139 time=507.2s + ttt_chunk [481/1893] bpb=1.096815 time=517.9s + ttt_chunk [491/1893] bpb=1.096482 time=528.7s + ttt_chunk [501/1893] bpb=1.095931 time=539.5s + ttt_chunk [511/1893] bpb=1.095307 time=550.2s + ttt_chunk [521/1893] bpb=1.094171 time=561.0s + ttt_chunk [531/1893] bpb=1.094322 time=571.8s + ttt_chunk [541/1893] bpb=1.094259 time=582.5s + ttt_chunk [551/1893] bpb=1.093065 time=593.3s + ttt_chunk [561/1893] bpb=1.093609 time=604.1s + ttt_chunk [571/1893] bpb=1.092916 time=614.9s + ttt_chunk [581/1893] bpb=1.092390 time=625.6s + ttt_chunk [591/1893] bpb=1.091737 time=636.4s + ttt_chunk [601/1893] bpb=1.092400 time=647.2s + ttt_chunk [611/1893] bpb=1.091962 time=657.9s + ttt_chunk [621/1893] bpb=1.091869 time=668.7s + ttt_chunk [631/1893] bpb=1.092283 time=679.5s + ttt_chunk [641/1893] bpb=1.092061 time=690.2s + ttt_chunk [651/1893] bpb=1.092058 time=701.0s + ttt_chunk [661/1893] bpb=1.091920 time=711.8s + ttt_chunk [671/1893] bpb=1.091539 time=722.5s + ttt_chunk [681/1893] bpb=1.091798 time=733.3s + ttt_chunk [691/1893] bpb=1.092461 time=744.1s + ttt_chunk [701/1893] bpb=1.091694 time=754.8s + ttt_chunk [711/1893] bpb=1.092247 time=765.6s + ttt_chunk [721/1893] bpb=1.091871 time=776.4s + ttt_chunk [731/1893] bpb=1.092268 time=787.1s + ttt_chunk [741/1893] bpb=1.092192 time=797.9s + ttt_chunk [751/1893] bpb=1.091850 time=808.7s + ttt_chunk [761/1893] bpb=1.091656 time=819.4s + ttt_chunk [771/1893] bpb=1.091412 time=830.2s + ttt_chunk [781/1893] bpb=1.091931 time=841.0s + ttt_chunk [791/1893] bpb=1.091605 time=851.7s + ttt_chunk [801/1893] bpb=1.091663 time=862.5s + ttt_chunk [811/1893] bpb=1.091183 time=873.3s + ttt_chunk [821/1893] bpb=1.090985 time=884.0s + ttt_chunk [831/1893] bpb=1.090538 time=894.8s + ttt_chunk [841/1893] bpb=1.089913 time=905.6s + ttt_chunk [851/1893] bpb=1.089767 time=916.3s + ttt_chunk [861/1893] bpb=1.089917 time=927.1s + ttt_chunk [871/1893] bpb=1.089943 time=937.9s + ttt_chunk [881/1893] bpb=1.089928 time=948.6s + ttt_chunk [891/1893] bpb=1.089727 time=959.4s + ttt_chunk [901/1893] bpb=1.089709 time=970.2s + ttt_chunk [911/1893] bpb=1.089704 time=980.9s + ttt_chunk [921/1893] bpb=1.090064 time=991.7s + ttt_chunk [931/1893] bpb=1.089907 time=1002.5s + ttt_chunk [941/1893] bpb=1.089783 time=1013.2s + ttt_chunk [951/1893] bpb=1.089754 time=1024.0s + ttt_chunk [961/1893] bpb=1.089474 time=1034.8s + ttt_chunk [971/1893] bpb=1.090185 time=1045.5s + ttt_chunk [981/1893] bpb=1.090325 time=1056.3s + ttt_chunk [991/1893] bpb=1.090220 time=1067.1s + ttt_chunk [1001/1893] bpb=1.090379 time=1077.8s + ttt_chunk [1011/1893] bpb=1.090640 time=1088.6s + ttt_chunk [1021/1893] bpb=1.090797 time=1099.4s + ttt_chunk [1031/1893] bpb=1.091342 time=1110.1s + ttt_chunk [1041/1893] bpb=1.091014 time=1120.9s + ttt_chunk [1051/1893] bpb=1.090723 time=1131.7s + ttt_chunk [1061/1893] bpb=1.090939 time=1142.4s + ttt_chunk [1071/1893] bpb=1.091387 time=1153.2s + ttt_chunk [1081/1893] bpb=1.091387 time=1164.0s + ttt_chunk [1091/1893] bpb=1.091788 time=1174.8s + ttt_chunk [1101/1893] bpb=1.091905 time=1185.5s + ttt_chunk [1111/1893] bpb=1.091667 time=1196.3s + ttt_chunk [1121/1893] bpb=1.091605 time=1207.1s + ttt_chunk [1131/1893] bpb=1.091489 time=1217.8s + ttt_chunk [1141/1893] bpb=1.091350 time=1228.6s + ttt_chunk [1151/1893] bpb=1.091391 time=1239.4s + ttt_chunk [1161/1893] bpb=1.090617 time=1250.1s + ttt_chunk [1171/1893] bpb=1.091199 time=1260.9s + ttt_chunk [1181/1893] bpb=1.090693 time=1271.7s + ttt_chunk [1191/1893] bpb=1.090401 time=1282.4s + ttt_chunk [1201/1893] bpb=1.090988 time=1293.2s + ttt_chunk [1211/1893] bpb=1.090391 time=1304.0s + ttt_chunk [1221/1893] bpb=1.090100 time=1314.7s + ttt_chunk [1231/1893] bpb=1.090028 time=1325.5s + ttt_chunk [1241/1893] bpb=1.089832 time=1336.3s + ttt_chunk [1251/1893] bpb=1.089583 time=1347.0s + ttt_chunk [1261/1893] bpb=1.089519 time=1357.8s + ttt_chunk [1271/1893] bpb=1.089310 time=1368.6s + ttt_chunk [1281/1893] bpb=1.089091 time=1379.4s + ttt_chunk [1291/1893] bpb=1.088909 time=1390.1s + ttt_chunk [1301/1893] bpb=1.088533 time=1400.9s + ttt_chunk [1311/1893] bpb=1.088238 time=1411.6s + ttt_chunk [1321/1893] bpb=1.088068 time=1422.4s + ttt_chunk [1331/1893] bpb=1.087985 time=1433.2s + ttt_chunk [1341/1893] bpb=1.087864 time=1444.0s + ttt_chunk [1351/1893] bpb=1.087802 time=1454.7s + ttt_chunk [1361/1893] bpb=1.087999 time=1465.5s + ttt_chunk [1371/1893] bpb=1.087859 time=1476.3s + ttt_chunk [1381/1893] bpb=1.087780 time=1487.0s + ttt_chunk [1391/1893] bpb=1.087256 time=1497.8s + ttt_chunk [1401/1893] bpb=1.087279 time=1508.6s + ttt_chunk [1411/1893] bpb=1.087307 time=1519.3s + ttt_chunk [1421/1893] bpb=1.087553 time=1530.1s + ttt_chunk [1431/1893] bpb=1.087406 time=1540.9s + ttt_chunk [1441/1893] bpb=1.088071 time=1551.6s + ttt_chunk [1451/1893] bpb=1.088177 time=1562.4s + ttt_chunk [1461/1893] bpb=1.087911 time=1573.2s + ttt_chunk [1471/1893] bpb=1.088820 time=1584.0s + ttt_chunk [1481/1893] bpb=1.088634 time=1594.7s + ttt_chunk [1491/1893] bpb=1.088608 time=1605.5s + ttt_chunk [1501/1893] bpb=1.088753 time=1616.3s + ttt_chunk [1511/1893] bpb=1.088811 time=1627.1s + ttt_chunk [1521/1893] bpb=1.088829 time=1637.8s + ttt_chunk [1531/1893] bpb=1.088633 time=1648.6s + ttt_chunk [1541/1893] bpb=1.088603 time=1659.4s + ttt_chunk [1551/1893] bpb=1.088943 time=1670.1s + ttt_chunk [1561/1893] bpb=1.089066 time=1680.9s + ttt_chunk [1571/1893] bpb=1.089191 time=1691.7s + ttt_chunk [1581/1893] bpb=1.089298 time=1702.4s + ttt_chunk [1591/1893] bpb=1.089250 time=1713.2s + ttt_chunk [1601/1893] bpb=1.089432 time=1724.0s + ttt_chunk [1611/1893] bpb=1.089515 time=1734.8s + ttt_chunk [1621/1893] bpb=1.089385 time=1745.5s + ttt_chunk [1631/1893] bpb=1.089559 time=1756.3s + ttt_chunk [1641/1893] bpb=1.089427 time=1767.1s + ttt_chunk [1651/1893] bpb=1.089333 time=1777.9s + ttt_chunk [1661/1893] bpb=1.089227 time=1788.6s + ttt_chunk [1671/1893] bpb=1.089572 time=1799.4s + ttt_chunk [1681/1893] bpb=1.089825 time=1810.2s + ttt_chunk [1691/1893] bpb=1.089799 time=1821.0s + ttt_chunk [1701/1893] bpb=1.089796 time=1831.7s + ttt_chunk [1711/1893] bpb=1.089627 time=1842.5s + ttt_chunk [1721/1893] bpb=1.089486 time=1853.3s + ttt_chunk [1731/1893] bpb=1.089433 time=1864.0s + ttt_chunk [1741/1893] bpb=1.089195 time=1874.8s + ttt_chunk [1751/1893] bpb=1.089058 time=1885.6s + ttt_chunk [1761/1893] bpb=1.089140 time=1896.4s + ttt_chunk [1771/1893] bpb=1.089074 time=1907.1s + ttt_chunk [1781/1893] bpb=1.089025 time=1917.9s + ttt_chunk [1791/1893] bpb=1.088638 time=1928.7s + ttt_chunk [1801/1893] bpb=1.088642 time=1939.4s + ttt_chunk [1811/1893] bpb=1.088479 time=1950.2s + ttt_chunk [1821/1893] bpb=1.088509 time=1961.0s + ttt_chunk [1831/1893] bpb=1.088110 time=1971.8s + ttt_chunk [1841/1893] bpb=1.088069 time=1982.5s + ttt_chunk [1851/1893] bpb=1.087847 time=1993.3s + ttt_chunk [1861/1893] bpb=1.087387 time=2004.1s + ttt_chunk [1871/1893] bpb=1.087241 time=2014.9s + ttt_chunk [1881/1893] bpb=1.086870 time=2025.6s + ttt_chunk [1891/1893] bpb=1.086702 time=2036.4s + ttt_chunk [1893/1893] bpb=1.086723 time=2038.1s +ttt_sliding:done val_loss=1.833146 val_bpb=1.085695 elapsed=2038.2s +legal_ttt val_loss:1.8331 val_bpb:1.0857 eval_time:2038883ms +legal_ttt_exact val_loss:1.83314639 val_bpb:1.08569542 + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA expects (B, H, T, D), FA3 uses (B, T, H, D) + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + # With MIG fix above, each process only sees its assigned GPU as device 0 + device = torch.device("cuda", 0) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + # Estimate artifact size (int6+lzma) at each validation step + try: + sd_snap = {k: v.detach().cpu() for k, v in base_model.state_dict().items() + if "mtp_heads" not in k} + unbanked_snap = _unbank_state_dict(sd_snap, args.num_layers) + q_snap, m_snap = mixed_quantize_int6(unbanked_snap, {"mlp", "attn"}) + snap_buf = io.BytesIO() + torch.save({"w": q_snap, "m": m_snap}, snap_buf) + snap_compressed = len(lzma.compress(snap_buf.getvalue(), preset=6)) + snap_code_bytes = len(code.encode("utf-8")) + snap_total = snap_compressed + snap_code_bytes + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_bytes:{snap_total} model_int6_lzma:{snap_compressed} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + except Exception as e: + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"artifact_size:error({e}) " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.10.14 (main, May 6 2024, 19:42:50) [GCC 11.2.0] +Running PyTorch 2.10.0+cu128 +Sun Apr 5 22:24:02 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 550.127.08 Driver Version: 550.127.08 CUDA Version: 12.4 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA A100-SXM4-80GB On | 00000000:2A:00.0 Off | On | +| N/A 33C P0 96W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA A100-SXM4-80GB On | 00000000:51:00.0 Off | On | +| N/A 33C P0 96W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA A100-SXM4-80GB On | 00000000:57:00.0 Off | On | +| N/A 37C P0 98W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA A100-SXM4-80GB On | 00000000:A4:00.0 Off | On | +| N/A 34C P0 97W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA A100-SXM4-80GB On | 00000000:C7:00.0 Off | On | +| N/A 33C P0 96W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA A100-SXM4-80GB On | 00000000:CA:00.0 Off | On | +| N/A 37C P0 93W / 400W | N/A | N/A Default | +| | | Enabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| MIG devices: | ++------------------+----------------------------------+-----------+-----------------------+ +| GPU GI CI MIG | Memory-Usage | Vol| Shared | +| ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG | +| | | ECC| | +|==================+==================================+===========+=======================| +| 0 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 0 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 1 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 2 2 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 3 1 0 0 | 355MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 2MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 4 1 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 1 0 0 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ +| 5 2 0 1 | 38MiB / 40192MiB | 42 0 | 3 0 2 0 0 | +| | 0MiB / 65535MiB | | | ++------------------+----------------------------------+-----------+-----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 573 C /opt/conda/bin/python3 0MiB | +| 1 N/A N/A 574 C /opt/conda/bin/python3 0MiB | +| 2 N/A N/A 575 C /opt/conda/bin/python3 0MiB | +| 3 N/A N/A 576 C /opt/conda/bin/python3 0MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26928220 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:4 grad_accum_steps:2 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:50000 warmup_steps:20 max_wallclock_seconds:0.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/50000 val_loss:6.9293 val_bpb:4.1039 artifact_bytes:4578194 model_int6_lzma:4486692 train_time:0ms step_avg:0.01ms +step:1/50000 train_loss:6.9305 train_time:915ms step_avg:914.86ms +step:2/50000 train_loss:8.6548 train_time:1459ms step_avg:729.33ms +step:3/50000 train_loss:7.6769 train_time:2309ms step_avg:769.55ms +step:4/50000 train_loss:7.1975 train_time:3160ms step_avg:789.92ms +step:5/50000 train_loss:7.1772 train_time:4010ms step_avg:802.07ms +step:6/50000 train_loss:7.1089 train_time:4861ms step_avg:810.12ms +step:7/50000 train_loss:7.1578 train_time:5711ms step_avg:815.90ms +step:8/50000 train_loss:6.9943 train_time:6562ms step_avg:820.22ms +step:9/50000 train_loss:6.6552 train_time:7412ms step_avg:823.56ms +step:10/50000 train_loss:6.2954 train_time:8263ms step_avg:826.26ms +step:500/50000 train_loss:2.3628 train_time:418023ms step_avg:836.05ms +step:1000/50000 train_loss:2.2661 train_time:839366ms step_avg:839.37ms +step:1500/50000 train_loss:2.1786 train_time:1263023ms step_avg:842.02ms +step:2000/50000 train_loss:2.0345 train_time:1689029ms step_avg:844.51ms +step:2500/50000 train_loss:2.1572 train_time:2114326ms step_avg:845.73ms +step:2500/50000 val_loss:2.1187 val_bpb:1.2548 artifact_bytes:13056334 model_int6_lzma:12964832 train_time:2114678ms step_avg:845.87ms +step:3000/50000 train_loss:2.1194 train_time:2539037ms step_avg:846.35ms +step:3500/50000 train_loss:2.1855 train_time:2963217ms step_avg:846.63ms +step:4000/50000 train_loss:2.1617 train_time:3387141ms step_avg:846.79ms +step:4500/50000 train_loss:2.1300 train_time:3811231ms step_avg:846.94ms +step:5000/50000 train_loss:2.0913 train_time:4234820ms step_avg:846.96ms +step:5000/50000 val_loss:2.0752 val_bpb:1.2291 artifact_bytes:13885946 model_int6_lzma:13794444 train_time:4235166ms step_avg:847.03ms +step:5500/50000 train_loss:2.1352 train_time:4660884ms step_avg:847.43ms +step:6000/50000 train_loss:2.0313 train_time:5085107ms step_avg:847.52ms +step:6500/50000 train_loss:2.2975 train_time:5508951ms step_avg:847.53ms +step:7000/50000 train_loss:1.9296 train_time:5933455ms step_avg:847.64ms +step:7500/50000 train_loss:2.0541 train_time:6354607ms step_avg:847.28ms +step:7500/50000 val_loss:2.0583 val_bpb:1.2190 artifact_bytes:15486174 model_int6_lzma:15394672 train_time:6354962ms step_avg:847.33ms +step:8000/50000 train_loss:2.0065 train_time:6778752ms step_avg:847.34ms +step:8500/50000 train_loss:1.9858 train_time:7202637ms step_avg:847.37ms +step:9000/50000 train_loss:2.0977 train_time:7626364ms step_avg:847.37ms +step:9500/50000 train_loss:2.2507 train_time:8052455ms step_avg:847.63ms +step:10000/50000 train_loss:2.0620 train_time:8478882ms step_avg:847.89ms +step:10000/50000 val_loss:2.0289 val_bpb:1.2016 artifact_bytes:17216690 model_int6_lzma:17125188 train_time:8479234ms step_avg:847.92ms +step:10500/50000 train_loss:2.0994 train_time:8895727ms step_avg:847.21ms +step:11000/50000 train_loss:1.9885 train_time:9308890ms step_avg:846.26ms +step:11500/50000 train_loss:1.9475 train_time:9722041ms step_avg:845.39ms +step:12000/50000 train_loss:1.9648 train_time:10135212ms step_avg:844.60ms +step:12500/50000 train_loss:1.9724 train_time:10548368ms step_avg:843.87ms +step:12500/50000 val_loss:2.0248 val_bpb:1.1992 artifact_bytes:17234558 model_int6_lzma:17143056 train_time:10548712ms step_avg:843.90ms +step:13000/50000 train_loss:1.9881 train_time:10961705ms step_avg:843.21ms +step:13500/50000 train_loss:2.1075 train_time:11374805ms step_avg:842.58ms +step:14000/50000 train_loss:1.9118 train_time:11787919ms step_avg:841.99ms +step:14500/50000 train_loss:2.0919 train_time:12201017ms step_avg:841.45ms +step:15000/50000 train_loss:2.0424 train_time:12614116ms step_avg:840.94ms +step:15000/50000 val_loss:2.0156 val_bpb:1.1937 artifact_bytes:17224954 model_int6_lzma:17133452 train_time:12614468ms step_avg:840.96ms +step:15500/50000 train_loss:2.0146 train_time:13027262ms step_avg:840.47ms +step:16000/50000 train_loss:2.0999 train_time:13440439ms step_avg:840.03ms +step:16500/50000 train_loss:2.0789 train_time:13853591ms step_avg:839.61ms +step:17000/50000 train_loss:2.0820 train_time:14266752ms step_avg:839.22ms +step:17500/50000 train_loss:2.0686 train_time:14679921ms step_avg:838.85ms +step:17500/50000 val_loss:2.0127 val_bpb:1.1920 artifact_bytes:17211154 model_int6_lzma:17119652 train_time:14680273ms step_avg:838.87ms +step:18000/50000 train_loss:1.9524 train_time:15093131ms step_avg:838.51ms +step:18500/50000 train_loss:1.9213 train_time:15506326ms step_avg:838.18ms +step:19000/50000 train_loss:1.9063 train_time:15919452ms step_avg:837.87ms +step:19500/50000 train_loss:2.0999 train_time:16332601ms step_avg:837.57ms +step:20000/50000 train_loss:2.0165 train_time:16745754ms step_avg:837.29ms +step:20000/50000 val_loss:2.0099 val_bpb:1.1904 artifact_bytes:17211686 model_int6_lzma:17120184 train_time:16746098ms step_avg:837.30ms +step:20500/50000 train_loss:2.0790 train_time:17158955ms step_avg:837.02ms +step:21000/50000 train_loss:1.9808 train_time:17572067ms step_avg:836.77ms +step:21500/50000 train_loss:2.0100 train_time:17985182ms step_avg:836.52ms +step:22000/50000 train_loss:2.0250 train_time:18398314ms step_avg:836.29ms +step:22500/50000 train_loss:2.0465 train_time:18811473ms step_avg:836.07ms +step:22500/50000 val_loss:2.0089 val_bpb:1.1898 artifact_bytes:17203554 model_int6_lzma:17112052 train_time:18811825ms step_avg:836.08ms +step:23000/50000 train_loss:1.9704 train_time:19224740ms step_avg:835.86ms +step:23500/50000 train_loss:2.0687 train_time:19637845ms step_avg:835.65ms +step:24000/50000 train_loss:1.9694 train_time:20050933ms step_avg:835.46ms +step:24500/50000 train_loss:2.0109 train_time:20464050ms step_avg:835.27ms +step:25000/50000 train_loss:2.0203 train_time:20877142ms step_avg:835.09ms +step:25000/50000 val_loss:2.0083 val_bpb:1.1894 artifact_bytes:17197970 model_int6_lzma:17106468 train_time:20877502ms step_avg:835.10ms +step:25500/50000 train_loss:2.0304 train_time:21290451ms step_avg:834.92ms +step:26000/50000 train_loss:2.0734 train_time:21703553ms step_avg:834.75ms +step:26500/50000 train_loss:2.0779 train_time:22116656ms step_avg:834.59ms +step:27000/50000 train_loss:1.9940 train_time:22529783ms step_avg:834.44ms +step:27500/50000 train_loss:1.9872 train_time:22942904ms step_avg:834.29ms +step:27500/50000 val_loss:2.0049 val_bpb:1.1874 artifact_bytes:17187270 model_int6_lzma:17095768 train_time:22943247ms step_avg:834.30ms +step:28000/50000 train_loss:1.8291 train_time:23356073ms step_avg:834.15ms +step:28500/50000 train_loss:2.0221 train_time:23769189ms step_avg:834.01ms +step:29000/50000 train_loss:2.0353 train_time:24182357ms step_avg:833.87ms +step:29500/50000 train_loss:2.0412 train_time:24595502ms step_avg:833.75ms +step:30000/50000 train_loss:1.9789 train_time:25008691ms step_avg:833.62ms +step:30000/50000 val_loss:2.0028 val_bpb:1.1861 artifact_bytes:17180426 model_int6_lzma:17088924 train_time:25009042ms step_avg:833.63ms +step:30500/50000 train_loss:2.0251 train_time:25422171ms step_avg:833.51ms +step:31000/50000 train_loss:2.0238 train_time:25835309ms step_avg:833.40ms +step:31500/50000 train_loss:1.9553 train_time:26248470ms step_avg:833.28ms +step:32000/50000 train_loss:2.0435 train_time:26661656ms step_avg:833.18ms +step:32500/50000 train_loss:1.9911 train_time:27074838ms step_avg:833.07ms +step:32500/50000 val_loss:1.9930 val_bpb:1.1803 artifact_bytes:17036114 model_int6_lzma:16944612 train_time:27075190ms step_avg:833.08ms +step:33000/50000 train_loss:1.9705 train_time:27488204ms step_avg:832.98ms +step:33500/50000 train_loss:1.9551 train_time:27901355ms step_avg:832.88ms +step:34000/50000 train_loss:2.0590 train_time:28314513ms step_avg:832.78ms +step:34500/50000 train_loss:1.9835 train_time:28727649ms step_avg:832.69ms +step:35000/50000 train_loss:1.9767 train_time:29140764ms step_avg:832.59ms +step:35000/50000 val_loss:1.9818 val_bpb:1.1737 artifact_bytes:16778662 model_int6_lzma:16687160 train_time:29141107ms step_avg:832.60ms +step:35500/50000 train_loss:2.0147 train_time:29553964ms step_avg:832.51ms +step:36000/50000 train_loss:1.9922 train_time:29967104ms step_avg:832.42ms +step:36500/50000 train_loss:2.0225 train_time:30380305ms step_avg:832.34ms +step:37000/50000 train_loss:1.8803 train_time:30793462ms step_avg:832.26ms +step:37500/50000 train_loss:2.0098 train_time:31206646ms step_avg:832.18ms +step:37500/50000 val_loss:1.9707 val_bpb:1.1671 artifact_bytes:16452362 model_int6_lzma:16360860 train_time:31206997ms step_avg:832.19ms +step:38000/50000 train_loss:2.0002 train_time:31619858ms step_avg:832.10ms +step:38500/50000 train_loss:1.9651 train_time:32033002ms step_avg:832.03ms +step:39000/50000 train_loss:1.8694 train_time:32446145ms step_avg:831.95ms +step:39500/50000 train_loss:1.9330 train_time:32859246ms step_avg:831.88ms +step:40000/50000 train_loss:1.9198 train_time:33272339ms step_avg:831.81ms +step:40000/50000 val_loss:1.9555 val_bpb:1.1582 artifact_bytes:16058426 model_int6_lzma:15966924 train_time:33272691ms step_avg:831.82ms +step:40500/50000 train_loss:2.3043 train_time:33685481ms step_avg:831.74ms +step:41000/50000 train_loss:1.8533 train_time:34098619ms step_avg:831.67ms +step:41500/50000 train_loss:1.7756 train_time:34511772ms step_avg:831.61ms +step:42000/50000 train_loss:1.9655 train_time:34924894ms step_avg:831.55ms +step:42500/50000 train_loss:1.9346 train_time:35338023ms step_avg:831.48ms +step:42500/50000 val_loss:1.9397 val_bpb:1.1488 artifact_bytes:15590046 model_int6_lzma:15498544 train_time:35338366ms step_avg:831.49ms +step:43000/50000 train_loss:2.0274 train_time:35751720ms step_avg:831.44ms +step:43500/50000 train_loss:2.1275 train_time:36164908ms step_avg:831.38ms +step:44000/50000 train_loss:1.8064 train_time:36578068ms step_avg:831.32ms +step:44500/50000 train_loss:1.8508 train_time:36991266ms step_avg:831.26ms +step:45000/50000 train_loss:1.9490 train_time:37404434ms step_avg:831.21ms +step:45000/50000 val_loss:1.9152 val_bpb:1.1343 artifact_bytes:15147478 model_int6_lzma:15055976 train_time:37404786ms step_avg:831.22ms +step:45500/50000 train_loss:1.9782 train_time:37817815ms step_avg:831.16ms +step:46000/50000 train_loss:1.8995 train_time:38231032ms step_avg:831.11ms +swa:start step:46150 +step:46500/50000 train_loss:1.8161 train_time:38644896ms step_avg:831.07ms +step:47000/50000 train_loss:1.9230 train_time:39058825ms step_avg:831.04ms +late_qat:enabled step:47076 scale:0.1499 +step:47500/50000 train_loss:1.8891 train_time:39472823ms step_avg:831.01ms +step:47500/50000 val_loss:1.8860 val_bpb:1.1170 artifact_bytes:14621598 model_int6_lzma:14530096 train_time:39473219ms step_avg:831.02ms +step:48000/50000 train_loss:1.8819 train_time:39886757ms step_avg:830.97ms +step:48500/50000 train_loss:1.8381 train_time:40300754ms step_avg:830.94ms +step:49000/50000 train_loss:1.7874 train_time:40714693ms step_avg:830.91ms +step:49500/50000 train_loss:1.8217 train_time:41128655ms step_avg:830.88ms +step:50000/50000 train_loss:1.8393 train_time:41542652ms step_avg:830.85ms +step:50000/50000 val_loss:1.8480 val_bpb:1.0945 artifact_bytes:14206210 model_int6_lzma:14114708 train_time:41543044ms step_avg:830.86ms +peak memory allocated: 22092 MiB reserved: 22360 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8474 val_bpb:1.0941 eval_time:19647ms +Serialized model: 106027446 bytes +Code size: 91502 bytes +Serialized model int6+lzma: 14275888 bytes +Total submission size int6+lzma: 14367390 bytes +final_int6_roundtrip val_loss:1.8956 val_bpb:1.1227 eval_time:37404ms +final_int6_roundtrip_exact val_loss:1.89563091 val_bpb:1.12269940 +final_int6_sliding_window val_loss:1.8561 val_bpb:1.0993 stride:64 eval_time:671805ms +final_int6_sliding_window_exact val_loss:1.85611057 val_bpb:1.09929614 +final_int8_zlib_roundtrip_exact val_loss:1.85611057 val_bpb:1.09929614 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26928220 frozen=0 + ttt_chunk [1/1893] bpb=1.148732 time=1.1s + ttt_chunk [11/1893] bpb=1.103345 time=11.9s + ttt_chunk [21/1893] bpb=1.106382 time=22.7s + ttt_chunk [31/1893] bpb=1.109180 time=33.4s + ttt_chunk [41/1893] bpb=1.096909 time=44.2s + ttt_chunk [51/1893] bpb=1.092659 time=54.9s + ttt_chunk [61/1893] bpb=1.097005 time=65.7s + ttt_chunk [71/1893] bpb=1.093883 time=76.5s + ttt_chunk [81/1893] bpb=1.093465 time=87.3s + ttt_chunk [91/1893] bpb=1.091886 time=98.0s + ttt_chunk [101/1893] bpb=1.094456 time=108.8s + ttt_chunk [111/1893] bpb=1.095400 time=119.5s + ttt_chunk [121/1893] bpb=1.091789 time=130.3s + ttt_chunk [131/1893] bpb=1.091323 time=141.1s + ttt_chunk [141/1893] bpb=1.090902 time=151.9s + ttt_chunk [151/1893] bpb=1.093933 time=162.6s + ttt_chunk [161/1893] bpb=1.095461 time=173.4s + ttt_chunk [171/1893] bpb=1.096016 time=184.2s + ttt_chunk [181/1893] bpb=1.096039 time=194.9s + ttt_chunk [191/1893] bpb=1.099175 time=205.7s + ttt_chunk [201/1893] bpb=1.099375 time=216.5s + ttt_chunk [211/1893] bpb=1.096758 time=227.2s + ttt_chunk [221/1893] bpb=1.098693 time=238.0s + ttt_chunk [231/1893] bpb=1.098051 time=248.8s + ttt_chunk [241/1893] bpb=1.097751 time=259.5s + ttt_chunk [251/1893] bpb=1.096160 time=270.3s + ttt_chunk [261/1893] bpb=1.094537 time=281.1s + ttt_chunk [271/1893] bpb=1.093177 time=291.8s + ttt_chunk [281/1893] bpb=1.095466 time=302.6s + ttt_chunk [291/1893] bpb=1.096133 time=313.4s + ttt_chunk [301/1893] bpb=1.096660 time=324.1s + ttt_chunk [311/1893] bpb=1.098199 time=334.9s + ttt_chunk [321/1893] bpb=1.099626 time=345.7s + ttt_chunk [331/1893] bpb=1.099448 time=356.4s + ttt_chunk [341/1893] bpb=1.099686 time=367.2s + ttt_chunk [351/1893] bpb=1.100967 time=378.0s + ttt_chunk [361/1893] bpb=1.102133 time=388.7s + ttt_chunk [371/1893] bpb=1.101643 time=399.5s + ttt_chunk [381/1893] bpb=1.101562 time=410.3s + ttt_chunk [391/1893] bpb=1.101139 time=421.0s + ttt_chunk [401/1893] bpb=1.099731 time=431.8s + ttt_chunk [411/1893] bpb=1.098607 time=442.6s + ttt_chunk [421/1893] bpb=1.098052 time=453.3s + ttt_chunk [431/1893] bpb=1.098693 time=464.1s + ttt_chunk [441/1893] bpb=1.098511 time=474.9s + ttt_chunk [451/1893] bpb=1.098270 time=485.6s + ttt_chunk [461/1893] bpb=1.097540 time=496.4s + ttt_chunk [471/1893] bpb=1.097139 time=507.2s + ttt_chunk [481/1893] bpb=1.096815 time=517.9s + ttt_chunk [491/1893] bpb=1.096482 time=528.7s + ttt_chunk [501/1893] bpb=1.095931 time=539.5s + ttt_chunk [511/1893] bpb=1.095307 time=550.2s + ttt_chunk [521/1893] bpb=1.094171 time=561.0s + ttt_chunk [531/1893] bpb=1.094322 time=571.8s + ttt_chunk [541/1893] bpb=1.094259 time=582.5s + ttt_chunk [551/1893] bpb=1.093065 time=593.3s + ttt_chunk [561/1893] bpb=1.093609 time=604.1s + ttt_chunk [571/1893] bpb=1.092916 time=614.9s + ttt_chunk [581/1893] bpb=1.092390 time=625.6s + ttt_chunk [591/1893] bpb=1.091737 time=636.4s + ttt_chunk [601/1893] bpb=1.092400 time=647.2s + ttt_chunk [611/1893] bpb=1.091962 time=657.9s + ttt_chunk [621/1893] bpb=1.091869 time=668.7s + ttt_chunk [631/1893] bpb=1.092283 time=679.5s + ttt_chunk [641/1893] bpb=1.092061 time=690.2s + ttt_chunk [651/1893] bpb=1.092058 time=701.0s + ttt_chunk [661/1893] bpb=1.091920 time=711.8s + ttt_chunk [671/1893] bpb=1.091539 time=722.5s + ttt_chunk [681/1893] bpb=1.091798 time=733.3s + ttt_chunk [691/1893] bpb=1.092461 time=744.1s + ttt_chunk [701/1893] bpb=1.091694 time=754.8s + ttt_chunk [711/1893] bpb=1.092247 time=765.6s + ttt_chunk [721/1893] bpb=1.091871 time=776.4s + ttt_chunk [731/1893] bpb=1.092268 time=787.1s + ttt_chunk [741/1893] bpb=1.092192 time=797.9s + ttt_chunk [751/1893] bpb=1.091850 time=808.7s + ttt_chunk [761/1893] bpb=1.091656 time=819.4s + ttt_chunk [771/1893] bpb=1.091412 time=830.2s + ttt_chunk [781/1893] bpb=1.091931 time=841.0s + ttt_chunk [791/1893] bpb=1.091605 time=851.7s + ttt_chunk [801/1893] bpb=1.091663 time=862.5s + ttt_chunk [811/1893] bpb=1.091183 time=873.3s + ttt_chunk [821/1893] bpb=1.090985 time=884.0s + ttt_chunk [831/1893] bpb=1.090538 time=894.8s + ttt_chunk [841/1893] bpb=1.089913 time=905.6s + ttt_chunk [851/1893] bpb=1.089767 time=916.3s + ttt_chunk [861/1893] bpb=1.089917 time=927.1s + ttt_chunk [871/1893] bpb=1.089943 time=937.9s + ttt_chunk [881/1893] bpb=1.089928 time=948.6s + ttt_chunk [891/1893] bpb=1.089727 time=959.4s + ttt_chunk [901/1893] bpb=1.089709 time=970.2s + ttt_chunk [911/1893] bpb=1.089704 time=980.9s + ttt_chunk [921/1893] bpb=1.090064 time=991.7s + ttt_chunk [931/1893] bpb=1.089907 time=1002.5s + ttt_chunk [941/1893] bpb=1.089783 time=1013.2s + ttt_chunk [951/1893] bpb=1.089754 time=1024.0s + ttt_chunk [961/1893] bpb=1.089474 time=1034.8s + ttt_chunk [971/1893] bpb=1.090185 time=1045.5s + ttt_chunk [981/1893] bpb=1.090325 time=1056.3s + ttt_chunk [991/1893] bpb=1.090220 time=1067.1s + ttt_chunk [1001/1893] bpb=1.090379 time=1077.8s + ttt_chunk [1011/1893] bpb=1.090640 time=1088.6s + ttt_chunk [1021/1893] bpb=1.090797 time=1099.4s + ttt_chunk [1031/1893] bpb=1.091342 time=1110.1s + ttt_chunk [1041/1893] bpb=1.091014 time=1120.9s + ttt_chunk [1051/1893] bpb=1.090723 time=1131.7s + ttt_chunk [1061/1893] bpb=1.090939 time=1142.4s + ttt_chunk [1071/1893] bpb=1.091387 time=1153.2s + ttt_chunk [1081/1893] bpb=1.091387 time=1164.0s + ttt_chunk [1091/1893] bpb=1.091788 time=1174.8s + ttt_chunk [1101/1893] bpb=1.091905 time=1185.5s + ttt_chunk [1111/1893] bpb=1.091667 time=1196.3s + ttt_chunk [1121/1893] bpb=1.091605 time=1207.1s + ttt_chunk [1131/1893] bpb=1.091489 time=1217.8s + ttt_chunk [1141/1893] bpb=1.091350 time=1228.6s + ttt_chunk [1151/1893] bpb=1.091391 time=1239.4s + ttt_chunk [1161/1893] bpb=1.090617 time=1250.1s + ttt_chunk [1171/1893] bpb=1.091199 time=1260.9s + ttt_chunk [1181/1893] bpb=1.090693 time=1271.7s + ttt_chunk [1191/1893] bpb=1.090401 time=1282.4s + ttt_chunk [1201/1893] bpb=1.090988 time=1293.2s + ttt_chunk [1211/1893] bpb=1.090391 time=1304.0s + ttt_chunk [1221/1893] bpb=1.090100 time=1314.7s + ttt_chunk [1231/1893] bpb=1.090028 time=1325.5s + ttt_chunk [1241/1893] bpb=1.089832 time=1336.3s + ttt_chunk [1251/1893] bpb=1.089583 time=1347.0s + ttt_chunk [1261/1893] bpb=1.089519 time=1357.8s + ttt_chunk [1271/1893] bpb=1.089310 time=1368.6s + ttt_chunk [1281/1893] bpb=1.089091 time=1379.4s + ttt_chunk [1291/1893] bpb=1.088909 time=1390.1s + ttt_chunk [1301/1893] bpb=1.088533 time=1400.9s + ttt_chunk [1311/1893] bpb=1.088238 time=1411.6s + ttt_chunk [1321/1893] bpb=1.088068 time=1422.4s + ttt_chunk [1331/1893] bpb=1.087985 time=1433.2s + ttt_chunk [1341/1893] bpb=1.087864 time=1444.0s + ttt_chunk [1351/1893] bpb=1.087802 time=1454.7s + ttt_chunk [1361/1893] bpb=1.087999 time=1465.5s + ttt_chunk [1371/1893] bpb=1.087859 time=1476.3s + ttt_chunk [1381/1893] bpb=1.087780 time=1487.0s + ttt_chunk [1391/1893] bpb=1.087256 time=1497.8s + ttt_chunk [1401/1893] bpb=1.087279 time=1508.6s + ttt_chunk [1411/1893] bpb=1.087307 time=1519.3s + ttt_chunk [1421/1893] bpb=1.087553 time=1530.1s + ttt_chunk [1431/1893] bpb=1.087406 time=1540.9s + ttt_chunk [1441/1893] bpb=1.088071 time=1551.6s + ttt_chunk [1451/1893] bpb=1.088177 time=1562.4s + ttt_chunk [1461/1893] bpb=1.087911 time=1573.2s + ttt_chunk [1471/1893] bpb=1.088820 time=1584.0s + ttt_chunk [1481/1893] bpb=1.088634 time=1594.7s + ttt_chunk [1491/1893] bpb=1.088608 time=1605.5s + ttt_chunk [1501/1893] bpb=1.088753 time=1616.3s + ttt_chunk [1511/1893] bpb=1.088811 time=1627.1s + ttt_chunk [1521/1893] bpb=1.088829 time=1637.8s + ttt_chunk [1531/1893] bpb=1.088633 time=1648.6s + ttt_chunk [1541/1893] bpb=1.088603 time=1659.4s + ttt_chunk [1551/1893] bpb=1.088943 time=1670.1s + ttt_chunk [1561/1893] bpb=1.089066 time=1680.9s + ttt_chunk [1571/1893] bpb=1.089191 time=1691.7s + ttt_chunk [1581/1893] bpb=1.089298 time=1702.4s + ttt_chunk [1591/1893] bpb=1.089250 time=1713.2s + ttt_chunk [1601/1893] bpb=1.089432 time=1724.0s + ttt_chunk [1611/1893] bpb=1.089515 time=1734.8s + ttt_chunk [1621/1893] bpb=1.089385 time=1745.5s + ttt_chunk [1631/1893] bpb=1.089559 time=1756.3s + ttt_chunk [1641/1893] bpb=1.089427 time=1767.1s + ttt_chunk [1651/1893] bpb=1.089333 time=1777.9s + ttt_chunk [1661/1893] bpb=1.089227 time=1788.6s + ttt_chunk [1671/1893] bpb=1.089572 time=1799.4s + ttt_chunk [1681/1893] bpb=1.089825 time=1810.2s + ttt_chunk [1691/1893] bpb=1.089799 time=1821.0s + ttt_chunk [1701/1893] bpb=1.089796 time=1831.7s + ttt_chunk [1711/1893] bpb=1.089627 time=1842.5s + ttt_chunk [1721/1893] bpb=1.089486 time=1853.3s + ttt_chunk [1731/1893] bpb=1.089433 time=1864.0s + ttt_chunk [1741/1893] bpb=1.089195 time=1874.8s + ttt_chunk [1751/1893] bpb=1.089058 time=1885.6s + ttt_chunk [1761/1893] bpb=1.089140 time=1896.4s + ttt_chunk [1771/1893] bpb=1.089074 time=1907.1s + ttt_chunk [1781/1893] bpb=1.089025 time=1917.9s + ttt_chunk [1791/1893] bpb=1.088638 time=1928.7s + ttt_chunk [1801/1893] bpb=1.088642 time=1939.4s + ttt_chunk [1811/1893] bpb=1.088479 time=1950.2s + ttt_chunk [1821/1893] bpb=1.088509 time=1961.0s + ttt_chunk [1831/1893] bpb=1.088110 time=1971.8s + ttt_chunk [1841/1893] bpb=1.088069 time=1982.5s + ttt_chunk [1851/1893] bpb=1.087847 time=1993.3s + ttt_chunk [1861/1893] bpb=1.087387 time=2004.1s + ttt_chunk [1871/1893] bpb=1.087241 time=2014.9s + ttt_chunk [1881/1893] bpb=1.086870 time=2025.6s + ttt_chunk [1891/1893] bpb=1.086702 time=2036.4s + ttt_chunk [1893/1893] bpb=1.086723 time=2038.1s +ttt_sliding:done val_loss=1.833146 val_bpb=1.085695 elapsed=2038.2s +legal_ttt val_loss:1.8331 val_bpb:1.0857 eval_time:2038883ms +legal_ttt_exact val_loss:1.83314639 val_bpb:1.08569542 From e73d565a36ff869ce7c190444a5c1fee2158beb5 Mon Sep 17 00:00:00 2001 From: OnlyJundong Date: Mon, 6 Apr 2026 14:54:43 -0400 Subject: [PATCH 3/3] added 20k pr link --- .../README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md index 1b45425ab5..774e9713d8 100644 --- a/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md +++ b/records/track_non_record_16mb/2026-04-06_ExtendedCompute_50K_12hours_4xA100MIG_ScalingAnalysis/README.md @@ -4,7 +4,7 @@ ## Summary -This submission is a **non-record submission**. It extends the [20K scaling run](../2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/) to 50K steps (~12 hours), using the same architecture and code from ([PR #549](https://github.com/openai/parameter-golf/pull/549) by @abaybektursun). Training runs on 4×A100 MIG instances (approximately 10× slower per step than 8×H100 SXM). +This submission is a **non-record submission**. It extends the [20K scaling run](https://github.com/openai/parameter-golf/pull/1407) to 50K steps (~12 hours), using the same architecture and code from ([PR #549](https://github.com/openai/parameter-golf/pull/549) by @abaybektursun). Training runs on 4×A100 MIG instances (approximately 10× slower per step than 8×H100 SXM). Key findings: - **50K steps achieves 1.0858 BPB post-TTT** (3-seed mean, std 0.0005) — improvement of **−0.0102 BPB** over the 20K run (1.0960)