From 93c3ebf72a3fff3a73445c13a0738e7f9ace85de Mon Sep 17 00:00:00 2001 From: Campbellb Date: Wed, 1 Apr 2026 22:02:37 -0400 Subject: [PATCH] Record: Scylla + n-gram + legal TTT (val_bpb=1.0903, 3-seed) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Three-seed validated result on 8×H100: - Seed 1337: legal_ttt 1.09042, 15,316,209 bytes - Seed 42: legal_ttt 1.09064, 15,329,825 bytes - Seed 2025: legal_ttt 1.08985, 14,945,965 bytes - Mean: 1.09030 ± 0.00040, all under 16MB cap --- .../2026-04-01_scylla-tuned-ttt/README.md | 69 + .../submission.json | 11 + .../2026-04-01_scylla-tuned-ttt/train_gpt.py | 2425 +++++++++++++++++ .../train_seed1337.log | 338 +++ .../train_seed2025.log | 338 +++ .../train_seed42.log | 338 +++ 6 files changed, 3519 insertions(+) create mode 100644 records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/README.md create mode 100644 records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/submission.json create mode 100644 records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_gpt.py create mode 100644 records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed2025.log create mode 100644 records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed42.log diff --git a/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/README.md b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/README.md new file mode 100644 index 0000000000..349c4cb5e8 --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/README.md @@ -0,0 +1,69 @@ +# Record: Scylla + n-gram + legal TTT — val_bpb 1.0903 (3-seed mean) + +**val_bpb: 1.0903** (3-seed mean, std 0.00040) | ≤15.4 MB | 8×H100 SXM | ~94ms/step | ~6160 steps + +Applies the Scylla tokenizer (PR #1143, @simon-marcus) with n-gram rescoring and tuned legal score-first TTT. + +## 3-Seed Results + +| Seed | legal_ttt BPB | bytes_total | Steps | Wall-clock | +|------|---------------|-------------|-------|------------| +| 1337 | 1.09042 | 15,316,209 | ~6151 | 600s | +| 42 | 1.09064 | 15,329,825 | ~6162 | 600s | +| 2025 | 1.08985 | 14,945,965 | ~6170 | 600s | +| **Mean ± Std** | **1.09030 ± 0.00040** | | | | + +All seeds stopped by wallclock cap. All artifacts under 16,000,000 bytes. + +## Technique Stack + +- **Scylla tokenizer** — ~998 token TokenMonster vocabulary (PR #1143) +- **N-gram rescoring** — orders 2–16, two-pass eval (neural pass then cache rescore) +- **Legal TTT** — score-first SGD, LR=0.005, 3 epochs, 32768 tokens/chunk +- **int6 quantization** — per-row with lzma, `clip_range=20` for stable byte compliance +- **11-layer cyclic shared blocks**, BigramHash embeddings (10240 vocab, 128-dim) +- **Parallel Muon** — matrix LR=0.025, momentum=0.99, WD=0.04; EMA decay=0.997 + +`clip_range` was tuned from 31→20 to eliminate seed-dependent artifact size variance. Cost: ~+0.006 BPB. All seeds now produce artifacts between 14.9–15.4 MB. + +## Legality + +**Training (≤600s on 8×H100):** Standard. No validation data accessed during training. + +**Evaluation — TTT (score-first):** Each chunk scored under `torch.inference_mode()` before any parameter update. The reported BPB is always computed before adaptation. TTT runs ~427s. + +**N-gram metric:** The `legal_ttt` score above uses only the TTT-adapted neural model. N-gram rescoring (`ngram_two_pass`, ~0.195 BPB) is reported separately and not used as the submission metric. + +## Reproduction + +```bash +# Retokenize (one-time, ~77 min on CPU) +python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 0 --with-docs +python3 data/retokenize_scylla.py \ + --docs ./data/docs_selected.jsonl \ + --vocab ./data/tokenizers/scylla/candidate.vocab \ + --out ./data/datasets/fineweb_scylla --train-shards 80 + +# Train (replace SEED with 1337, 42, or 2025) +DATA_PATH=./data/datasets/fineweb_scylla \ +TOKENIZER_PATH=./data/tokenizers/scylla/candidate.vocab \ +TOKENIZER_META_PATH=./data/tokenizers/scylla/candidate.meta.npz \ +NGRAM_ENABLED=1 NGRAM_MAX_ORDER=16 TTT_ENABLED=1 TTT_LR=0.005 XSA_LAST_N=4 \ +SEED=1337 NUM_LAYERS=11 NUM_SHARED_BLOCKS=11 SHARED_BLOCK_LAYOUT=cyclic \ +MLP_MULT=3 VOCAB_SIZE=1024 INT6_QUANT=1 SMEAR_GATE=1 BIGRAM_VOCAB_SIZE=10240 \ +ORTHO_INIT=1 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=1500 \ +MUON_WD=0.04 GRAD_CLIP_NORM=0.3 TRAIN_SEQ_LEN=2048 TRAIN_BATCH_TOKENS=786432 \ +ITERATIONS=9000 WARMDOWN_ITERS=3500 MAX_WALLCLOCK_SECONDS=600 \ +EVAL_STRIDE=64 SWA_INTERVAL=50 SWA_LR_THRESHOLD=0.2 EMA_DECAY=0.997 \ +TTT_EPOCHS=3 TTT_MAX_SEQS=64 TTT_CHUNK_TOKENS=32768 TTT_GRAD_CLIP=1.0 \ +ROPE_DIMS=16 LN_SCALE=1 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +Requires 8×H100, PyTorch 2.9+, CUDA 12.8, tokenmonster. + +## Credits + +- Scylla tokenizer: @simon-marcus (PR #1143) +- Training base: @abaybektursun (PR #549) diff --git a/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/submission.json b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/submission.json new file mode 100644 index 0000000000..bd46d1da1b --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/submission.json @@ -0,0 +1,11 @@ +{ + "author": "Campbell Barton", + "github_id": "Campbellb", + "name": "Scylla tokenizer + n-gram + tuned legal TTT", + "blurb": "Scylla ~998-token TokenMonster vocab with n-gram (order 2-16) eval and score-first legal TTT (LR=0.005, 3 epochs). 11-layer 512-dim transformer with cyclic shared blocks, BigramHash embeddings, int6-per-row + lzma quantization (clip_range=20). Three seeds validated on 8xH100, all artifacts under 15.4MB.", + "date": "2026-04-01", + "val_bpb": 1.09030, + "val_loss": 1.96466, + "bytes_total": 15329825, + "bytes_code": 115885 +} diff --git a/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_gpt.py b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_gpt.py new file mode 100644 index 0000000000..1e4d61dabb --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_gpt.py @@ -0,0 +1,2425 @@ +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 +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 +except ImportError: + def flash_attn_3_func(q, k, v, causal=True): + q2 = q.transpose(1, 2) + k2 = k.transpose(1, 2) + v2 = v.transpose(1, 2) + if k2.size(1) != q2.size(1): + groups = q2.size(1) // k2.size(1) + k2 = k2.repeat_interleave(groups, dim=1) + v2 = v2.repeat_interleave(groups, dim=1) + o = F.scaled_dot_product_attention(q2, k2, v2, is_causal=causal) + return o.transpose(1, 2) +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") + tokenizer_meta_path = os.environ.get("TOKENIZER_META_PATH", "") + tokenizer_meta_validate = bool(int(os.environ.get("TOKENIZER_META_VALIDATE", "0"))) + # N-gram eval + ngram_enabled = bool(int(os.environ.get("NGRAM_ENABLED", "0"))) + ngram_min_order = int(os.environ.get("NGRAM_MIN_ORDER", 2)) + ngram_max_order = int(os.environ.get("NGRAM_MAX_ORDER", 12)) + ngram_num_buckets = int(os.environ.get("NGRAM_NUM_BUCKETS", 16_777_216)) + ngram_alpha_min = float(os.environ.get("NGRAM_ALPHA_MIN", 0.05)) + ngram_alpha_max = float(os.environ.get("NGRAM_ALPHA_MAX", 0.70)) + ngram_entropy_center = float(os.environ.get("NGRAM_ENTROPY_CENTER", 3.0)) + ngram_entropy_scale = float(os.environ.get("NGRAM_ENTROPY_SCALE", 2.0)) + ngram_min_count = int(os.environ.get("NGRAM_MIN_COUNT", 2)) + ngram_leave_one_out = bool(int(os.environ.get("NGRAM_LEAVE_ONE_OUT", "1"))) + 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)) + ttt_adaptive = bool(int(os.environ.get("TTT_ADAPTIVE", "0"))) + gptq_hessian = bool(int(os.environ.get("GPTQ_HESSIAN", "0"))) + gptq_calib_samples = int(os.environ.get("GPTQ_CALIB_SAMPLES", "64")) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", "128")) + gptq_damp_pct = float(os.environ.get("GPTQ_DAMP_PCT", "0.01")) + +# --- 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 --- + +TOKENIZER_META_FORMAT_VERSION = 1 +TOKENIZER_META_SUFFIX = ".meta.npz" + +def _derive_tokenizer_meta_path(tokenizer_path: str) -> Path: + tokenizer = Path(tokenizer_path) + if tokenizer.suffix == ".model": + return tokenizer.with_suffix(TOKENIZER_META_SUFFIX) + return tokenizer.with_name(tokenizer.name + TOKENIZER_META_SUFFIX) + +def build_sentencepiece_luts_np( + sp: spm.SentencePieceProcessor, vocab_size: int +) -> tuple[np.ndarray, np.ndarray, np.ndarray]: + 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 base_bytes_np, has_leading_space_np, is_boundary_token_np + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + base_bytes_np, has_leading_space_np, is_boundary_token_np = build_sentencepiece_luts_np(sp, vocab_size) + 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_tokenizer_meta_luts_np( + meta_path: Path, vocab_size: int +) -> tuple[np.ndarray, np.ndarray, np.ndarray, dict]: + def _scalar(value): + arr = np.asarray(value) + return arr.item() if arr.ndim == 0 else arr.reshape(-1)[0].item() + with np.load(meta_path, allow_pickle=False) as data: + fmt = int(_scalar(data["format_version"])) + if fmt != TOKENIZER_META_FORMAT_VERSION: + raise ValueError(f"Unsupported tokenizer meta format_version={fmt}") + meta_vocab_size = int(_scalar(data["vocab_size"])) + tokenizer_kind = str(_scalar(data["tokenizer_kind"])) + source_model_name = str(_scalar(data["source_model_name"])) + base_bytes_np = np.asarray(data["base_bytes"], dtype=np.int16) + has_leading_space_np = np.asarray(data["has_leading_space"], dtype=np.bool_) + is_boundary_token_np = np.asarray(data["is_boundary_token"], dtype=np.bool_) + table_size = max(meta_vocab_size, vocab_size) + if base_bytes_np.shape[0] < table_size: + pb = np.zeros((table_size,), dtype=np.int16) + ph = np.zeros((table_size,), dtype=np.bool_) + pi = np.ones((table_size,), dtype=np.bool_) + pb[:base_bytes_np.shape[0]] = base_bytes_np + ph[:has_leading_space_np.shape[0]] = has_leading_space_np + pi[:is_boundary_token_np.shape[0]] = is_boundary_token_np + base_bytes_np, has_leading_space_np, is_boundary_token_np = pb, ph, pi + metadata = {"format_version": fmt, "tokenizer_kind": tokenizer_kind, + "source_model_name": source_model_name, "vocab_size": meta_vocab_size, + "meta_path": str(meta_path)} + return base_bytes_np, has_leading_space_np, is_boundary_token_np, metadata + +def load_tokenizer_luts( + tokenizer_path: str, tokenizer_meta_path: str, vocab_size: int, device: torch.device, + *, validate_meta: bool = False, +) -> tuple[tuple[Tensor, Tensor, Tensor], dict]: + meta_path = Path(tokenizer_meta_path) if tokenizer_meta_path else _derive_tokenizer_meta_path(tokenizer_path) + if meta_path.exists(): + base_bytes_np, has_leading_space_np, is_boundary_token_np, metadata = load_tokenizer_meta_luts_np(meta_path, vocab_size) + if validate_meta and str(tokenizer_path).endswith(".model"): + sp = spm.SentencePieceProcessor(model_file=tokenizer_path) + sp_luts = build_sentencepiece_luts_np(sp, vocab_size) + if not (np.array_equal(base_bytes_np, sp_luts[0]) and + np.array_equal(has_leading_space_np, sp_luts[1]) and + np.array_equal(is_boundary_token_np, sp_luts[2])): + raise ValueError(f"Tokenizer metadata mismatch for {meta_path}") + 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), + ), metadata + if not str(tokenizer_path).endswith(".model"): + raise FileNotFoundError(f"TOKENIZER_META_PATH does not exist: {meta_path}") + sp = spm.SentencePieceProcessor(model_file=tokenizer_path) + return build_sentencepiece_luts(sp, vocab_size, device), { + "tokenizer_kind": "sentencepiece", "source_model_name": str(tokenizer_path), + "vocab_size": int(sp.vocab_size()), "meta_path": None, + } +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 + cos = self._cos_cached.to(dtype=dtype) + sin = self._sin_cached.to(dtype=dtype) + if torch.is_grad_enabled(): + cos, sin = cos.clone(), sin.clone() + return cos, sin +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)) + self._gptq_collect = False + self._gptq_h_o: Tensor | None = None # Hessian for out-projection (input = attn output) + 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] + y = flash_attn_3_func(q, k, v, causal=True) + 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) + if self._gptq_collect: + yf = y.reshape(-1, dim).float().detach() + h = yf.T @ yf + self._gptq_h_o = (self._gptq_h_o + h) if self._gptq_h_o is not None else h + 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 + self._gptq_collect = False + self._gptq_h_down: Tensor | None = None # Hessian for down-projection (input = post-activation) + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + h = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + h_sq = h.square() + if self._gptq_collect: + hf = h_sq.reshape(-1, h_sq.size(-1)).float().detach() + hess = hf.T @ hf + self._gptq_h_down = (self._gptq_h_down + hess) if self._gptq_h_down is not None else hess + return F.linear(h_sq, 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 + self._gptq_collect = False + self._gptq_h_qkv: Tensor | None = None # Hessian for Q/K/V projections (shared input) + self._gptq_h_up: Tensor | None = None # Hessian for MLP up-projection + 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 + x_attn_in = self.attn_norm(x_in) * self.ln_scale_factor + if self._gptq_collect: + xf = x_attn_in.reshape(-1, x_attn_in.size(-1)).float().detach() + h = xf.T @ xf + self._gptq_h_qkv = (self._gptq_h_qkv + h) if self._gptq_h_qkv is not None else h + attn_out, raw_v = self.attn(x_attn_in, 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_mlp_in = self.mlp_norm(x_out) * self.ln_scale_factor + if self._gptq_collect: + xf2 = x_mlp_in.reshape(-1, x_mlp_in.size(-1)).float().detach() + h2 = xf2.T @ xf2 + self._gptq_h_up = (self._gptq_h_up + h2) if self._gptq_h_up is not None else h2 + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(x_mlp_in, 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} adaptive={args.ttt_adaptive}") + + 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() + chunk_avg_losses: list[float] = [] # per-chunk avg NLL for adaptive LR scaling + + 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] + + chunk_nll_local = torch.zeros((), device=device, dtype=torch.float64) + chunk_tok_local = torch.zeros((), device=device, dtype=torch.float64) + + base_model.eval() + with torch.no_grad(): + 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) + chunk_nll_local += scored_nll.sum() + chunk_tok_local += 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() + + # Sync chunk loss across ranks for accurate difficulty estimate + if dist.is_available() and dist.is_initialized(): + sync_buf = torch.stack([chunk_nll_local, chunk_tok_local]) + dist.all_reduce(sync_buf, op=dist.ReduceOp.SUM) + chunk_avg_loss = (sync_buf[0] / sync_buf[1].clamp(min=1)).item() + else: + chunk_avg_loss = (chunk_nll_local / chunk_tok_local.clamp(min=1)).item() + chunk_avg_losses.append(chunk_avg_loss) + + # --- 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))) + if args.ttt_adaptive and len(chunk_avg_losses) >= 3: + median_loss = float(np.median(chunk_avg_losses)) + difficulty = chunk_avg_loss / max(median_loss, 1e-8) + cos_lr *= max(0.5, min(2.0, difficulty)) + 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 + diff_str = f" chunk_loss={chunk_avg_loss:.4f}" if args.ttt_adaptive else "" + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s{diff_str}") + + 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 = 20) -> 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 + +# --- Full GPTQ (Frantar et al., ICLR 2023) --- + +def gptq_quantize_weight( + W: Tensor, + H: Tensor, + clip_range: int = 31, + block_size: int = 128, + damp_pct: float = 0.01, +) -> tuple[Tensor, Tensor]: + """GPTQ with Hessian-guided error compensation. + W: [out_dim, in_dim] float. H: [in_dim, in_dim] float (X^T X from calibration data). + Returns (int8 quantized [out, in], float16 per-row scale [out]).""" + W = W.float().clone() + n_out, n_in = W.shape + + # Damp diagonal to handle dead/near-zero columns + H = H.float().clone() + damp = damp_pct * H.diag().mean().clamp(min=1e-8) + H.diagonal().add_(damp) + dead = H.diag() == 0 + H[dead, :] = 0; H[:, dead] = 0; H[dead, dead] = 1 + + # Activation order: quantize highest-curvature columns first + perm = torch.argsort(H.diag(), descending=True) + W = W[:, perm] + H = H[perm][:, perm] + + # Determine per-row scales on permuted W (best of 5 percentiles) + best_s, best_err = None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + row_clip = torch.quantile(W.abs(), pct, dim=1) if pct < 1.0 else W.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(W / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + err = (W - q.float() * s.float()[:, None]).pow(2).mean().item() + if err < best_err: + best_s, best_err = s, err + + # Cholesky inverse of Hessian for error propagation + try: + Hinv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + Hinv_upper = torch.linalg.cholesky(Hinv, upper=True) + except torch.linalg.LinAlgError: + # Degenerate Hessian: fall back to simple quantization + inv_perm = torch.argsort(perm) + return quantize_int6_per_row(W[:, inv_perm], clip_range) + + s_f = best_s.float()[:, None] # [out, 1] — fixed per-row scale + Q = torch.zeros(n_out, n_in, dtype=torch.int8, device=W.device) + + for i1 in range(0, n_in, block_size): + i2 = min(i1 + block_size, n_in) + W_block = W[:, i1:i2].clone() + Err_block = torch.zeros(n_out, i2 - i1, device=W.device) + Hinv_blk = Hinv_upper[i1:i2, i1:i2] + + for j in range(i2 - i1): + w_col = W_block[:, j] + d = Hinv_blk[j, j].clamp(min=1e-8) + q_col = torch.clamp(torch.round(w_col / s_f[:, 0]), -clip_range, clip_range) + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - q_col * s_f[:, 0]) / d + Err_block[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_blk[j, j + 1:].unsqueeze(0) + + if i2 < n_in: + W[:, i2:] -= Err_block @ Hinv_upper[i1:i2, i2:] + + inv_perm = torch.argsort(perm) + Q = Q[:, inv_perm] + return Q, best_s + + +def gptq_collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, world_size: int +) -> dict[str, Tensor]: + """Run calibration forward passes to collect per-layer Hessians for GPTQ. + Returns dict mapping unbanked weight name → Hessian [in_dim, in_dim].""" + # Reset and enable collection on all blocks + for block in model.blocks: + block._gptq_collect = True + block._gptq_h_qkv = None + block._gptq_h_up = None + block.attn._gptq_collect = True + block.attn._gptq_h_o = None + block.mlp._gptq_collect = True + block.mlp._gptq_h_down = None + + model.eval() + with torch.no_grad(): + for _ in range(args.gptq_calib_samples): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, 1) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model(x, y) + + # Disable collection + for block in model.blocks: + block._gptq_collect = False + block.attn._gptq_collect = False + block.mlp._gptq_collect = False + model.train() + + # All-reduce Hessians across ranks + hessians: dict[str, Tensor] = {} + n = args.num_layers + for i, block in enumerate(model.blocks): + for key, h in [ + (f"blocks.{i}.attn.c_q.weight", block._gptq_h_qkv), + (f"blocks.{i}.attn.c_k.weight", block._gptq_h_qkv), + (f"blocks.{i}.attn.c_v.weight", block._gptq_h_qkv), + (f"blocks.{i}.attn.proj.weight", block.attn._gptq_h_o), + (f"blocks.{i}.mlp.fc.weight", block._gptq_h_up), + (f"blocks.{i}.mlp.proj.weight", block.mlp._gptq_h_down), + ]: + if h is not None: + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(h, op=dist.ReduceOp.SUM) + h = h / world_size + hessians[key] = h.cpu() + + return hessians + + +def mixed_quantize_int6( + state_dict: dict[str, Tensor], + int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, + gptq_damp_pct: float = 0.01, +): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = gptq_quantize_weight(t, H, block_size=gptq_block_size, damp_pct=gptq_damp_pct) + meta[name] = {"type": "int6", "gptq": True} + else: + q, s = quantize_int6_per_row(t) + meta[name] = {"type": "int6"} + result[name + ".q"] = q + result[name + ".scale"] = s + 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 + +# === N-GRAM EVAL CACHE + TWO-PASS RESCORE === + +_NGRAM_PRIMES = np.array([ + 36313, 27191, 51647, 81929, 131071, 174763, 233017, 283721, + 347237, 411527, 479909, 557927, 646333, 746773, 862319, 992353, +], dtype=np.int64) + +_ORDER_MULTS = np.array([0.30, 0.30, 0.97, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0, 2.0], dtype=np.float32) + +class NgramCache: + def __init__(self, min_order: int = 2, max_order: int = 12, num_buckets: int = 16_777_216): + self.min_order = min_order + self.max_order = max_order + self.num_orders = max_order - min_order + 1 + self.num_buckets = num_buckets + self.bucket_mask = np.int64(num_buckets - 1) + self.ctx_tables = [np.zeros(num_buckets, dtype=np.int32) for _ in range(self.num_orders)] + self.full_tables = [np.zeros(num_buckets, dtype=np.int32) for _ in range(self.num_orders)] + + def _compute_hashes(self, tokens_np, start, end, order_idx): + n = self.min_order + order_idx + valid_start = max(start, n - 1) + N = end - valid_start + if N <= 0: + return None, None, valid_start + h = np.zeros(N, dtype=np.int64) + for k in range(n - 1): + offset = valid_start - (n - 1) + k + h ^= tokens_np[offset:offset + N].astype(np.int64) * _NGRAM_PRIMES[k % len(_NGRAM_PRIMES)] + ctx_h = h & self.bucket_mask + target_prime = _NGRAM_PRIMES[min(n - 1, len(_NGRAM_PRIMES) - 1)] + full_h = (h ^ (tokens_np[valid_start:end].astype(np.int64) * target_prime)) & self.bucket_mask + return ctx_h, full_h, valid_start + + def _bincount_add(self, table, indices): + counts = np.bincount(indices.astype(np.intp), minlength=self.num_buckets) + table += counts[:self.num_buckets].astype(table.dtype) + + def build_full(self, tokens_np): + for oi in range(self.num_orders): + ctx_h, full_h, _ = self._compute_hashes(tokens_np, 0, len(tokens_np), oi) + if ctx_h is None: + continue + self._bincount_add(self.ctx_tables[oi], ctx_h) + self._bincount_add(self.full_tables[oi], full_h) + + def score_positions(self, tokens_np, positions, min_count=2, leave_one_out=False): + N = len(positions) + ngram_prob = np.zeros(N, dtype=np.float32) + matched_order = np.full(N, -1, dtype=np.int32) + matched = np.zeros(N, dtype=bool) + if N == 0: + return ngram_prob, matched_order + positions = positions.astype(np.int64, copy=False) + for oi in range(self.num_orders - 1, -1, -1): + n = self.min_order + oi + ctx_h_all, full_h_all, valid_start = self._compute_hashes(tokens_np, 0, len(tokens_np), oi) + if ctx_h_all is None: + continue + remaining_idx = np.where(~matched)[0] + if remaining_idx.size == 0: + break + pos_sub = positions[remaining_idx] + valid_mask = pos_sub >= valid_start + if not np.any(valid_mask): + continue + valid_idx = remaining_idx[valid_mask] + lookup = (pos_sub[valid_mask] - valid_start).astype(np.int64) + ctx_h = ctx_h_all[lookup] + full_h = full_h_all[lookup] + ctx_counts = self.ctx_tables[oi][ctx_h].astype(np.int64) + full_counts = self.full_tables[oi][full_h].astype(np.int64) + if leave_one_out: + ctx_counts = np.maximum(ctx_counts - 1, 0) + full_counts = np.maximum(full_counts - 1, 0) + full_counts = np.minimum(full_counts, ctx_counts) + eligible = (ctx_counts >= min_count) & (full_counts > 0) + if not np.any(eligible): + continue + out_idx = valid_idx[eligible] + prob = full_counts[eligible].astype(np.float32) / np.maximum(ctx_counts[eligible].astype(np.float32), 1.0) + ngram_prob[out_idx] = prob + matched_order[out_idx] = n + matched[out_idx] = True + return ngram_prob, matched_order + + +def eval_val_sliding_store( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride, batch_seqs=32, log0=print, +): + seq_len = 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] + my_s = (len(window_starts) * rank) // world_size + my_e = (len(window_starts) * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + model_p_list, entropy_list, bytes_list, position_list, nll_list = [], [], [], [], [] + 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 = [] + for i, ws in enumerate(batch_ws): + end_pos = min(ws + seq_len, total_tokens) + wlen = end_pos - ws + wlens.append(wlen) + chunk = val_tokens[ws:end_pos + 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) + logits_f = logits.float() + log_probs = F.log_softmax(logits_f, dim=-1) + probs = log_probs.exp() + nll_all = F.cross_entropy(logits_f.reshape(-1, logits_f.size(-1)), + y_batch.reshape(-1), reduction="none").reshape(bsz, seq_len) + mp = probs.gather(2, y_batch.unsqueeze(-1)).squeeze(-1) + ent = -(probs * log_probs).sum(dim=-1) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + positions = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + position_list.append(positions) + model_p_list.append(mp[i, s:wlen].cpu().numpy().astype(np.float32)) + entropy_list.append(ent[i, s:wlen].cpu().numpy().astype(np.float32)) + nll_list.append(nll_all[i, s:wlen].cpu().numpy().astype(np.float64)) + 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) + bytes_list.append(tb.cpu().numpy()) + all_positions = np.concatenate(position_list) if position_list else np.array([], dtype=np.int64) + all_model_p = np.concatenate(model_p_list) if model_p_list else np.array([], dtype=np.float32) + all_entropy = np.concatenate(entropy_list) if entropy_list else np.array([], dtype=np.float32) + all_nll = np.concatenate(nll_list) if nll_list else np.array([], dtype=np.float64) + all_bytes = np.concatenate(bytes_list) if bytes_list else np.array([], dtype=np.float64) + loss_sum_t = torch.tensor(all_nll.sum(), device=device, dtype=torch.float64) + token_count_t = torch.tensor(float(len(all_nll)), device=device, dtype=torch.float64) + byte_count_t = torch.tensor(all_bytes.sum(), device=device, dtype=torch.float64) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum_t, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count_t, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count_t, op=dist.ReduceOp.SUM) + val_loss = (loss_sum_t / token_count_t).item() + val_bpb = val_loss / math.log(2.0) * (token_count_t.item() / byte_count_t.item()) + base_model.train() + return all_model_p, all_entropy, all_bytes, all_positions, val_loss, val_bpb + + +def ngram_rescore(args, tokens_np, cache, model_p, entropy, token_bytes, positions, + rank, world_size, device, log0=print): + N = len(positions) + if N == 0: + return 0.0, 0.0 + ngram_prob, matched_order = cache.score_positions( + tokens_np, positions, min_count=args.ngram_min_count, + leave_one_out=args.ngram_leave_one_out, + ) + matched = matched_order >= 0 + alpha = np.zeros(N, dtype=np.float32) + if np.any(matched): + order_idx = (matched_order[matched] - cache.min_order).astype(np.int32) + centers = args.ngram_entropy_center - 0.25 * order_idx.astype(np.float32) + sig = 1.0 / (1.0 + np.exp(-args.ngram_entropy_scale * (entropy[matched] - centers))) + raw_alpha = args.ngram_alpha_min + (args.ngram_alpha_max - args.ngram_alpha_min) * sig + mults = _ORDER_MULTS[np.minimum(order_idx, len(_ORDER_MULTS) - 1)] + raw_alpha *= mults + alpha[matched] = np.clip(raw_alpha, 0.0, 0.95) + p_blend = np.maximum((1.0 - alpha) * model_p + alpha * ngram_prob, 1e-10) + p_blend[~matched] = np.maximum(model_p[~matched], 1e-10) + nll = -np.log(p_blend).astype(np.float64) + loss_sum_t = torch.tensor(nll.sum(), device=device, dtype=torch.float64) + token_count_t = torch.tensor(float(N), device=device, dtype=torch.float64) + byte_count_t = torch.tensor(token_bytes.sum(), device=device, dtype=torch.float64) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum_t, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count_t, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count_t, op=dist.ReduceOp.SUM) + val_loss = (loss_sum_t / token_count_t).item() + val_bpb = val_loss / math.log(2.0) * (token_count_t.item() / byte_count_t.item()) + n_matched = int(matched.sum()) + log0(f"ngram_rescore: matched={n_matched}/{N} ({100*n_matched/max(N,1):.1f}%) " + f"mean_alpha={alpha[matched].mean():.3f}" if n_matched > 0 else f"ngram_rescore: no matches") + return val_loss, val_bpb + + +def eval_ngram_two_pass( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride, batch_seqs=32, log0=print, +): + t0 = time.perf_counter() + log0("ngram_two_pass: starting Pass 1 (sliding-window neural eval)") + model_p, entropy, token_bytes, positions, pass1_loss, pass1_bpb = eval_val_sliding_store( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=stride, batch_seqs=batch_seqs, log0=log0, + ) + t_pass1 = time.perf_counter() + log0(f"ngram_two_pass: Pass 1 done val_bpb={pass1_bpb:.6f} tokens_scored={len(positions)} time={t_pass1-t0:.1f}s") + log0(f"ngram_two_pass: building cache orders={args.ngram_min_order}-{args.ngram_max_order} buckets={args.ngram_num_buckets}") + tokens_np = val_tokens.numpy().astype(np.int16) + cache = NgramCache(min_order=args.ngram_min_order, max_order=args.ngram_max_order, num_buckets=args.ngram_num_buckets) + cache.build_full(tokens_np) + t_cache = time.perf_counter() + log0(f"ngram_two_pass: cache built in {t_cache-t_pass1:.1f}s") + log0("ngram_two_pass: starting Pass 2 (n-gram rescore)") + val_loss, val_bpb = ngram_rescore( + args, tokens_np, cache, model_p, entropy, token_bytes, positions, + rank, world_size, device, log0=log0, + ) + t_pass2 = time.perf_counter() + log0(f"ngram_two_pass: Pass 2 done val_bpb={val_bpb:.6f} improvement={pass1_bpb-val_bpb:.6f} time={t_pass2-t_cache:.1f}s") + log0(f"ngram_two_pass: total time={t_pass2-t0:.1f}s") + return val_loss, val_bpb + +# --- 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") + device = torch.device("cuda", local_rank) + 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) + 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), tokenizer_meta = load_tokenizer_luts( + args.tokenizer_path, args.tokenizer_meta_path, args.vocab_size, device, + validate_meta=args.tokenizer_meta_validate, + ) + log0(f"val_bpb:enabled tokenizer_kind={tokenizer_meta['tokenizer_kind']} 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) + # Full GPTQ: collect Hessians from calibration passes, then quantize with error compensation + gptq_hessians = None + if args.gptq_hessian: + log0(f"gptq:collecting hessians calib_samples={args.gptq_calib_samples}") + t_gptq = time.perf_counter() + gptq_hessians = gptq_collect_hessians(base_model, train_loader, args, device, world_size) + log0(f"gptq:hessians collected keys={len(gptq_hessians)} time={time.perf_counter()-t_gptq:.1f}s") + quant_result, quant_meta = mixed_quantize_int6( + unbanked_sd, {"mlp", "attn"}, + hessians=gptq_hessians, + gptq_block_size=args.gptq_block_size, + gptq_damp_pct=args.gptq_damp_pct, + ) + 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 args.ngram_enabled: + torch.cuda.synchronize() + t_ng = time.perf_counter() + ng_val_loss, ng_val_bpb = eval_ngram_two_pass( + 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"ngram_two_pass val_loss:{ng_val_loss:.4f} val_bpb:{ng_val_bpb:.4f} " + f"eval_time:{1000.0*(time.perf_counter()-t_ng):.0f}ms") + log0(f"ngram_two_pass_exact val_loss:{ng_val_loss:.8f} val_bpb:{ng_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed1337.log b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed1337.log new file mode 100644 index 0000000000..406675eb1c --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed1337.log @@ -0,0 +1,338 @@ +=== PRE-FLIGHT === +python_check_ok +Warning: Permanently added '[103.207.149.67]:12547' (ED25519) to the list of known hosts. + +=== BOOTSTRAP + TRAINING === +Warning: Permanently added '[103.207.149.67]:12547' (ED25519) to the list of known hosts. +====================================================================== +bootstrap_scylla_full_run.sh + SEED=1337 NPROC=8 + DATA_PATH=./data/datasets/fineweb_scylla +====================================================================== + +=== Step 0: Python environment === +/usr/local/lib/python3.12/dist-packages/torch/_subclasses/functional_tensor.py:279: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.) + cpu = _conversion_method_template(device=torch.device("cpu")) +python3 -> Python 3.12.12, torch=2.9.1+cu128 + +=== Step 1: Install dependencies === +WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning. + +[notice] A new release of pip is available: 25.3 -> 26.0.1 +[notice] To update, run: python3 -m pip install --upgrade pip +Dependencies installed. + +=== Step 2: Verify tokenizer artifacts === + candidate.vocab: vocab_size=998 OK + candidate.meta.npz: keys=base_bytes,format_version,has_leading_space,is_boundary_token,source_model_name,tokenizer_kind,vocab_size OK + +=== Step 3: Acquire Scylla shards === + Existing shards: train=0 +0 val=0 +0 + SHARD_SOURCE_URL=https://m9a3es9ein9yh4-19123.proxy.runpod.net + Downloading pre-tokenized Scylla shards from HTTP source... +scripts/bootstrap_scylla_full_run.sh: line 130: [: 0 +0: integer expression expected + Download complete in 219s train_shards=80 + +=== Step 5: Verify shard outputs === + Train shards: 80 / 80 + Val shards: 1 / 1 + fineweb_train_000000.bin: magic=OK version=1 ntokens=100,000,175 size=200.0MB + fineweb_val_000000.bin: magic=OK version=1 ntokens=62,365,136 size=124.7MB + Shard verification: PASSED + +=== Step 6: Launch training (SEED=1337, NPROC=8) === + Config: locked Scylla full-run + DATA_PATH=./data/datasets/fineweb_scylla + TOKENIZER_PATH=./data/tokenizers/scylla/candidate.vocab + TOKENIZER_META_PATH=./data/tokenizers/scylla/candidate.meta.npz +W0401 23:53:58.943000 2398 torch/distributed/run.py:803] +W0401 23:53:58.943000 2398 torch/distributed/run.py:803] ***************************************** +W0401 23:53:58.943000 2398 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0401 23:53:58.943000 2398 torch/distributed/run.py:803] ***************************************** +logs/c721c885-c4f2-42d6-8f3b-2ab714dfa4c8.txt +val_bpb:enabled tokenizer_kind=tokenmonster tokenizer_path=./data/tokenizers/scylla/candidate.vocab +train_loader:dataset:fineweb_scylla train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb_scylla/fineweb_val_*.bin tokens:62363648 +model_params:28042332 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +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:9000 warmup_steps:20 max_wallclock_seconds:600.000 +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:0/9000 val_loss:6.9196 val_bpb:3.8684 train_time:0ms step_avg:0.01ms +step:1/9000 train_loss:6.9177 train_time:152ms step_avg:152.48ms +step:2/9000 train_loss:9.1754 train_time:188ms step_avg:93.76ms +step:3/9000 train_loss:8.9028 train_time:281ms step_avg:93.77ms +step:4/9000 train_loss:8.3455 train_time:376ms step_avg:94.04ms +step:5/9000 train_loss:7.6653 train_time:471ms step_avg:94.28ms +step:6/9000 train_loss:7.3441 train_time:567ms step_avg:94.44ms +step:7/9000 train_loss:6.9563 train_time:661ms step_avg:94.45ms +step:8/9000 train_loss:6.9140 train_time:757ms step_avg:94.67ms +step:9/9000 train_loss:6.5448 train_time:851ms step_avg:94.61ms +step:10/9000 train_loss:6.3127 train_time:947ms step_avg:94.65ms +step:500/9000 train_loss:2.3993 train_time:48389ms step_avg:96.78ms +step:1000/9000 train_loss:2.2809 train_time:97235ms step_avg:97.23ms +step:1500/9000 train_loss:2.0851 train_time:146033ms step_avg:97.36ms +step:2000/9000 train_loss:2.1999 train_time:194822ms step_avg:97.41ms +step:2500/9000 train_loss:2.1855 train_time:243591ms step_avg:97.44ms +step:3000/9000 train_loss:2.1184 train_time:292399ms step_avg:97.47ms +step:3500/9000 train_loss:2.0059 train_time:341124ms step_avg:97.46ms +step:4000/9000 train_loss:2.0291 train_time:389834ms step_avg:97.46ms +step:4000/9000 val_loss:2.0756 val_bpb:1.1603 train_time:389903ms step_avg:97.48ms +step:4500/9000 train_loss:1.9333 train_time:438550ms step_avg:97.46ms +step:5000/9000 train_loss:2.1720 train_time:487217ms step_avg:97.44ms +swa:start step:5500 +step:5500/9000 train_loss:1.9618 train_time:535877ms step_avg:97.43ms +late_qat:enabled step:5630 scale:0.1500 +step:6000/9000 train_loss:1.9286 train_time:585231ms step_avg:97.54ms +step:6151/9000 val_loss:1.9727 val_bpb:1.1028 train_time:600237ms step_avg:97.58ms +stopping_early: wallclock_cap train_time:600237ms step:6151/9000 +peak memory allocated: 22928 MiB reserved: 22948 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9712 val_bpb:1.1020 eval_time:2337ms +Serialized model: 108255670 bytes +Code size: 115885 bytes +Serialized model int6+lzma: 15200324 bytes +Total submission size int6+lzma: 15316209 bytes +final_int6_roundtrip val_loss:2.0039 val_bpb:1.1203 eval_time:19857ms +final_int6_roundtrip_exact val_loss:2.00388982 val_bpb:1.12027810 +final_int6_sliding_window val_loss:1.9640 val_bpb:1.0980 stride:64 eval_time:107310ms +final_int6_sliding_window_exact val_loss:1.96402724 val_bpb:1.09798465 +final_int8_zlib_roundtrip_exact val_loss:1.96402724 val_bpb:1.09798465 +ttt_sliding:start chunks=1904 chunk_tokens=32768 total_windows=974432 stride=64 ttt_lr=0.005 ttt_epochs=3 freeze_blocks=2 adaptive=False +ttt_sliding:params unfrozen=28038220 frozen=4112 + ttt_chunk [1/1904] bpb=1.054894 time=0.5s + ttt_chunk [11/1904] bpb=1.055610 time=2.8s + ttt_chunk [21/1904] bpb=1.082887 time=5.0s + ttt_chunk [31/1904] bpb=1.087592 time=7.3s + ttt_chunk [41/1904] bpb=1.118241 time=9.5s + ttt_chunk [51/1904] bpb=1.116396 time=11.7s + ttt_chunk [61/1904] bpb=1.117567 time=14.0s + ttt_chunk [71/1904] bpb=1.109776 time=16.2s + ttt_chunk [81/1904] bpb=1.123131 time=18.4s + ttt_chunk [91/1904] bpb=1.114731 time=20.7s + ttt_chunk [101/1904] bpb=1.112598 time=22.9s + ttt_chunk [111/1904] bpb=1.107152 time=25.2s + ttt_chunk [121/1904] bpb=1.108043 time=27.4s + ttt_chunk [131/1904] bpb=1.107816 time=29.6s + ttt_chunk [141/1904] bpb=1.103947 time=31.9s + ttt_chunk [151/1904] bpb=1.100255 time=34.1s + ttt_chunk [161/1904] bpb=1.099649 time=36.3s + ttt_chunk [171/1904] bpb=1.097074 time=38.6s + ttt_chunk [181/1904] bpb=1.096028 time=40.8s + ttt_chunk [191/1904] bpb=1.096203 time=43.1s + ttt_chunk [201/1904] bpb=1.097756 time=45.3s + ttt_chunk [211/1904] bpb=1.096074 time=47.6s + ttt_chunk [221/1904] bpb=1.091891 time=49.8s + ttt_chunk [231/1904] bpb=1.089406 time=52.1s + ttt_chunk [241/1904] bpb=1.092005 time=54.3s + ttt_chunk [251/1904] bpb=1.090863 time=56.6s + ttt_chunk [261/1904] bpb=1.087582 time=58.8s + ttt_chunk [271/1904] bpb=1.087240 time=61.0s + ttt_chunk [281/1904] bpb=1.086764 time=63.3s + ttt_chunk [291/1904] bpb=1.085729 time=65.5s + ttt_chunk [301/1904] bpb=1.086570 time=67.8s + ttt_chunk [311/1904] bpb=1.087497 time=70.0s + ttt_chunk [321/1904] bpb=1.087488 time=72.3s + ttt_chunk [331/1904] bpb=1.087417 time=74.5s + ttt_chunk [341/1904] bpb=1.086031 time=76.7s + ttt_chunk [351/1904] bpb=1.085312 time=79.0s + ttt_chunk [361/1904] bpb=1.086841 time=81.2s + ttt_chunk [371/1904] bpb=1.086096 time=83.5s + ttt_chunk [381/1904] bpb=1.085486 time=85.7s + ttt_chunk [391/1904] bpb=1.084632 time=88.0s + ttt_chunk [401/1904] bpb=1.084362 time=90.2s + ttt_chunk [411/1904] bpb=1.083136 time=92.5s + ttt_chunk [421/1904] bpb=1.081497 time=94.7s + ttt_chunk [431/1904] bpb=1.079821 time=96.9s + ttt_chunk [441/1904] bpb=1.079107 time=99.2s + ttt_chunk [451/1904] bpb=1.076907 time=101.4s + ttt_chunk [461/1904] bpb=1.076710 time=103.7s + ttt_chunk [471/1904] bpb=1.075427 time=105.9s + ttt_chunk [481/1904] bpb=1.075757 time=108.1s + ttt_chunk [491/1904] bpb=1.076231 time=110.3s + ttt_chunk [501/1904] bpb=1.076831 time=112.6s + ttt_chunk [511/1904] bpb=1.077767 time=114.8s + ttt_chunk [521/1904] bpb=1.080749 time=117.0s + ttt_chunk [531/1904] bpb=1.080722 time=119.3s + ttt_chunk [541/1904] bpb=1.082046 time=121.5s + ttt_chunk [551/1904] bpb=1.081395 time=123.7s + ttt_chunk [561/1904] bpb=1.081296 time=126.0s + ttt_chunk [571/1904] bpb=1.081679 time=128.2s + ttt_chunk [581/1904] bpb=1.082895 time=130.5s + ttt_chunk [591/1904] bpb=1.083581 time=132.7s + ttt_chunk [601/1904] bpb=1.083631 time=135.0s + ttt_chunk [611/1904] bpb=1.083254 time=137.2s + ttt_chunk [621/1904] bpb=1.084947 time=139.4s + ttt_chunk [631/1904] bpb=1.085482 time=141.7s + ttt_chunk [641/1904] bpb=1.086112 time=143.9s + ttt_chunk [651/1904] bpb=1.085911 time=146.1s + ttt_chunk [661/1904] bpb=1.086411 time=148.4s + ttt_chunk [671/1904] bpb=1.087078 time=150.6s + ttt_chunk [681/1904] bpb=1.087155 time=152.8s + ttt_chunk [691/1904] bpb=1.087187 time=155.1s + ttt_chunk [701/1904] bpb=1.087881 time=157.3s + ttt_chunk [711/1904] bpb=1.088333 time=159.5s + ttt_chunk [721/1904] bpb=1.087666 time=161.8s + ttt_chunk [731/1904] bpb=1.087562 time=164.0s + ttt_chunk [741/1904] bpb=1.087439 time=166.2s + ttt_chunk [751/1904] bpb=1.086691 time=168.5s + ttt_chunk [761/1904] bpb=1.086538 time=170.7s + ttt_chunk [771/1904] bpb=1.085989 time=172.9s + ttt_chunk [781/1904] bpb=1.085819 time=175.2s + ttt_chunk [791/1904] bpb=1.085675 time=177.4s + ttt_chunk [801/1904] bpb=1.086011 time=179.7s + ttt_chunk [811/1904] bpb=1.087074 time=181.9s + ttt_chunk [821/1904] bpb=1.086378 time=184.1s + ttt_chunk [831/1904] bpb=1.087082 time=186.4s + ttt_chunk [841/1904] bpb=1.087021 time=188.6s + ttt_chunk [851/1904] bpb=1.087372 time=190.9s + ttt_chunk [861/1904] bpb=1.087830 time=193.1s + ttt_chunk [871/1904] bpb=1.088134 time=195.3s + ttt_chunk [881/1904] bpb=1.088265 time=197.5s + ttt_chunk [891/1904] bpb=1.089295 time=199.7s + ttt_chunk [901/1904] bpb=1.089975 time=202.0s + ttt_chunk [911/1904] bpb=1.090712 time=204.2s + ttt_chunk [921/1904] bpb=1.091622 time=206.4s + ttt_chunk [931/1904] bpb=1.092122 time=208.7s + ttt_chunk [941/1904] bpb=1.091715 time=210.9s + ttt_chunk [951/1904] bpb=1.091911 time=213.1s + ttt_chunk [961/1904] bpb=1.091917 time=215.3s + ttt_chunk [971/1904] bpb=1.092524 time=217.6s + ttt_chunk [981/1904] bpb=1.092918 time=219.8s + ttt_chunk [991/1904] bpb=1.092711 time=222.1s + ttt_chunk [1001/1904] bpb=1.092846 time=224.3s + ttt_chunk [1011/1904] bpb=1.093511 time=226.6s + ttt_chunk [1021/1904] bpb=1.094133 time=228.8s + ttt_chunk [1031/1904] bpb=1.093783 time=231.0s + ttt_chunk [1041/1904] bpb=1.093591 time=233.3s + ttt_chunk [1051/1904] bpb=1.093217 time=235.5s + ttt_chunk [1061/1904] bpb=1.092486 time=237.7s + ttt_chunk [1071/1904] bpb=1.092339 time=240.0s + ttt_chunk [1081/1904] bpb=1.092128 time=242.2s + ttt_chunk [1091/1904] bpb=1.092001 time=244.4s + ttt_chunk [1101/1904] bpb=1.092208 time=246.6s + ttt_chunk [1111/1904] bpb=1.092181 time=248.9s + ttt_chunk [1121/1904] bpb=1.092484 time=251.1s + ttt_chunk [1131/1904] bpb=1.092416 time=253.3s + ttt_chunk [1141/1904] bpb=1.092756 time=255.6s + ttt_chunk [1151/1904] bpb=1.092325 time=257.8s + ttt_chunk [1161/1904] bpb=1.092349 time=260.0s + ttt_chunk [1171/1904] bpb=1.092047 time=262.3s + ttt_chunk [1181/1904] bpb=1.092121 time=264.5s + ttt_chunk [1191/1904] bpb=1.091601 time=266.7s + ttt_chunk [1201/1904] bpb=1.091936 time=268.9s + ttt_chunk [1211/1904] bpb=1.092087 time=271.2s + ttt_chunk [1221/1904] bpb=1.092161 time=273.4s + ttt_chunk [1231/1904] bpb=1.092952 time=275.7s + ttt_chunk [1241/1904] bpb=1.093045 time=277.9s + ttt_chunk [1251/1904] bpb=1.093643 time=280.1s + ttt_chunk [1261/1904] bpb=1.093420 time=282.4s + ttt_chunk [1271/1904] bpb=1.093560 time=284.6s + ttt_chunk [1281/1904] bpb=1.093789 time=286.9s + ttt_chunk [1291/1904] bpb=1.094141 time=289.1s + ttt_chunk [1301/1904] bpb=1.094773 time=291.3s + ttt_chunk [1311/1904] bpb=1.095029 time=293.6s + ttt_chunk [1321/1904] bpb=1.095595 time=295.8s + ttt_chunk [1331/1904] bpb=1.095527 time=298.1s + ttt_chunk [1341/1904] bpb=1.095892 time=300.3s + ttt_chunk [1351/1904] bpb=1.096343 time=302.5s + ttt_chunk [1361/1904] bpb=1.096751 time=304.8s + ttt_chunk [1371/1904] bpb=1.096533 time=307.0s + ttt_chunk [1381/1904] bpb=1.095908 time=309.2s + ttt_chunk [1391/1904] bpb=1.096146 time=311.5s + ttt_chunk [1401/1904] bpb=1.095667 time=313.7s + ttt_chunk [1411/1904] bpb=1.095635 time=316.0s + ttt_chunk [1421/1904] bpb=1.095545 time=318.2s + ttt_chunk [1431/1904] bpb=1.095562 time=320.4s + ttt_chunk [1441/1904] bpb=1.095593 time=322.7s + ttt_chunk [1451/1904] bpb=1.095629 time=324.9s + ttt_chunk [1461/1904] bpb=1.095886 time=327.1s + ttt_chunk [1471/1904] bpb=1.095631 time=329.4s + ttt_chunk [1481/1904] bpb=1.095338 time=331.6s + ttt_chunk [1491/1904] bpb=1.094517 time=333.8s + ttt_chunk [1501/1904] bpb=1.095074 time=336.1s + ttt_chunk [1511/1904] bpb=1.095000 time=338.3s + ttt_chunk [1521/1904] bpb=1.094789 time=340.5s + ttt_chunk [1531/1904] bpb=1.094410 time=342.8s + ttt_chunk [1541/1904] bpb=1.094796 time=345.1s + ttt_chunk [1551/1904] bpb=1.094866 time=347.3s + ttt_chunk [1561/1904] bpb=1.094985 time=349.6s + ttt_chunk [1571/1904] bpb=1.095108 time=351.8s + ttt_chunk [1581/1904] bpb=1.094970 time=354.0s + ttt_chunk [1591/1904] bpb=1.094646 time=356.2s + ttt_chunk [1601/1904] bpb=1.094974 time=358.5s + ttt_chunk [1611/1904] bpb=1.094930 time=360.7s + ttt_chunk [1621/1904] bpb=1.095099 time=362.9s + ttt_chunk [1631/1904] bpb=1.095130 time=365.2s + ttt_chunk [1641/1904] bpb=1.095004 time=367.4s + ttt_chunk [1651/1904] bpb=1.095186 time=369.6s + ttt_chunk [1661/1904] bpb=1.095366 time=371.9s + ttt_chunk [1671/1904] bpb=1.094845 time=374.1s + ttt_chunk [1681/1904] bpb=1.095056 time=376.3s + ttt_chunk [1691/1904] bpb=1.094855 time=378.5s + ttt_chunk [1701/1904] bpb=1.094916 time=380.8s + ttt_chunk [1711/1904] bpb=1.095448 time=383.0s + ttt_chunk [1721/1904] bpb=1.095527 time=385.3s + ttt_chunk [1731/1904] bpb=1.096169 time=387.5s + ttt_chunk [1741/1904] bpb=1.096064 time=389.8s + ttt_chunk [1751/1904] bpb=1.096179 time=392.0s + ttt_chunk [1761/1904] bpb=1.096028 time=394.2s + ttt_chunk [1771/1904] bpb=1.096003 time=396.5s + ttt_chunk [1781/1904] bpb=1.095826 time=398.7s + ttt_chunk [1791/1904] bpb=1.095615 time=400.9s + ttt_chunk [1801/1904] bpb=1.095525 time=403.2s + ttt_chunk [1811/1904] bpb=1.095529 time=405.4s + ttt_chunk [1821/1904] bpb=1.095235 time=407.7s + ttt_chunk [1831/1904] bpb=1.095377 time=409.9s + ttt_chunk [1841/1904] bpb=1.095274 time=412.2s + ttt_chunk [1851/1904] bpb=1.095062 time=414.4s + ttt_chunk [1861/1904] bpb=1.094945 time=416.7s + ttt_chunk [1871/1904] bpb=1.094863 time=418.9s + ttt_chunk [1881/1904] bpb=1.095022 time=421.1s + ttt_chunk [1891/1904] bpb=1.094469 time=423.4s + ttt_chunk [1901/1904] bpb=1.094760 time=425.6s + ttt_chunk [1904/1904] bpb=1.094770 time=426.1s +ttt_sliding:done val_loss=1.950487 val_bpb=1.090415 elapsed=426.1s +legal_ttt val_loss:1.9505 val_bpb:1.0904 eval_time:426473ms +legal_ttt_exact val_loss:1.95048726 val_bpb:1.09041515 +ngram_two_pass: starting Pass 1 (sliding-window neural eval) +ngram_two_pass: Pass 1 done val_bpb=1.088839 tokens_scored=7797440 time=99.8s +ngram_two_pass: building cache orders=2-16 buckets=16777216 +ngram_two_pass: cache built in 61.5s +ngram_two_pass: starting Pass 2 (n-gram rescore) +ngram_rescore: matched=7797440/7797440 (100.0%) mean_alpha=0.948 +ngram_two_pass: Pass 2 done val_bpb=0.195180 improvement=0.893659 time=39.1s +ngram_two_pass: total time=200.4s +ngram_two_pass val_loss:0.3491 val_bpb:0.1952 eval_time:200551ms +ngram_two_pass_exact val_loss:0.34912988 val_bpb:0.19518021 + +====================================================================== +Training complete — SEED=1337 +====================================================================== diff --git a/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed2025.log b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed2025.log new file mode 100644 index 0000000000..1f4e48b393 --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed2025.log @@ -0,0 +1,338 @@ +=== PRE-FLIGHT === +python_check_ok +Warning: Permanently added '[103.207.149.67]:16173' (ED25519) to the list of known hosts. + +=== BOOTSTRAP + TRAINING === +Warning: Permanently added '[103.207.149.67]:16173' (ED25519) to the list of known hosts. +====================================================================== +bootstrap_scylla_full_run.sh + SEED=2025 NPROC=8 + DATA_PATH=./data/datasets/fineweb_scylla +====================================================================== + +=== Step 0: Python environment === +/usr/local/lib/python3.12/dist-packages/torch/_subclasses/functional_tensor.py:279: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.) + cpu = _conversion_method_template(device=torch.device("cpu")) +python3 -> Python 3.12.12, torch=2.9.1+cu128 + +=== Step 1: Install dependencies === +WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning. + +[notice] A new release of pip is available: 25.3 -> 26.0.1 +[notice] To update, run: python3 -m pip install --upgrade pip +Dependencies installed. + +=== Step 2: Verify tokenizer artifacts === + candidate.vocab: vocab_size=998 OK + candidate.meta.npz: keys=base_bytes,format_version,has_leading_space,is_boundary_token,source_model_name,tokenizer_kind,vocab_size OK + +=== Step 3: Acquire Scylla shards === + Existing shards: train=0 +0 val=0 +0 + SHARD_SOURCE_URL=https://m9a3es9ein9yh4-19123.proxy.runpod.net + Downloading pre-tokenized Scylla shards from HTTP source... +scripts/bootstrap_scylla_full_run.sh: line 130: [: 0 +0: integer expression expected + Download complete in 207s train_shards=80 + +=== Step 5: Verify shard outputs === + Train shards: 80 / 80 + Val shards: 1 / 1 + fineweb_train_000000.bin: magic=OK version=1 ntokens=100,000,175 size=200.0MB + fineweb_val_000000.bin: magic=OK version=1 ntokens=62,365,136 size=124.7MB + Shard verification: PASSED + +=== Step 6: Launch training (SEED=2025, NPROC=8) === + Config: locked Scylla full-run + DATA_PATH=./data/datasets/fineweb_scylla + TOKENIZER_PATH=./data/tokenizers/scylla/candidate.vocab + TOKENIZER_META_PATH=./data/tokenizers/scylla/candidate.meta.npz +W0401 22:52:19.793000 2363 torch/distributed/run.py:803] +W0401 22:52:19.793000 2363 torch/distributed/run.py:803] ***************************************** +W0401 22:52:19.793000 2363 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0401 22:52:19.793000 2363 torch/distributed/run.py:803] ***************************************** +logs/f7c0102e-680a-41de-a670-4b02130dc47e.txt +val_bpb:enabled tokenizer_kind=tokenmonster tokenizer_path=./data/tokenizers/scylla/candidate.vocab +train_loader:dataset:fineweb_scylla train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb_scylla/fineweb_val_*.bin tokens:62363648 +model_params:28042332 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +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:9000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:2025 +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:0/9000 val_loss:6.9148 val_bpb:3.8657 train_time:0ms step_avg:0.02ms +step:1/9000 train_loss:6.9131 train_time:151ms step_avg:151.11ms +step:2/9000 train_loss:9.0225 train_time:182ms step_avg:90.96ms +step:3/9000 train_loss:9.0262 train_time:276ms step_avg:91.93ms +step:4/9000 train_loss:8.5687 train_time:370ms step_avg:92.53ms +step:5/9000 train_loss:7.8871 train_time:466ms step_avg:93.27ms +step:6/9000 train_loss:7.3884 train_time:561ms step_avg:93.42ms +step:7/9000 train_loss:6.9636 train_time:655ms step_avg:93.64ms +step:8/9000 train_loss:6.8882 train_time:750ms step_avg:93.75ms +step:9/9000 train_loss:6.5444 train_time:845ms step_avg:93.89ms +step:10/9000 train_loss:6.3197 train_time:940ms step_avg:93.99ms +step:500/9000 train_loss:2.4049 train_time:48214ms step_avg:96.43ms +step:1000/9000 train_loss:2.2857 train_time:96895ms step_avg:96.89ms +step:1500/9000 train_loss:2.0842 train_time:145555ms step_avg:97.04ms +step:2000/9000 train_loss:2.2012 train_time:194157ms step_avg:97.08ms +step:2500/9000 train_loss:2.1855 train_time:242748ms step_avg:97.10ms +step:3000/9000 train_loss:2.1196 train_time:291353ms step_avg:97.12ms +step:3500/9000 train_loss:2.0045 train_time:339885ms step_avg:97.11ms +step:4000/9000 train_loss:2.0282 train_time:388527ms step_avg:97.13ms +step:4000/9000 val_loss:2.0760 val_bpb:1.1606 train_time:388594ms step_avg:97.15ms +step:4500/9000 train_loss:1.9367 train_time:437059ms step_avg:97.12ms +step:5000/9000 train_loss:2.1682 train_time:485588ms step_avg:97.12ms +swa:start step:5500 +step:5500/9000 train_loss:1.9559 train_time:534094ms step_avg:97.11ms +late_qat:enabled step:5650 scale:0.1499 +step:6000/9000 train_loss:1.9313 train_time:583327ms step_avg:97.22ms +step:6170/9000 val_loss:1.9714 val_bpb:1.1021 train_time:600128ms step_avg:97.27ms +stopping_early: wallclock_cap train_time:600128ms step:6170/9000 +peak memory allocated: 22928 MiB reserved: 22948 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9699 val_bpb:1.1013 eval_time:2336ms +Serialized model: 108255670 bytes +Code size: 115885 bytes +Serialized model int6+lzma: 14830080 bytes +Total submission size int6+lzma: 14945965 bytes +final_int6_roundtrip val_loss:2.0063 val_bpb:1.1216 eval_time:20990ms +final_int6_roundtrip_exact val_loss:2.00630171 val_bpb:1.12162648 +final_int6_sliding_window val_loss:1.9661 val_bpb:1.0992 stride:64 eval_time:108937ms +final_int6_sliding_window_exact val_loss:1.96613210 val_bpb:1.09916136 +final_int8_zlib_roundtrip_exact val_loss:1.96613210 val_bpb:1.09916136 +ttt_sliding:start chunks=1904 chunk_tokens=32768 total_windows=974432 stride=64 ttt_lr=0.005 ttt_epochs=3 freeze_blocks=2 adaptive=False +ttt_sliding:params unfrozen=28038220 frozen=4112 + ttt_chunk [1/1904] bpb=1.057617 time=0.5s + ttt_chunk [11/1904] bpb=1.057336 time=2.7s + ttt_chunk [21/1904] bpb=1.084272 time=5.0s + ttt_chunk [31/1904] bpb=1.088110 time=7.2s + ttt_chunk [41/1904] bpb=1.118325 time=9.4s + ttt_chunk [51/1904] bpb=1.116715 time=11.7s + ttt_chunk [61/1904] bpb=1.117770 time=13.9s + ttt_chunk [71/1904] bpb=1.110174 time=16.2s + ttt_chunk [81/1904] bpb=1.123224 time=18.4s + ttt_chunk [91/1904] bpb=1.114779 time=20.7s + ttt_chunk [101/1904] bpb=1.112509 time=22.9s + ttt_chunk [111/1904] bpb=1.107212 time=25.1s + ttt_chunk [121/1904] bpb=1.108004 time=27.4s + ttt_chunk [131/1904] bpb=1.107557 time=29.6s + ttt_chunk [141/1904] bpb=1.103460 time=31.9s + ttt_chunk [151/1904] bpb=1.099967 time=34.1s + ttt_chunk [161/1904] bpb=1.099428 time=36.4s + ttt_chunk [171/1904] bpb=1.096886 time=38.6s + ttt_chunk [181/1904] bpb=1.096124 time=40.8s + ttt_chunk [191/1904] bpb=1.096142 time=43.1s + ttt_chunk [201/1904] bpb=1.097662 time=45.3s + ttt_chunk [211/1904] bpb=1.096065 time=47.6s + ttt_chunk [221/1904] bpb=1.091886 time=49.8s + ttt_chunk [231/1904] bpb=1.089306 time=52.1s + ttt_chunk [241/1904] bpb=1.092019 time=54.3s + ttt_chunk [251/1904] bpb=1.090761 time=56.6s + ttt_chunk [261/1904] bpb=1.087535 time=58.8s + ttt_chunk [271/1904] bpb=1.087241 time=61.1s + ttt_chunk [281/1904] bpb=1.086652 time=63.3s + ttt_chunk [291/1904] bpb=1.085686 time=65.6s + ttt_chunk [301/1904] bpb=1.086511 time=67.8s + ttt_chunk [311/1904] bpb=1.087499 time=70.0s + ttt_chunk [321/1904] bpb=1.087469 time=72.3s + ttt_chunk [331/1904] bpb=1.087337 time=74.5s + ttt_chunk [341/1904] bpb=1.086033 time=76.8s + ttt_chunk [351/1904] bpb=1.085282 time=79.0s + ttt_chunk [361/1904] bpb=1.086803 time=81.2s + ttt_chunk [371/1904] bpb=1.086003 time=83.5s + ttt_chunk [381/1904] bpb=1.085258 time=85.7s + ttt_chunk [391/1904] bpb=1.084345 time=87.9s + ttt_chunk [401/1904] bpb=1.084105 time=90.2s + ttt_chunk [411/1904] bpb=1.082865 time=92.4s + ttt_chunk [421/1904] bpb=1.081225 time=94.6s + ttt_chunk [431/1904] bpb=1.079591 time=96.9s + ttt_chunk [441/1904] bpb=1.078856 time=99.1s + ttt_chunk [451/1904] bpb=1.076699 time=101.3s + ttt_chunk [461/1904] bpb=1.076441 time=103.6s + ttt_chunk [471/1904] bpb=1.075102 time=105.8s + ttt_chunk [481/1904] bpb=1.075514 time=108.1s + ttt_chunk [491/1904] bpb=1.076034 time=110.3s + ttt_chunk [501/1904] bpb=1.076637 time=112.5s + ttt_chunk [511/1904] bpb=1.077577 time=114.8s + ttt_chunk [521/1904] bpb=1.080582 time=117.1s + ttt_chunk [531/1904] bpb=1.080507 time=119.3s + ttt_chunk [541/1904] bpb=1.081836 time=121.6s + ttt_chunk [551/1904] bpb=1.081148 time=123.8s + ttt_chunk [561/1904] bpb=1.081044 time=126.0s + ttt_chunk [571/1904] bpb=1.081422 time=128.3s + ttt_chunk [581/1904] bpb=1.082585 time=130.5s + ttt_chunk [591/1904] bpb=1.083261 time=132.8s + ttt_chunk [601/1904] bpb=1.083374 time=135.0s + ttt_chunk [611/1904] bpb=1.082954 time=137.2s + ttt_chunk [621/1904] bpb=1.084630 time=139.4s + ttt_chunk [631/1904] bpb=1.085182 time=141.6s + ttt_chunk [641/1904] bpb=1.085798 time=143.9s + ttt_chunk [651/1904] bpb=1.085591 time=146.1s + ttt_chunk [661/1904] bpb=1.086132 time=148.4s + ttt_chunk [671/1904] bpb=1.086828 time=150.6s + ttt_chunk [681/1904] bpb=1.086907 time=152.9s + ttt_chunk [691/1904] bpb=1.086913 time=155.1s + ttt_chunk [701/1904] bpb=1.087630 time=157.4s + ttt_chunk [711/1904] bpb=1.088028 time=159.6s + ttt_chunk [721/1904] bpb=1.087360 time=161.9s + ttt_chunk [731/1904] bpb=1.087262 time=164.1s + ttt_chunk [741/1904] bpb=1.087125 time=166.4s + ttt_chunk [751/1904] bpb=1.086378 time=168.6s + ttt_chunk [761/1904] bpb=1.086224 time=170.9s + ttt_chunk [771/1904] bpb=1.085681 time=173.1s + ttt_chunk [781/1904] bpb=1.085497 time=175.4s + ttt_chunk [791/1904] bpb=1.085345 time=177.6s + ttt_chunk [801/1904] bpb=1.085686 time=179.9s + ttt_chunk [811/1904] bpb=1.086758 time=182.2s + ttt_chunk [821/1904] bpb=1.086067 time=184.5s + ttt_chunk [831/1904] bpb=1.086764 time=186.7s + ttt_chunk [841/1904] bpb=1.086677 time=188.9s + ttt_chunk [851/1904] bpb=1.087021 time=191.2s + ttt_chunk [861/1904] bpb=1.087490 time=193.4s + ttt_chunk [871/1904] bpb=1.087793 time=195.6s + ttt_chunk [881/1904] bpb=1.087940 time=197.9s + ttt_chunk [891/1904] bpb=1.088956 time=200.1s + ttt_chunk [901/1904] bpb=1.089606 time=202.3s + ttt_chunk [911/1904] bpb=1.090334 time=204.6s + ttt_chunk [921/1904] bpb=1.091237 time=206.8s + ttt_chunk [931/1904] bpb=1.091702 time=209.0s + ttt_chunk [941/1904] bpb=1.091291 time=211.3s + ttt_chunk [951/1904] bpb=1.091506 time=213.5s + ttt_chunk [961/1904] bpb=1.091523 time=215.7s + ttt_chunk [971/1904] bpb=1.092095 time=218.0s + ttt_chunk [981/1904] bpb=1.092479 time=220.2s + ttt_chunk [991/1904] bpb=1.092262 time=222.4s + ttt_chunk [1001/1904] bpb=1.092398 time=224.7s + ttt_chunk [1011/1904] bpb=1.093041 time=226.9s + ttt_chunk [1021/1904] bpb=1.093675 time=229.1s + ttt_chunk [1031/1904] bpb=1.093290 time=231.4s + ttt_chunk [1041/1904] bpb=1.093101 time=233.6s + ttt_chunk [1051/1904] bpb=1.092725 time=235.9s + ttt_chunk [1061/1904] bpb=1.092001 time=238.1s + ttt_chunk [1071/1904] bpb=1.091825 time=240.4s + ttt_chunk [1081/1904] bpb=1.091606 time=242.6s + ttt_chunk [1091/1904] bpb=1.091471 time=244.8s + ttt_chunk [1101/1904] bpb=1.091680 time=247.0s + ttt_chunk [1111/1904] bpb=1.091640 time=249.3s + ttt_chunk [1121/1904] bpb=1.091949 time=251.5s + ttt_chunk [1131/1904] bpb=1.091891 time=253.8s + ttt_chunk [1141/1904] bpb=1.092212 time=256.0s + ttt_chunk [1151/1904] bpb=1.091762 time=258.3s + ttt_chunk [1161/1904] bpb=1.091772 time=260.5s + ttt_chunk [1171/1904] bpb=1.091469 time=262.7s + ttt_chunk [1181/1904] bpb=1.091569 time=265.0s + ttt_chunk [1191/1904] bpb=1.091059 time=267.2s + ttt_chunk [1201/1904] bpb=1.091386 time=269.4s + ttt_chunk [1211/1904] bpb=1.091549 time=271.7s + ttt_chunk [1221/1904] bpb=1.091632 time=273.9s + ttt_chunk [1231/1904] bpb=1.092419 time=276.1s + ttt_chunk [1241/1904] bpb=1.092527 time=278.4s + ttt_chunk [1251/1904] bpb=1.093126 time=280.6s + ttt_chunk [1261/1904] bpb=1.092890 time=282.9s + ttt_chunk [1271/1904] bpb=1.093048 time=285.1s + ttt_chunk [1281/1904] bpb=1.093271 time=287.4s + ttt_chunk [1291/1904] bpb=1.093619 time=289.6s + ttt_chunk [1301/1904] bpb=1.094244 time=291.9s + ttt_chunk [1311/1904] bpb=1.094491 time=294.1s + ttt_chunk [1321/1904] bpb=1.095073 time=296.4s + ttt_chunk [1331/1904] bpb=1.095019 time=298.6s + ttt_chunk [1341/1904] bpb=1.095379 time=300.8s + ttt_chunk [1351/1904] bpb=1.095844 time=303.1s + ttt_chunk [1361/1904] bpb=1.096220 time=305.3s + ttt_chunk [1371/1904] bpb=1.096015 time=307.5s + ttt_chunk [1381/1904] bpb=1.095378 time=309.8s + ttt_chunk [1391/1904] bpb=1.095611 time=312.0s + ttt_chunk [1401/1904] bpb=1.095136 time=314.2s + ttt_chunk [1411/1904] bpb=1.095113 time=316.5s + ttt_chunk [1421/1904] bpb=1.095022 time=318.7s + ttt_chunk [1431/1904] bpb=1.095040 time=320.9s + ttt_chunk [1441/1904] bpb=1.095078 time=323.1s + ttt_chunk [1451/1904] bpb=1.095096 time=325.4s + ttt_chunk [1461/1904] bpb=1.095353 time=327.6s + ttt_chunk [1471/1904] bpb=1.095104 time=329.8s + ttt_chunk [1481/1904] bpb=1.094819 time=332.1s + ttt_chunk [1491/1904] bpb=1.094004 time=334.3s + ttt_chunk [1501/1904] bpb=1.094567 time=336.6s + ttt_chunk [1511/1904] bpb=1.094483 time=338.8s + ttt_chunk [1521/1904] bpb=1.094280 time=341.0s + ttt_chunk [1531/1904] bpb=1.093902 time=343.3s + ttt_chunk [1541/1904] bpb=1.094278 time=345.5s + ttt_chunk [1551/1904] bpb=1.094355 time=347.7s + ttt_chunk [1561/1904] bpb=1.094472 time=350.0s + ttt_chunk [1571/1904] bpb=1.094625 time=352.2s + ttt_chunk [1581/1904] bpb=1.094495 time=354.5s + ttt_chunk [1591/1904] bpb=1.094159 time=356.7s + ttt_chunk [1601/1904] bpb=1.094460 time=359.0s + ttt_chunk [1611/1904] bpb=1.094421 time=361.2s + ttt_chunk [1621/1904] bpb=1.094600 time=363.5s + ttt_chunk [1631/1904] bpb=1.094599 time=365.7s + ttt_chunk [1641/1904] bpb=1.094474 time=367.9s + ttt_chunk [1651/1904] bpb=1.094623 time=370.2s + ttt_chunk [1661/1904] bpb=1.094786 time=372.4s + ttt_chunk [1671/1904] bpb=1.094281 time=374.6s + ttt_chunk [1681/1904] bpb=1.094493 time=376.9s + ttt_chunk [1691/1904] bpb=1.094292 time=379.1s + ttt_chunk [1701/1904] bpb=1.094342 time=381.4s + ttt_chunk [1711/1904] bpb=1.094865 time=383.6s + ttt_chunk [1721/1904] bpb=1.094956 time=385.8s + ttt_chunk [1731/1904] bpb=1.095599 time=388.1s + ttt_chunk [1741/1904] bpb=1.095466 time=390.3s + ttt_chunk [1751/1904] bpb=1.095582 time=392.5s + ttt_chunk [1761/1904] bpb=1.095420 time=394.8s + ttt_chunk [1771/1904] bpb=1.095412 time=397.0s + ttt_chunk [1781/1904] bpb=1.095226 time=399.2s + ttt_chunk [1791/1904] bpb=1.095020 time=401.5s + ttt_chunk [1801/1904] bpb=1.094945 time=403.7s + ttt_chunk [1811/1904] bpb=1.094945 time=405.9s + ttt_chunk [1821/1904] bpb=1.094652 time=408.2s + ttt_chunk [1831/1904] bpb=1.094794 time=410.4s + ttt_chunk [1841/1904] bpb=1.094689 time=412.6s + ttt_chunk [1851/1904] bpb=1.094473 time=414.9s + ttt_chunk [1861/1904] bpb=1.094346 time=417.1s + ttt_chunk [1871/1904] bpb=1.094269 time=419.4s + ttt_chunk [1881/1904] bpb=1.094428 time=421.6s + ttt_chunk [1891/1904] bpb=1.093891 time=423.8s + ttt_chunk [1901/1904] bpb=1.094184 time=426.1s + ttt_chunk [1904/1904] bpb=1.094188 time=426.6s +ttt_sliding:done val_loss=1.949476 val_bpb=1.089850 elapsed=426.6s +legal_ttt val_loss:1.9495 val_bpb:1.0898 eval_time:427036ms +legal_ttt_exact val_loss:1.94947593 val_bpb:1.08984977 +ngram_two_pass: starting Pass 1 (sliding-window neural eval) +ngram_two_pass: Pass 1 done val_bpb=1.088180 tokens_scored=7797440 time=99.5s +ngram_two_pass: building cache orders=2-16 buckets=16777216 +ngram_two_pass: cache built in 58.7s +ngram_two_pass: starting Pass 2 (n-gram rescore) +ngram_rescore: matched=7797440/7797440 (100.0%) mean_alpha=0.948 +ngram_two_pass: Pass 2 done val_bpb=0.195170 improvement=0.893010 time=40.1s +ngram_two_pass: total time=198.3s +ngram_two_pass val_loss:0.3491 val_bpb:0.1952 eval_time:198338ms +ngram_two_pass_exact val_loss:0.34911233 val_bpb:0.19517040 + +====================================================================== +Training complete — SEED=2025 +====================================================================== diff --git a/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed42.log b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed42.log new file mode 100644 index 0000000000..0880c34cff --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_scylla-tuned-ttt/train_seed42.log @@ -0,0 +1,338 @@ +=== PRE-FLIGHT === +python_check_ok +Warning: Permanently added '[103.207.149.67]:16831' (ED25519) to the list of known hosts. + +=== BOOTSTRAP + TRAINING === +Warning: Permanently added '[103.207.149.67]:16831' (ED25519) to the list of known hosts. +====================================================================== +bootstrap_scylla_full_run.sh + SEED=42 NPROC=8 + DATA_PATH=./data/datasets/fineweb_scylla +====================================================================== + +=== Step 0: Python environment === +/usr/local/lib/python3.12/dist-packages/torch/_subclasses/functional_tensor.py:279: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.) + cpu = _conversion_method_template(device=torch.device("cpu")) +python3 -> Python 3.12.12, torch=2.9.1+cu128 + +=== Step 1: Install dependencies === +WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning. + +[notice] A new release of pip is available: 25.3 -> 26.0.1 +[notice] To update, run: python3 -m pip install --upgrade pip +Dependencies installed. + +=== Step 2: Verify tokenizer artifacts === + candidate.vocab: vocab_size=998 OK + candidate.meta.npz: keys=base_bytes,format_version,has_leading_space,is_boundary_token,source_model_name,tokenizer_kind,vocab_size OK + +=== Step 3: Acquire Scylla shards === + Existing shards: train=0 +0 val=0 +0 + SHARD_SOURCE_URL=https://m9a3es9ein9yh4-19123.proxy.runpod.net + Downloading pre-tokenized Scylla shards from HTTP source... +scripts/bootstrap_scylla_full_run.sh: line 130: [: 0 +0: integer expression expected + Download complete in 200s train_shards=80 + +=== Step 5: Verify shard outputs === + Train shards: 80 / 80 + Val shards: 1 / 1 + fineweb_train_000000.bin: magic=OK version=1 ntokens=100,000,175 size=200.0MB + fineweb_val_000000.bin: magic=OK version=1 ntokens=62,365,136 size=124.7MB + Shard verification: PASSED + +=== Step 6: Launch training (SEED=42, NPROC=8) === + Config: locked Scylla full-run + DATA_PATH=./data/datasets/fineweb_scylla + TOKENIZER_PATH=./data/tokenizers/scylla/candidate.vocab + TOKENIZER_META_PATH=./data/tokenizers/scylla/candidate.meta.npz +W0401 23:23:09.355000 2359 torch/distributed/run.py:803] +W0401 23:23:09.355000 2359 torch/distributed/run.py:803] ***************************************** +W0401 23:23:09.355000 2359 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0401 23:23:09.355000 2359 torch/distributed/run.py:803] ***************************************** +logs/d1bfa21f-4e5d-4d56-8aa8-6ef12f2ca5eb.txt +val_bpb:enabled tokenizer_kind=tokenmonster tokenizer_path=./data/tokenizers/scylla/candidate.vocab +train_loader:dataset:fineweb_scylla train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb_scylla/fineweb_val_*.bin tokens:62363648 +model_params:28042332 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +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:9000 warmup_steps:20 max_wallclock_seconds:600.000 +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:0/9000 val_loss:6.9186 val_bpb:3.8678 train_time:0ms step_avg:0.02ms +step:1/9000 train_loss:6.9178 train_time:149ms step_avg:149.00ms +step:2/9000 train_loss:9.0404 train_time:180ms step_avg:89.97ms +step:3/9000 train_loss:8.8774 train_time:274ms step_avg:91.23ms +step:4/9000 train_loss:8.4242 train_time:370ms step_avg:92.44ms +step:5/9000 train_loss:7.7514 train_time:464ms step_avg:92.76ms +step:6/9000 train_loss:7.4517 train_time:558ms step_avg:93.01ms +step:7/9000 train_loss:7.0215 train_time:653ms step_avg:93.26ms +step:8/9000 train_loss:6.9108 train_time:749ms step_avg:93.57ms +step:9/9000 train_loss:6.5038 train_time:845ms step_avg:93.91ms +step:10/9000 train_loss:6.3147 train_time:940ms step_avg:94.03ms +step:500/9000 train_loss:2.3977 train_time:48267ms step_avg:96.53ms +step:1000/9000 train_loss:2.2860 train_time:96970ms step_avg:96.97ms +step:1500/9000 train_loss:2.0853 train_time:145724ms step_avg:97.15ms +step:2000/9000 train_loss:2.2014 train_time:194448ms step_avg:97.22ms +step:2500/9000 train_loss:2.1841 train_time:243097ms step_avg:97.24ms +step:3000/9000 train_loss:2.1185 train_time:291726ms step_avg:97.24ms +step:3500/9000 train_loss:2.0038 train_time:340396ms step_avg:97.26ms +step:4000/9000 train_loss:2.0304 train_time:389022ms step_avg:97.26ms +step:4000/9000 val_loss:2.0766 val_bpb:1.1609 train_time:389089ms step_avg:97.27ms +step:4500/9000 train_loss:1.9356 train_time:437647ms step_avg:97.25ms +step:5000/9000 train_loss:2.1710 train_time:486231ms step_avg:97.25ms +swa:start step:5500 +step:5500/9000 train_loss:1.9625 train_time:534813ms step_avg:97.24ms +late_qat:enabled step:5643 scale:0.1497 +step:6000/9000 train_loss:1.9312 train_time:584052ms step_avg:97.34ms +step:6162/9000 val_loss:1.9729 val_bpb:1.1030 train_time:600098ms step_avg:97.39ms +stopping_early: wallclock_cap train_time:600098ms step:6162/9000 +peak memory allocated: 22928 MiB reserved: 22948 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9714 val_bpb:1.1021 eval_time:2342ms +Serialized model: 108255670 bytes +Code size: 115885 bytes +Serialized model int6+lzma: 15213940 bytes +Total submission size int6+lzma: 15329825 bytes +final_int6_roundtrip val_loss:2.0059 val_bpb:1.1214 eval_time:19863ms +final_int6_roundtrip_exact val_loss:2.00592433 val_bpb:1.12141550 +final_int6_sliding_window val_loss:1.9659 val_bpb:1.0991 stride:64 eval_time:107335ms +final_int6_sliding_window_exact val_loss:1.96593534 val_bpb:1.09905136 +final_int8_zlib_roundtrip_exact val_loss:1.96593534 val_bpb:1.09905136 +ttt_sliding:start chunks=1904 chunk_tokens=32768 total_windows=974432 stride=64 ttt_lr=0.005 ttt_epochs=3 freeze_blocks=2 adaptive=False +ttt_sliding:params unfrozen=28038220 frozen=4112 + ttt_chunk [1/1904] bpb=1.055464 time=0.5s + ttt_chunk [11/1904] bpb=1.056895 time=2.8s + ttt_chunk [21/1904] bpb=1.084718 time=5.1s + ttt_chunk [31/1904] bpb=1.087796 time=7.3s + ttt_chunk [41/1904] bpb=1.118556 time=9.6s + ttt_chunk [51/1904] bpb=1.116595 time=11.9s + ttt_chunk [61/1904] bpb=1.117724 time=14.1s + ttt_chunk [71/1904] bpb=1.110518 time=16.4s + ttt_chunk [81/1904] bpb=1.123795 time=18.7s + ttt_chunk [91/1904] bpb=1.115442 time=21.0s + ttt_chunk [101/1904] bpb=1.113184 time=23.2s + ttt_chunk [111/1904] bpb=1.107888 time=25.5s + ttt_chunk [121/1904] bpb=1.108851 time=27.8s + ttt_chunk [131/1904] bpb=1.108481 time=30.0s + ttt_chunk [141/1904] bpb=1.104455 time=32.3s + ttt_chunk [151/1904] bpb=1.100842 time=34.6s + ttt_chunk [161/1904] bpb=1.100292 time=36.8s + ttt_chunk [171/1904] bpb=1.097726 time=39.1s + ttt_chunk [181/1904] bpb=1.096905 time=41.4s + ttt_chunk [191/1904] bpb=1.096973 time=43.6s + ttt_chunk [201/1904] bpb=1.098541 time=45.9s + ttt_chunk [211/1904] bpb=1.096899 time=48.2s + ttt_chunk [221/1904] bpb=1.092749 time=50.4s + ttt_chunk [231/1904] bpb=1.090189 time=52.7s + ttt_chunk [241/1904] bpb=1.092861 time=55.0s + ttt_chunk [251/1904] bpb=1.091640 time=57.2s + ttt_chunk [261/1904] bpb=1.088392 time=59.5s + ttt_chunk [271/1904] bpb=1.087918 time=61.8s + ttt_chunk [281/1904] bpb=1.087321 time=64.0s + ttt_chunk [291/1904] bpb=1.086393 time=66.3s + ttt_chunk [301/1904] bpb=1.087280 time=68.6s + ttt_chunk [311/1904] bpb=1.088221 time=70.9s + ttt_chunk [321/1904] bpb=1.088113 time=73.1s + ttt_chunk [331/1904] bpb=1.087984 time=75.4s + ttt_chunk [341/1904] bpb=1.086615 time=77.7s + ttt_chunk [351/1904] bpb=1.085887 time=79.9s + ttt_chunk [361/1904] bpb=1.087351 time=82.2s + ttt_chunk [371/1904] bpb=1.086584 time=84.5s + ttt_chunk [381/1904] bpb=1.085960 time=86.7s + ttt_chunk [391/1904] bpb=1.085000 time=89.0s + ttt_chunk [401/1904] bpb=1.084714 time=91.3s + ttt_chunk [411/1904] bpb=1.083404 time=93.5s + ttt_chunk [421/1904] bpb=1.081750 time=95.8s + ttt_chunk [431/1904] bpb=1.080134 time=98.0s + ttt_chunk [441/1904] bpb=1.079401 time=100.3s + ttt_chunk [451/1904] bpb=1.077249 time=102.5s + ttt_chunk [461/1904] bpb=1.077013 time=104.8s + ttt_chunk [471/1904] bpb=1.075711 time=107.1s + ttt_chunk [481/1904] bpb=1.076096 time=109.4s + ttt_chunk [491/1904] bpb=1.076546 time=111.7s + ttt_chunk [501/1904] bpb=1.077164 time=113.9s + ttt_chunk [511/1904] bpb=1.078083 time=116.2s + ttt_chunk [521/1904] bpb=1.081108 time=118.4s + ttt_chunk [531/1904] bpb=1.081064 time=120.7s + ttt_chunk [541/1904] bpb=1.082381 time=123.0s + ttt_chunk [551/1904] bpb=1.081750 time=125.2s + ttt_chunk [561/1904] bpb=1.081644 time=127.5s + ttt_chunk [571/1904] bpb=1.082039 time=129.8s + ttt_chunk [581/1904] bpb=1.083217 time=132.0s + ttt_chunk [591/1904] bpb=1.083872 time=134.3s + ttt_chunk [601/1904] bpb=1.083935 time=136.6s + ttt_chunk [611/1904] bpb=1.083606 time=138.9s + ttt_chunk [621/1904] bpb=1.085308 time=141.2s + ttt_chunk [631/1904] bpb=1.085869 time=143.4s + ttt_chunk [641/1904] bpb=1.086485 time=145.7s + ttt_chunk [651/1904] bpb=1.086277 time=147.9s + ttt_chunk [661/1904] bpb=1.086773 time=150.1s + ttt_chunk [671/1904] bpb=1.087446 time=152.4s + ttt_chunk [681/1904] bpb=1.087552 time=154.6s + ttt_chunk [691/1904] bpb=1.087563 time=156.9s + ttt_chunk [701/1904] bpb=1.088293 time=159.1s + ttt_chunk [711/1904] bpb=1.088754 time=161.4s + ttt_chunk [721/1904] bpb=1.088070 time=163.6s + ttt_chunk [731/1904] bpb=1.087962 time=165.8s + ttt_chunk [741/1904] bpb=1.087837 time=168.1s + ttt_chunk [751/1904] bpb=1.087085 time=170.3s + ttt_chunk [761/1904] bpb=1.086919 time=172.6s + ttt_chunk [771/1904] bpb=1.086354 time=174.8s + ttt_chunk [781/1904] bpb=1.086173 time=177.1s + ttt_chunk [791/1904] bpb=1.086024 time=179.3s + ttt_chunk [801/1904] bpb=1.086355 time=181.6s + ttt_chunk [811/1904] bpb=1.087384 time=183.8s + ttt_chunk [821/1904] bpb=1.086672 time=186.0s + ttt_chunk [831/1904] bpb=1.087371 time=188.3s + ttt_chunk [841/1904] bpb=1.087304 time=190.6s + ttt_chunk [851/1904] bpb=1.087687 time=192.8s + ttt_chunk [861/1904] bpb=1.088151 time=195.1s + ttt_chunk [871/1904] bpb=1.088468 time=197.3s + ttt_chunk [881/1904] bpb=1.088622 time=199.6s + ttt_chunk [891/1904] bpb=1.089639 time=201.8s + ttt_chunk [901/1904] bpb=1.090284 time=204.1s + ttt_chunk [911/1904] bpb=1.091001 time=206.3s + ttt_chunk [921/1904] bpb=1.091924 time=208.6s + ttt_chunk [931/1904] bpb=1.092394 time=210.8s + ttt_chunk [941/1904] bpb=1.092007 time=213.0s + ttt_chunk [951/1904] bpb=1.092237 time=215.3s + ttt_chunk [961/1904] bpb=1.092253 time=217.5s + ttt_chunk [971/1904] bpb=1.092857 time=219.8s + ttt_chunk [981/1904] bpb=1.093264 time=222.0s + ttt_chunk [991/1904] bpb=1.093056 time=224.3s + ttt_chunk [1001/1904] bpb=1.093184 time=226.5s + ttt_chunk [1011/1904] bpb=1.093815 time=228.8s + ttt_chunk [1021/1904] bpb=1.094447 time=231.0s + ttt_chunk [1031/1904] bpb=1.094077 time=233.3s + ttt_chunk [1041/1904] bpb=1.093885 time=235.5s + ttt_chunk [1051/1904] bpb=1.093499 time=237.8s + ttt_chunk [1061/1904] bpb=1.092753 time=240.0s + ttt_chunk [1071/1904] bpb=1.092607 time=242.3s + ttt_chunk [1081/1904] bpb=1.092381 time=244.5s + ttt_chunk [1091/1904] bpb=1.092224 time=246.8s + ttt_chunk [1101/1904] bpb=1.092433 time=249.0s + ttt_chunk [1111/1904] bpb=1.092403 time=251.3s + ttt_chunk [1121/1904] bpb=1.092705 time=253.5s + ttt_chunk [1131/1904] bpb=1.092630 time=255.8s + ttt_chunk [1141/1904] bpb=1.092948 time=258.0s + ttt_chunk [1151/1904] bpb=1.092490 time=260.3s + ttt_chunk [1161/1904] bpb=1.092510 time=262.5s + ttt_chunk [1171/1904] bpb=1.092214 time=264.8s + ttt_chunk [1181/1904] bpb=1.092306 time=267.0s + ttt_chunk [1191/1904] bpb=1.091785 time=269.3s + ttt_chunk [1201/1904] bpb=1.092107 time=271.5s + ttt_chunk [1211/1904] bpb=1.092271 time=273.8s + ttt_chunk [1221/1904] bpb=1.092337 time=276.0s + ttt_chunk [1231/1904] bpb=1.093133 time=278.2s + ttt_chunk [1241/1904] bpb=1.093232 time=280.5s + ttt_chunk [1251/1904] bpb=1.093811 time=282.7s + ttt_chunk [1261/1904] bpb=1.093586 time=285.0s + ttt_chunk [1271/1904] bpb=1.093747 time=287.3s + ttt_chunk [1281/1904] bpb=1.093955 time=289.5s + ttt_chunk [1291/1904] bpb=1.094301 time=291.7s + ttt_chunk [1301/1904] bpb=1.094940 time=294.0s + ttt_chunk [1311/1904] bpb=1.095182 time=296.2s + ttt_chunk [1321/1904] bpb=1.095771 time=298.5s + ttt_chunk [1331/1904] bpb=1.095686 time=300.7s + ttt_chunk [1341/1904] bpb=1.096068 time=303.0s + ttt_chunk [1351/1904] bpb=1.096532 time=305.2s + ttt_chunk [1361/1904] bpb=1.096941 time=307.5s + ttt_chunk [1371/1904] bpb=1.096742 time=309.7s + ttt_chunk [1381/1904] bpb=1.096104 time=312.0s + ttt_chunk [1391/1904] bpb=1.096329 time=314.3s + ttt_chunk [1401/1904] bpb=1.095853 time=316.5s + ttt_chunk [1411/1904] bpb=1.095804 time=318.8s + ttt_chunk [1421/1904] bpb=1.095718 time=321.0s + ttt_chunk [1431/1904] bpb=1.095746 time=323.3s + ttt_chunk [1441/1904] bpb=1.095772 time=325.5s + ttt_chunk [1451/1904] bpb=1.095813 time=327.7s + ttt_chunk [1461/1904] bpb=1.096051 time=330.0s + ttt_chunk [1471/1904] bpb=1.095813 time=332.2s + ttt_chunk [1481/1904] bpb=1.095526 time=334.4s + ttt_chunk [1491/1904] bpb=1.094697 time=336.7s + ttt_chunk [1501/1904] bpb=1.095272 time=338.9s + ttt_chunk [1511/1904] bpb=1.095192 time=341.2s + ttt_chunk [1521/1904] bpb=1.094988 time=343.4s + ttt_chunk [1531/1904] bpb=1.094612 time=345.7s + ttt_chunk [1541/1904] bpb=1.094996 time=347.9s + ttt_chunk [1551/1904] bpb=1.095058 time=350.2s + ttt_chunk [1561/1904] bpb=1.095167 time=352.4s + ttt_chunk [1571/1904] bpb=1.095299 time=354.6s + ttt_chunk [1581/1904] bpb=1.095171 time=356.9s + ttt_chunk [1591/1904] bpb=1.094858 time=359.2s + ttt_chunk [1601/1904] bpb=1.095184 time=361.4s + ttt_chunk [1611/1904] bpb=1.095143 time=363.7s + ttt_chunk [1621/1904] bpb=1.095328 time=365.9s + ttt_chunk [1631/1904] bpb=1.095357 time=368.1s + ttt_chunk [1641/1904] bpb=1.095248 time=370.4s + ttt_chunk [1651/1904] bpb=1.095423 time=372.7s + ttt_chunk [1661/1904] bpb=1.095592 time=374.9s + ttt_chunk [1671/1904] bpb=1.095084 time=377.2s + ttt_chunk [1681/1904] bpb=1.095292 time=379.5s + ttt_chunk [1691/1904] bpb=1.095096 time=381.7s + ttt_chunk [1701/1904] bpb=1.095146 time=384.0s + ttt_chunk [1711/1904] bpb=1.095671 time=386.2s + ttt_chunk [1721/1904] bpb=1.095755 time=388.5s + ttt_chunk [1731/1904] bpb=1.096405 time=390.7s + ttt_chunk [1741/1904] bpb=1.096298 time=393.0s + ttt_chunk [1751/1904] bpb=1.096404 time=395.2s + ttt_chunk [1761/1904] bpb=1.096274 time=397.5s + ttt_chunk [1771/1904] bpb=1.096250 time=399.7s + ttt_chunk [1781/1904] bpb=1.096063 time=401.9s + ttt_chunk [1791/1904] bpb=1.095853 time=404.2s + ttt_chunk [1801/1904] bpb=1.095761 time=406.4s + ttt_chunk [1811/1904] bpb=1.095779 time=408.7s + ttt_chunk [1821/1904] bpb=1.095492 time=411.0s + ttt_chunk [1831/1904] bpb=1.095631 time=413.2s + ttt_chunk [1841/1904] bpb=1.095530 time=415.5s + ttt_chunk [1851/1904] bpb=1.095317 time=417.7s + ttt_chunk [1861/1904] bpb=1.095191 time=420.0s + ttt_chunk [1871/1904] bpb=1.095091 time=422.2s + ttt_chunk [1881/1904] bpb=1.095239 time=424.5s + ttt_chunk [1891/1904] bpb=1.094705 time=426.7s + ttt_chunk [1901/1904] bpb=1.094996 time=429.0s + ttt_chunk [1904/1904] bpb=1.094998 time=429.5s +ttt_sliding:done val_loss=1.950891 val_bpb=1.090641 elapsed=429.5s +legal_ttt val_loss:1.9509 val_bpb:1.0906 eval_time:429875ms +legal_ttt_exact val_loss:1.95089092 val_bpb:1.09064082 +ngram_two_pass: starting Pass 1 (sliding-window neural eval) +ngram_two_pass: Pass 1 done val_bpb=1.088995 tokens_scored=7797440 time=99.7s +ngram_two_pass: building cache orders=2-16 buckets=16777216 +ngram_two_pass: cache built in 59.8s +ngram_two_pass: starting Pass 2 (n-gram rescore) +ngram_rescore: matched=7797440/7797440 (100.0%) mean_alpha=0.948 +ngram_two_pass: Pass 2 done val_bpb=0.195173 improvement=0.893823 time=38.4s +ngram_two_pass: total time=197.9s +ngram_two_pass val_loss:0.3491 val_bpb:0.1952 eval_time:198058ms +ngram_two_pass_exact val_loss:0.34911648 val_bpb:0.19517272 + +====================================================================== +Training complete — SEED=42 +======================================================================