diff options
| author | Nan Liu <45443761+nanlliu@users.noreply.github.com> | 2026-03-19 15:26:46 -0700 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2026-03-19 15:26:46 -0700 |
| commit | 9ac12c26d550481a1a486ce2b450b1ffed60b832 (patch) | |
| tree | fa30460ec2e96320f9f8761c31df31f798490f94 | |
| parent | ae882089b58c74d37a02eda8358219f41cd5f4e1 (diff) | |
Record: 10L Mixed Precision: val_bpb=1.2147 (10 layers + int6 middle layers) (#39)
* Add Lower LR submission: val_bpb=1.2230 (MATRIX_LR=0.02)
Systematic LR sweep showed default Muon/Adam learning rates (0.04) were
too high. MATRIX_LR=0.02, SCALAR_LR=0.02, TIED_EMBED_LR=0.03 gives
consistent improvement. Same 9L/512d architecture, no other changes.
* Add 10L Mixed Precision submission: val_bpb=1.2147
10 transformer layers (vs baseline 9) with mixed int8/int6 compression:
- Full int8 for first/last 3 layers (precision-sensitive)
- Int6 (step=4 rounding) for middle layers 3-6 (compression-friendly)
- Lower LR: MATRIX_LR=0.02, SCALAR_LR=0.02, TIED_EMBED_LR=0.03
- Artifact: 15,928,974 bytes (under 16MB cap)
- Improvement: 0.0097 bpb / 0.0217 nats over baseline (1.2244)
Also adds PRUNE_RATIO and INT4_LAYERS/INT4_STEP support to train_gpt.py
for mixed-precision post-training quantization.
* Revert root train_gpt.py to upstream baseline
The root script should remain the baseline. Submission-specific
modifications (PRUNE_RATIO, INT4_LAYERS, INT4_STEP) only belong
in the records/ folder copy.
8 files changed, 5699 insertions, 0 deletions
diff --git a/records/track_10min_16mb/2026-03-18_LowerLR/README.md b/records/track_10min_16mb/2026-03-18_LowerLR/README.md new file mode 100644 index 0000000..de55954 --- /dev/null +++ b/records/track_10min_16mb/2026-03-18_LowerLR/README.md @@ -0,0 +1,66 @@ +This record captures the `Lower LR` submission. + +## Summary + +Same baseline architecture (9x512, SP-1024, 4 KV heads, tied embeddings, relu^2 MLP) with lower learning rates. A systematic LR sweep over 8 experiments showed the default Muon/Adam learning rates (MATRIX_LR=0.04, SCALAR_LR=0.04, TIED_EMBED_LR=0.05) were too high. Optimal is approximately half the default. + +## Changes from baseline +- `MATRIX_LR=0.02` (default: 0.04) +- `SCALAR_LR=0.02` (default: 0.04) +- `TIED_EMBED_LR=0.03` (default: 0.05) + +No architecture, schedule, or other hyperparameter changes. + +## LR sweep results (8-GPU H200, 600s) + +| MATRIX_LR | val_bpb (post-quant) | Delta vs baseline | +|---|---|---| +| 0.06 | 1.2445 | +0.0159 (much worse) | +| 0.04 (default) | 1.2286 | — | +| 0.03 | 1.2279 | -0.0007 | +| 0.025 | 1.2250 | -0.0036 | +| **0.02** | **1.2230** | **-0.0056** | +| 0.015 | 1.2234 | -0.0052 | + +Configuration: +- Layout: `VOCAB_SIZE=1024 NUM_LAYERS=9 MODEL_DIM=512 NUM_HEADS=8 NUM_KV_HEADS=4 MLP_MULT=2` +- Tied output/input embeddings: `TIE_EMBEDDINGS=1` +- Learning rates: `MATRIX_LR=0.02 SCALAR_LR=0.02 TIED_EMBED_LR=0.03` +- Batching: `TRAIN_BATCH_TOKENS=524288 TRAIN_SEQ_LEN=1024` + +Command: +```bash +RUN_ID=exp25_lr_0.02 \ +DATA_PATH=./data/datasets/fineweb10B_sp1024/ \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 \ +MAX_WALLCLOCK_SECONDS=600 \ +VAL_LOSS_EVERY=200 \ +TRAIN_LOG_EVERY=50 \ +MATRIX_LR=0.02 \ +SCALAR_LR=0.02 \ +TIED_EMBED_LR=0.03 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +Key metrics (from `train.log`): +- Timed training stopped at `14421/20000` steps due to the wallclock cap. +- Pre-quant eval at stop: `val_loss:2.0571`, `val_bpb:1.2183` +- Post-quant roundtrip eval: `val_loss:2.0649`, `val_bpb:1.2230` +- Exact printed metric: `final_int8_zlib_roundtrip_exact val_bpb:1.22296644` +- Train time: `599847ms` (`step_avg:41.60ms`) +- Peak memory: `10246 MiB allocated`, `10310 MiB reserved` +- Serialized model int8+zlib: `15803327 bytes` +- Code size: `50919 bytes` +- Total submission size int8+zlib: `15854246 bytes` + +Training volume: +- Global batch: `524288` tokens/step +- Total train tokens seen: `7560609792` + +Note: Run performed on 8xH200 (141GB HBM3e). Step time (41.60ms) is comparable to 8xH100 baseline (43.54ms), and memory usage (10.2 GiB) is well within H100's 80GB limit. The ~5% faster step time on H200 yields ~400 extra steps, which may account for a small portion of the improvement. + +Included files: +- `train_gpt.py` (code snapshot used for the run) +- `train.log` (exact training log) +- `submission.json` (leaderboard metadata) diff --git a/records/track_10min_16mb/2026-03-18_LowerLR/submission.json b/records/track_10min_16mb/2026-03-18_LowerLR/submission.json new file mode 100644 index 0000000..42ec327 --- /dev/null +++ b/records/track_10min_16mb/2026-03-18_LowerLR/submission.json @@ -0,0 +1,11 @@ +{ + "author": "Nan Liu", + "github_id": "nanlliu", + "name": "Lower LR", + "blurb": "Same 9x512 SP-1024 KV4 tied-embedding baseline architecture with lower Muon/Adam learning rates (MATRIX_LR=0.02, SCALAR_LR=0.02, TIED_EMBED_LR=0.03). Systematic LR sweep showed default 0.04 was too high; optimal is ~0.02.", + "date": "2026-03-18T22:30:00Z", + "val_loss": 2.06492760, + "val_bpb": 1.22296644, + "bytes_total": 15854246, + "bytes_code": 50919 +} diff --git a/records/track_10min_16mb/2026-03-18_LowerLR/train.log b/records/track_10min_16mb/2026-03-18_LowerLR/train.log new file mode 100644 index 0000000..f13b904 --- /dev/null +++ b/records/track_10min_16mb/2026-03-18_LowerLR/train.log @@ -0,0 +1,1667 @@ +""" +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: `train_gpt.py` and `train_gpt_mlx.py` must never be 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)) + lr_warmup_steps = int(os.environ.get("LR_WARMUP_STEPS", 200)) + lr_schedule = os.environ.get("LR_SCHEDULE", "linear") # "linear" (original) or "cosine" + 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)) + qat_start_frac = float(os.environ.get("QAT_START_FRAC", 0.0)) # fraction of training to start QAT (0=disabled, 0.8=last 20%) + 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)) + mlp_type = os.environ.get("MLP_TYPE", "relu2") # "relu2" (original) or "swiglu" + 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("▁"): + 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("<i4").itemsize + token_bytes = np.dtype("<u2").itemsize + header = np.fromfile(file, dtype="<i4", count=256) + # SHARD HEADER INTS & SHARD_MAGIC + if header.size != 256 or int(header[0]) != 20240520 or int(header[1]) != 1: + raise ValueError(f"Unexpected shard header for {file}") + num_tokens = int(header[2]) + expected_size = header_bytes + num_tokens * token_bytes + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype="<u2", count=num_tokens, offset=header_bytes) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + + +class TokenStream: + # Reads shards sequentially and wraps around forever. The training loop therefore + # has deterministic, simple streaming behavior with no sampling or workers. + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # 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) + + +_QAT_ENABLED = False # Set by training loop during QAT phase + +def _fake_quantize_int8_ste(w: Tensor) -> Tensor: + """Simulate per-row int8 quantization with straight-through estimator.""" + with torch.no_grad(): + w32 = w.float() + if w32.ndim == 2: + amax = w32.abs().amax(dim=1, keepdim=True).clamp_min(1e-7) + scale = amax / 127.0 + q = torch.clamp(torch.round(w32 / scale), -127, 127) + w_q = (q * scale).to(w.dtype) + else: + amax = w32.abs().amax().clamp_min(1e-7) + scale = amax / 127.0 + q = torch.clamp(torch.round(w32 / scale), -127, 127) + w_q = (q * scale).to(w.dtype) + # STE: forward uses quantized weights, backward uses original gradients + return w + (w_q - w).detach() + + +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: + w = self.weight.to(x.dtype) + if _QAT_ENABLED and self.weight.numel() > INT8_KEEP_FLOAT_MAX_NUMEL: + w = _fake_quantize_int8_ste(w) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # 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 SwiGLUMLP(nn.Module): + # SwiGLU MLP (LLaMA-style). Parameter-matched to relu^2 MLP: + # relu2 has 2*dim*hidden params; swiglu has 3*dim*hidden, so hidden = 2/3 * mlp_mult * dim + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(2 * mlp_mult * dim / 3) + # Round to nearest multiple of 8 for efficiency + hidden = ((hidden + 7) // 8) * 8 + self.gate = CastedLinear(dim, hidden, bias=False) + self.up = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + return self.proj(F.silu(self.gate(x)) * self.up(x)) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + mlp_type: str, + 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 = SwiGLUMLP(dim, mlp_mult) if mlp_type == "swiglu" else 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, + mlp_type: str, + 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, + mlp_type, + 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, + mlp_type=args.mlp_type, + 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: + # LR warmup: linear ramp from 0 to 1 over lr_warmup_steps + warmup_mul = min(step / max(args.lr_warmup_steps, 1), 1.0) if args.lr_warmup_steps > 0 else 1.0 + + if args.lr_schedule == "cosine": + # Cosine decay: estimate total steps from wallclock, decay over full run + if max_wallclock_ms is not None and step > 0: + step_ms = elapsed_ms / step + estimated_total = int(max_wallclock_ms / step_ms) + else: + estimated_total = args.iterations + progress = min(step / max(estimated_total, 1), 1.0) + decay_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + return warmup_mul * max(decay_mul, 0.0) + + # Original linear warmdown schedule + if args.warmdown_iters <= 0: + return warmup_mul + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + decay_mul = max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + return warmup_mul * decay_mul + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + decay_mul = remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + return warmup_mul * decay_mul + + # 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) + # Enable QAT for the last fraction of training + global _QAT_ENABLED + if args.qat_start_frac > 0 and max_wallclock_ms is not None: + _QAT_ENABLED = elapsed_ms >= args.qat_start_frac * max_wallclock_ms + 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() + +==================================================================================================== +Running Python 3.11.9 (main, Nov 10 2025, 02:08:09) [GCC 11.4.0] +Running PyTorch 2.8.0+cu128 +Wed Mar 18 21:58:25 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H200 Off | 00000002:00:01.0 Off | 0 | +| N/A 34C P0 123W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H200 Off | 00000002:00:02.0 Off | 0 | +| N/A 38C P0 121W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H200 Off | 00000002:00:03.0 Off | 0 | +| N/A 39C P0 124W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H200 Off | 00000002:00:04.0 Off | 0 | +| N/A 34C P0 122W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H200 Off | 00000003:00:01.0 Off | 0 | +| N/A 34C P0 122W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H200 Off | 00000003:00:02.0 Off | 0 | +| N/A 41C P0 123W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H200 Off | 00000003:00:03.0 Off | 0 | +| N/A 38C P0 119W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H200 Off | 00000003:00:04.0 Off | 0 | +| N/A 33C P0 120W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 3930892 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 1 N/A N/A 3930893 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 2 N/A N/A 3930894 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 3 N/A N/A 3930895 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 4 N/A N/A 3930896 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 5 N/A N/A 3930897 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 6 N/A N/A 3930898 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 7 N/A N/A 3930899 C ...nv/versions/3.11.9/bin/python 1506MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:180 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059912 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.03 head_lr:0.0 matrix_lr:0.02 scalar_lr:0.02 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9357 val_bpb:4.1077 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9370 train_time:27ms step_avg:26.75ms +step:2/20000 train_loss:6.9395 train_time:67ms step_avg:33.75ms +step:3/20000 train_loss:6.9196 train_time:108ms step_avg:36.00ms +step:4/20000 train_loss:6.8807 train_time:158ms step_avg:39.57ms +step:5/20000 train_loss:6.8231 train_time:200ms step_avg:40.03ms +step:6/20000 train_loss:6.7488 train_time:242ms step_avg:40.37ms +step:7/20000 train_loss:6.6228 train_time:285ms step_avg:40.68ms +step:8/20000 train_loss:6.4791 train_time:317ms step_avg:39.57ms +step:9/20000 train_loss:6.3144 train_time:358ms step_avg:39.73ms +step:10/20000 train_loss:6.1475 train_time:399ms step_avg:39.88ms +step:50/20000 train_loss:5.1576 train_time:2023ms step_avg:40.47ms +step:100/20000 train_loss:4.0009 train_time:4063ms step_avg:40.63ms +step:150/20000 train_loss:3.3581 train_time:6099ms step_avg:40.66ms +step:200/20000 train_loss:2.9814 train_time:8294ms step_avg:41.47ms +step:200/20000 val_loss:2.9791 val_bpb:1.7644 train_time:8306ms step_avg:41.53ms +step:250/20000 train_loss:2.8186 train_time:10326ms step_avg:41.30ms +step:300/20000 train_loss:2.5205 train_time:12366ms step_avg:41.22ms +step:350/20000 train_loss:2.7077 train_time:14401ms step_avg:41.15ms +step:400/20000 train_loss:2.3745 train_time:16576ms step_avg:41.44ms +step:400/20000 val_loss:2.5855 val_bpb:1.5313 train_time:16595ms step_avg:41.49ms +step:450/20000 train_loss:2.5282 train_time:18609ms step_avg:41.35ms +step:500/20000 train_loss:2.5144 train_time:20646ms step_avg:41.29ms +step:550/20000 train_loss:2.4087 train_time:22687ms step_avg:41.25ms +step:600/20000 train_loss:2.5538 train_time:24921ms step_avg:41.53ms +step:600/20000 val_loss:2.4592 val_bpb:1.4565 train_time:24937ms step_avg:41.56ms +step:650/20000 train_loss:2.3962 train_time:26951ms step_avg:41.46ms +step:700/20000 train_loss:2.4482 train_time:28984ms step_avg:41.41ms +step:750/20000 train_loss:2.2825 train_time:31030ms step_avg:41.37ms +step:800/20000 train_loss:2.3037 train_time:33204ms step_avg:41.50ms +step:800/20000 val_loss:2.3866 val_bpb:1.4135 train_time:33225ms step_avg:41.53ms +step:850/20000 train_loss:2.7243 train_time:35241ms step_avg:41.46ms +step:900/20000 train_loss:2.3437 train_time:37284ms step_avg:41.43ms +step:950/20000 train_loss:2.4026 train_time:39326ms step_avg:41.40ms +step:1000/20000 train_loss:2.3816 train_time:41503ms step_avg:41.50ms +step:1000/20000 val_loss:2.3386 val_bpb:1.3851 train_time:41516ms step_avg:41.52ms +step:1050/20000 train_loss:2.4882 train_time:43527ms step_avg:41.45ms +step:1100/20000 train_loss:2.2661 train_time:45563ms step_avg:41.42ms +step:1150/20000 train_loss:2.2548 train_time:47747ms step_avg:41.52ms +step:1200/20000 train_loss:2.3863 train_time:49794ms step_avg:41.50ms +step:1200/20000 val_loss:2.3043 val_bpb:1.3648 train_time:49797ms step_avg:41.50ms +step:1250/20000 train_loss:2.2116 train_time:51819ms step_avg:41.46ms +step:1300/20000 train_loss:2.3616 train_time:53857ms step_avg:41.43ms +step:1350/20000 train_loss:2.2752 train_time:56018ms step_avg:41.49ms +step:1400/20000 train_loss:2.4304 train_time:58072ms step_avg:41.48ms +step:1400/20000 val_loss:2.2823 val_bpb:1.3517 train_time:58082ms step_avg:41.49ms +step:1450/20000 train_loss:2.2339 train_time:60097ms step_avg:41.45ms +step:1500/20000 train_loss:2.2223 train_time:62137ms step_avg:41.42ms +step:1550/20000 train_loss:2.1548 train_time:64312ms step_avg:41.49ms +step:1600/20000 train_loss:2.0963 train_time:66353ms step_avg:41.47ms +step:1600/20000 val_loss:2.2670 val_bpb:1.3426 train_time:66361ms step_avg:41.48ms +step:1650/20000 train_loss:2.2300 train_time:68378ms step_avg:41.44ms +step:1700/20000 train_loss:2.1717 train_time:70418ms step_avg:41.42ms +step:1750/20000 train_loss:2.2492 train_time:72603ms step_avg:41.49ms +step:1800/20000 train_loss:2.1983 train_time:74635ms step_avg:41.46ms +step:1800/20000 val_loss:2.2502 val_bpb:1.3327 train_time:74657ms step_avg:41.48ms +step:1850/20000 train_loss:2.3093 train_time:76681ms step_avg:41.45ms +step:1900/20000 train_loss:2.1858 train_time:78723ms step_avg:41.43ms +step:1950/20000 train_loss:2.2117 train_time:80921ms step_avg:41.50ms +step:2000/20000 train_loss:2.2537 train_time:82947ms step_avg:41.47ms +step:2000/20000 val_loss:2.2354 val_bpb:1.3239 train_time:82970ms step_avg:41.48ms +step:2050/20000 train_loss:2.2511 train_time:84982ms step_avg:41.45ms +step:2100/20000 train_loss:2.2674 train_time:87184ms step_avg:41.52ms +step:2150/20000 train_loss:2.1881 train_time:89227ms step_avg:41.50ms +step:2200/20000 train_loss:2.0740 train_time:91265ms step_avg:41.48ms +step:2200/20000 val_loss:2.2266 val_bpb:1.3187 train_time:91285ms step_avg:41.49ms +step:2250/20000 train_loss:2.1619 train_time:93305ms step_avg:41.47ms +step:2300/20000 train_loss:2.3779 train_time:95499ms step_avg:41.52ms +step:2350/20000 train_loss:2.2010 train_time:97535ms step_avg:41.50ms +step:2400/20000 train_loss:2.2015 train_time:99573ms step_avg:41.49ms +step:2400/20000 val_loss:2.2165 val_bpb:1.3127 train_time:99594ms step_avg:41.50ms +step:2450/20000 train_loss:2.2036 train_time:101613ms step_avg:41.47ms +step:2500/20000 train_loss:2.1252 train_time:103791ms step_avg:41.52ms +step:2550/20000 train_loss:2.1366 train_time:105846ms step_avg:41.51ms +step:2600/20000 train_loss:2.4066 train_time:107889ms step_avg:41.50ms +step:2600/20000 val_loss:2.2161 val_bpb:1.3125 train_time:107909ms step_avg:41.50ms +step:2650/20000 train_loss:2.2435 train_time:109929ms step_avg:41.48ms +step:2700/20000 train_loss:2.1554 train_time:112113ms step_avg:41.52ms +step:2750/20000 train_loss:2.3645 train_time:114133ms step_avg:41.50ms +step:2800/20000 train_loss:2.2346 train_time:116174ms step_avg:41.49ms +step:2800/20000 val_loss:2.2012 val_bpb:1.3037 train_time:116194ms step_avg:41.50ms +step:2850/20000 train_loss:2.1841 train_time:118209ms step_avg:41.48ms +step:2900/20000 train_loss:2.1789 train_time:120386ms step_avg:41.51ms +step:2950/20000 train_loss:2.2378 train_time:122425ms step_avg:41.50ms +step:3000/20000 train_loss:2.2245 train_time:124465ms step_avg:41.49ms +step:3000/20000 val_loss:2.1939 val_bpb:1.2994 train_time:124485ms step_avg:41.49ms +step:3050/20000 train_loss:2.1679 train_time:126506ms step_avg:41.48ms +step:3100/20000 train_loss:2.2084 train_time:128674ms step_avg:41.51ms +step:3150/20000 train_loss:2.1561 train_time:130715ms step_avg:41.50ms +step:3200/20000 train_loss:2.1857 train_time:132756ms step_avg:41.49ms +step:3200/20000 val_loss:2.1893 val_bpb:1.2966 train_time:132776ms step_avg:41.49ms +step:3250/20000 train_loss:2.0908 train_time:134931ms step_avg:41.52ms +step:3300/20000 train_loss:2.2346 train_time:136978ms step_avg:41.51ms +step:3350/20000 train_loss:2.0918 train_time:139011ms step_avg:41.50ms +step:3400/20000 train_loss:2.1562 train_time:141051ms step_avg:41.49ms +step:3400/20000 val_loss:2.1852 val_bpb:1.2942 train_time:141073ms step_avg:41.49ms +step:3450/20000 train_loss:2.1047 train_time:143351ms step_avg:41.55ms +step:3500/20000 train_loss:2.2490 train_time:145398ms step_avg:41.54ms +step:3550/20000 train_loss:2.3891 train_time:147432ms step_avg:41.53ms +step:3600/20000 train_loss:2.1173 train_time:149471ms step_avg:41.52ms +step:3600/20000 val_loss:2.1779 val_bpb:1.2899 train_time:149492ms step_avg:41.53ms +step:3650/20000 train_loss:2.2191 train_time:151652ms step_avg:41.55ms +step:3700/20000 train_loss:2.1520 train_time:153697ms step_avg:41.54ms +step:3750/20000 train_loss:2.1464 train_time:155743ms step_avg:41.53ms +step:3800/20000 train_loss:2.2206 train_time:157776ms step_avg:41.52ms +step:3800/20000 val_loss:2.1740 val_bpb:1.2876 train_time:157798ms step_avg:41.53ms +step:3850/20000 train_loss:2.1749 train_time:159987ms step_avg:41.56ms +step:3900/20000 train_loss:1.9899 train_time:162037ms step_avg:41.55ms +step:3950/20000 train_loss:2.1256 train_time:164070ms step_avg:41.54ms +step:4000/20000 train_loss:2.1611 train_time:166109ms step_avg:41.53ms +step:4000/20000 val_loss:2.1691 val_bpb:1.2847 train_time:166123ms step_avg:41.53ms +step:4050/20000 train_loss:2.0968 train_time:168299ms step_avg:41.56ms +step:4100/20000 train_loss:2.1860 train_time:170355ms step_avg:41.55ms +step:4150/20000 train_loss:2.3258 train_time:172391ms step_avg:41.54ms +step:4200/20000 train_loss:2.1748 train_time:174581ms step_avg:41.57ms +step:4200/20000 val_loss:2.1651 val_bpb:1.2823 train_time:174599ms step_avg:41.57ms +step:4250/20000 train_loss:2.1285 train_time:176628ms step_avg:41.56ms +step:4300/20000 train_loss:2.0212 train_time:178664ms step_avg:41.55ms +step:4350/20000 train_loss:2.2053 train_time:180796ms step_avg:41.56ms +step:4400/20000 train_loss:2.1137 train_time:182978ms step_avg:41.59ms +step:4400/20000 val_loss:2.1651 val_bpb:1.2823 train_time:183003ms step_avg:41.59ms +step:4450/20000 train_loss:2.0669 train_time:185022ms step_avg:41.58ms +step:4500/20000 train_loss:2.2530 train_time:187065ms step_avg:41.57ms +step:4550/20000 train_loss:2.0511 train_time:189106ms step_avg:41.56ms +step:4600/20000 train_loss:1.9698 train_time:191295ms step_avg:41.59ms +step:4600/20000 val_loss:2.1604 val_bpb:1.2795 train_time:191314ms step_avg:41.59ms +step:4650/20000 train_loss:2.0762 train_time:193332ms step_avg:41.58ms +step:4700/20000 train_loss:2.2670 train_time:195381ms step_avg:41.57ms +step:4750/20000 train_loss:1.9802 train_time:197412ms step_avg:41.56ms +step:4800/20000 train_loss:2.2632 train_time:199621ms step_avg:41.59ms +step:4800/20000 val_loss:2.1566 val_bpb:1.2773 train_time:199631ms step_avg:41.59ms +step:4850/20000 train_loss:2.1477 train_time:201656ms step_avg:41.58ms +step:4900/20000 train_loss:2.1638 train_time:203691ms step_avg:41.57ms +step:4950/20000 train_loss:2.3388 train_time:205733ms step_avg:41.56ms +step:5000/20000 train_loss:2.0292 train_time:207937ms step_avg:41.59ms +step:5000/20000 val_loss:2.1522 val_bpb:1.2746 train_time:207949ms step_avg:41.59ms +step:5050/20000 train_loss:2.2033 train_time:209966ms step_avg:41.58ms +step:5100/20000 train_loss:2.0231 train_time:212016ms step_avg:41.57ms +step:5150/20000 train_loss:2.2743 train_time:214202ms step_avg:41.59ms +step:5200/20000 train_loss:2.1705 train_time:216243ms step_avg:41.59ms +step:5200/20000 val_loss:2.1518 val_bpb:1.2744 train_time:216266ms step_avg:41.59ms +step:5250/20000 train_loss:2.1210 train_time:218283ms step_avg:41.58ms +step:5300/20000 train_loss:2.2122 train_time:220327ms step_avg:41.57ms +step:5350/20000 train_loss:2.1381 train_time:222499ms step_avg:41.59ms +step:5400/20000 train_loss:2.1869 train_time:224549ms step_avg:41.58ms +step:5400/20000 val_loss:2.1478 val_bpb:1.2721 train_time:224570ms step_avg:41.59ms +step:5450/20000 train_loss:2.1970 train_time:226597ms step_avg:41.58ms +step:5500/20000 train_loss:2.1369 train_time:228630ms step_avg:41.57ms +step:5550/20000 train_loss:2.1072 train_time:230827ms step_avg:41.59ms +step:5600/20000 train_loss:2.1806 train_time:232867ms step_avg:41.58ms +step:5600/20000 val_loss:2.1472 val_bpb:1.2717 train_time:232888ms step_avg:41.59ms +step:5650/20000 train_loss:2.0527 train_time:234911ms step_avg:41.58ms +step:5700/20000 train_loss:2.1755 train_time:236963ms step_avg:41.57ms +step:5750/20000 train_loss:2.2142 train_time:239137ms step_avg:41.59ms +step:5800/20000 train_loss:2.1436 train_time:241182ms step_avg:41.58ms +step:5800/20000 val_loss:2.1447 val_bpb:1.2702 train_time:241200ms step_avg:41.59ms +step:5850/20000 train_loss:2.1794 train_time:243224ms step_avg:41.58ms +step:5900/20000 train_loss:2.0950 train_time:245271ms step_avg:41.57ms +step:5950/20000 train_loss:2.1320 train_time:247462ms step_avg:41.59ms +step:6000/20000 train_loss:2.2187 train_time:249501ms step_avg:41.58ms +step:6000/20000 val_loss:2.1419 val_bpb:1.2686 train_time:249521ms step_avg:41.59ms +step:6050/20000 train_loss:2.1237 train_time:251549ms step_avg:41.58ms +step:6100/20000 train_loss:2.1189 train_time:253580ms step_avg:41.57ms +step:6150/20000 train_loss:2.1001 train_time:255763ms step_avg:41.59ms +step:6200/20000 train_loss:2.0846 train_time:257796ms step_avg:41.58ms +step:6200/20000 val_loss:2.1402 val_bpb:1.2676 train_time:257819ms step_avg:41.58ms +step:6250/20000 train_loss:2.1530 train_time:259841ms step_avg:41.57ms +step:6300/20000 train_loss:2.0335 train_time:262015ms step_avg:41.59ms +step:6350/20000 train_loss:2.0226 train_time:264055ms step_avg:41.58ms +step:6400/20000 train_loss:2.1600 train_time:266098ms step_avg:41.58ms +step:6400/20000 val_loss:2.1375 val_bpb:1.2660 train_time:266113ms step_avg:41.58ms +step:6450/20000 train_loss:2.0737 train_time:268131ms step_avg:41.57ms +step:6500/20000 train_loss:2.0776 train_time:270315ms step_avg:41.59ms +step:6550/20000 train_loss:2.2121 train_time:272357ms step_avg:41.58ms +step:6600/20000 train_loss:2.1218 train_time:274402ms step_avg:41.58ms +step:6600/20000 val_loss:2.1340 val_bpb:1.2639 train_time:274418ms step_avg:41.58ms +step:6650/20000 train_loss:2.2922 train_time:276435ms step_avg:41.57ms +step:6700/20000 train_loss:2.1544 train_time:278630ms step_avg:41.59ms +step:6750/20000 train_loss:2.3298 train_time:280669ms step_avg:41.58ms +step:6800/20000 train_loss:2.1900 train_time:282710ms step_avg:41.58ms +step:6800/20000 val_loss:2.1321 val_bpb:1.2627 train_time:282732ms step_avg:41.58ms +step:6850/20000 train_loss:2.0224 train_time:284745ms step_avg:41.57ms +step:6900/20000 train_loss:2.0966 train_time:286927ms step_avg:41.58ms +step:6950/20000 train_loss:2.1716 train_time:288970ms step_avg:41.58ms +step:7000/20000 train_loss:2.2194 train_time:291023ms step_avg:41.57ms +step:7000/20000 val_loss:2.1303 val_bpb:1.2617 train_time:291032ms step_avg:41.58ms +step:7050/20000 train_loss:2.2478 train_time:293055ms step_avg:41.57ms +step:7100/20000 train_loss:2.0619 train_time:295238ms step_avg:41.58ms +step:7150/20000 train_loss:2.1470 train_time:297279ms step_avg:41.58ms +step:7200/20000 train_loss:2.1975 train_time:299325ms step_avg:41.57ms +step:7200/20000 val_loss:2.1302 val_bpb:1.2616 train_time:299345ms step_avg:41.58ms +step:7250/20000 train_loss:2.0975 train_time:301520ms step_avg:41.59ms +step:7300/20000 train_loss:2.0807 train_time:303560ms step_avg:41.58ms +step:7350/20000 train_loss:2.1775 train_time:305601ms step_avg:41.58ms +step:7400/20000 train_loss:2.1169 train_time:307642ms step_avg:41.57ms +step:7400/20000 val_loss:2.1272 val_bpb:1.2599 train_time:307663ms step_avg:41.58ms +step:7450/20000 train_loss:2.1137 train_time:309830ms step_avg:41.59ms +step:7500/20000 train_loss:2.1093 train_time:311864ms step_avg:41.58ms +step:7550/20000 train_loss:2.1671 train_time:313901ms step_avg:41.58ms +step:7600/20000 train_loss:1.9977 train_time:315946ms step_avg:41.57ms +step:7600/20000 val_loss:2.1262 val_bpb:1.2593 train_time:315967ms step_avg:41.57ms +step:7650/20000 train_loss:2.2816 train_time:318120ms step_avg:41.58ms +step:7700/20000 train_loss:2.0852 train_time:320179ms step_avg:41.58ms +step:7750/20000 train_loss:2.1093 train_time:322212ms step_avg:41.58ms +step:7800/20000 train_loss:2.1463 train_time:324257ms step_avg:41.57ms +step:7800/20000 val_loss:2.1232 val_bpb:1.2575 train_time:324277ms step_avg:41.57ms +step:7850/20000 train_loss:1.9975 train_time:326453ms step_avg:41.59ms +step:7900/20000 train_loss:2.1335 train_time:328489ms step_avg:41.58ms +step:7950/20000 train_loss:2.0959 train_time:330530ms step_avg:41.58ms +step:8000/20000 train_loss:2.1129 train_time:332566ms step_avg:41.57ms +step:8000/20000 val_loss:2.1209 val_bpb:1.2561 train_time:332587ms step_avg:41.57ms +step:8050/20000 train_loss:2.0770 train_time:334758ms step_avg:41.58ms +step:8100/20000 train_loss:2.1438 train_time:336797ms step_avg:41.58ms +step:8150/20000 train_loss:2.2537 train_time:338838ms step_avg:41.58ms +step:8200/20000 train_loss:2.1884 train_time:340881ms step_avg:41.57ms +step:8200/20000 val_loss:2.1198 val_bpb:1.2555 train_time:340896ms step_avg:41.57ms +step:8250/20000 train_loss:2.1417 train_time:343054ms step_avg:41.58ms +step:8300/20000 train_loss:2.1192 train_time:345096ms step_avg:41.58ms +step:8350/20000 train_loss:2.2272 train_time:347133ms step_avg:41.57ms +step:8400/20000 train_loss:2.1348 train_time:349312ms step_avg:41.58ms +step:8400/20000 val_loss:2.1193 val_bpb:1.2552 train_time:349332ms step_avg:41.59ms +step:8450/20000 train_loss:2.2258 train_time:351351ms step_avg:41.58ms +step:8500/20000 train_loss:2.1283 train_time:353392ms step_avg:41.58ms +step:8550/20000 train_loss:2.1952 train_time:355438ms step_avg:41.57ms +step:8600/20000 train_loss:2.1338 train_time:357628ms step_avg:41.58ms +step:8600/20000 val_loss:2.1167 val_bpb:1.2536 train_time:357649ms step_avg:41.59ms +step:8650/20000 train_loss:2.0988 train_time:359663ms step_avg:41.58ms +step:8700/20000 train_loss:2.0309 train_time:361710ms step_avg:41.58ms +step:8750/20000 train_loss:2.1958 train_time:363752ms step_avg:41.57ms +step:8800/20000 train_loss:2.1006 train_time:365952ms step_avg:41.59ms +step:8800/20000 val_loss:2.1161 val_bpb:1.2532 train_time:365973ms step_avg:41.59ms +step:8850/20000 train_loss:2.3085 train_time:367999ms step_avg:41.58ms +step:8900/20000 train_loss:2.1983 train_time:370035ms step_avg:41.58ms +step:8950/20000 train_loss:2.1584 train_time:372078ms step_avg:41.57ms +step:9000/20000 train_loss:2.0222 train_time:374260ms step_avg:41.58ms +step:9000/20000 val_loss:2.1159 val_bpb:1.2532 train_time:374284ms step_avg:41.59ms +step:9050/20000 train_loss:2.0589 train_time:376310ms step_avg:41.58ms +step:9100/20000 train_loss:2.3027 train_time:378344ms step_avg:41.58ms +step:9150/20000 train_loss:1.9951 train_time:380383ms step_avg:41.57ms +step:9200/20000 train_loss:2.0824 train_time:382571ms step_avg:41.58ms +step:9200/20000 val_loss:2.1141 val_bpb:1.2521 train_time:382592ms step_avg:41.59ms +step:9250/20000 train_loss:2.1997 train_time:384620ms step_avg:41.58ms +step:9300/20000 train_loss:2.1245 train_time:386656ms step_avg:41.58ms +step:9350/20000 train_loss:2.2276 train_time:388840ms step_avg:41.59ms +step:9400/20000 train_loss:2.1315 train_time:390881ms step_avg:41.58ms +step:9400/20000 val_loss:2.1122 val_bpb:1.2510 train_time:390901ms step_avg:41.59ms +step:9450/20000 train_loss:2.1611 train_time:392928ms step_avg:41.58ms +step:9500/20000 train_loss:2.2559 train_time:394969ms step_avg:41.58ms +step:9550/20000 train_loss:2.1978 train_time:397134ms step_avg:41.58ms +step:9600/20000 train_loss:2.1446 train_time:399182ms step_avg:41.58ms +step:9600/20000 val_loss:2.1113 val_bpb:1.2504 train_time:399195ms step_avg:41.58ms +step:9650/20000 train_loss:2.0912 train_time:401213ms step_avg:41.58ms +step:9700/20000 train_loss:2.1043 train_time:403258ms step_avg:41.57ms +step:9750/20000 train_loss:2.0616 train_time:405438ms step_avg:41.58ms +step:9800/20000 train_loss:2.0748 train_time:407485ms step_avg:41.58ms +step:9800/20000 val_loss:2.1112 val_bpb:1.2503 train_time:407507ms step_avg:41.58ms +step:9850/20000 train_loss:2.0331 train_time:409526ms step_avg:41.58ms +step:9900/20000 train_loss:2.1455 train_time:411576ms step_avg:41.57ms +step:9950/20000 train_loss:2.0212 train_time:413748ms step_avg:41.58ms +step:10000/20000 train_loss:2.1121 train_time:415797ms step_avg:41.58ms +step:10000/20000 val_loss:2.1106 val_bpb:1.2500 train_time:415810ms step_avg:41.58ms +step:10050/20000 train_loss:2.1000 train_time:417829ms step_avg:41.58ms +step:10100/20000 train_loss:2.0911 train_time:419866ms step_avg:41.57ms +step:10150/20000 train_loss:2.0633 train_time:422044ms step_avg:41.58ms +step:10200/20000 train_loss:2.0664 train_time:424080ms step_avg:41.58ms +step:10200/20000 val_loss:2.1073 val_bpb:1.2481 train_time:424100ms step_avg:41.58ms +step:10250/20000 train_loss:2.0637 train_time:426124ms step_avg:41.57ms +step:10300/20000 train_loss:2.1898 train_time:428338ms step_avg:41.59ms +step:10350/20000 train_loss:2.1276 train_time:430363ms step_avg:41.58ms +step:10400/20000 train_loss:2.0965 train_time:432406ms step_avg:41.58ms +step:10400/20000 val_loss:2.1071 val_bpb:1.2480 train_time:432426ms step_avg:41.58ms +step:10450/20000 train_loss:2.0837 train_time:434443ms step_avg:41.57ms +step:10500/20000 train_loss:1.9827 train_time:436639ms step_avg:41.58ms +step:10550/20000 train_loss:2.0099 train_time:438689ms step_avg:41.58ms +step:10600/20000 train_loss:1.9746 train_time:440730ms step_avg:41.58ms +step:10600/20000 val_loss:2.1071 val_bpb:1.2479 train_time:440744ms step_avg:41.58ms +step:10650/20000 train_loss:2.1979 train_time:442761ms step_avg:41.57ms +step:10700/20000 train_loss:2.0719 train_time:444932ms step_avg:41.58ms +step:10750/20000 train_loss:2.1335 train_time:446980ms step_avg:41.58ms +step:10800/20000 train_loss:2.1824 train_time:449018ms step_avg:41.58ms +step:10800/20000 val_loss:2.1050 val_bpb:1.2467 train_time:449043ms step_avg:41.58ms +step:10850/20000 train_loss:2.1403 train_time:451055ms step_avg:41.57ms +step:10900/20000 train_loss:2.1497 train_time:453278ms step_avg:41.59ms +step:10950/20000 train_loss:2.1130 train_time:455306ms step_avg:41.58ms +step:11000/20000 train_loss:2.1196 train_time:457343ms step_avg:41.58ms +step:11000/20000 val_loss:2.1038 val_bpb:1.2460 train_time:457364ms step_avg:41.58ms +step:11050/20000 train_loss:2.0790 train_time:459384ms step_avg:41.57ms +step:11100/20000 train_loss:2.0557 train_time:461577ms step_avg:41.58ms +step:11150/20000 train_loss:2.1606 train_time:463610ms step_avg:41.58ms +step:11200/20000 train_loss:2.0688 train_time:465649ms step_avg:41.58ms +step:11200/20000 val_loss:2.1029 val_bpb:1.2454 train_time:465671ms step_avg:41.58ms +step:11250/20000 train_loss:1.9445 train_time:467691ms step_avg:41.57ms +step:11300/20000 train_loss:1.9937 train_time:469892ms step_avg:41.58ms +step:11350/20000 train_loss:1.9798 train_time:471919ms step_avg:41.58ms +step:11400/20000 train_loss:2.0541 train_time:473961ms step_avg:41.58ms +step:11400/20000 val_loss:2.1033 val_bpb:1.2457 train_time:473981ms step_avg:41.58ms +step:11450/20000 train_loss:2.0449 train_time:476142ms step_avg:41.58ms +step:11500/20000 train_loss:2.1067 train_time:478185ms step_avg:41.58ms +step:11550/20000 train_loss:2.1091 train_time:480227ms step_avg:41.58ms +step:11600/20000 train_loss:2.0575 train_time:482265ms step_avg:41.57ms +step:11600/20000 val_loss:2.1010 val_bpb:1.2444 train_time:482285ms step_avg:41.58ms +step:11650/20000 train_loss:2.1794 train_time:484478ms step_avg:41.59ms +step:11700/20000 train_loss:2.2088 train_time:486515ms step_avg:41.58ms +step:11750/20000 train_loss:2.1199 train_time:488561ms step_avg:41.58ms +step:11800/20000 train_loss:2.0924 train_time:490596ms step_avg:41.58ms +step:11800/20000 val_loss:2.1000 val_bpb:1.2437 train_time:490617ms step_avg:41.58ms +step:11850/20000 train_loss:2.1360 train_time:492795ms step_avg:41.59ms +step:11900/20000 train_loss:2.0570 train_time:494848ms step_avg:41.58ms +step:11950/20000 train_loss:2.0831 train_time:496888ms step_avg:41.58ms +step:12000/20000 train_loss:2.0670 train_time:498923ms step_avg:41.58ms +step:12000/20000 val_loss:2.0985 val_bpb:1.2428 train_time:498946ms step_avg:41.58ms +step:12050/20000 train_loss:2.0843 train_time:501106ms step_avg:41.59ms +step:12100/20000 train_loss:2.1068 train_time:503144ms step_avg:41.58ms +step:12150/20000 train_loss:2.2622 train_time:505197ms step_avg:41.58ms +step:12200/20000 train_loss:2.2202 train_time:507228ms step_avg:41.58ms +step:12200/20000 val_loss:2.0984 val_bpb:1.2428 train_time:507250ms step_avg:41.58ms +step:12250/20000 train_loss:1.9104 train_time:509409ms step_avg:41.58ms +step:12300/20000 train_loss:2.1068 train_time:511453ms step_avg:41.58ms +step:12350/20000 train_loss:2.1745 train_time:513496ms step_avg:41.58ms +step:12400/20000 train_loss:1.8565 train_time:515722ms step_avg:41.59ms +step:12400/20000 val_loss:2.0982 val_bpb:1.2427 train_time:515743ms step_avg:41.59ms +step:12450/20000 train_loss:2.0249 train_time:517767ms step_avg:41.59ms +step:12500/20000 train_loss:2.3661 train_time:519800ms step_avg:41.58ms +step:12550/20000 train_loss:2.1391 train_time:521843ms step_avg:41.58ms +step:12600/20000 train_loss:2.0869 train_time:524014ms step_avg:41.59ms +step:12600/20000 val_loss:2.0987 val_bpb:1.2430 train_time:524036ms step_avg:41.59ms +step:12650/20000 train_loss:2.0534 train_time:526069ms step_avg:41.59ms +step:12700/20000 train_loss:2.0893 train_time:528110ms step_avg:41.58ms +step:12750/20000 train_loss:2.1019 train_time:530150ms step_avg:41.58ms +step:12800/20000 train_loss:2.1097 train_time:532313ms step_avg:41.59ms +step:12800/20000 val_loss:2.0972 val_bpb:1.2421 train_time:532334ms step_avg:41.59ms +step:12850/20000 train_loss:2.0167 train_time:534356ms step_avg:41.58ms +step:12900/20000 train_loss:2.1447 train_time:536388ms step_avg:41.58ms +step:12950/20000 train_loss:2.0178 train_time:538439ms step_avg:41.58ms +step:13000/20000 train_loss:2.1921 train_time:540633ms step_avg:41.59ms +step:13000/20000 val_loss:2.0971 val_bpb:1.2420 train_time:540655ms step_avg:41.59ms +step:13050/20000 train_loss:2.1280 train_time:542674ms step_avg:41.58ms +step:13100/20000 train_loss:2.0189 train_time:544714ms step_avg:41.58ms +step:13150/20000 train_loss:2.0806 train_time:546755ms step_avg:41.58ms +step:13200/20000 train_loss:2.2015 train_time:548947ms step_avg:41.59ms +step:13200/20000 val_loss:2.0942 val_bpb:1.2403 train_time:548959ms step_avg:41.59ms +step:13250/20000 train_loss:2.2855 train_time:550972ms step_avg:41.58ms +step:13300/20000 train_loss:2.0595 train_time:553013ms step_avg:41.58ms +step:13350/20000 train_loss:2.0810 train_time:555056ms step_avg:41.58ms +step:13400/20000 train_loss:2.0734 train_time:557234ms step_avg:41.58ms +step:13400/20000 val_loss:2.0899 val_bpb:1.2378 train_time:557254ms step_avg:41.59ms +step:13450/20000 train_loss:2.1027 train_time:559274ms step_avg:41.58ms +step:13500/20000 train_loss:2.0397 train_time:561311ms step_avg:41.58ms +step:13550/20000 train_loss:2.1138 train_time:563503ms step_avg:41.59ms +step:13600/20000 train_loss:1.9325 train_time:565549ms step_avg:41.58ms +step:13600/20000 val_loss:2.0847 val_bpb:1.2347 train_time:565564ms step_avg:41.59ms +step:13650/20000 train_loss:2.1098 train_time:567585ms step_avg:41.58ms +step:13700/20000 train_loss:2.1089 train_time:569632ms step_avg:41.58ms +step:13750/20000 train_loss:2.1835 train_time:571799ms step_avg:41.59ms +step:13800/20000 train_loss:2.0195 train_time:573851ms step_avg:41.58ms +step:13800/20000 val_loss:2.0755 val_bpb:1.2292 train_time:573871ms step_avg:41.58ms +step:13850/20000 train_loss:2.0397 train_time:575897ms step_avg:41.58ms +step:13900/20000 train_loss:2.1022 train_time:577930ms step_avg:41.58ms +step:13950/20000 train_loss:1.9046 train_time:580197ms step_avg:41.59ms +step:14000/20000 train_loss:2.0752 train_time:582232ms step_avg:41.59ms +step:14000/20000 val_loss:2.0691 val_bpb:1.2254 train_time:582252ms step_avg:41.59ms +step:14050/20000 train_loss:1.9633 train_time:584270ms step_avg:41.59ms +step:14100/20000 train_loss:2.0442 train_time:586464ms step_avg:41.59ms +step:14150/20000 train_loss:2.1136 train_time:588646ms step_avg:41.60ms +step:14200/20000 train_loss:2.1544 train_time:590693ms step_avg:41.60ms +step:14200/20000 val_loss:2.0628 val_bpb:1.2217 train_time:590705ms step_avg:41.60ms +step:14250/20000 train_loss:2.2636 train_time:592724ms step_avg:41.59ms +step:14300/20000 train_loss:2.1059 train_time:594763ms step_avg:41.59ms +step:14350/20000 train_loss:2.0329 train_time:596936ms step_avg:41.60ms +step:14400/20000 train_loss:2.0489 train_time:598976ms step_avg:41.60ms +step:14400/20000 val_loss:2.0573 val_bpb:1.2184 train_time:598996ms step_avg:41.60ms +step:14421/20000 val_loss:2.0571 val_bpb:1.2183 train_time:599847ms step_avg:41.60ms +stopping_early: wallclock_cap train_time:599847ms step:14421/20000 +peak memory allocated: 10246 MiB reserved: 10310 MiB +Serialized model: 67224983 bytes +Code size: 50919 bytes +Total submission size: 67275902 bytes +Serialized model int8+zlib: 15803327 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x) +Total submission size int8+zlib: 15854246 bytes +final_int8_zlib_roundtrip val_loss:2.0649 val_bpb:1.2230 eval_time:1284ms +final_int8_zlib_roundtrip_exact val_loss:2.06492760 val_bpb:1.22296644 diff --git a/records/track_10min_16mb/2026-03-18_LowerLR/train_gpt.py b/records/track_10min_16mb/2026-03-18_LowerLR/train_gpt.py new file mode 100644 index 0000000..0deb056 --- /dev/null +++ b/records/track_10min_16mb/2026-03-18_LowerLR/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: `train_gpt.py` and `train_gpt_mlx.py` must never be 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("▁"): + 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("<i4").itemsize + token_bytes = np.dtype("<u2").itemsize + header = np.fromfile(file, dtype="<i4", count=256) + # SHARD HEADER INTS & SHARD_MAGIC + if header.size != 256 or int(header[0]) != 20240520 or int(header[1]) != 1: + raise ValueError(f"Unexpected shard header for {file}") + num_tokens = int(header[2]) + expected_size = header_bytes + num_tokens * token_bytes + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype="<u2", count=num_tokens, offset=header_bytes) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + + +class TokenStream: + # Reads shards sequentially and wraps around forever. The training loop therefore + # has deterministic, simple streaming behavior with no sampling or workers. + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # 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_10min_16mb/2026-03-19_10L_MixedPrecision/README.md b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/README.md new file mode 100644 index 0000000..ce9815d --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/README.md @@ -0,0 +1,75 @@ +This record captures the `10L Mixed Precision` submission. + +## Summary + +Two key improvements over the baseline: + +1. **10 transformer layers** instead of 9 — adds depth for better language modeling +2. **Lower learning rates** — MATRIX_LR=0.02, SCALAR_LR=0.02, TIED_EMBED_LR=0.03 (vs default 0.04/0.04/0.05) +3. **Mixed int8/int6 compression** — middle layers (3,4,5,6) use int6 precision (round int8 to nearest 4) for better zlib compression, while first/last layers keep full int8 + +The 10-layer model at dim=512 has 18.9M params which compresses to 17.6MB with standard int8+zlib — 1.6MB over the 16MB cap. By reducing precision on the 4 middle layers to int6 (64 quantization levels instead of 256), the compressed size drops to 15.9MB with only 0.0018 bpb quality loss from quantization. + +## Configuration + +- Layout: `VOCAB_SIZE=1024 NUM_LAYERS=10 MODEL_DIM=512 NUM_HEADS=8 NUM_KV_HEADS=4 MLP_MULT=2` +- Tied output/input embeddings: `TIE_EMBEDDINGS=1` +- Learning rates: `MATRIX_LR=0.02 SCALAR_LR=0.02 TIED_EMBED_LR=0.03` +- Mixed precision: `INT4_LAYERS=3,4,5,6 INT4_STEP=4` (int6 for middle layers) +- Batching: `TRAIN_BATCH_TOKENS=524288 TRAIN_SEQ_LEN=1024` + +## Command + +```bash +RUN_ID=exp45_10L_int6_mid \ +DATA_PATH=./data/datasets/fineweb10B_sp1024/ \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 \ +MAX_WALLCLOCK_SECONDS=600 \ +VAL_LOSS_EVERY=200 \ +TRAIN_LOG_EVERY=50 \ +NUM_LAYERS=10 \ +MATRIX_LR=0.02 \ +SCALAR_LR=0.02 \ +TIED_EMBED_LR=0.03 \ +INT4_LAYERS=3,4,5,6 \ +INT4_STEP=4 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## Key Metrics + +- Pre-quant eval: `val_loss:2.0480`, `val_bpb:1.2129` +- Post-quant (int8/int6 mixed + zlib): `val_loss:2.0510`, `val_bpb:1.2147` +- Exact: `final_int8_zlib_roundtrip_exact val_bpb:1.21474500` +- Quantization gap: 0.0018 bpb (vs baseline's 0.0093) +- Train time: `599732ms` (`step_avg:45.78ms`) +- Steps: 13,100/20,000 (wallclock limited) +- Peak memory: 11,389 MiB allocated +- Artifact: 15,928,974 bytes (code: 48,917 + model: 15,880,057) + +## Compression Analysis + +| Layer Group | Precision | Reason | +|---|---|---| +| Layers 0-2 (early) | int8 (256 levels) | Critical for input processing | +| Layers 3-6 (middle) | int6 (64 levels) | Less sensitive, saves ~1.6MB | +| Layers 7-9 (late) | int8 (256 levels) | Critical for output quality | + +## LR Sweep Results + +Systematic sweep showed default LR (0.04) was too high: +| MATRIX_LR | val_bpb (9L baseline) | +|---|---| +| 0.04 (default) | 1.2286 | +| 0.02 (optimal) | 1.2230 | + +## Note on Hardware + +Run performed on 8xH200 (step_avg: 45.78ms). H100 baseline was 43.54ms/step for 9 layers; 10 layers would be ~47-48ms on H100, yielding ~12,500-12,700 steps. Results should be comparable. + +## Included Files + +- `train_gpt.py` (code snapshot) +- `train.log` (training log) +- `submission.json` (metadata) diff --git a/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/submission.json b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/submission.json new file mode 100644 index 0000000..0f3c947 --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/submission.json @@ -0,0 +1,11 @@ +{ + "author": "Nan Liu", + "github_id": "nanlliu", + "name": "10L Mixed Precision", + "blurb": "10-layer 512-dim model with lower LR (MATRIX_LR=0.02) and mixed int8/int6 compression: full int8 for first/last 3 layers, int6 (step=4 rounding) for middle layers 3-6. Fits 16MB via better compression while gaining an extra transformer layer over baseline.", + "date": "2026-03-19T03:30:00Z", + "val_loss": 2.05104604, + "val_bpb": 1.21474500, + "bytes_total": 15928974, + "bytes_code": 48917 +} diff --git a/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/train.log b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/train.log new file mode 100644 index 0000000..c40e941 --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/train.log @@ -0,0 +1,1591 @@ +""" +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: `train_gpt.py` and `train_gpt_mlx.py` must never be 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)) + prune_ratio = float(os.environ.get("PRUNE_RATIO", 0.0)) # fraction of int8 range to prune (e.g. 0.1 = zero out |val| <= 12) + int4_layers = os.environ.get("INT4_LAYERS", "") # comma-separated layer indices for reduced precision (e.g. "3,4,5,6") + int4_step = int(os.environ.get("INT4_STEP", 16)) # rounding step: 2=int7, 4=int6, 8=int5, 16=int4 + + # 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("▁"): + 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("<i4").itemsize + token_bytes = np.dtype("<u2").itemsize + header = np.fromfile(file, dtype="<i4", count=256) + # SHARD HEADER INTS & SHARD_MAGIC + if header.size != 256 or int(header[0]) != 20240520 or int(header[1]) != 1: + raise ValueError(f"Unexpected shard header for {file}") + num_tokens = int(header[2]) + expected_size = header_bytes + num_tokens * token_bytes + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype="<u2", count=num_tokens, offset=header_bytes) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + + +class TokenStream: + # Reads shards sequentially and wraps around forever. The training loop therefore + # has deterministic, simple streaming behavior with no sampling or workers. + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # 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()) + # Optional post-quantization pruning: zero out small int8 values for better compression + if args.prune_ratio > 0: + threshold = int(127 * args.prune_ratio) + for name in list(quant_obj.get("quantized", {}).keys()): + t = quant_obj["quantized"][name] + t[t.abs() <= threshold] = 0 + # Optional mixed-precision: round middle layers to int4 (16 levels) for better compression + if args.int4_layers: + int4_set = set(int(x) for x in args.int4_layers.split(",") if x.strip()) + for name in list(quant_obj.get("quantized", {}).keys()): + layer_num = -1 + if "blocks." in name: + try: + layer_num = int(name.split("blocks.")[1].split(".")[0]) + except (ValueError, IndexError): + pass + if layer_num in int4_set: + t = quant_obj["quantized"][name] + step = args.int4_step + quant_obj["quantized"][name] = ((t.float() / step).round() * step).clamp(-127, 127).to(torch.int8) + 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() + +==================================================================================================== +Running Python 3.11.9 (main, Nov 10 2025, 02:08:09) [GCC 11.4.0] +Running PyTorch 2.8.0+cu128 +Thu Mar 19 02:57:29 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H200 Off | 00000002:00:01.0 Off | 0 | +| N/A 32C P0 122W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H200 Off | 00000002:00:02.0 Off | 0 | +| N/A 34C P0 118W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H200 Off | 00000002:00:03.0 Off | 0 | +| N/A 35C P0 121W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H200 Off | 00000002:00:04.0 Off | 0 | +| N/A 32C P0 121W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H200 Off | 00000003:00:01.0 Off | 0 | +| N/A 32C P0 120W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H200 Off | 00000003:00:02.0 Off | 0 | +| N/A 37C P0 121W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H200 Off | 00000003:00:03.0 Off | 0 | +| N/A 35C P0 117W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H200 Off | 00000003:00:04.0 Off | 0 | +| N/A 31C P0 119W / 700W | 1516MiB / 143771MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 301282 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 1 N/A N/A 301283 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 2 N/A N/A 301284 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 3 N/A N/A 301285 C ...nv/versions/3.11.9/bin/python 1530MiB | +| 4 N/A N/A 301286 C ...nv/versions/3.11.9/bin/python 1534MiB | +| 5 N/A N/A 301287 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 6 N/A N/A 301288 C ...nv/versions/3.11.9/bin/python 1506MiB | +| 7 N/A N/A 301289 C ...nv/versions/3.11.9/bin/python 1506MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:180 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.03 head_lr:0.0 matrix_lr:0.02 scalar_lr:0.02 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9363 val_bpb:4.1080 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9355 train_time:31ms step_avg:31.28ms +step:2/20000 train_loss:12.1231 train_time:73ms step_avg:36.68ms +step:3/20000 train_loss:7.2165 train_time:125ms step_avg:41.81ms +step:4/20000 train_loss:6.4176 train_time:165ms step_avg:41.19ms +step:5/20000 train_loss:6.8785 train_time:213ms step_avg:42.51ms +step:6/20000 train_loss:7.6248 train_time:256ms step_avg:42.66ms +step:7/20000 train_loss:6.8402 train_time:303ms step_avg:43.24ms +step:8/20000 train_loss:6.4526 train_time:348ms step_avg:43.54ms +step:9/20000 train_loss:6.2574 train_time:394ms step_avg:43.79ms +step:10/20000 train_loss:6.1774 train_time:441ms step_avg:44.06ms +step:50/20000 train_loss:4.0508 train_time:2226ms step_avg:44.51ms +step:100/20000 train_loss:3.2461 train_time:4465ms step_avg:44.65ms +step:150/20000 train_loss:2.9349 train_time:6704ms step_avg:44.69ms +step:200/20000 train_loss:2.7576 train_time:9077ms step_avg:45.38ms +step:200/20000 val_loss:2.7303 val_bpb:1.6170 train_time:9103ms step_avg:45.52ms +step:250/20000 train_loss:2.6695 train_time:11324ms step_avg:45.30ms +step:300/20000 train_loss:2.4316 train_time:13571ms step_avg:45.24ms +step:350/20000 train_loss:2.6215 train_time:15818ms step_avg:45.19ms +step:400/20000 train_loss:2.3094 train_time:18198ms step_avg:45.49ms +step:400/20000 val_loss:2.5186 val_bpb:1.4917 train_time:18221ms step_avg:45.55ms +step:450/20000 train_loss:2.4622 train_time:20441ms step_avg:45.42ms +step:500/20000 train_loss:2.4606 train_time:22696ms step_avg:45.39ms +step:550/20000 train_loss:2.3666 train_time:24931ms step_avg:45.33ms +step:600/20000 train_loss:2.5197 train_time:27345ms step_avg:45.57ms +step:600/20000 val_loss:2.4169 val_bpb:1.4315 train_time:27361ms step_avg:45.60ms +step:650/20000 train_loss:2.3557 train_time:29586ms step_avg:45.52ms +step:700/20000 train_loss:2.4120 train_time:31830ms step_avg:45.47ms +step:750/20000 train_loss:2.2413 train_time:34073ms step_avg:45.43ms +step:800/20000 train_loss:2.2639 train_time:36456ms step_avg:45.57ms +step:800/20000 val_loss:2.3534 val_bpb:1.3938 train_time:36482ms step_avg:45.60ms +step:850/20000 train_loss:2.6951 train_time:38715ms step_avg:45.55ms +step:900/20000 train_loss:2.3127 train_time:40946ms step_avg:45.50ms +step:950/20000 train_loss:2.3720 train_time:43193ms step_avg:45.47ms +step:1000/20000 train_loss:2.3507 train_time:45593ms step_avg:45.59ms +step:1000/20000 val_loss:2.3092 val_bpb:1.3677 train_time:45617ms step_avg:45.62ms +step:1050/20000 train_loss:2.4618 train_time:47845ms step_avg:45.57ms +step:1100/20000 train_loss:2.2360 train_time:50109ms step_avg:45.55ms +step:1150/20000 train_loss:2.2286 train_time:52507ms step_avg:45.66ms +step:1200/20000 train_loss:2.3629 train_time:54759ms step_avg:45.63ms +step:1200/20000 val_loss:2.2795 val_bpb:1.3501 train_time:54781ms step_avg:45.65ms +step:1250/20000 train_loss:2.1861 train_time:57017ms step_avg:45.61ms +step:1300/20000 train_loss:2.3350 train_time:59267ms step_avg:45.59ms +step:1350/20000 train_loss:2.2513 train_time:61658ms step_avg:45.67ms +step:1400/20000 train_loss:2.4046 train_time:63916ms step_avg:45.65ms +step:1400/20000 val_loss:2.2574 val_bpb:1.3370 train_time:63932ms step_avg:45.67ms +step:1450/20000 train_loss:2.2156 train_time:66161ms step_avg:45.63ms +step:1500/20000 train_loss:2.1999 train_time:68424ms step_avg:45.62ms +step:1550/20000 train_loss:2.1331 train_time:70797ms step_avg:45.68ms +step:1600/20000 train_loss:2.0744 train_time:73040ms step_avg:45.65ms +step:1600/20000 val_loss:2.2430 val_bpb:1.3284 train_time:73065ms step_avg:45.67ms +step:1650/20000 train_loss:2.2082 train_time:75294ms step_avg:45.63ms +step:1700/20000 train_loss:2.1512 train_time:77538ms step_avg:45.61ms +step:1750/20000 train_loss:2.2288 train_time:79928ms step_avg:45.67ms +step:1800/20000 train_loss:2.1758 train_time:82184ms step_avg:45.66ms +step:1800/20000 val_loss:2.2278 val_bpb:1.3194 train_time:82201ms step_avg:45.67ms +step:1850/20000 train_loss:2.2862 train_time:84429ms step_avg:45.64ms +step:1900/20000 train_loss:2.1669 train_time:86684ms step_avg:45.62ms +step:1950/20000 train_loss:2.1909 train_time:89069ms step_avg:45.68ms +step:2000/20000 train_loss:2.2319 train_time:91324ms step_avg:45.66ms +step:2000/20000 val_loss:2.2129 val_bpb:1.3106 train_time:91351ms step_avg:45.68ms +step:2050/20000 train_loss:2.2279 train_time:93580ms step_avg:45.65ms +step:2100/20000 train_loss:2.2434 train_time:95987ms step_avg:45.71ms +step:2150/20000 train_loss:2.1649 train_time:98238ms step_avg:45.69ms +step:2200/20000 train_loss:2.0522 train_time:100494ms step_avg:45.68ms +step:2200/20000 val_loss:2.2044 val_bpb:1.3055 train_time:100513ms step_avg:45.69ms +step:2250/20000 train_loss:2.1400 train_time:102739ms step_avg:45.66ms +step:2300/20000 train_loss:2.3533 train_time:105126ms step_avg:45.71ms +step:2350/20000 train_loss:2.1788 train_time:107379ms step_avg:45.69ms +step:2400/20000 train_loss:2.1812 train_time:109631ms step_avg:45.68ms +step:2400/20000 val_loss:2.1944 val_bpb:1.2997 train_time:109653ms step_avg:45.69ms +step:2450/20000 train_loss:2.1842 train_time:111889ms step_avg:45.67ms +step:2500/20000 train_loss:2.0987 train_time:114285ms step_avg:45.71ms +step:2550/20000 train_loss:2.1142 train_time:116542ms step_avg:45.70ms +step:2600/20000 train_loss:2.3890 train_time:118794ms step_avg:45.69ms +step:2600/20000 val_loss:2.1926 val_bpb:1.2986 train_time:118817ms step_avg:45.70ms +step:2650/20000 train_loss:2.2227 train_time:121046ms step_avg:45.68ms +step:2700/20000 train_loss:2.1354 train_time:123439ms step_avg:45.72ms +step:2750/20000 train_loss:2.3391 train_time:125689ms step_avg:45.71ms +step:2800/20000 train_loss:2.2132 train_time:127942ms step_avg:45.69ms +step:2800/20000 val_loss:2.1801 val_bpb:1.2912 train_time:127965ms step_avg:45.70ms +step:2850/20000 train_loss:2.1652 train_time:130191ms step_avg:45.68ms +step:2900/20000 train_loss:2.1553 train_time:132603ms step_avg:45.73ms +step:2950/20000 train_loss:2.2221 train_time:134847ms step_avg:45.71ms +step:3000/20000 train_loss:2.2030 train_time:137103ms step_avg:45.70ms +step:3000/20000 val_loss:2.1722 val_bpb:1.2865 train_time:137127ms step_avg:45.71ms +step:3050/20000 train_loss:2.1472 train_time:139356ms step_avg:45.69ms +step:3100/20000 train_loss:2.1888 train_time:141742ms step_avg:45.72ms +step:3150/20000 train_loss:2.1400 train_time:143993ms step_avg:45.71ms +step:3200/20000 train_loss:2.1645 train_time:146242ms step_avg:45.70ms +step:3200/20000 val_loss:2.1671 val_bpb:1.2835 train_time:146268ms step_avg:45.71ms +step:3250/20000 train_loss:2.0671 train_time:148637ms step_avg:45.73ms +step:3300/20000 train_loss:2.2137 train_time:150885ms step_avg:45.72ms +step:3350/20000 train_loss:2.0731 train_time:153128ms step_avg:45.71ms +step:3400/20000 train_loss:2.1360 train_time:155380ms step_avg:45.70ms +step:3400/20000 val_loss:2.1642 val_bpb:1.2818 train_time:155406ms step_avg:45.71ms +step:3450/20000 train_loss:2.0890 train_time:157762ms step_avg:45.73ms +step:3500/20000 train_loss:2.2285 train_time:160022ms step_avg:45.72ms +step:3550/20000 train_loss:2.3672 train_time:162275ms step_avg:45.71ms +step:3600/20000 train_loss:2.0920 train_time:164524ms step_avg:45.70ms +step:3600/20000 val_loss:2.1558 val_bpb:1.2768 train_time:164550ms step_avg:45.71ms +step:3650/20000 train_loss:2.2009 train_time:166952ms step_avg:45.74ms +step:3700/20000 train_loss:2.1262 train_time:169204ms step_avg:45.73ms +step:3750/20000 train_loss:2.1242 train_time:171454ms step_avg:45.72ms +step:3800/20000 train_loss:2.1979 train_time:173690ms step_avg:45.71ms +step:3800/20000 val_loss:2.1522 val_bpb:1.2747 train_time:173716ms step_avg:45.71ms +step:3850/20000 train_loss:2.1522 train_time:176098ms step_avg:45.74ms +step:3900/20000 train_loss:1.9648 train_time:178363ms step_avg:45.73ms +step:3950/20000 train_loss:2.1063 train_time:180605ms step_avg:45.72ms +step:4000/20000 train_loss:2.1404 train_time:182855ms step_avg:45.71ms +step:4000/20000 val_loss:2.1483 val_bpb:1.2723 train_time:182880ms step_avg:45.72ms +step:4050/20000 train_loss:2.0791 train_time:185248ms step_avg:45.74ms +step:4100/20000 train_loss:2.1671 train_time:187504ms step_avg:45.73ms +step:4150/20000 train_loss:2.3022 train_time:189754ms step_avg:45.72ms +step:4200/20000 train_loss:2.1544 train_time:192160ms step_avg:45.75ms +step:4200/20000 val_loss:2.1438 val_bpb:1.2697 train_time:192179ms step_avg:45.76ms +step:4250/20000 train_loss:2.1045 train_time:194411ms step_avg:45.74ms +step:4300/20000 train_loss:2.0035 train_time:196665ms step_avg:45.74ms +step:4350/20000 train_loss:2.1856 train_time:198915ms step_avg:45.73ms +step:4400/20000 train_loss:2.0900 train_time:201320ms step_avg:45.75ms +step:4400/20000 val_loss:2.1441 val_bpb:1.2698 train_time:201344ms step_avg:45.76ms +step:4450/20000 train_loss:2.0453 train_time:203578ms step_avg:45.75ms +step:4500/20000 train_loss:2.2388 train_time:205817ms step_avg:45.74ms +step:4550/20000 train_loss:2.0374 train_time:208069ms step_avg:45.73ms +step:4600/20000 train_loss:1.9508 train_time:210467ms step_avg:45.75ms +step:4600/20000 val_loss:2.1397 val_bpb:1.2672 train_time:210490ms step_avg:45.76ms +step:4650/20000 train_loss:2.0559 train_time:212730ms step_avg:45.75ms +step:4700/20000 train_loss:2.2390 train_time:214973ms step_avg:45.74ms +step:4750/20000 train_loss:1.9568 train_time:217228ms step_avg:45.73ms +step:4800/20000 train_loss:2.2412 train_time:219613ms step_avg:45.75ms +step:4800/20000 val_loss:2.1358 val_bpb:1.2649 train_time:219633ms step_avg:45.76ms +step:4850/20000 train_loss:2.1316 train_time:221864ms step_avg:45.75ms +step:4900/20000 train_loss:2.1431 train_time:224117ms step_avg:45.74ms +step:4950/20000 train_loss:2.3209 train_time:226367ms step_avg:45.73ms +step:5000/20000 train_loss:2.0024 train_time:228775ms step_avg:45.75ms +step:5000/20000 val_loss:2.1310 val_bpb:1.2621 train_time:228799ms step_avg:45.76ms +step:5050/20000 train_loss:2.1805 train_time:231025ms step_avg:45.75ms +step:5100/20000 train_loss:2.0028 train_time:233279ms step_avg:45.74ms +step:5150/20000 train_loss:2.2555 train_time:235682ms step_avg:45.76ms +step:5200/20000 train_loss:2.1495 train_time:237949ms step_avg:45.76ms +step:5200/20000 val_loss:2.1311 val_bpb:1.2621 train_time:237972ms step_avg:45.76ms +step:5250/20000 train_loss:2.1005 train_time:240194ms step_avg:45.75ms +step:5300/20000 train_loss:2.1919 train_time:242450ms step_avg:45.75ms +step:5350/20000 train_loss:2.1187 train_time:244835ms step_avg:45.76ms +step:5400/20000 train_loss:2.1622 train_time:247089ms step_avg:45.76ms +step:5400/20000 val_loss:2.1266 val_bpb:1.2595 train_time:247105ms step_avg:45.76ms +step:5450/20000 train_loss:2.1771 train_time:249340ms step_avg:45.75ms +step:5500/20000 train_loss:2.1191 train_time:251580ms step_avg:45.74ms +step:5550/20000 train_loss:2.0819 train_time:253964ms step_avg:45.76ms +step:5600/20000 train_loss:2.1589 train_time:256223ms step_avg:45.75ms +step:5600/20000 val_loss:2.1262 val_bpb:1.2593 train_time:256248ms step_avg:45.76ms +step:5650/20000 train_loss:2.0342 train_time:258475ms step_avg:45.75ms +step:5700/20000 train_loss:2.1558 train_time:260721ms step_avg:45.74ms +step:5750/20000 train_loss:2.1967 train_time:263132ms step_avg:45.76ms +step:5800/20000 train_loss:2.1182 train_time:265385ms step_avg:45.76ms +step:5800/20000 val_loss:2.1233 val_bpb:1.2575 train_time:265407ms step_avg:45.76ms +step:5850/20000 train_loss:2.1585 train_time:267633ms step_avg:45.75ms +step:5900/20000 train_loss:2.0729 train_time:269888ms step_avg:45.74ms +step:5950/20000 train_loss:2.1111 train_time:272265ms step_avg:45.76ms +step:6000/20000 train_loss:2.2017 train_time:274533ms step_avg:45.76ms +step:6000/20000 val_loss:2.1214 val_bpb:1.2564 train_time:274549ms step_avg:45.76ms +step:6050/20000 train_loss:2.1040 train_time:276778ms step_avg:45.75ms +step:6100/20000 train_loss:2.0972 train_time:279041ms step_avg:45.74ms +step:6150/20000 train_loss:2.0780 train_time:281419ms step_avg:45.76ms +step:6200/20000 train_loss:2.0625 train_time:283675ms step_avg:45.75ms +step:6200/20000 val_loss:2.1192 val_bpb:1.2551 train_time:283696ms step_avg:45.76ms +step:6250/20000 train_loss:2.1312 train_time:285929ms step_avg:45.75ms +step:6300/20000 train_loss:2.0109 train_time:288345ms step_avg:45.77ms +step:6350/20000 train_loss:1.9996 train_time:290597ms step_avg:45.76ms +step:6400/20000 train_loss:2.1393 train_time:292855ms step_avg:45.76ms +step:6400/20000 val_loss:2.1165 val_bpb:1.2535 train_time:292870ms step_avg:45.76ms +step:6450/20000 train_loss:2.0561 train_time:295104ms step_avg:45.75ms +step:6500/20000 train_loss:2.0563 train_time:297478ms step_avg:45.77ms +step:6550/20000 train_loss:2.1868 train_time:299733ms step_avg:45.76ms +step:6600/20000 train_loss:2.1016 train_time:301990ms step_avg:45.76ms +step:6600/20000 val_loss:2.1126 val_bpb:1.2512 train_time:302015ms step_avg:45.76ms +step:6650/20000 train_loss:2.2676 train_time:304249ms step_avg:45.75ms +step:6700/20000 train_loss:2.1344 train_time:306638ms step_avg:45.77ms +step:6750/20000 train_loss:2.3069 train_time:308884ms step_avg:45.76ms +step:6800/20000 train_loss:2.1721 train_time:311133ms step_avg:45.75ms +step:6800/20000 val_loss:2.1120 val_bpb:1.2508 train_time:311158ms step_avg:45.76ms +step:6850/20000 train_loss:2.0027 train_time:313387ms step_avg:45.75ms +step:6900/20000 train_loss:2.0708 train_time:315782ms step_avg:45.77ms +step:6950/20000 train_loss:2.1526 train_time:318017ms step_avg:45.76ms +step:7000/20000 train_loss:2.2026 train_time:320274ms step_avg:45.75ms +step:7000/20000 val_loss:2.1100 val_bpb:1.2497 train_time:320294ms step_avg:45.76ms +step:7050/20000 train_loss:2.2315 train_time:322525ms step_avg:45.75ms +step:7100/20000 train_loss:2.0491 train_time:325043ms step_avg:45.78ms +step:7150/20000 train_loss:2.1278 train_time:327297ms step_avg:45.78ms +step:7200/20000 train_loss:2.1779 train_time:329554ms step_avg:45.77ms +step:7200/20000 val_loss:2.1091 val_bpb:1.2492 train_time:329578ms step_avg:45.77ms +step:7250/20000 train_loss:2.0836 train_time:331941ms step_avg:45.78ms +step:7300/20000 train_loss:2.0657 train_time:334194ms step_avg:45.78ms +step:7350/20000 train_loss:2.1627 train_time:336441ms step_avg:45.77ms +step:7400/20000 train_loss:2.0957 train_time:338691ms step_avg:45.77ms +step:7400/20000 val_loss:2.1063 val_bpb:1.2475 train_time:338718ms step_avg:45.77ms +step:7450/20000 train_loss:2.0930 train_time:341085ms step_avg:45.78ms +step:7500/20000 train_loss:2.0893 train_time:343339ms step_avg:45.78ms +step:7550/20000 train_loss:2.1512 train_time:345594ms step_avg:45.77ms +step:7600/20000 train_loss:1.9759 train_time:347849ms step_avg:45.77ms +step:7600/20000 val_loss:2.1051 val_bpb:1.2467 train_time:347873ms step_avg:45.77ms +step:7650/20000 train_loss:2.2603 train_time:350250ms step_avg:45.78ms +step:7700/20000 train_loss:2.0707 train_time:352506ms step_avg:45.78ms +step:7750/20000 train_loss:2.0895 train_time:354756ms step_avg:45.78ms +step:7800/20000 train_loss:2.1256 train_time:357009ms step_avg:45.77ms +step:7800/20000 val_loss:2.1026 val_bpb:1.2453 train_time:357034ms step_avg:45.77ms +step:7850/20000 train_loss:1.9783 train_time:359400ms step_avg:45.78ms +step:7900/20000 train_loss:2.1132 train_time:361649ms step_avg:45.78ms +step:7950/20000 train_loss:2.0734 train_time:363898ms step_avg:45.77ms +step:8000/20000 train_loss:2.0955 train_time:366149ms step_avg:45.77ms +step:8000/20000 val_loss:2.1000 val_bpb:1.2437 train_time:366174ms step_avg:45.77ms +step:8050/20000 train_loss:2.0608 train_time:368554ms step_avg:45.78ms +step:8100/20000 train_loss:2.1254 train_time:370808ms step_avg:45.78ms +step:8150/20000 train_loss:2.2333 train_time:373056ms step_avg:45.77ms +step:8200/20000 train_loss:2.1677 train_time:375301ms step_avg:45.77ms +step:8200/20000 val_loss:2.0991 val_bpb:1.2432 train_time:375326ms step_avg:45.77ms +step:8250/20000 train_loss:2.1263 train_time:377716ms step_avg:45.78ms +step:8300/20000 train_loss:2.0939 train_time:379966ms step_avg:45.78ms +step:8350/20000 train_loss:2.2056 train_time:382213ms step_avg:45.77ms +step:8400/20000 train_loss:2.1145 train_time:384602ms step_avg:45.79ms +step:8400/20000 val_loss:2.0986 val_bpb:1.2429 train_time:384627ms step_avg:45.79ms +step:8450/20000 train_loss:2.2060 train_time:386859ms step_avg:45.78ms +step:8500/20000 train_loss:2.1064 train_time:389125ms step_avg:45.78ms +step:8550/20000 train_loss:2.1724 train_time:391359ms step_avg:45.77ms +step:8600/20000 train_loss:2.1130 train_time:393767ms step_avg:45.79ms +step:8600/20000 val_loss:2.0957 val_bpb:1.2412 train_time:393779ms step_avg:45.79ms +step:8650/20000 train_loss:2.0786 train_time:396002ms step_avg:45.78ms +step:8700/20000 train_loss:2.0084 train_time:398259ms step_avg:45.78ms +step:8750/20000 train_loss:2.1722 train_time:400508ms step_avg:45.77ms +step:8800/20000 train_loss:2.0755 train_time:402921ms step_avg:45.79ms +step:8800/20000 val_loss:2.0949 val_bpb:1.2407 train_time:402947ms step_avg:45.79ms +step:8850/20000 train_loss:2.2856 train_time:405174ms step_avg:45.78ms +step:8900/20000 train_loss:2.1794 train_time:407424ms step_avg:45.78ms +step:8950/20000 train_loss:2.1362 train_time:409675ms step_avg:45.77ms +step:9000/20000 train_loss:2.0033 train_time:412068ms step_avg:45.79ms +step:9000/20000 val_loss:2.0952 val_bpb:1.2409 train_time:412090ms step_avg:45.79ms +step:9050/20000 train_loss:2.0378 train_time:414315ms step_avg:45.78ms +step:9100/20000 train_loss:2.2843 train_time:416568ms step_avg:45.78ms +step:9150/20000 train_loss:1.9771 train_time:418810ms step_avg:45.77ms +step:9200/20000 train_loss:2.0639 train_time:421205ms step_avg:45.78ms +step:9200/20000 val_loss:2.0933 val_bpb:1.2398 train_time:421229ms step_avg:45.79ms +step:9250/20000 train_loss:2.1723 train_time:423454ms step_avg:45.78ms +step:9300/20000 train_loss:2.1050 train_time:425706ms step_avg:45.77ms +step:9350/20000 train_loss:2.2069 train_time:428101ms step_avg:45.79ms +step:9400/20000 train_loss:2.1070 train_time:430349ms step_avg:45.78ms +step:9400/20000 val_loss:2.0910 val_bpb:1.2384 train_time:430374ms step_avg:45.78ms +step:9450/20000 train_loss:2.1418 train_time:432606ms step_avg:45.78ms +step:9500/20000 train_loss:2.2405 train_time:434854ms step_avg:45.77ms +step:9550/20000 train_loss:2.1759 train_time:437256ms step_avg:45.79ms +step:9600/20000 train_loss:2.1233 train_time:439511ms step_avg:45.78ms +step:9600/20000 val_loss:2.0904 val_bpb:1.2381 train_time:439538ms step_avg:45.79ms +step:9650/20000 train_loss:2.0687 train_time:441764ms step_avg:45.78ms +step:9700/20000 train_loss:2.0842 train_time:444016ms step_avg:45.77ms +step:9750/20000 train_loss:2.0422 train_time:446408ms step_avg:45.79ms +step:9800/20000 train_loss:2.0480 train_time:448658ms step_avg:45.78ms +step:9800/20000 val_loss:2.0920 val_bpb:1.2390 train_time:448685ms step_avg:45.78ms +step:9850/20000 train_loss:2.0115 train_time:450915ms step_avg:45.78ms +step:9900/20000 train_loss:2.1271 train_time:453166ms step_avg:45.77ms +step:9950/20000 train_loss:2.0026 train_time:455566ms step_avg:45.79ms +step:10000/20000 train_loss:2.0928 train_time:457811ms step_avg:45.78ms +step:10000/20000 val_loss:2.0901 val_bpb:1.2379 train_time:457836ms step_avg:45.78ms +step:10050/20000 train_loss:2.0821 train_time:460064ms step_avg:45.78ms +step:10100/20000 train_loss:2.0749 train_time:462314ms step_avg:45.77ms +step:10150/20000 train_loss:2.0414 train_time:464692ms step_avg:45.78ms +step:10200/20000 train_loss:2.0427 train_time:466956ms step_avg:45.78ms +step:10200/20000 val_loss:2.0864 val_bpb:1.2357 train_time:466975ms step_avg:45.78ms +step:10250/20000 train_loss:2.0468 train_time:469202ms step_avg:45.78ms +step:10300/20000 train_loss:2.1732 train_time:471628ms step_avg:45.79ms +step:10350/20000 train_loss:2.1053 train_time:473869ms step_avg:45.78ms +step:10400/20000 train_loss:2.0771 train_time:476113ms step_avg:45.78ms +step:10400/20000 val_loss:2.0866 val_bpb:1.2358 train_time:476129ms step_avg:45.78ms +step:10450/20000 train_loss:2.0673 train_time:478356ms step_avg:45.78ms +step:10500/20000 train_loss:1.9587 train_time:480740ms step_avg:45.78ms +step:10550/20000 train_loss:1.9904 train_time:482992ms step_avg:45.78ms +step:10600/20000 train_loss:1.9568 train_time:485256ms step_avg:45.78ms +step:10600/20000 val_loss:2.0864 val_bpb:1.2357 train_time:485269ms step_avg:45.78ms +step:10650/20000 train_loss:2.1711 train_time:487496ms step_avg:45.77ms +step:10700/20000 train_loss:2.0553 train_time:489875ms step_avg:45.78ms +step:10750/20000 train_loss:2.1130 train_time:492156ms step_avg:45.78ms +step:10800/20000 train_loss:2.1674 train_time:494388ms step_avg:45.78ms +step:10800/20000 val_loss:2.0843 val_bpb:1.2345 train_time:494408ms step_avg:45.78ms +step:10850/20000 train_loss:2.1129 train_time:496643ms step_avg:45.77ms +step:10900/20000 train_loss:2.1275 train_time:499051ms step_avg:45.78ms +step:10950/20000 train_loss:2.0914 train_time:501302ms step_avg:45.78ms +step:11000/20000 train_loss:2.0934 train_time:503554ms step_avg:45.78ms +step:11000/20000 val_loss:2.0832 val_bpb:1.2338 train_time:503578ms step_avg:45.78ms +step:11050/20000 train_loss:2.0554 train_time:505806ms step_avg:45.77ms +step:11100/20000 train_loss:2.0394 train_time:508195ms step_avg:45.78ms +step:11150/20000 train_loss:2.1367 train_time:510449ms step_avg:45.78ms +step:11200/20000 train_loss:2.0480 train_time:512705ms step_avg:45.78ms +step:11200/20000 val_loss:2.0825 val_bpb:1.2333 train_time:512727ms step_avg:45.78ms +step:11250/20000 train_loss:1.9284 train_time:514958ms step_avg:45.77ms +step:11300/20000 train_loss:1.9753 train_time:517364ms step_avg:45.78ms +step:11350/20000 train_loss:1.9595 train_time:519613ms step_avg:45.78ms +step:11400/20000 train_loss:2.0328 train_time:521860ms step_avg:45.78ms +step:11400/20000 val_loss:2.0824 val_bpb:1.2333 train_time:521885ms step_avg:45.78ms +step:11450/20000 train_loss:2.0225 train_time:524252ms step_avg:45.79ms +step:11500/20000 train_loss:2.0859 train_time:526498ms step_avg:45.78ms +step:11550/20000 train_loss:2.0876 train_time:528748ms step_avg:45.78ms +step:11600/20000 train_loss:2.0380 train_time:531003ms step_avg:45.78ms +step:11600/20000 val_loss:2.0805 val_bpb:1.2322 train_time:531027ms step_avg:45.78ms +step:11650/20000 train_loss:2.1590 train_time:533404ms step_avg:45.79ms +step:11700/20000 train_loss:2.1872 train_time:535658ms step_avg:45.78ms +step:11750/20000 train_loss:2.0971 train_time:537908ms step_avg:45.78ms +step:11800/20000 train_loss:2.0735 train_time:540163ms step_avg:45.78ms +step:11800/20000 val_loss:2.0792 val_bpb:1.2314 train_time:540177ms step_avg:45.78ms +step:11850/20000 train_loss:2.1114 train_time:542551ms step_avg:45.78ms +step:11900/20000 train_loss:2.0381 train_time:544803ms step_avg:45.78ms +step:11950/20000 train_loss:2.0598 train_time:547043ms step_avg:45.78ms +step:12000/20000 train_loss:2.0470 train_time:549297ms step_avg:45.77ms +step:12000/20000 val_loss:2.0757 val_bpb:1.2293 train_time:549321ms step_avg:45.78ms +step:12050/20000 train_loss:2.0646 train_time:551692ms step_avg:45.78ms +step:12100/20000 train_loss:2.0815 train_time:553944ms step_avg:45.78ms +step:12150/20000 train_loss:2.2381 train_time:556195ms step_avg:45.78ms +step:12200/20000 train_loss:2.1889 train_time:558450ms step_avg:45.77ms +step:12200/20000 val_loss:2.0693 val_bpb:1.2256 train_time:558473ms step_avg:45.78ms +step:12250/20000 train_loss:1.8816 train_time:560835ms step_avg:45.78ms +step:12300/20000 train_loss:2.0743 train_time:563087ms step_avg:45.78ms +step:12350/20000 train_loss:2.1383 train_time:565338ms step_avg:45.78ms +step:12400/20000 train_loss:1.8282 train_time:567731ms step_avg:45.78ms +step:12400/20000 val_loss:2.0625 val_bpb:1.2215 train_time:567756ms step_avg:45.79ms +step:12450/20000 train_loss:1.9936 train_time:569994ms step_avg:45.78ms +step:12500/20000 train_loss:2.3264 train_time:572244ms step_avg:45.78ms +step:12550/20000 train_loss:2.1008 train_time:574498ms step_avg:45.78ms +step:12600/20000 train_loss:2.0522 train_time:576894ms step_avg:45.79ms +step:12600/20000 val_loss:2.0554 val_bpb:1.2174 train_time:576920ms step_avg:45.79ms +step:12650/20000 train_loss:2.0168 train_time:579149ms step_avg:45.78ms +step:12700/20000 train_loss:2.0485 train_time:581399ms step_avg:45.78ms +step:12750/20000 train_loss:2.0600 train_time:583649ms step_avg:45.78ms +step:12800/20000 train_loss:2.0687 train_time:586059ms step_avg:45.79ms +step:12800/20000 val_loss:2.0480 val_bpb:1.2129 train_time:586089ms step_avg:45.79ms +step:12850/20000 train_loss:1.9659 train_time:588321ms step_avg:45.78ms +step:12900/20000 train_loss:2.0918 train_time:590577ms step_avg:45.78ms +step:12950/20000 train_loss:1.9700 train_time:592831ms step_avg:45.78ms +step:13000/20000 train_loss:2.1380 train_time:595226ms step_avg:45.79ms +step:13000/20000 val_loss:2.0416 val_bpb:1.2092 train_time:595253ms step_avg:45.79ms +step:13050/20000 train_loss:2.0707 train_time:597483ms step_avg:45.78ms +step:13100/20000 train_loss:1.9696 train_time:599748ms step_avg:45.78ms +step:13101/20000 val_loss:2.0396 val_bpb:1.2080 train_time:599812ms step_avg:45.78ms +stopping_early: wallclock_cap train_time:599812ms step:13101/20000 +peak memory allocated: 11389 MiB reserved: 11704 MiB +Serialized model: 74578915 bytes +Code size: 49058 bytes +Total submission size: 74627973 bytes +Serialized model int8+zlib: 15879916 bytes (payload:19030336 raw_torch:19080377 payload_ratio:3.92x) +Total submission size int8+zlib: 15928974 bytes +final_int8_zlib_roundtrip val_loss:2.0510 val_bpb:1.2147 eval_time:1432ms +final_int8_zlib_roundtrip_exact val_loss:2.05104604 val_bpb:1.21474500 diff --git a/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/train_gpt.py b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/train_gpt.py new file mode 100644 index 0000000..f06bd07 --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_10L_MixedPrecision/train_gpt.py @@ -0,0 +1,1152 @@ +""" +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: `train_gpt.py` and `train_gpt_mlx.py` must never be 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)) + prune_ratio = float(os.environ.get("PRUNE_RATIO", 0.0)) # fraction of int8 range to prune (e.g. 0.1 = zero out |val| <= 12) + int4_layers = os.environ.get("INT4_LAYERS", "") # comma-separated layer indices for reduced precision (e.g. "3,4,5,6") + int4_step = int(os.environ.get("INT4_STEP", 16)) # rounding step: 2=int7, 4=int6, 8=int5, 16=int4 + + # 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("▁"): + 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("<i4").itemsize + token_bytes = np.dtype("<u2").itemsize + header = np.fromfile(file, dtype="<i4", count=256) + # SHARD HEADER INTS & SHARD_MAGIC + if header.size != 256 or int(header[0]) != 20240520 or int(header[1]) != 1: + raise ValueError(f"Unexpected shard header for {file}") + num_tokens = int(header[2]) + expected_size = header_bytes + num_tokens * token_bytes + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype="<u2", count=num_tokens, offset=header_bytes) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + + +class TokenStream: + # Reads shards sequentially and wraps around forever. The training loop therefore + # has deterministic, simple streaming behavior with no sampling or workers. + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # 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()) + # Optional post-quantization pruning: zero out small int8 values for better compression + if args.prune_ratio > 0: + threshold = int(127 * args.prune_ratio) + for name in list(quant_obj.get("quantized", {}).keys()): + t = quant_obj["quantized"][name] + t[t.abs() <= threshold] = 0 + # Optional mixed-precision: round middle layers to int4 (16 levels) for better compression + if args.int4_layers: + int4_set = set(int(x) for x in args.int4_layers.split(",") if x.strip()) + for name in list(quant_obj.get("quantized", {}).keys()): + layer_num = -1 + if "blocks." in name: + try: + layer_num = int(name.split("blocks.")[1].split(".")[0]) + except (ValueError, IndexError): + pass + if layer_num in int4_set: + t = quant_obj["quantized"][name] + step = args.int4_step + quant_obj["quantized"][name] = ((t.float() / step).round() * step).clamp(-127, 127).to(torch.int8) + 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() |