diff --git a/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/README.md b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/README.md new file mode 100644 index 0000000000..67a0c5d19e --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/README.md @@ -0,0 +1,50 @@ +# R08 Higher-LR Matrix/Scalar GPT 17M A40 + +- Date: 2026-04-02 +- Track: non_record_16mb +- Author: Siddhardha Nanda (SID-6921) +- Reported val_bpb: 2.1827 + +## Summary + +Increased matrix and scalar parameter learning rates relative to the stock baseline. Setting `MATRIX_LR=0.05` and `SCALAR_LR=0.04` (with matching `TIED_EMBED_LR=0.05`) while also halving the batch size to 1M tokens and using a 400-step warmdown schedule achieved a significant improvement from the stock baseline. + +## What Changed + +- `MATRIX_LR=0.05` (increased from stock default ~0.04) +- `SCALAR_LR=0.04` (adjusted) +- `TIED_EMBED_LR=0.05` (matched to MATRIX_LR) +- `TRAIN_BATCH_TOKENS=1048576` (halved from 2M default) +- `WARMDOWN_ITERS=400` (reduced from 1200 default) +- `ITERATIONS=60` (doubled from baseline 30) +- Architecture unchanged: 17M params, GQA (8 heads, 4 KV heads), ReLU², sp_bpe_1024 tokenizer + +## Repro Command + +```bash +export DATA_PATH=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024 \ + TOKENIZER_PATH=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model \ + VOCAB_SIZE=1024 \ + TRAIN_BATCH_TOKENS=1048576 \ + WARMDOWN_ITERS=400 \ + MATRIX_LR=0.05 \ + SCALAR_LR=0.04 \ + TIED_EMBED_LR=0.05 \ + ITERATIONS=60 \ + MAX_WALLCLOCK_SECONDS=900 +torchrun --standalone --nproc_per_node=1 train_gpt.py +``` + +## Results + +- val_bpb: 2.18271188 (int8+zlib roundtrip) +- val_loss: 3.68541758 (int8+zlib roundtrip) +- pre_quant_val_bpb: 2.1795 +- pre_quant_val_loss: 3.6800 +- compressed_bytes: 9,897,284 bytes (~9.4 MB, well under 16 MB cap) +- wallclock_seconds: ~169s +- GPU: 1× NVIDIA A40 + +## Notes + +This was the best run (#1 of 10) from an automated campaign (`04_non_record_a40_campaign.sh`) testing batch sizes, warmdown schedules, QK gain values, and learning rate combinations. Higher LR for matrix parameters combined with a smaller batch size appears to be the key driver of improvement over the stock baseline. diff --git a/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/requirements.txt b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/requirements.txt new file mode 100644 index 0000000000..911b0e52f0 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/requirements.txt @@ -0,0 +1,10 @@ +numpy +tqdm +torch +huggingface-hub +kernels +setuptools +typing-extensions==4.15.0 +datasets +tiktoken +sentencepiece \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/submission.json b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/submission.json new file mode 100644 index 0000000000..af9bfa142e --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/submission.json @@ -0,0 +1,18 @@ +{ + "author": "Siddhardha Nanda", + "github_id": "SID-6921", + "name": "R08 Higher-LR Matrix/Scalar GPT 17M A40", + "blurb": "Increased matrix and scalar learning rates (MATRIX_LR=0.05, SCALAR_LR=0.04, embed_lr=0.05) vs stock defaults, with 1M-token batch and 400-step warmdown. 60 steps on single A40. Improves val_bpb from baseline 3.2686 to 2.1827, submission size 9.4MB well under 16MB cap.", + "date": "2026-04-02T00:00:00Z", + "track": "non_record_16mb", + "val_loss": 3.68541758, + "val_bpb": 2.18271188, + "pre_quant_val_loss": 3.6800, + "pre_quant_val_bpb": 2.1795, + "step_stop": 60, + "wallclock_seconds": 169, + "bytes_total": 9897284, + "bytes_model_int8_zlib": 9849598, + "bytes_code": 47686, + "gpu": "1xA40" +} diff --git a/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/train.log b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/train.log new file mode 100644 index 0000000000..0d01f90592 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/train.log @@ -0,0 +1,66 @@ +logs/R08_lr_matrix_up.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:1 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059912 +world_size:1 grad_accum_steps:8 +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:0.05 head_lr:0.0 matrix_lr:0.05 scalar_lr:0.04 +train_batch_tokens:1048576 train_seq_len:1024 iterations:60 warmup_steps:20 max_wallclock_seconds:900.000 +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:0/60 val_loss:6.9357 val_bpb:4.1077 train_time:0ms step_avg:0.02ms +step:1/60 train_loss:6.9358 train_time:2748ms step_avg:2748.41ms +step:2/60 train_loss:16.6092 train_time:5554ms step_avg:2776.77ms +step:3/60 train_loss:9.2487 train_time:8362ms step_avg:2787.44ms +step:4/60 train_loss:6.5276 train_time:11167ms step_avg:2791.67ms +step:5/60 train_loss:6.5161 train_time:13971ms step_avg:2794.21ms +step:6/60 train_loss:6.5023 train_time:16792ms step_avg:2798.62ms +step:7/60 train_loss:6.3541 train_time:19599ms step_avg:2799.79ms +step:8/60 train_loss:6.1645 train_time:22414ms step_avg:2801.79ms +step:9/60 train_loss:6.0548 train_time:25223ms step_avg:2802.61ms +step:10/60 train_loss:5.9123 train_time:28032ms step_avg:2803.21ms +step:10/60 val_loss:5.8557 val_bpb:3.4681 train_time:28037ms step_avg:2803.72ms +step:15/60 train_loss:5.6224 train_time:42108ms step_avg:2807.19ms +step:20/60 train_loss:5.0835 train_time:56189ms step_avg:2809.44ms +step:20/60 val_loss:5.0103 val_bpb:2.9674 train_time:56209ms step_avg:2810.44ms +step:25/60 train_loss:4.7200 train_time:70286ms step_avg:2811.43ms +step:30/60 train_loss:4.5004 train_time:84364ms step_avg:2812.13ms +step:30/60 val_loss:4.4625 val_bpb:2.6429 train_time:84378ms step_avg:2812.59ms +step:35/60 train_loss:4.3816 train_time:98437ms step_avg:2812.49ms +step:40/60 train_loss:4.2317 train_time:112498ms step_avg:2812.44ms +step:40/60 val_loss:4.1971 val_bpb:2.4858 train_time:112516ms step_avg:2812.89ms +step:45/60 train_loss:4.0509 train_time:126570ms step_avg:2812.66ms +step:50/60 train_loss:3.9271 train_time:140629ms step_avg:2812.58ms +step:50/60 val_loss:3.8891 val_bpb:2.3033 train_time:140646ms step_avg:2812.92ms +step:55/60 train_loss:3.7612 train_time:154697ms step_avg:2812.67ms +step:60/60 train_loss:3.7153 train_time:168755ms step_avg:2812.58ms +step:60/60 val_loss:3.6800 val_bpb:2.1795 train_time:168786ms step_avg:2813.11ms +peak memory allocated: 20066 MiB reserved: 20776 MiB +Serialized model: 67224983 bytes +Code size: 47686 bytes +Total submission size: 67272669 bytes +Serialized model int8+zlib: 9849598 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x) +Total submission size int8+zlib: 9897284 bytes +final_int8_zlib_roundtrip val_loss:3.6854 val_bpb:2.1827 eval_time:54545ms +final_int8_zlib_roundtrip_exact val_loss:3.68541758 val_bpb:2.18271188 diff --git a/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/train_gpt.py b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/train_gpt.py new file mode 100644 index 0000000000..8b5b80a4ce --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_R08_LRMatrixUp_A40_17M/train_gpt.py @@ -0,0 +1,1126 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + 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 = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + 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", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + 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", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + 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: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Scale correction from Muon reference implementations. + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("Γûü"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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, +) -> tuple[float, float]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_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}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_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) + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +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", + ).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: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + 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() + + # Vectors / scalars use a simpler per-tensor scale. + 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]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + 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 + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + 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) + # Broadcast the saved row scale back across trailing dimensions. + 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(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + 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: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + 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): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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 + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + 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, + ): + 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") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + 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()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +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, + ): + super().__init__() + 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.tok_emb = nn.Embedding(vocab_size, 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)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + 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._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) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, 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) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + 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 + + # Fast math knobs + 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) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + 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) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + 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, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + 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}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + 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 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + 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): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + 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() + 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() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + 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) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + 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) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + 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" + ) + + # Needed to sync whether we've reached the wallclock cap. + 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" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "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") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_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, + ) + torch.cuda.synchronize() + log0( + f"final_int8_zlib_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_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/README.md b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/README.md new file mode 100644 index 0000000000..81b45c9c15 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/README.md @@ -0,0 +1,54 @@ +# Stock Baseline GPT 17M sp1024 — A40 10min + +**Track:** non_record_16mb +**GPU:** 1× A40 +**val_bpb (int8+zlib roundtrip):** 3.2686 +**Submission size:** 5,596,364 bytes (~5.3 MB) + +## Summary + +Unmodified stock `train_gpt.py` run on a single A40 GPU using the `sp_bpe_1024` tokenizer (vocab size = 1024). Serves as an A40 baseline reference point. + +## Run Configuration + +| Parameter | Value | +|---|---| +| VOCAB_SIZE | 1024 | +| ITERATIONS | 30 | +| TRAIN_BATCH_TOKENS | 2,097,152 | +| VAL_LOSS_EVERY | 10 | +| MAX_WALLCLOCK_SECONDS | 600 | +| Model params | 17,059,912 | +| GPU | 1× NVIDIA A40 | + +## Training Command + +```bash +export DATA_PATH=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024 \ + TOKENIZER_PATH=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model \ + VOCAB_SIZE=1024 \ + ITERATIONS=30 \ + TRAIN_BATCH_TOKENS=2097152 \ + VAL_LOSS_EVERY=10 \ + MAX_WALLCLOCK_SECONDS=600 +torchrun --standalone --nproc_per_node=1 train_gpt.py +``` + +## Results + +| Step | val_loss | val_bpb | +|---|---|---| +| 0/30 | 6.9357 | 4.1077 | +| 10/30 | 7.1298 | 4.2226 | +| 20/30 | 5.6383 | 3.3393 | +| 30/30 | 5.5269 | 3.2734 | +| **int8+zlib roundtrip** | **5.5190** | **3.2686** | + +Exact: `val_loss=5.51896729 val_bpb=3.26864330` + +## Artifact Sizes + +- Serialized model (fp32): 67,224,983 bytes +- Model int8+zlib: 5,548,678 bytes +- Code: 47,686 bytes +- **Total int8+zlib submission: 5,596,364 bytes** diff --git a/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/requirements.txt b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/requirements.txt new file mode 100644 index 0000000000..911b0e52f0 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/requirements.txt @@ -0,0 +1,10 @@ +numpy +tqdm +torch +huggingface-hub +kernels +setuptools +typing-extensions==4.15.0 +datasets +tiktoken +sentencepiece \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/submission.json b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/submission.json new file mode 100644 index 0000000000..363c334faf --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/submission.json @@ -0,0 +1,18 @@ +{ + "author": "Siddhardha Nanda", + "github_id": "SID-6921", + "name": "Stock Baseline GPT 17M sp1024 A40", + "blurb": "Stock unmodified baseline run using sp_bpe_1024 tokenizer (vocab=1024) on a single A40 GPU. 17M parameter model (~17M params), 30 training steps with 2M token batches (~600s wall-clock budget). Final int8+zlib roundtrip val_bpb: 3.2686, submission size 5.3MB well under the 16MB cap.", + "date": "2026-04-02T00:00:00Z", + "track": "non_record_16mb", + "val_loss": 5.51896729, + "val_bpb": 3.26864330, + "pre_quant_val_loss": 5.5269, + "pre_quant_val_bpb": 3.2734, + "step_stop": 30, + "wallclock_seconds": 166, + "bytes_total": 5596364, + "bytes_model_int8_zlib": 5548678, + "bytes_code": 47686, + "gpu": "1xA40" +} diff --git a/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/train_gpt.py b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/train_gpt.py new file mode 100644 index 0000000000..8b5b80a4ce --- /dev/null +++ b/records/track_non_record_16mb/2026-04-02_StockBaseline_A40_17M/train_gpt.py @@ -0,0 +1,1126 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + 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 = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + 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", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + 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", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + 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: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Scale correction from Muon reference implementations. + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("Γûü"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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, +) -> tuple[float, float]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_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}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_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) + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +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", + ).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: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + 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() + + # Vectors / scalars use a simpler per-tensor scale. + 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]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + 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 + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + 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) + # Broadcast the saved row scale back across trailing dimensions. + 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(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + 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: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + 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): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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 + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + 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, + ): + 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") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + 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()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +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, + ): + super().__init__() + 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.tok_emb = nn.Embedding(vocab_size, 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)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + 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._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) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, 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) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + 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 + + # Fast math knobs + 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) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + 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) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + 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, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + 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}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + 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 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + 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): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + 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() + 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() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + 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) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + 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) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + 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" + ) + + # Needed to sync whether we've reached the wallclock cap. + 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" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "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") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_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, + ) + torch.cuda.synchronize() + log0( + f"final_int8_zlib_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_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/starter_kit/START_HERE.md b/starter_kit/START_HERE.md new file mode 100644 index 0000000000..a759886578 --- /dev/null +++ b/starter_kit/START_HERE.md @@ -0,0 +1,54 @@ +# Parameter Golf Starter Kit + +This folder is a low-budget workflow to get from first run to a valid non-record PR. + +## 1) Fork + set your remote + +From your local repo root: + +```bash +git remote rename origin upstream +git remote add origin https://github.com/YOUR_GITHUB_USERNAME/parameter-golf.git +git fetch upstream +git checkout -b exp/first-runs upstream/main +git push -u origin exp/first-runs +``` + +## 2) On RunPod: first smoke run + +Use scripts in `starter_kit/scripts`: + +1. `01_runpod_bootstrap.sh` +2. `02_smoke_run.sh` + +## 3) Promote to serious run + +Run `03_full_run.sh` once smoke logs look healthy. + +Optional: run the full 10-run A40 non-record campaign and auto-rank outputs: + +```bash +bash starter_kit/scripts/04_non_record_a40_campaign.sh +``` + +## 4) Prepare a PR-ready records folder + +Run: + +```bash +python starter_kit/scripts/prepare_submission.py \ + --track non_record_16mb \ + --run-name my_first_non_record \ + --author-name "Your Name" \ + --github-id "your_github" \ + --val-bpb 1.1999 +``` + +Then copy your real train log into the generated folder and edit README details. + +## 5) Submission checklist + +- Folder only adds one new path under `records/track_non_record_16mb/` or `records/track_10min_16mb/`. +- Includes `README.md`, `submission.json`, `train_gpt.py`, and train log. +- Repro steps are explicit and complete. +- No validation-data leakage or rule violations. diff --git a/starter_kit/notes/EXPERIMENT_LOG_TEMPLATE.md b/starter_kit/notes/EXPERIMENT_LOG_TEMPLATE.md new file mode 100644 index 0000000000..d5803ee6eb --- /dev/null +++ b/starter_kit/notes/EXPERIMENT_LOG_TEMPLATE.md @@ -0,0 +1,32 @@ +# Experiment Log Template + +## Run Metadata + +- run_id: +- date: +- gpu: +- cost_estimate_usd: +- dataset_variant: +- train_shards: +- max_wallclock_seconds: + +## Config Delta + +- base_commit: +- branch: +- changed_hparams: +- changed_code_paths: + +## Outcomes + +- val_loss: +- val_bpb: +- final_int8_zlib_roundtrip_bytes: +- step_count: +- runtime_seconds: + +## Decision + +- keep / drop: +- reason: +- next test: diff --git a/starter_kit/scripts/01_runpod_bootstrap.sh b/starter_kit/scripts/01_runpod_bootstrap.sh new file mode 100644 index 0000000000..db2a49b041 --- /dev/null +++ b/starter_kit/scripts/01_runpod_bootstrap.sh @@ -0,0 +1,24 @@ +#!/usr/bin/env bash +set -euo pipefail + +# Usage: +# bash starter_kit/scripts/01_runpod_bootstrap.sh https://github.com/YOUR_GITHUB_USERNAME/parameter-golf.git + +FORK_URL="${1:-}" +if [[ -z "$FORK_URL" ]]; then + echo "Provide your fork URL as first arg." + exit 1 +fi + +cd /workspace +if [[ ! -d parameter-golf ]]; then + git clone "$FORK_URL" parameter-golf +fi + +cd parameter-golf +git remote -v + +echo "Downloading small dataset slice for low-cost iteration..." +python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 1 + +echo "Bootstrap complete. Run: bash starter_kit/scripts/02_smoke_run.sh" diff --git a/starter_kit/scripts/02_smoke_run.sh b/starter_kit/scripts/02_smoke_run.sh new file mode 100644 index 0000000000..d2a65eb299 --- /dev/null +++ b/starter_kit/scripts/02_smoke_run.sh @@ -0,0 +1,19 @@ +#!/usr/bin/env bash +set -euo pipefail + +# Quick low-cost run (~4 minutes max) +cd /workspace/parameter-golf + +RUN_ID="${RUN_ID:-smoke_sp1024_$(date +%Y%m%d_%H%M%S)}" +export RUN_ID +export DATA_PATH=./data/datasets/fineweb10B_sp1024/ +export TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model +export VOCAB_SIZE=1024 +export MAX_WALLCLOCK_SECONDS=240 +export VAL_LOSS_EVERY=0 + +mkdir -p logs + +torchrun --standalone --nproc_per_node=1 train_gpt.py | tee "logs/${RUN_ID}.log" + +echo "Smoke run done: logs/${RUN_ID}.log" diff --git a/starter_kit/scripts/03_full_run.sh b/starter_kit/scripts/03_full_run.sh new file mode 100644 index 0000000000..3035395f9b --- /dev/null +++ b/starter_kit/scripts/03_full_run.sh @@ -0,0 +1,19 @@ +#!/usr/bin/env bash +set -euo pipefail + +# Full baseline-style run (~10 minutes) +cd /workspace/parameter-golf + +RUN_ID="${RUN_ID:-full_sp1024_$(date +%Y%m%d_%H%M%S)}" +export RUN_ID +export DATA_PATH=./data/datasets/fineweb10B_sp1024/ +export TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model +export VOCAB_SIZE=1024 +export MAX_WALLCLOCK_SECONDS=600 +export VAL_LOSS_EVERY=200 + +mkdir -p logs + +torchrun --standalone --nproc_per_node=1 train_gpt.py | tee "logs/${RUN_ID}.log" + +echo "Full run done: logs/${RUN_ID}.log" diff --git a/starter_kit/scripts/04_non_record_a40_campaign.sh b/starter_kit/scripts/04_non_record_a40_campaign.sh new file mode 100644 index 0000000000..56db659458 --- /dev/null +++ b/starter_kit/scripts/04_non_record_a40_campaign.sh @@ -0,0 +1,54 @@ +#!/usr/bin/env bash +set -euo pipefail + +# 10-run non-record campaign focused on improving the stock A40 baseline. +# Run from repo root: bash starter_kit/scripts/04_non_record_a40_campaign.sh + +cd /workspace/parameter-golf +mkdir -p logs + +BASE_ENV=( + DATA_PATH=./data/datasets/fineweb10B_sp1024 + TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model + VOCAB_SIZE=1024 + VAL_LOSS_EVERY=10 + TRAIN_LOG_EVERY=5 + WARMUP_STEPS=20 + ITERATIONS=60 + MAX_WALLCLOCK_SECONDS=900 +) + +# Format: RUN_ID|SPACE_SEPARATED_ENV_OVERRIDES +RUNS=( + "R01_baseline_longer|TRAIN_BATCH_TOKENS=2097152 WARMDOWN_ITERS=1200" + "R02_warmdown_short|TRAIN_BATCH_TOKENS=2097152 WARMDOWN_ITERS=200" + "R03_warmdown_medium|TRAIN_BATCH_TOKENS=2097152 WARMDOWN_ITERS=600" + "R04_batch_half|TRAIN_BATCH_TOKENS=1048576 WARMDOWN_ITERS=400" + "R05_batch_quarter|TRAIN_BATCH_TOKENS=524288 WARMDOWN_ITERS=300" + "R06_qk_gain_low|TRAIN_BATCH_TOKENS=1048576 WARMDOWN_ITERS=400 QK_GAIN_INIT=1.2" + "R07_qk_gain_high|TRAIN_BATCH_TOKENS=1048576 WARMDOWN_ITERS=400 QK_GAIN_INIT=1.8" + "R08_lr_matrix_up|TRAIN_BATCH_TOKENS=1048576 WARMDOWN_ITERS=400 MATRIX_LR=0.05 SCALAR_LR=0.04 TIED_EMBED_LR=0.05" + "R09_lr_matrix_down|TRAIN_BATCH_TOKENS=1048576 WARMDOWN_ITERS=400 MATRIX_LR=0.03 SCALAR_LR=0.035 TIED_EMBED_LR=0.045" + "R10_capacity_bump|TRAIN_BATCH_TOKENS=524288 WARMDOWN_ITERS=500 MODEL_DIM=640 NUM_HEADS=8 NUM_KV_HEADS=4 MLP_MULT=2" +) + +for entry in "${RUNS[@]}"; do + IFS='|' read -r run_id override_str <<< "$entry" + read -r -a override_env <<< "$override_str" + + log_path="logs/${run_id}.log" + echo "============================================================" + echo "Starting ${run_id}" + echo "Log: ${log_path}" + + env RUN_ID="$run_id" "${BASE_ENV[@]}" "${override_env[@]}" \ + torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | tee "$log_path" + + metric_line=$(grep -E "final_int8_zlib_roundtrip_exact|final_int8_zlib_roundtrip " "$log_path" | tail -1 || true) + size_line=$(grep -E "Total submission size int8\+zlib:" "$log_path" | tail -1 || true) + echo "Completed ${run_id}" + echo " ${metric_line}" + echo " ${size_line}" +done + +python starter_kit/scripts/rank_campaign_results.py --logs-glob "logs/R*.log" diff --git a/starter_kit/scripts/05_non_record_a40_top_chase.sh b/starter_kit/scripts/05_non_record_a40_top_chase.sh new file mode 100644 index 0000000000..0d4d70f4e6 --- /dev/null +++ b/starter_kit/scripts/05_non_record_a40_top_chase.sh @@ -0,0 +1,51 @@ +#!/usr/bin/env bash +set -euo pipefail + +# Long non-record push focused on leaderboard-chasing configs. +# Run from repo root: bash starter_kit/scripts/05_non_record_a40_top_chase.sh + +cd /workspace/parameter-golf +mkdir -p logs + +BASE_ENV=( + DATA_PATH=./data/datasets/fineweb10B_sp1024 + TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model + VOCAB_SIZE=1024 + ENABLE_COMPILE=0 + TRAIN_SEQ_LEN=1024 + VAL_LOSS_EVERY=1000 + TRAIN_LOG_EVERY=100 + WARMUP_STEPS=20 + ITERATIONS=500000 + MAX_WALLCLOCK_SECONDS=14400 + WARMDOWN_FRAC=0.2 +) + +# Format: RUN_ID|SPACE_SEPARATED_ENV_OVERRIDES +RUNS=( + "A01_swiglu640_qtrbatch|MLP_ACTIVATION=swiglu MLP_HIDDEN=640 TRAIN_BATCH_TOKENS=131072 GRAD_ACCUM_STEPS=2 MATRIX_LR=0.05 SCALAR_LR=0.04 TIED_EMBED_LR=0.05" + "A02_swiglu704_halfbatch|MLP_ACTIVATION=swiglu MLP_HIDDEN=704 TRAIN_BATCH_TOKENS=262144 GRAD_ACCUM_STEPS=2 MATRIX_LR=0.05 SCALAR_LR=0.04 TIED_EMBED_LR=0.05" + "A03_swiglu640_qkgain|MLP_ACTIVATION=swiglu MLP_HIDDEN=640 TRAIN_BATCH_TOKENS=131072 GRAD_ACCUM_STEPS=2 QK_GAIN_INIT=1.8 MATRIX_LR=0.055 SCALAR_LR=0.04 TIED_EMBED_LR=0.05" +) + +for entry in "${RUNS[@]}"; do + IFS='|' read -r run_id override_str <<< "$entry" + read -r -a override_env <<< "$override_str" + + log_path="logs/${run_id}.log" + echo "============================================================" + echo "Starting ${run_id}" + echo "Log: ${log_path}" + + env RUN_ID="$run_id" "${BASE_ENV[@]}" "${override_env[@]}" \ + torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | tee "$log_path" + + metric_line=$(grep -E "final_int8_zlib_roundtrip_exact|final_int8_zlib_roundtrip " "$log_path" | tail -1 || true) + size_line=$(grep -E "Total submission size int8\+zlib:" "$log_path" | tail -1 || true) + echo "Completed ${run_id}" + echo " ${metric_line}" + echo " ${size_line}" + +done + +python starter_kit/scripts/rank_campaign_results.py --logs-glob "logs/A*.log" diff --git a/starter_kit/scripts/06_non_record_h100_top_chase.sh b/starter_kit/scripts/06_non_record_h100_top_chase.sh new file mode 100644 index 0000000000..ff15021721 --- /dev/null +++ b/starter_kit/scripts/06_non_record_h100_top_chase.sh @@ -0,0 +1,49 @@ +#!/usr/bin/env bash +set -euo pipefail + +# H100 top-chase campaign (non-record): long run(s) targeting leaderboard performance. +# Run from repo root: bash starter_kit/scripts/06_non_record_h100_top_chase.sh + +cd /workspace/parameter-golf +mkdir -p logs + +BASE_ENV=( + DATA_PATH=./data/datasets/fineweb10B_sp1024 + TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model + VOCAB_SIZE=1024 + ENABLE_COMPILE=0 + TRAIN_SEQ_LEN=1024 + VAL_LOSS_EVERY=2000 + TRAIN_LOG_EVERY=200 + WARMUP_STEPS=20 + ITERATIONS=500000 + MAX_WALLCLOCK_SECONDS=14400 + WARMDOWN_FRAC=0.2 +) + +# Format: RUN_ID|SPACE_SEPARATED_ENV_OVERRIDES +RUNS=( + "H01_swiglu640_qtrbatch|MLP_ACTIVATION=swiglu MLP_HIDDEN=640 TRAIN_BATCH_TOKENS=131072 GRAD_ACCUM_STEPS=2 MATRIX_LR=0.05 SCALAR_LR=0.04 TIED_EMBED_LR=0.05" + "H02_quasi10b_recipe|MLP_ACTIVATION=relu2 TRAIN_BATCH_TOKENS=524288 GRAD_ACCUM_STEPS=8 MATRIX_LR=0.04 SCALAR_LR=0.04 TIED_EMBED_LR=0.05" +) + +for entry in "${RUNS[@]}"; do + IFS='|' read -r run_id override_str <<< "$entry" + read -r -a override_env <<< "$override_str" + + log_path="logs/${run_id}.log" + echo "============================================================" + echo "Starting ${run_id}" + echo "Log: ${log_path}" + + env RUN_ID="$run_id" "${BASE_ENV[@]}" "${override_env[@]}" \ + torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | tee "$log_path" + + metric_line=$(grep -E "final_int8_zlib_roundtrip_exact|final_int8_zlib_roundtrip " "$log_path" | tail -1 || true) + size_line=$(grep -E "Total submission size int8\+zlib:" "$log_path" | tail -1 || true) + echo "Completed ${run_id}" + echo " ${metric_line}" + echo " ${size_line}" +done + +python starter_kit/scripts/rank_campaign_results.py --logs-glob "logs/H*.log" diff --git a/starter_kit/scripts/prepare_submission.py b/starter_kit/scripts/prepare_submission.py new file mode 100644 index 0000000000..e48ffcc433 --- /dev/null +++ b/starter_kit/scripts/prepare_submission.py @@ -0,0 +1,86 @@ +#!/usr/bin/env python3 +import argparse +import datetime as dt +import json +import re +from pathlib import Path +import shutil + + +def main() -> None: + parser = argparse.ArgumentParser(description="Create a PR-ready records folder.") + parser.add_argument("--track", choices=["10min_16mb", "non_record_16mb"], required=True) + parser.add_argument("--run-name", required=True) + parser.add_argument("--author-name", required=True) + parser.add_argument("--github-id", required=True) + parser.add_argument("--val-bpb", type=float, required=True) + parser.add_argument("--source-train-script", default="train_gpt.py") + args = parser.parse_args() + + repo_root = Path(__file__).resolve().parents[2] + resolved_repo_root = repo_root.resolve() + date = dt.datetime.now().strftime("%Y-%m-%d") + display_run_name = args.run_name.strip() + if not display_run_name: + display_run_name = "run" + # Ensure run name is safe for use as a single path component. + safe_run_name = re.sub(r"[^A-Za-z0-9._-]", "_", args.run_name).strip("._") + if not safe_run_name: + safe_run_name = "run" + slug = f"{date}_{safe_run_name}" + + if args.track == "10min_16mb": + track_dir = repo_root / "records" / "track_10min_16mb" + else: + track_dir = repo_root / "records" / "track_non_record_16mb" + + out_dir = track_dir / slug + # Prevent path traversal by enforcing final location under selected track dir. + if track_dir.resolve() not in out_dir.resolve().parents: + raise ValueError(f"Unsafe output path resolved outside track dir: {out_dir}") + out_dir.mkdir(parents=True, exist_ok=False) + + template_dir = repo_root / "starter_kit" / "templates" + readme_tpl = (template_dir / "README_submission_template.md").read_text(encoding="utf-8") + readme = ( + readme_tpl + .replace("{{RUN_NAME}}", display_run_name) + .replace("{{DATE}}", date) + .replace("{{TRACK}}", args.track) + .replace("{{AUTHOR_NAME}}", args.author_name) + .replace("{{GITHUB_ID}}", args.github_id) + .replace("{{VAL_BPB}}", f"{args.val_bpb:.4f}") + ) + (out_dir / "README.md").write_text(readme, encoding="utf-8") + + submission = { + "author": args.author_name, + "name": display_run_name, + "blurb": "Fill out details and attach train logs.", + "github_id": args.github_id, + "track": args.track, + "val_bpb": round(args.val_bpb, 4), + "date": date + } + (out_dir / "submission.json").write_text(json.dumps(submission, indent=2) + "\n", encoding="utf-8") + + source_script_arg = Path(args.source_train_script) + if source_script_arg.is_absolute(): + raise ValueError("--source-train-script must be a relative path within the repository") + source_script = (resolved_repo_root / source_script_arg).resolve() + if source_script != resolved_repo_root and resolved_repo_root not in source_script.parents: + raise ValueError( + f"Unsafe source train script resolved outside repository: {args.source_train_script}" + ) + if not source_script.exists(): + raise FileNotFoundError(f"Could not find train script: {source_script}") + shutil.copy2(source_script, out_dir / "train_gpt.py") + + (out_dir / "train.log").write_text("# Paste or copy real run logs here\n", encoding="utf-8") + + print(f"Created: {out_dir}") + print("Next: copy your actual log into train.log and complete README details.") + + +if __name__ == "__main__": + main() diff --git a/starter_kit/scripts/rank_campaign_results.py b/starter_kit/scripts/rank_campaign_results.py new file mode 100644 index 0000000000..5d21ba6bc0 --- /dev/null +++ b/starter_kit/scripts/rank_campaign_results.py @@ -0,0 +1,55 @@ +#!/usr/bin/env python3 +import argparse +import glob +import re +from pathlib import Path + +ROUNDTRIP_RE = re.compile(r"final_int8_zlib_roundtrip(?:_exact)?\s+val_loss:([0-9.]+)\s+val_bpb:([0-9.]+)") +SIZE_RE = re.compile(r"Total submission size int8\+zlib:\s*([0-9]+)\s*bytes") + + +def parse_log(path: Path): + text = path.read_text(encoding="utf-8", errors="ignore") + roundtrip_matches = ROUNDTRIP_RE.findall(text) + if not roundtrip_matches: + return None + val_loss_s, val_bpb_s = roundtrip_matches[-1] + size_matches = SIZE_RE.findall(text) + total_size = int(size_matches[-1]) if size_matches else None + return { + "run_id": path.stem, + "val_loss": float(val_loss_s), + "val_bpb": float(val_bpb_s), + "bytes_total": total_size, + "log_path": str(path), + } + + +def main() -> None: + parser = argparse.ArgumentParser(description="Rank campaign runs by final roundtrip val_bpb.") + parser.add_argument("--logs-glob", default="logs/R*.log") + args = parser.parse_args() + + rows = [] + for p in sorted(glob.glob(args.logs_glob)): + parsed = parse_log(Path(p)) + if parsed is not None: + rows.append(parsed) + + if not rows: + print("No completed run logs with final_int8_zlib_roundtrip found.") + return + + rows.sort(key=lambda r: r["val_bpb"]) + + print("Ranked campaign results (lower val_bpb is better):") + for i, r in enumerate(rows, start=1): + size_txt = str(r["bytes_total"]) if r["bytes_total"] is not None else "n/a" + print( + f"{i:02d}. {r['run_id']} val_bpb={r['val_bpb']:.6f} " + f"val_loss={r['val_loss']:.6f} bytes_total={size_txt}" + ) + + +if __name__ == "__main__": + main() diff --git a/starter_kit/templates/README_submission_template.md b/starter_kit/templates/README_submission_template.md new file mode 100644 index 0000000000..b06bb2d6d6 --- /dev/null +++ b/starter_kit/templates/README_submission_template.md @@ -0,0 +1,37 @@ +# {{RUN_NAME}} + +- Date: {{DATE}} +- Track: {{TRACK}} +- Author: {{AUTHOR_NAME}} ({{GITHUB_ID}}) +- Reported val_bpb: {{VAL_BPB}} + +## Summary + +Short summary of the idea and why it may help. + +## What Changed + +- List architecture changes. +- List optimization and schedule changes. +- List quantization or eval changes. + +## Repro Command + +```bash +RUN_ID={{RUN_NAME}} \ +DATA_PATH=./data/datasets/fineweb10B_sp1024/ \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 \ +torchrun --standalone --nproc_per_node=1 train_gpt.py +``` + +## Results + +- val_bpb: +- val_loss: +- compressed_bytes: +- wallclock_seconds: + +## Notes + +Any caveats, negative findings, or follow-up experiments. diff --git a/starter_kit/templates/submission.json.template b/starter_kit/templates/submission.json.template new file mode 100644 index 0000000000..f808c869c1 --- /dev/null +++ b/starter_kit/templates/submission.json.template @@ -0,0 +1,10 @@ +{ + "author": "Your Name", + "name": "your_run_name", + "blurb": "Fill with concise methodology and constraints.", + "github_id": "your_github", + "track": "non_record_16mb", + "val_bpb": 1.2000, + "date": "YYYY-MM-DD", + "seed": 1337 +} diff --git a/train_gpt.py b/train_gpt.py index 651beb2b89..013e71ce75 100644 --- a/train_gpt.py +++ b/train_gpt.py @@ -53,6 +53,7 @@ class Hyperparameters: # Training length. iterations = int(os.environ.get("ITERATIONS", 20000)) warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmdown_frac = float(os.environ.get("WARMDOWN_FRAC", 0.0)) warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) @@ -66,6 +67,8 @@ class Hyperparameters: model_dim = int(os.environ.get("MODEL_DIM", 512)) num_heads = int(os.environ.get("NUM_HEADS", 8)) mlp_mult = int(os.environ.get("MLP_MULT", 2)) + mlp_hidden = int(os.environ.get("MLP_HIDDEN", 0)) + mlp_activation = os.environ.get("MLP_ACTIVATION", "relu2").lower() 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)) @@ -85,6 +88,7 @@ class Hyperparameters: 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", 0.0)) + enable_compile = bool(int(os.environ.get("ENABLE_COMPILE", "1"))) # ----------------------------- # MUON OPTIMIZER @@ -591,30 +595,55 @@ def forward(self, x: Tensor) -> Tensor: q = apply_rotary_emb(q, cos, sin) k = apply_rotary_emb(k, cos, sin) q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] - y = F.scaled_dot_product_attention( - q, - k, - v, - attn_mask=None, - is_causal=True, - enable_gqa=(self.num_kv_heads != self.num_heads), - ) + use_gqa = self.num_kv_heads != self.num_heads + try: + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=use_gqa, + ) + except TypeError: + # Older PyTorch builds do not expose enable_gqa; emulate by repeating KV heads. + if use_gqa: + repeat_factor = self.num_heads // self.num_kv_heads + k = k.repeat_interleave(repeat_factor, dim=1) + v = v.repeat_interleave(repeat_factor, dim=1) + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + ) y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) return self.proj(y) class MLP(nn.Module): - # relu^2 MLP from the original modded-nanogpt setup - def __init__(self, dim: int, mlp_mult: int): + # Supports relu^2 baseline and optional SwiGLU gating. + def __init__(self, dim: int, mlp_mult: int, mlp_hidden: int = 0, activation: str = "relu2"): super().__init__() - hidden = mlp_mult * dim - self.fc = CastedLinear(dim, hidden, bias=False) + hidden = mlp_hidden if mlp_hidden > 0 else mlp_mult * dim + self.activation = activation + if self.activation == "swiglu": + self.fc = CastedLinear(dim, 2 * hidden, bias=False) + else: + self.fc = CastedLinear(dim, hidden, bias=False) self.proj = CastedLinear(hidden, dim, bias=False) self.proj._zero_init = True def forward(self, x: Tensor) -> Tensor: - x = torch.relu(self.fc(x)) - return self.proj(x.square()) + h = self.fc(x) + if self.activation == "swiglu": + gate, value = h.chunk(2, dim=-1) + return self.proj(F.silu(gate) * value) + if self.activation == "relu2": + h = torch.relu(h) + return self.proj(h.square()) + raise ValueError(f"Unsupported MLP_ACTIVATION: {self.activation}") class Block(nn.Module): @@ -624,6 +653,8 @@ def __init__( num_heads: int, num_kv_heads: int, mlp_mult: int, + mlp_hidden: int, + mlp_activation: str, rope_base: float, qk_gain_init: float, ): @@ -631,7 +662,7 @@ def __init__( self.attn_norm = RMSNorm() self.mlp_norm = RMSNorm() self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) - self.mlp = MLP(dim, mlp_mult) + self.mlp = MLP(dim, mlp_mult, mlp_hidden=mlp_hidden, activation=mlp_activation) 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()) @@ -654,6 +685,8 @@ def __init__( num_heads: int, num_kv_heads: int, mlp_mult: int, + mlp_hidden: int, + mlp_activation: str, tie_embeddings: bool, tied_embed_init_std: float, logit_softcap: float, @@ -678,6 +711,8 @@ def __init__( num_heads, num_kv_heads, mlp_mult, + mlp_hidden, + mlp_activation, rope_base, qk_gain_init, ) @@ -733,7 +768,8 @@ def main() -> None: code = Path(__file__).read_text(encoding="utf-8") args = Hyperparameters() - zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + if args.enable_compile: + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) # ----------------------------- # DISTRIBUTED + CUDA SETUP @@ -745,9 +781,13 @@ def main() -> None: 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_accum_steps_env = int(os.environ.get("GRAD_ACCUM_STEPS", "0")) + if grad_accum_steps_env > 0: + grad_accum_steps = grad_accum_steps_env + else: + 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") @@ -830,6 +870,8 @@ def log0(msg: str, console: bool = True) -> None: num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + mlp_hidden=args.mlp_hidden, + mlp_activation=args.mlp_activation, tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, logit_softcap=args.logit_softcap, @@ -840,8 +882,8 @@ def log0(msg: str, console: bool = True) -> None: if isinstance(module, CastedLinear): module.float() restore_low_dim_params_to_fp32(base_model) - compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) - model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + run_model = torch.compile(base_model, dynamic=False, fullgraph=True) if args.enable_compile else base_model + model: nn.Module = DDP(run_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else run_model # Optimizer split: # - token embedding (Adam) uses EMBED_LR @@ -897,6 +939,8 @@ def log0(msg: str, console: bool = True) -> None: 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}") + hidden_width = args.mlp_hidden if args.mlp_hidden > 0 else args.mlp_mult * args.model_dim + log0(f"mlp_activation:{args.mlp_activation} mlp_hidden:{hidden_width}") 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} " @@ -907,6 +951,7 @@ def log0(msg: str, console: bool = True) -> None: f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" ) + log0(f"enable_compile:{args.enable_compile}") log0(f"seed:{args.seed}") # ----------------------------- @@ -922,6 +967,10 @@ def zero_grad_all() -> None: 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_frac > 0 and max_wallclock_ms is not None: + warmdown_ms = max_wallclock_ms * args.warmdown_frac + 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.warmdown_iters <= 0: return 1.0 if max_wallclock_ms is None: