diff --git a/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/README.md b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/README.md new file mode 100644 index 0000000000..0fb7e9c77f --- /dev/null +++ b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/README.md @@ -0,0 +1,79 @@ +# Non-Record Submission: Saliency-Guided Local 5090 + +This record documents a local 1xRTX 5090 saliency-guided sweep around the 9x512 GQA baseline. The run sequence was: + +1. the original 5B-token full-eval run on March 26, 2026 +2. a 24-hour full-eval continuation on April 8, 2026 +3. a 1-hour cosine-schedule proxy run on April 9, 2026 + +The folder is intentionally normalized to the standard record layout: + +- `README.md` for the write-up +- `results.tsv` for the run table +- `submission.json` for the best full-eval run metadata +- `train_gpt.py` for the experiment-local code snapshot + +`results.tsv` is transcribed from the local `wandb/` run directories, not from ad hoc notes. + +## Best Result + +The best full-eval run in this sweep is the 24-hour continuation: + +- run: `saliency_24h_30gb_legacy_seed2025` +- started: `2026-04-08 11:21:09Z` / `2026-04-08 20:21 KST` +- pre-quant `val_bpb`: `1.21953852` +- post-quant `val_bpb`: `1.22864374` +- counted artifact bytes: `15,850,915` +- stop step: `44,628` +- GPU: `1xRTX5090` + +This is the run described in `submission.json`. + +## Sweep Summary + +All three recorded runs use the same core saliency recipe: + +- 9 layers, width 512, 8 heads, 4 KV heads, MLP mult 2 +- SentencePiece vocab 1024, sequence length 1024 +- saliency token prior on +- saliency dynamic correction on +- saliency phrase term on +- saliency attention bias on +- saliency bigram off +- `TRAIN_BATCH_TOKENS=262144` + +The changes across the sweep were: + +- 5B run: fixed-token budget, `50` train shards, legacy flat-plus-tail warmdown schedule +- 24h run: same legacy schedule and saliency settings, but extended to `125` train shards and a 24-hour wallclock cap +- 1h run: same saliency settings, but reduced to `8` train shards and switched to cosine LR with `LR_WARMUP_STEPS=64`, `LR_DECAY_START_FRAC=0.65`, `LR_MIN_SCALE=0.15` + +## Run Chronology + +| Run | Start | Eval | Key result | +| --- | --- | --- | --- | +| 5B | 2026-03-26 18:01 KST | full | post-quant `1.24582822` | +| 24h | 2026-04-08 20:21 KST | full | post-quant `1.22864374` | +| 1h | 2026-04-09 23:26 KST | proxy (4,194,304 val tokens) | post-quant `1.35156821` | + +## Results Source + +The values in `results.tsv` were read from these local W&B run folders: + +- `wandb/run-20260326_180143-wzspgrg0` +- `wandb/run-20260408_202109-vajvzlaj` +- `wandb/run-20260409_232646-hgygstyg` + +The recorded numbers come from `wandb-summary.json`, `config.yaml`, and `output.log` in those directories. + +## Included Files + +- `train_gpt.py`: experiment-local trainer snapshot kept with this record +- `results.tsv`: normalized run table from the local W&B artifacts +- `submission.json`: metadata for the best full-eval run + +## Notes + +- This folder date, `2026-04-13`, is the record write-up date, not the original training date. +- The included `train_gpt.py` is the experiment-local code snapshot kept for recordkeeping and reruns. The reported metrics themselves come from the recorded W&B runs listed above. +- The 1-hour run is intentionally marked as proxy-only because it used `VAL_MAX_TOKENS=4194304` with `LOCAL_PROXY_EVAL=1`. diff --git a/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/results.tsv b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/results.tsv new file mode 100644 index 0000000000..ba294befab --- /dev/null +++ b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/results.tsv @@ -0,0 +1,4 @@ +label run_id wandb_run_dir started_at_utc eval_split train_shards lr_schedule step_stop train_time_s pre_quant_val_bpb post_quant_val_bpb bytes_total notes +5b_full_eval saliency_5b_30gb_seed2025 wandb/run-20260326_180143-wzspgrg0 2026-03-26T09:01:43.917794Z full 50 flat_warmdown 19073 39633.45 1.23881099 1.24582822 15858012 original 5B saliency winner +24h_full_eval saliency_24h_30gb_legacy_seed2025 wandb/run-20260408_202109-vajvzlaj 2026-04-08T11:21:09.972987Z full 125 flat_warmdown 44628 86401.05 1.21953852 1.22864374 15850915 best full-eval continuation +1h_proxy_eval saliency_1h_30gb_legacy_seed2025 wandb/run-20260409_232646-hgygstyg 2026-04-09T14:26:46.332733Z proxy_4194304_tokens 8 cosine 1844 3601.39 1.34998024 1.35156821 15303442 1-hour scheduler probe diff --git a/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/submission.json b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/submission.json new file mode 100644 index 0000000000..ef8cf7121b --- /dev/null +++ b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/submission.json @@ -0,0 +1,18 @@ +{ + "author": "", + "github_id": "", + "name": "Saliency-Guided Local 5090", + "blurb": "Single-RTX5090 non-record saliency-guided sweep documenting the original 5B run, a 24-hour continuation, and a 1-hour cosine-schedule probe. Best full-eval post-quant score is 1.22864374 val_bpb at 15,850,915 counted bytes.", + "date": "2026-04-08T11:21:09.972987Z", + "track": "non-record-16mb", + "val_loss": 2.074513482238166, + "val_bpb": 1.2286437358795264, + "pre_quant_val_loss": 2.0591396984249735, + "pre_quant_val_bpb": 1.2195385151419558, + "step_stop": 44628, + "wallclock_seconds": 86400, + "bytes_total": 15850915, + "bytes_model_int8_zlib": 15781469, + "bytes_code": 69446, + "gpu": "1xRTX5090" +} diff --git a/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/train_gpt.py b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/train_gpt.py new file mode 100644 index 0000000000..ef92a3a656 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-13_SaliencyGuidedLocal5090/train_gpt.py @@ -0,0 +1,1364 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +WORD_BOUNDARY = "\u2581" +OFFICIAL_CAP_BYTES = 16_000_000 +def getenv_alias(name: str, aliases: tuple[str, ...] = ()) -> str | None: + value = os.environ.get(name) + if value is not None: + return value + for alias in aliases: + value = os.environ.get(alias) + if value is not None: + return value + return None +def env_flag(name: str, default: bool, aliases: tuple[str, ...] = ()) -> bool: + raw = getenv_alias(name, aliases) + if raw is None: + return default + normalized = raw.strip().lower() + if normalized in {"true", "yes", "on"}: + return True + if normalized in {"false", "no", "off"}: + return False + return float(normalized) != 0.0 +def env_int(name: str, default: int, aliases: tuple[str, ...] = ()) -> int: + raw = getenv_alias(name, aliases) + return default if raw is None else int(raw) +def env_float(name: str, default: float, aliases: tuple[str, ...] = ()) -> float: + raw = getenv_alias(name, aliases) + return default if raw is None else float(raw) +# ----------------------------- +# 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 = env_int("VAL_BATCH_SIZE", 524_288) + val_loss_every = env_int("VAL_LOSS_EVERY", 1000) + val_max_tokens = env_int("VAL_MAX_TOKENS", 0) + local_proxy_eval = env_flag("LOCAL_PROXY_EVAL", False) + train_log_every = env_int("TRAIN_LOG_EVERY", 200) + # Training length. + iterations = env_int("ITERATIONS", 20000) + warmdown_iters = env_int("WARMDOWN_ITERS", 1200) + warmup_steps = env_int("WARMUP_STEPS", 20) + lr_schedule = os.environ.get("LR_SCHEDULE", "flat_warmdown").strip().lower() + lr_warmup_steps = env_int("LR_WARMUP_STEPS", 0) + lr_decay_start_frac = env_float("LR_DECAY_START_FRAC", 1.0) + lr_min_scale = env_float("LR_MIN_SCALE", 0.0) + train_batch_tokens = env_int("TRAIN_BATCH_TOKENS", 524_288) + train_seq_len = env_int("TRAIN_SEQ_LEN", 1024) + max_wallclock_seconds = env_float("MAX_WALLCLOCK_SECONDS", 600.0) + qk_gain_init = env_float("QK_GAIN_INIT", 1.5) + # Model shape. + vocab_size = env_int("VOCAB_SIZE", 1024) + num_layers = env_int("NUM_LAYERS", 9) + num_kv_heads = env_int("NUM_KV_HEADS", 4) + model_dim = env_int("MODEL_DIM", 512) + num_heads = env_int("NUM_HEADS", 8) + mlp_mult = env_int("MLP_MULT", 2) + tie_embeddings = env_flag("TIE_EMBEDDINGS", True) + rope_base = env_float("ROPE_BASE", 10000.0) + logit_softcap = env_float("LOGIT_SOFTCAP", 30.0) + # Optimizer hyperparameters. + embed_lr = env_float("EMBED_LR", 0.6) + head_lr = env_float("HEAD_LR", 0.008) + tied_embed_lr = env_float("TIED_EMBED_LR", 0.05) + tied_embed_init_std = env_float("TIED_EMBED_INIT_STD", 0.005) + matrix_lr = env_float("MATRIX_LR", 0.04) + scalar_lr = env_float("SCALAR_LR", 0.04) + muon_momentum = env_float("MUON_MOMENTUM", 0.95) + muon_backend_steps = env_int("MUON_BACKEND_STEPS", 5) + muon_momentum_warmup_start = env_float("MUON_MOMENTUM_WARMUP_START", 0.85) + muon_momentum_warmup_steps = env_int("MUON_MOMENTUM_WARMUP_STEPS", 500) + beta1 = env_float("BETA1", 0.9) + beta2 = env_float("BETA2", 0.95) + adam_eps = env_float("ADAM_EPS", 1e-8) + grad_clip_norm = env_float("GRAD_CLIP_NORM", 0.0) + use_compile = env_flag("USE_COMPILE", False) + artifact_budget_bytes = env_int("ARTIFACT_BUDGET_BYTES", 15_950_000) + saliency_enable = env_flag("SALIENCY_ENABLE", False, aliases=("SALIENCY_ENABLED",)) + saliency_token_prior = env_flag("SALIENCY_TOKEN_PRIOR", True, aliases=("SALIENCY_STATIC_PRIOR",)) + saliency_bigram = env_flag("SALIENCY_BIGRAM", True) + saliency_dynamic = env_flag("SALIENCY_DYNAMIC", True) + saliency_phrase = env_flag("SALIENCY_PHRASE", True) + saliency_attn_bias = env_flag("SALIENCY_ATTN_BIAS", True) + saliency_value_scale = env_flag("SALIENCY_VALUE_SCALE", False) + saliency_min = env_float("SALIENCY_MIN", 0.90) + saliency_max = env_float("SALIENCY_MAX", 1.12) + saliency_lambda = env_float("SALIENCY_LAMBDA", 0.10, aliases=("SALIENCY_BIAS_SCALE",)) + saliency_gamma = env_float("SALIENCY_GAMMA", 0.20) + saliency_bigram_buckets = env_int("SALIENCY_BIGRAM_BUCKETS", 4096) + saliency_token_scale = env_float("SALIENCY_TOKEN_SCALE", 0.04, aliases=("SALIENCY_STATIC_SCALE",)) + saliency_bigram_scale = env_float("SALIENCY_BIGRAM_SCALE", 0.05) + saliency_dynamic_scale = env_float("SALIENCY_DYNAMIC_SCALE", 0.08, aliases=("SALIENCY_LEARNED_SCALE",)) + saliency_phrase_scale = env_float("SALIENCY_PHRASE_SCALE", 0.03, aliases=("SALIENCY_SPAN_SCALE",)) + saliency_reg_weight = env_float("SALIENCY_REG_WEIGHT", 0.0) + saliency_smooth_weight = env_float("SALIENCY_SMOOTH_WEIGHT", 0.5) +# ----------------------------- +# 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(WORD_BOUNDARY): + 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, max_tokens: int = 0) -> 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 max_tokens > 0: + usable = min(usable, (max_tokens // seq_len) * seq_len) + if usable <= 0: + raise ValueError( + f"Validation split is too short for TRAIN_SEQ_LEN={seq_len} " + f"and VAL_MAX_TOKENS={max_tokens}" + ) + 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) +def measure_counted_artifact( + script_path: Path, + compressed_model_path: str, + soft_budget_bytes: int, +) -> dict[str, int]: + train_gpt_utf8_bytes = len(script_path.read_text(encoding="utf-8").encode("utf-8")) + compressed_model_bytes = os.path.getsize(compressed_model_path) + counted_total_bytes = train_gpt_utf8_bytes + compressed_model_bytes + return { + "train_gpt_utf8_bytes": train_gpt_utf8_bytes, + "compressed_model_bytes": compressed_model_bytes, + "counted_total_bytes": counted_total_bytes, + "remaining_to_official_cap": OFFICIAL_CAP_BYTES - counted_total_bytes, + "remaining_to_soft_budget": soft_budget_bytes - counted_total_bytes, + } +# ----------------------------- +# 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,saliency", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +# ----------------------------- +# DATA LOADING +# ----------------------------- +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + should_refresh_cache = ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ) + if should_refresh_cache and torch.is_inference_mode_enabled(): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + cos = freqs.cos()[None, None, :, :] + sin = freqs.sin()[None, None, :, :] + return cos.to(dtype=dtype), sin.to(dtype=dtype) + if should_refresh_cache: + 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 SaliencyGuidance(nn.Module): + # Computes a bounded importance multiplier m_t from token-level, hashed bigram, + # and cheap local corrections. This softly biases attention without dropping tokens. + def __init__( + self, + dim: int, + vocab_size: int, + bigram_buckets: int, + use_token_prior: bool, + use_bigram: bool, + use_dynamic: bool, + use_phrase: bool, + use_attn_bias: bool, + use_value_scale: bool, + importance_min: float, + importance_max: float, + lambda_scale: float, + gamma_scale: float, + token_scale: float, + bigram_scale: float, + dynamic_scale: float, + phrase_scale: float, + smooth_weight: float, + ): + super().__init__() + self.use_token_prior = use_token_prior + self.use_bigram = use_bigram + self.use_dynamic = use_dynamic + self.use_phrase = use_phrase + self.use_attn_bias = use_attn_bias + self.use_value_scale = use_value_scale + if use_bigram and bigram_buckets < 2: + raise ValueError(f"SALIENCY_BIGRAM_BUCKETS must be >= 2, got {bigram_buckets}") + self.bigram_buckets = bigram_buckets + self.importance_min = importance_min + self.importance_max = importance_max + self.lambda_scale = lambda_scale + self.gamma_scale = gamma_scale + self.token_scale = token_scale + self.bigram_scale = bigram_scale + self.dynamic_scale = dynamic_scale + self.phrase_scale = phrase_scale + self.smooth_weight = smooth_weight + self.token_prior = nn.Parameter(torch.zeros(vocab_size, dtype=torch.float32)) + self.bigram_prior = nn.Parameter(torch.zeros(bigram_buckets, dtype=torch.float32)) + self.saliency_proj = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + self.saliency_bias = nn.Parameter(torch.zeros((), dtype=torch.float32)) + self.local_mix = nn.Parameter(torch.ones(3, dtype=torch.float32)) + def _causal_avg(self, x: Tensor, window: int) -> Tensor: + if window <= 1: + return x + return F.avg_pool1d(F.pad(x.unsqueeze(1), (window - 1, 0)), kernel_size=window, stride=1).squeeze(1) + def _bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_buckets - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, input_ids: Tensor, x: Tensor) -> tuple[Tensor | None, Tensor | None, Tensor]: + token_prior = torch.tanh(self.token_prior[input_ids]).float() if self.use_token_prior else torch.zeros_like(input_ids, dtype=torch.float32) + bigram_prior = ( + torch.tanh(self.bigram_prior[self._bigram_hash(input_ids)]).float() + if self.use_bigram + else torch.zeros_like(input_ids, dtype=torch.float32) + ) + dynamic_correction = torch.zeros_like(token_prior) + if self.use_dynamic: + dynamic_base = ( + F.rms_norm(x, (x.size(-1),)) * self.saliency_proj.to(dtype=x.dtype)[None, None, :] + ).sum(dim=-1).float() + mix = self.local_mix.float() + dynamic_local = ( + mix[0] * dynamic_base + + mix[1] * self._causal_avg(dynamic_base, 2) + + mix[2] * self._causal_avg(dynamic_base, 4) + ) / mix.abs().sum().clamp_min(1.0) + dynamic_correction = torch.tanh(dynamic_local + self.saliency_bias.float()) + phrase_saliency = torch.zeros_like(token_prior) + if self.use_phrase: + phrase_source = token_prior + bigram_prior + dynamic_correction + phrase_saliency = torch.tanh( + ( + self._causal_avg(phrase_source, 2) + + self._causal_avg(phrase_source, 3) + + self._causal_avg(phrase_source, 4) + ) + / 3.0 + ) + raw_saliency = ( + self.token_scale * token_prior + + self.bigram_scale * bigram_prior + + self.dynamic_scale * dynamic_correction + + self.phrase_scale * phrase_saliency + ) + importance = torch.clamp(torch.exp(raw_saliency), min=self.importance_min, max=self.importance_max) + attn_log_bias = self.lambda_scale * torch.log(importance) if self.use_attn_bias else None + value_multiplier = 1.0 + self.gamma_scale * (importance - 1.0) if self.use_value_scale else None + center_reg = (importance - 1.0).square().mean() + smooth_reg = ( + (importance[:, 1:] - importance[:, :-1]).square().mean() + if importance.size(1) > 1 + else center_reg.new_zeros(()) + ) + reg_loss = center_reg + self.smooth_weight * smooth_reg + bias_out = attn_log_bias[:, None, None, :].to(dtype=x.dtype) if attn_log_bias is not None else None + value_out = value_multiplier[:, None, :, None].to(dtype=x.dtype) if value_multiplier is not None else None + return bias_out, value_out, reg_loss +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) + self._attn_mask_seq_len_cached = 0 + self._attn_mask_cached: Tensor | None = None + def _causal_attn_mask(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> Tensor: + if ( + self._attn_mask_cached is None + or self._attn_mask_seq_len_cached != seq_len + or self._attn_mask_cached.device != device + ): + mask = torch.full((seq_len, seq_len), float("-inf"), device=device, dtype=torch.float32) + self._attn_mask_cached = torch.triu(mask, diagonal=1)[None, None, :, :] + self._attn_mask_seq_len_cached = seq_len + return self._attn_mask_cached.to(dtype=dtype) + def forward( + self, + x: Tensor, + saliency_log_bias: Tensor | None = None, + saliency_value_multiplier: Tensor | None = None, + ) -> 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] + if saliency_value_multiplier is not None: + v = v * saliency_value_multiplier.to(dtype=v.dtype) + attn_mask = None + is_causal = True + if saliency_log_bias is not None: + attn_mask = self._causal_attn_mask(seqlen, x.device, q.dtype) + saliency_log_bias.to(dtype=q.dtype) + is_causal = False + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=attn_mask, + is_causal=is_causal, + 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, + saliency_log_bias: Tensor | None = None, + saliency_value_multiplier: Tensor | None = None, + ) -> 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), saliency_log_bias, saliency_value_multiplier) + 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, + saliency_enable: bool, + saliency_token_prior: bool, + saliency_bigram: bool, + saliency_dynamic: bool, + saliency_phrase: bool, + saliency_attn_bias: bool, + saliency_value_scale: bool, + saliency_min: float, + saliency_max: float, + saliency_lambda: float, + saliency_gamma: float, + saliency_bigram_buckets: int, + saliency_token_scale: float, + saliency_bigram_scale: float, + saliency_dynamic_scale: float, + saliency_phrase_scale: float, + saliency_smooth_weight: 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.saliency_reg_loss: Tensor | None = None + 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.saliency = None + if saliency_enable: + self.saliency = SaliencyGuidance( + model_dim, + vocab_size, + saliency_bigram_buckets, + saliency_token_prior, + saliency_bigram, + saliency_dynamic, + saliency_phrase, + saliency_attn_bias, + saliency_value_scale, + saliency_min, + saliency_max, + saliency_lambda, + saliency_gamma, + saliency_token_scale, + saliency_bigram_scale, + saliency_dynamic_scale, + saliency_phrase_scale, + saliency_smooth_weight, + ) + 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 + saliency_log_bias = None + saliency_value_multiplier = None + self.saliency_reg_loss = x.new_zeros((), dtype=torch.float32) + if self.saliency is not None: + saliency_log_bias, saliency_value_multiplier, self.saliency_reg_loss = self.saliency(input_ids, x0) + 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, saliency_log_bias, saliency_value_multiplier) + 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, saliency_log_bias, saliency_value_multiplier) + 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() + if args.lr_schedule not in {"flat_warmdown", "cosine"}: + raise ValueError(f"Unsupported LR_SCHEDULE={args.lr_schedule}") + if args.lr_warmup_steps < 0: + raise ValueError(f"LR_WARMUP_STEPS must be >= 0, got {args.lr_warmup_steps}") + if not 0.0 < args.lr_decay_start_frac <= 1.0: + raise ValueError(f"LR_DECAY_START_FRAC must be in (0, 1], got {args.lr_decay_start_frac}") + if not 0.0 <= args.lr_min_scale <= 1.0: + raise ValueError(f"LR_MIN_SCALE must be in [0, 1], got {args.lr_min_scale}") + if args.use_compile: + 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(args.saliency_enable and args.saliency_attn_bias) + 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())}" + ) + if args.val_max_tokens > 0 and not args.local_proxy_eval: + raise ValueError( + "VAL_MAX_TOKENS is a local proxy-only evaluation shortcut. " + "Set LOCAL_PROXY_EVAL=1 to acknowledge non-challenge metrics." + ) + 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, args.val_max_tokens) + 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}") + if args.val_max_tokens > 0: + log0("local_proxy_eval:1 truncated_validation:enabled") + # ----------------------------- + # 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, + saliency_enable=args.saliency_enable, + saliency_token_prior=args.saliency_token_prior, + saliency_bigram=args.saliency_bigram, + saliency_dynamic=args.saliency_dynamic, + saliency_phrase=args.saliency_phrase, + saliency_attn_bias=args.saliency_attn_bias, + saliency_value_scale=args.saliency_value_scale, + saliency_min=args.saliency_min, + saliency_max=args.saliency_max, + saliency_lambda=args.saliency_lambda, + saliency_gamma=args.saliency_gamma, + saliency_bigram_buckets=args.saliency_bigram_buckets, + saliency_token_scale=args.saliency_token_scale, + saliency_bigram_scale=args.saliency_bigram_scale, + saliency_dynamic_scale=args.saliency_dynamic_scale, + saliency_phrase_scale=args.saliency_phrase_scale, + saliency_smooth_weight=args.saliency_smooth_weight, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + model_body: nn.Module = torch.compile(base_model, dynamic=False, fullgraph=True) if args.use_compile else base_model + model: nn.Module = DDP(model_body, device_ids=[local_rank], broadcast_buffers=False) if distributed else model_body + # 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( + f"sdp_backends:cudnn=False flash=True mem_efficient=False math:{int(args.saliency_enable and args.saliency_attn_bias)}" + ) + log0(f"use_compile:{int(args.use_compile)}") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + if args.saliency_enable: + log0( + "saliency:" + f"token_prior={int(args.saliency_token_prior)} bigram={int(args.saliency_bigram)} " + f"bigram_buckets={args.saliency_bigram_buckets} dynamic={int(args.saliency_dynamic)} " + f"phrase={int(args.saliency_phrase)} attn_bias={int(args.saliency_attn_bias)} " + f"value_scale={int(args.saliency_value_scale)} min={args.saliency_min:.3f} " + f"max={args.saliency_max:.3f} lambda={args.saliency_lambda:.3f} gamma={args.saliency_gamma:.3f} " + f"token_scale={args.saliency_token_scale:.3f} bigram_scale={args.saliency_bigram_scale:.3f} " + f"dynamic_scale={args.saliency_dynamic_scale:.3f} " + f"phrase_scale={args.saliency_phrase_scale:.3f} reg={args.saliency_reg_weight:.6f}" + ) + 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"local_proxy_eval:{int(args.local_proxy_eval)} val_max_tokens:{args.val_max_tokens} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0( + f"lr_schedule:{args.lr_schedule} lr_warmup_steps:{args.lr_warmup_steps} " + f"lr_decay_start_frac:{args.lr_decay_start_frac:.3f} lr_min_scale:{args.lr_min_scale:.3f}" + ) + log0(f"artifact_budget_bytes:{args.artifact_budget_bytes} official_cap_bytes:{OFFICIAL_CAP_BYTES}") + 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) + def add_saliency_regularizer(model_loss: Tensor) -> Tensor: + if not args.saliency_enable or args.saliency_reg_weight <= 0.0: + return model_loss + reg_loss = base_model.saliency_reg_loss + if reg_loss is None: + return model_loss + return model_loss + args.saliency_reg_weight * reg_loss + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def budget_progress(step: int, elapsed_ms: float) -> float: + if max_wallclock_ms is not None: + return min(elapsed_ms / max(max_wallclock_ms, 1e-9), 1.0) + return min(step / max(args.iterations, 1), 1.0) + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.lr_schedule == "cosine": + if args.lr_warmup_steps > 0 and step < args.lr_warmup_steps: + return (step + 1) / args.lr_warmup_steps + progress = budget_progress(step, elapsed_ms) + if progress <= args.lr_decay_start_frac: + return 1.0 + decay_progress = min( + max((progress - args.lr_decay_start_frac) / max(1.0 - args.lr_decay_start_frac, 1e-9), 0.0), + 1.0, + ) + cosine = 0.5 * (1.0 + math.cos(math.pi * decay_progress)) + return args.lr_min_scale + (1.0 - args.lr_min_scale) * cosine + 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 = add_saliency_regularizer(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() + val_metric_prefix = "proxy_val" if args.val_max_tokens > 0 else "val" + roundtrip_prefix = "final_int8_zlib_roundtrip_proxy" if args.val_max_tokens > 0 else "final_int8_zlib_roundtrip" + roundtrip_exact_prefix = ( + "final_int8_zlib_roundtrip_proxy_exact" if args.val_max_tokens > 0 else "final_int8_zlib_roundtrip_exact" + ) + 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_metric_prefix}_loss:{val_loss:.4f} " + f"{val_metric_prefix}_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): + model_loss = model(x, y) + loss = add_saliency_regularizer(model_loss) + train_loss += model_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"{roundtrip_prefix} {val_metric_prefix}_loss:{q_val_loss:.4f} {val_metric_prefix}_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"{roundtrip_exact_prefix} {val_metric_prefix}_loss:{q_val_loss:.8f} {val_metric_prefix}_bpb:{q_val_bpb:.8f}") + size_report = measure_counted_artifact(Path(__file__), "final_model.int8.ptz", args.artifact_budget_bytes) + log0(f"train_gpt_utf8_bytes:{size_report['train_gpt_utf8_bytes']}") + log0(f"compressed_model_bytes:{size_report['compressed_model_bytes']}") + log0(f"counted_total_bytes:{size_report['counted_total_bytes']}") + log0(f"remaining_to_official_cap:{size_report['remaining_to_official_cap']}") + log0(f"remaining_to_soft_budget:{size_report['remaining_to_soft_budget']}") + official_cap_ok = size_report["counted_total_bytes"] < OFFICIAL_CAP_BYTES + soft_budget_ok = size_report["counted_total_bytes"] < args.artifact_budget_bytes + log0(f"official_cap_check:{'PASS' if official_cap_ok else 'FAIL'} cap_bytes:{OFFICIAL_CAP_BYTES}") + log0(f"soft_budget_check:{'PASS' if soft_budget_ok else 'WARN'} budget_bytes:{args.artifact_budget_bytes}") + if not soft_budget_ok: + log0("WARNING: counted artifact is at or above ARTIFACT_BUDGET_BYTES soft budget.") + if distributed: + dist.destroy_process_group() + if master_process and not official_cap_ok: + raise RuntimeError( + f"Counted artifact exceeds official cap: {size_report['counted_total_bytes']} >= {OFFICIAL_CAP_BYTES}" + ) +if __name__ == "__main__": + main()