diff --git a/README.md b/README.md index 9c645dc4e..8d7422aa2 100644 --- a/README.md +++ b/README.md @@ -34,7 +34,7 @@ Happy training! | Muon WD + 10 layer | 1.1748 | notapplica | Includes prev. wins + Spectral embed init + resid mix | 2026-03-19 | [info](records/track_10min_16mb/2026-03-19_SlidingWindow_FP16Emb_10L_MuonWD_OvertoneInit/README.md) | | Sliding Window Eval | 1.1925 | Matthew Li | Sliding window evaluation at stride=64, increasing context for eval | 2026-03-19 | [info](records/track_10min_16mb/2026-03-19_SlidingWindowEval/README.md) | | Lora TTT | 1.1928 | samacqua | Test-time training with LORAs | 2026-03-19 | [info](records/track_10min_16mb/2026-03-17_LoRA_TTT/README.md) | -| 4k seq length| 1.2014 | Spokane Way | 4k seq length + better hypers | 2026-03-19 | [info](records/track_10min_16mb/2026-03-18_LongContextSeq2048/README.md) | +| 4k seq length| 1.2014 | Spokane Way | 4k seq length + better hypers | 2026-03-19 | [info](records/track_10min_16mb/2026-03-19_TrainingOptSeq4096/README.md) | | 2048 seq length | 1.206 | Spokane Way | 2048 seq length (train + val) | 2026-03-18 | [info](records/track_10min_16mb/2026-03-18_LongContextSeq2048/README.md) | | int6 mixed precision | 1.2147 | Nan Liu | 10 layers, mixed int8/int6 | 2026-03-18 | [info](records/track_10min_16mb/2026-03-19_10L_MixedPrecision/README.md) | | fp16 Embed | 1.2197 | Renier Velazco | FP16 Tied Embedding + LR/Warmdown Tuning | 2026-03-18 | [info](records/track_10min_16mb/2026-03-18_FP16Embed_WD3600/README.md) | diff --git a/modal_repro_longcontext.py b/modal_repro_longcontext.py new file mode 100644 index 000000000..78a55daa9 --- /dev/null +++ b/modal_repro_longcontext.py @@ -0,0 +1,220 @@ +from __future__ import annotations + +import json +import os +import re +import subprocess + +import modal + +APP_NAME = "parameter-golf-repro-longcontext-8h100" +REPO_REMOTE_PATH = "/workspace/parameter-golf" +TARGET_SCRIPT = "records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_gpt.py" + +STEP_RE = re.compile(r"step:(\d+)/(\d+).*step_avg:([0-9.]+)ms") +CONFIG_RE = re.compile( + r"train_batch_tokens:(\d+)\s+train_seq_len:(\d+)\s+iterations:(\d+)\s+warmup_steps:(\d+)\s+" + r"max_wallclock_seconds:([0-9.]+)" +) +FINAL_RE = re.compile(r"final_int8_zlib_roundtrip_exact\s+val_loss:([0-9.]+)\s+val_bpb:([0-9.]+)") +SIZE_RE = re.compile(r"Total submission size int8\+zlib:\s*([0-9]+)\s*bytes") +STOP_RE = re.compile(r"stopping_early: wallclock_cap .* step:([0-9]+)/([0-9]+)") + +app = modal.App(APP_NAME) +image = ( + # The devel image includes the CUDA toolchain pieces that torch.compile / Triton + # tend to expect on tuned boxes; the runtime image is more likely to underperform. + modal.Image.from_registry("pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel") + .apt_install("build-essential") + .pip_install( + "numpy", + "tqdm", + "huggingface-hub", + "kernels", + "setuptools", + "typing-extensions==4.15.0", + "datasets", + "tiktoken", + "sentencepiece", + ) + .add_local_dir(".", remote_path=REPO_REMOTE_PATH) +) + + +def _run_checked(cmd: list[str], *, env: dict[str, str] | None = None) -> None: + subprocess.run(cmd, check=True, env=env) + + +def _print_probe(command: list[str]) -> None: + result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, check=False) + joined = " ".join(command) + print(f"\n=== PROBE: {joined} ===") + print(result.stdout.rstrip()) + + +@app.function(image=image, gpu="H100:8", timeout=90 * 60, cpu=64, memory=196608) +def run( + run_id: str = "modal_longcontext_8h100_repro", + target_script: str = TARGET_SCRIPT, + data_variant: str = "sp1024", + max_wallclock_seconds: int = 600, + expected_train_seq_len: int = 2048, + expected_max_step_avg_ms: float = 60.0, + gate_check_step: int = 1000, + enable_throughput_gate: bool = False, + nccl_ib_disable: int | None = None, + extra_env_json: str = "{}", +) -> dict[str, object]: + os.chdir(REPO_REMOTE_PATH) + + _run_checked(["python", "data/cached_challenge_fineweb.py", "--variant", data_variant]) + + env = os.environ.copy() + env.update( + { + "RUN_ID": run_id, + "DATA_PATH": f"./data/datasets/fineweb10B_{data_variant}", + "TOKENIZER_PATH": "./data/tokenizers/fineweb_1024_bpe.model", + "VOCAB_SIZE": "1024", + "MAX_WALLCLOCK_SECONDS": str(max_wallclock_seconds), + "OMP_NUM_THREADS": "1", + "TORCH_NCCL_ASYNC_ERROR_HANDLING": "1", + "CC": "gcc", + "CXX": "g++", + } + ) + if nccl_ib_disable is None: + env.pop("NCCL_IB_DISABLE", None) + else: + env["NCCL_IB_DISABLE"] = str(int(nccl_ib_disable)) + + extra_env = json.loads(extra_env_json) + if not isinstance(extra_env, dict): + raise TypeError("extra_env_json must decode to a JSON object") + env.update({str(k): str(v) for k, v in extra_env.items()}) + + _print_probe(["python", "-c", "import torch; print(torch.__version__)"]) + _print_probe(["python", "-c", "import triton; print(triton.__version__)"]) + _print_probe(["bash", "-lc", "command -v ptxas || true"]) + _print_probe(["nvidia-smi", "topo", "-m"]) + _print_probe(["python", "-c", "import os; print(os.cpu_count())"]) + + print("\n=== REPRO ENV ===") + print( + { + key: env[key] + for key in ( + "RUN_ID", + "DATA_PATH", + "TOKENIZER_PATH", + "VOCAB_SIZE", + "MAX_WALLCLOCK_SECONDS", + "OMP_NUM_THREADS", + "TORCH_NCCL_ASYNC_ERROR_HANDLING", + "NCCL_IB_DISABLE", + "CC", + "CXX", + ) + if key in env + } + ) + + cmd = ["torchrun", "--standalone", "--nproc_per_node=8", target_script] + proc = subprocess.Popen(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1) + + lines: list[str] = [] + last_step = 0 + last_step_avg_ms = 0.0 + observed_train_seq_len: int | None = None + gate_checked = False + + assert proc.stdout is not None + for line in proc.stdout: + print(line, end="") + lines.append(line) + + step_match = STEP_RE.search(line) + if step_match: + last_step = int(step_match.group(1)) + last_step_avg_ms = float(step_match.group(3)) + if enable_throughput_gate and (not gate_checked) and last_step >= gate_check_step: + gate_checked = True + if last_step_avg_ms > expected_max_step_avg_ms: + proc.terminate() + try: + proc.wait(timeout=30) + except subprocess.TimeoutExpired: + proc.kill() + raise RuntimeError( + "Throughput gate failed: " + f"step={last_step} step_avg_ms={last_step_avg_ms:.2f} " + f"threshold_ms={expected_max_step_avg_ms:.2f}" + ) + + config_match = CONFIG_RE.search(line) + if config_match: + observed_train_seq_len = int(config_match.group(2)) + + rc = proc.wait() + if rc != 0: + raise RuntimeError(f"Training failed with exit code {rc}") + + if observed_train_seq_len is None: + raise RuntimeError("Could not parse train_seq_len from training log") + if observed_train_seq_len != expected_train_seq_len: + raise RuntimeError( + f"Unexpected TRAIN_SEQ_LEN in log: expected {expected_train_seq_len}, got {observed_train_seq_len}" + ) + if enable_throughput_gate and not gate_checked: + raise RuntimeError(f"Throughput gate was enabled but log never reached step {gate_check_step}") + + log = "".join(lines) + exact = FINAL_RE.search(log) + size = SIZE_RE.search(log) + stop = STOP_RE.search(log) + + out = { + "run_id": run_id, + "target_script": target_script, + "observed_train_seq_len": observed_train_seq_len, + "last_step_seen": last_step, + "last_step_avg_ms": last_step_avg_ms, + "val_loss": float(exact.group(1)) if exact else None, + "val_bpb": float(exact.group(2)) if exact else None, + "bytes_total_int8_zlib": int(size.group(1)) if size else None, + "steps_done": int(stop.group(1)) if stop else None, + "steps_target": int(stop.group(2)) if stop else None, + "nccl_ib_disable": env.get("NCCL_IB_DISABLE"), + } + print("\n=== REPRO SUMMARY ===") + print(out) + return out + + +@app.local_entrypoint() +def main( + run_id: str = "modal_longcontext_8h100_repro", + target_script: str = TARGET_SCRIPT, + data_variant: str = "sp1024", + max_wallclock_seconds: int = 600, + expected_train_seq_len: int = 2048, + expected_max_step_avg_ms: float = 60.0, + gate_check_step: int = 1000, + enable_throughput_gate: bool = False, + nccl_ib_disable: int | None = None, + extra_env_json: str = "{}", +) -> None: + print( + run.remote( + run_id=run_id, + target_script=target_script, + data_variant=data_variant, + max_wallclock_seconds=max_wallclock_seconds, + expected_train_seq_len=expected_train_seq_len, + expected_max_step_avg_ms=expected_max_step_avg_ms, + gate_check_step=gate_check_step, + enable_throughput_gate=enable_throughput_gate, + nccl_ib_disable=nccl_ib_disable, + extra_env_json=extra_env_json, + ) + ) diff --git a/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/README.md b/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/README.md new file mode 100644 index 000000000..386e420b6 --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/README.md @@ -0,0 +1,63 @@ +# WIP Placeholder Submission + +This is a working submission scaffold for the 10-minute / 16MB track. +Rename this folder later once you have final results. + +## Goal + +- Track: `track_10min_16mb` +- Objective: improve `val_bpb` while staying under the 16,000,000-byte artifact cap +- Budget: reproducible <= 10 minutes training on 8xH100 (SXM) + +## Current Status + +- Status: work in progress +- Baseline script source: root `train_gpt.py` copied into this folder +- Final metrics: pending + +## Planned Changes + +- [ ] Model/optimizer changes +- [ ] Data/tokenizer changes (if any) +- [ ] Eval method changes (if any) +- [ ] Compression/export changes (if any) + +## Run Command (Template) + +```bash +RUN_ID=wip_placeholder \ +DATA_PATH=./data/datasets/fineweb10B_sp1024 \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 \ +MAX_WALLCLOCK_SECONDS=600 \ +TRAIN_LOG_EVERY=200 \ +VAL_LOSS_EVERY=1000 \ +torchrun --standalone --nproc_per_node=8 \ + records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/train_gpt.py | tee records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/train.log +``` + +## Required Files Checklist + +- [x] `train_gpt.py` +- [ ] `train.log` (generate after run) +- [x] `README.md` +- [x] `submission.json` +- [ ] extra seed logs for SOTA significance (if needed) + +## Results (Fill In) + +Primary run: +- seed: +- steps reached in 600s: +- pre-quant `val_bpb`: +- post-quant `val_bpb` (`final_int8_zlib_roundtrip_exact`): +- `Total submission size int8+zlib`: + +Extra reproducibility runs (if claiming SOTA): +- `train_seedXXXX.log`: +- `train_seedYYYY.log`: + +## Notes + +- Keep logs and code in this folder so the PR is self-contained. +- If tokenizer or dataset changes are made, include proof that `val_bpb` is computed correctly. diff --git a/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/submission.json b/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/submission.json new file mode 100644 index 000000000..33e2dcc10 --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/submission.json @@ -0,0 +1,11 @@ +{ + "author": "YOUR_NAME", + "github_id": "YOUR_GITHUB_ID", + "name": "WIP Placeholder Submission", + "blurb": "Short summary of your approach and what changed from baseline.", + "date": "2026-03-19", + "val_loss": 0.0, + "val_bpb": 0.0, + "bytes_total": 0, + "bytes_code": 0 +} diff --git a/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/train_gpt.py b/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/train_gpt.py new file mode 100644 index 000000000..651beb2b8 --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER/train_gpt.py @@ -0,0 +1,1126 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Scale correction from Muon reference implementations. + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/README.md b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/README.md new file mode 100644 index 000000000..577c8b132 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/README.md @@ -0,0 +1,56 @@ +# Auto Precision Budget 10L (1xH100 exploratory) + +This folder is a small non-record experiment built on top of the `2026-03-20_10L_Int5MLP_MuonWD04_SWA50` recipe. + +The motivation is simple: the current best compressed models already rely on a hand-designed precision budget, with a few tensors kept at higher precision and the rest quantized more aggressively. That works, but it is still mostly heuristic. I wanted to try a more systematic version of the same idea by measuring which tensors are actually most sensitive to quantization and then spending bytes there. + +The core change in `train_gpt.py` is a calibration-driven precision allocator. After training, SWA, and pruning, the script: +- starts from the default mixed-precision export policy, +- evaluates a small set of candidate tensor promotions, +- measures the calibration impact after roundtripping through the compressed export path, and +- greedily accepts promotions that give the best improvement per added byte while staying under the 16,000,000-byte limit. + +The current candidate set includes: +- `tok_emb.weight` +- `bigram.proj.weight` +- `bigram.embed.weight` +- late-layer attention `c_k`, `c_v`, and `c_proj` weights + +I also added a rank-0-only calibration path for distributed runs so the same method can be reused in future 8xH100 experiments without repeatedly doing distributed calibration collectives. + +## Submitted run + +This submission is intentionally modest. It is a cheap 1xH100 Modal smoke run, not a leaderboard attempt. + +Configuration: +- GPU: `1xH100` +- Training data: `--train-shards 1` +- `MAX_WALLCLOCK_SECONDS=60` +- `ITERATIONS=150` +- `AUTO_CALIBRATION_WINDOWS=16` +- `FINAL_EVAL_MAX_WINDOWS=16` + +Because the final evaluation is capped to 16 sliding windows, the reported number is a smoke metric rather than a full-validation metric. I am submitting it as a concrete, working non-record experiment that motivates further runs, not as a strong score claim. + +## Result + +From `train.log`: +- training stopped at `87/150` steps because of the 60-second wallclock cap +- final exact metric: `val_loss:5.53668879`, `val_bpb:3.08435975` +- compressed model bytes: `15,771,560` +- total submission bytes: `15,836,818` +- selected promotions: `blocks.9.attn.c_k.weight`, `blocks.9.attn.c_v.weight` + +In earlier paired smoke testing on the same idea, the allocator also showed a small improvement over a fixed export policy, which is why I think this direction is still worth pursuing. + +## Why I think this is worth exploring + +Even if the gain here is small, I think the idea is useful for longer-term work: +- strong current recipes already depend on manual mixed-precision choices, +- a sensitivity-driven allocator should transfer better across future architecture changes, +- and even negative results help identify which tensors are real quantization bottlenecks versus harmless heuristics. + +Included files: +- `train_gpt.py` +- `train.log` +- `submission.json` diff --git a/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/submission.json b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/submission.json new file mode 100644 index 000000000..925041955 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/submission.json @@ -0,0 +1,15 @@ +{ + "author": "Sanjay Patnala", + "github_id": "spatnala18", + "name": "Auto Precision Budget 10L (1xH100 exploratory)", + "blurb": "Non-record 1xH100 exploratory run built on the 2026-03-20 10L recipe. Replaces hand-authored mixed-precision exceptions with a calibration-driven precision allocator that greedily promotes quantization-sensitive tensors under the 16MB cap. This submitted run is a short smoke experiment intended to validate the method and motivate longer follow-up experiments.", + "date": "2026-03-22", + "track": "non-record-16mb", + "val_loss": 5.53668879, + "val_bpb": 3.08435975, + "step_stop": 87, + "wallclock_seconds": 60.025, + "bytes_total": 15836818, + "bytes_model_int8_zlib": 15771560, + "bytes_code": 65258 +} diff --git a/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/train.log b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/train.log new file mode 100644 index 000000000..4f41b693e --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/train.log @@ -0,0 +1,111 @@ +logs/modal_autoprecision_nonrecord_1xh100_v1.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:1 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:1 grad_accum_steps:8 +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.03 matrix_lr:0.02 scalar_lr:0.02 +train_batch_tokens:786432 train_seq_len:2048 iterations:150 warmup_steps:20 max_wallclock_seconds:60.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:1/150 train_loss:6.9334 train_time:822ms step_avg:822.24ms +step:2/150 train_loss:8.1796 train_time:1509ms step_avg:754.37ms +step:3/150 train_loss:8.1744 train_time:2194ms step_avg:731.35ms +step:4/150 train_loss:8.1194 train_time:2882ms step_avg:720.40ms +step:5/150 train_loss:8.0174 train_time:3571ms step_avg:714.21ms +step:6/150 train_loss:7.9515 train_time:4260ms step_avg:709.96ms +step:7/150 train_loss:7.8315 train_time:4949ms step_avg:707.07ms +step:8/150 train_loss:7.7282 train_time:5634ms step_avg:704.30ms +step:9/150 train_loss:7.5965 train_time:6324ms step_avg:702.64ms +step:10/150 train_loss:7.4754 train_time:7010ms step_avg:700.98ms +swa:start step:50 +step:87/150 val_loss:5.5614 val_bpb:3.2938 train_time:60025ms step_avg:689.94ms +stopping_early: wallclock_cap train_time:60025ms step:87/150 +peak memory allocated: 19146 MiB reserved: 19218 MiB +Serialized model: 98437419 bytes +Code size: 65258 bytes +Total submission size: 98502677 bytes +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084380 +auto_precision:baseline calib_loss:5.536725 calib_bpb:3.084380 bytes:15311264 +auto_precision:skip name:tok_emb.weight to:fp16 reason:bytes +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084405 +auto_precision:candidate name:bigram.proj.weight to:fp16 calib_bpb:3.084405 delta:-0.000026 bytes:15563919 added:+252655 +auto_precision:skip name:bigram.embed.weight to:fp16 reason:bytes +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084367 +auto_precision:candidate name:blocks.6.attn.c_k.weight to:int8 calib_bpb:3.084367 delta:+0.000013 bytes:15578627 added:+267363 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084387 +auto_precision:candidate name:blocks.6.attn.c_v.weight to:int8 calib_bpb:3.084387 delta:-0.000007 bytes:15379754 added:+68490 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084398 +auto_precision:candidate name:blocks.7.attn.c_k.weight to:int8 calib_bpb:3.084398 delta:-0.000018 bytes:15601520 added:+290256 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084509 +auto_precision:candidate name:blocks.7.attn.c_v.weight to:int8 calib_bpb:3.084509 delta:-0.000129 bytes:15551702 added:+240438 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084394 +auto_precision:candidate name:blocks.8.attn.c_k.weight to:int8 calib_bpb:3.084394 delta:-0.000014 bytes:15313844 added:+2580 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084388 +auto_precision:candidate name:blocks.8.attn.c_v.weight to:int8 calib_bpb:3.084388 delta:-0.000008 bytes:15595099 added:+283835 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084368 +auto_precision:candidate name:blocks.9.attn.c_k.weight to:int8 calib_bpb:3.084368 delta:+0.000012 bytes:15358594 added:+47330 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084384 +auto_precision:candidate name:blocks.9.attn.c_v.weight to:int8 calib_bpb:3.084384 delta:-0.000004 bytes:15641180 added:+329916 +auto_precision:selected name:blocks.9.attn.c_k.weight to:int8 calib_bpb:3.084368 bytes:15358594 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084363 +auto_precision:candidate name:blocks.6.attn.c_k.weight to:int8 calib_bpb:3.084363 delta:+0.000005 bytes:15703326 added:+344732 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084399 +auto_precision:candidate name:blocks.6.attn.c_v.weight to:int8 calib_bpb:3.084399 delta:-0.000031 bytes:15504375 added:+145781 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084378 +auto_precision:candidate name:blocks.7.attn.c_k.weight to:int8 calib_bpb:3.084378 delta:-0.000010 bytes:15532550 added:+173956 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084493 +auto_precision:candidate name:blocks.7.attn.c_v.weight to:int8 calib_bpb:3.084493 delta:-0.000125 bytes:15505742 added:+147148 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084380 +auto_precision:candidate name:blocks.8.attn.c_k.weight to:int8 calib_bpb:3.084380 delta:-0.000012 bytes:15504094 added:+145500 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084378 +auto_precision:candidate name:blocks.8.attn.c_v.weight to:int8 calib_bpb:3.084378 delta:-0.000010 bytes:15610359 added:+251765 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084360 +auto_precision:candidate name:blocks.9.attn.c_v.weight to:int8 calib_bpb:3.084360 delta:+0.000008 bytes:15836818 added:+478224 +auto_precision:selected name:blocks.9.attn.c_v.weight to:int8 calib_bpb:3.084360 bytes:15836818 +auto_precision:sync_path logs/modal_autoprecision_nonrecord_1xh100_v1_auto_precision_policy.pt +auto_precision:final calib_bpb:3.084360 bytes:15836818 fp16:[] int8:[blocks.9.attn.c_k.weight,blocks.9.attn.c_v.weight] +Serialized model mixed+zstd: 15771560 bytes +Total submission size int8+zlib: 15836818 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 max_windows:16 +sliding_eval_config rank:0 world_size:1 total_windows:16 local_windows:16 max_windows:16 + sliding_eval [100.0%] 16/16 windows running_bpb=3.084360 +final_int8_zlib_roundtrip val_loss:5.5367 val_bpb:3.0844 eval_time:139ms +final_int8_zlib_roundtrip_exact val_loss:5.53668879 val_bpb:3.08435975 diff --git a/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/train_gpt.py b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/train_gpt.py new file mode 100644 index 000000000..ed90bb225 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_AutoPrecisionBudget_10L_1xH100/train_gpt.py @@ -0,0 +1,1531 @@ +""" +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 + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 42)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 10)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.04)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 32)) + final_eval_max_windows = int(os.environ.get("FINAL_EVAL_MAX_WINDOWS", 0)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 10240)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + auto_precision_policy = bool(int(os.environ.get("AUTO_PRECISION_POLICY", "1"))) + auto_calibration_windows = int(os.environ.get("AUTO_CALIBRATION_WINDOWS", 128)) + auto_calibration_rank0_only = bool(int(os.environ.get("AUTO_CALIBRATION_RANK0_ONLY", "1"))) + auto_fp16_topk = int(os.environ.get("AUTO_FP16_TOPK", 1)) + auto_int8_topk = int(os.environ.get("AUTO_INT8_TOPK", 2)) + auto_max_total_bytes = int(os.environ.get("AUTO_MAX_TOTAL_BYTES", 15_960_000)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + 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, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov, weight_decay=weight_decay), + ) + + @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) + 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) + + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + 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 (INT8 legacy + INT6 mixed) +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,bigram.scale", + ).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 quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if "bigram" in name: + return "bigram" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip_range).clamp_min(1e-12).to(torch.float16) + scale = scale.clamp_min(torch.finfo(torch.float16).tiny) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -(clip_range+1), clip_range).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range+1), clip_range).to(torch.int8) + return q, scale + +def default_precision_for_name(name: str, tensor: Tensor) -> str: + if not tensor.is_floating_point() or tensor.numel() <= 8192: + return "passthrough" + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + return "passthrough_ctrl" + cat = _classify_param(name) + if cat == "mlp" and tensor.ndim >= 1: + return "int5" + if cat in {"attn", "bigram"} and tensor.ndim >= 1: + return "int6" + return "int8" + + +def build_default_precision_policy(state_dict: dict[str, Tensor]) -> dict[str, str]: + return {name: default_precision_for_name(name, tensor) for name, tensor in state_dict.items()} + + +def build_candidate_promotions(args: Hyperparameters, state_dict: dict[str, Tensor]) -> dict[str, list[str]]: + fp16_candidates: list[str] = [] + for name in ("tok_emb.weight", "bigram.proj.weight", "bigram.embed.weight", "lm_head.weight"): + if name in state_dict and state_dict[name].is_floating_point(): + fp16_candidates.append(name) + + int8_candidates: list[str] = [] + start_block = max(args.num_layers - 4, 0) + for block_idx in range(start_block, args.num_layers): + for suffix in ("c_k.weight", "c_v.weight", "c_proj.weight"): + name = f"blocks.{block_idx}.attn.{suffix}" + if name in state_dict and state_dict[name].is_floating_point(): + int8_candidates.append(name) + return {"fp16": fp16_candidates, "int8": int8_candidates} + + +def quantize_state_dict_with_policy(state_dict: dict[str, Tensor], policy: dict[str, str]): + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + precision = policy.get(name, default_precision_for_name(name, t)) + if precision == "passthrough": + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if precision == "passthrough_ctrl": + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if precision == "fp16": + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if precision in {"int5", "int6"} and t.ndim >= 1: + clip = 15 if precision == "int5" else 31 + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": precision} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + + +def serialize_quantized_state(quant_result: dict[str, Tensor], quant_meta: dict[str, object]) -> bytes: + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + return zstandard.ZstdCompressor(level=22).compress(quant_raw) + return zlib.compress(quant_raw, 9) + + +def total_submission_bytes_from_quantized(quant_blob: bytes, code_bytes: int) -> int: + return len(quant_blob) + code_bytes + + +def dequantize_mixed_precision(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta[name] + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0): + 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): + def __init__(self, dim: int, mlp_mult: float): + super().__init__() + hidden = int(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 SmearGate(nn.Module): + """Blend each token's embedding with the previous token's embedding.""" + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + + +class BigramHashEmbedding(nn.Module): + """Hash consecutive token pairs into a learned embedding table.""" + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, 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: float, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + ): + 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.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + 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.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList( + [ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ 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) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + 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") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + 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) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + max_windows: int | None = None, + reduce_distributed: bool = True, +) -> tuple[float, float]: + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + if max_windows is not None: + window_starts = window_starts[:max_windows] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + if rank == 0: + print( + f"sliding_eval_config rank:{rank} world_size:{world_size} " + f"total_windows:{total_windows} local_windows:{len(my_windows)} max_windows:{max_windows}", + flush=True, + ) + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if rank == 0 and my_windows and (bi // batch_seqs) % 50 == 0: + done = min(bi + batch_seqs, len(my_windows)) + pct = done / len(my_windows) * 100 + running_bpb = 0.0 + if token_count.item() > 0: + rl = (loss_sum / token_count).item() + running_bpb = rl / math.log(2.0) * (token_count.item() / byte_count.item()) + print(f" sliding_eval [{pct:5.1f}%] {done}/{len(my_windows)} windows running_bpb={running_bpb:.6f}", flush=True) + + if reduce_distributed and dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + if token_count.item() == 0 or byte_count.item() == 0: + raise RuntimeError("Sliding-window evaluation saw zero calibration windows; increase AUTO_CALIBRATION_WINDOWS.") + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def evaluate_quantized_policy( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + state_dict: dict[str, Tensor], + policy: dict[str, str], + code_bytes: int, + eval_rank: int, + eval_world_size: int, + reduce_distributed: bool, +) -> dict[str, float] | None: + quant_result, quant_meta = quantize_state_dict_with_policy(state_dict, policy) + quant_blob = serialize_quantized_state(quant_result, quant_meta) + total_bytes = total_submission_bytes_from_quantized(quant_blob, code_bytes) + if total_bytes > args.auto_max_total_bytes: + return None + deq_state = dequantize_mixed_precision(quant_result, quant_meta, state_dict) + base_model.load_state_dict(deq_state, strict=True) + val_loss, val_bpb = eval_val_sliding( + args, + base_model, + eval_rank, + eval_world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + stride=args.eval_stride, + batch_seqs=args.eval_batch_seqs, + max_windows=args.auto_calibration_windows, + reduce_distributed=reduce_distributed, + ) + return { + "val_loss": float(val_loss), + "val_bpb": float(val_bpb), + "total_bytes": float(total_bytes), + "quant_bytes": float(len(quant_blob)), + } + + +def select_precision_policy( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + state_dict: dict[str, Tensor], + code_bytes: int, + log0, +) -> tuple[dict[str, str], dict[str, float]]: + distributed_eval = dist.is_available() and dist.is_initialized() + default_policy = build_default_precision_policy(state_dict) + + def run_policy_search( + eval_rank: int, + eval_world_size: int, + reduce_distributed: bool, + ) -> tuple[dict[str, str], dict[str, float]]: + if not args.auto_precision_policy: + metrics = evaluate_quantized_policy( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + state_dict, + default_policy, + code_bytes, + eval_rank, + eval_world_size, + reduce_distributed, + ) + if metrics is None: + raise RuntimeError("Default precision policy exceeds AUTO_MAX_TOTAL_BYTES.") + base_model.load_state_dict(state_dict, strict=True) + return default_policy, metrics + + candidates = build_candidate_promotions(args, state_dict) + current_policy = dict(default_policy) + current_metrics = evaluate_quantized_policy( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + state_dict, + current_policy, + code_bytes, + eval_rank, + eval_world_size, + reduce_distributed, + ) + if current_metrics is None: + raise RuntimeError("Default precision policy exceeds AUTO_MAX_TOTAL_BYTES.") + if not (math.isfinite(current_metrics["val_bpb"]) and 0.0 < current_metrics["val_bpb"] < 10.0): + raise RuntimeError(f"Default precision policy produced invalid calibration bpb: {current_metrics['val_bpb']}") + log0( + "auto_precision:baseline " + f"calib_loss:{current_metrics['val_loss']:.6f} " + f"calib_bpb:{current_metrics['val_bpb']:.6f} " + f"bytes:{int(current_metrics['total_bytes'])}" + ) + + for target_precision, topk in (("fp16", args.auto_fp16_topk), ("int8", args.auto_int8_topk)): + remaining = [name for name in candidates[target_precision] if current_policy.get(name) != target_precision] + for _ in range(topk): + best: dict[str, object] | None = None + for name in remaining: + trial_policy = dict(current_policy) + trial_policy[name] = target_precision + trial_metrics = evaluate_quantized_policy( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + state_dict, + trial_policy, + code_bytes, + eval_rank, + eval_world_size, + reduce_distributed, + ) + if trial_metrics is None: + log0(f"auto_precision:skip name:{name} to:{target_precision} reason:bytes") + continue + if not (math.isfinite(trial_metrics["val_bpb"]) and 0.0 < trial_metrics["val_bpb"] < 10.0): + log0(f"auto_precision:skip name:{name} to:{target_precision} reason:invalid_metric") + continue + gain = current_metrics["val_bpb"] - trial_metrics["val_bpb"] + added_bytes = int(trial_metrics["total_bytes"] - current_metrics["total_bytes"]) + score = gain / max(added_bytes, 1) if added_bytes > 0 else gain + 1e-9 + log0( + "auto_precision:candidate " + f"name:{name} to:{target_precision} " + f"calib_bpb:{trial_metrics['val_bpb']:.6f} delta:{gain:+.6f} " + f"bytes:{int(trial_metrics['total_bytes'])} added:{added_bytes:+d}" + ) + if gain <= 0.0: + continue + if best is None or score > best["score"]: + best = { + "name": name, + "policy": trial_policy, + "metrics": trial_metrics, + "score": score, + } + if best is None: + break + chosen_name = str(best["name"]) + current_policy = dict(best["policy"]) + current_metrics = dict(best["metrics"]) + remaining = [name for name in remaining if name != chosen_name] + log0( + "auto_precision:selected " + f"name:{chosen_name} to:{target_precision} " + f"calib_bpb:{current_metrics['val_bpb']:.6f} " + f"bytes:{int(current_metrics['total_bytes'])}" + ) + + base_model.load_state_dict(state_dict, strict=True) + return current_policy, current_metrics + + if distributed_eval and args.auto_calibration_rank0_only: + policy_sync_path = Path("logs") / f"{args.run_id}_auto_precision_policy.pt" + if rank == 0 and policy_sync_path.exists(): + policy_sync_path.unlink() + if distributed_eval: + dist.barrier() + if rank == 0: + selected_policy, current_metrics = run_policy_search(0, 1, False) + torch.save({"policy": selected_policy, "metrics": current_metrics}, policy_sync_path) + log0(f"auto_precision:sync_path {policy_sync_path}") + else: + deadline = time.time() + 1800.0 + while not policy_sync_path.exists(): + if time.time() > deadline: + raise RuntimeError(f"Timed out waiting for rank0 auto precision policy at {policy_sync_path}") + time.sleep(1.0) + payload = torch.load(policy_sync_path, map_location="cpu") + selected_policy = dict(payload["policy"]) + current_metrics = dict(payload["metrics"]) + if distributed_eval: + dist.barrier() + base_model.load_state_dict(state_dict, strict=True) + return selected_policy, current_metrics + + return run_policy_search(rank, world_size, True) + + +# ----------------------------- +# 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 = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + 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, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + ).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 + + 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) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.weight_decay, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=0.04, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.weight_decay, + 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"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"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 + + 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 + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + 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) + + # SWA: collect checkpoints during warmdown + if args.swa_enabled and scale < args.swa_start_frac and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # Apply SWA if collected + if args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + current_state = base_model.state_dict() + avg_state = { + name: (tensor / swa_count).to(dtype=current_state[name].dtype) + for name, tensor in swa_state.items() + } + base_model.load_state_dict(avg_state, strict=True) + + code_bytes = len(code.encode("utf-8")) + + # SERIALIZATION + ROUNDTRIP VALIDATION + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + # Magnitude pruning: zero out smallest weights to improve compression + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + threshold = torch.quantile(param.abs().float().flatten(), 0.03) + mask = param.abs() < threshold + param.masked_fill_(mask, 0.0) + + # Mixed-precision export with calibration-driven byte allocation + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + selected_policy, calib_metrics = select_precision_policy( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + sd_cpu, + code_bytes, + log0, + ) + promoted_fp16 = sorted(name for name, precision in selected_policy.items() if precision == "fp16") + promoted_int8 = sorted( + name for name, precision in selected_policy.items() + if precision == "int8" and default_precision_for_name(name, sd_cpu[name]) != "int8" + ) + log0( + "auto_precision:final " + f"calib_bpb:{calib_metrics['val_bpb']:.6f} " + f"bytes:{int(calib_metrics['total_bytes'])} " + f"fp16:[{','.join(promoted_fp16)}] " + f"int8:[{','.join(promoted_int8)}]" + ) + quant_result, quant_meta = quantize_state_dict_with_policy(sd_cpu, selected_policy) + quant_blob = serialize_quantized_state(quant_result, quant_meta) + quant_file_bytes = len(quant_blob) + if quant_file_bytes + code_bytes > 16_000_000: + raise RuntimeError( + f"Quantized artifact exceeds competition limit: {quant_file_bytes + code_bytes} bytes total." + ) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + log0(f"Serialized model mixed+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + quant_blob_disk = quant_blob + if master_process: + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + if _COMPRESSOR == "zstd": + decompressed = zstandard.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = zlib.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_precision(quant_state["w"], quant_state["m"], sd_cpu) + base_model.load_state_dict(deq_state, strict=True) + if distributed: + dist.barrier() + + # Sliding window eval on roundtripped mixed-precision weights + torch.cuda.synchronize() + t_qeval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0( + f"final_eval_mode:sliding_window stride:{args.eval_stride} " + f"batch_seqs:{args.eval_batch_seqs} max_windows:{args.final_eval_max_windows or 'full'}" + ) + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.eval_batch_seqs, + max_windows=args.final_eval_max_windows or None, + ) + else: + log0("final_eval_mode:standard") + 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.barrier() + dist.destroy_process_group() + + +if __name__ == "__main__": + main() +# fixes applied +# tuned diff --git a/scripts/_modal_repro_longcontext.py b/scripts/_modal_repro_longcontext.py new file mode 100644 index 000000000..4a8092860 --- /dev/null +++ b/scripts/_modal_repro_longcontext.py @@ -0,0 +1,74 @@ +import os +import re +import subprocess +import modal + +APP_NAME = "parameter-golf-repro-longcontext-8h100" +REPO_REMOTE_PATH = "/workspace/parameter-golf" +TARGET_SCRIPT = "records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_gpt.py" + +app = modal.App(APP_NAME) +image = ( + modal.Image.from_registry("pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime") + .pip_install( + "numpy", + "tqdm", + "huggingface-hub", + "kernels", + "setuptools", + "typing-extensions==4.15.0", + "datasets", + "tiktoken", + "sentencepiece", + ) + .add_local_dir(".", remote_path=REPO_REMOTE_PATH) +) + +@app.function(image=image, gpu="H100:8", timeout=90 * 60, cpu=32, memory=196608) +def run(): + os.chdir(REPO_REMOTE_PATH) + + subprocess.run(["python", "data/cached_challenge_fineweb.py", "--variant", "sp1024"], check=True) + + env = os.environ.copy() + env.update( + { + "RUN_ID": "modal_longcontext_8h100_repro", + "DATA_PATH": "./data/datasets/fineweb10B_sp1024", + "TOKENIZER_PATH": "./data/tokenizers/fineweb_1024_bpe.model", + "VOCAB_SIZE": "1024", + "MAX_WALLCLOCK_SECONDS": "600", + "NCCL_IB_DISABLE": "1", + } + ) + + cmd = ["torchrun", "--standalone", "--nproc_per_node=8", TARGET_SCRIPT] + proc = subprocess.Popen(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1) + lines = [] + assert proc.stdout is not None + for line in proc.stdout: + print(line, end="") + lines.append(line) + rc = proc.wait() + if rc != 0: + raise RuntimeError(f"Training failed with exit code {rc}") + + log = "".join(lines) + exact = re.search(r"final_int8_zlib_roundtrip_exact\\s+val_loss:([0-9.]+)\\s+val_bpb:([0-9.]+)", log) + size = re.search(r"Total submission size int8\\+zlib:\\s*([0-9]+)\\s*bytes", log) + step = re.search(r"stopping_early: wallclock_cap .* step:([0-9]+)/([0-9]+)", log) + + out = { + "val_loss": float(exact.group(1)) if exact else None, + "val_bpb": float(exact.group(2)) if exact else None, + "bytes_total_int8_zlib": int(size.group(1)) if size else None, + "steps_done": int(step.group(1)) if step else None, + "steps_target": int(step.group(2)) if step else None, + } + print("\\n=== REPRO SUMMARY ===") + print(out) + return out + +@app.local_entrypoint() +def main(): + print(run.remote()) diff --git a/scripts/modal_probe_gpu_variant.py b/scripts/modal_probe_gpu_variant.py new file mode 100644 index 000000000..9dec3e08a --- /dev/null +++ b/scripts/modal_probe_gpu_variant.py @@ -0,0 +1,28 @@ +import subprocess +import modal + +app = modal.App("parameter-golf-gpu-variant-probe") + +image = modal.Image.from_registry("pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime") + + +@app.function(image=image, gpu="H100:8", timeout=30 * 60, cpu=4, memory=8192) +def probe() -> None: + cmds = [ + ["bash", "-lc", "nvidia-smi -L"], + [ + "bash", + "-lc", + "nvidia-smi --query-gpu=name,gpu_bus_id,pci.bus_id,pci.device_id,driver_version --format=csv,noheader", + ], + ["bash", "-lc", "nvidia-smi topo -m"], + ] + for cmd in cmds: + print("\n===", " ".join(cmd), "===") + proc = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True) + print(proc.stdout) + + +@app.local_entrypoint() +def main() -> None: + probe.remote() diff --git a/scripts/modal_repro_longcontext.py b/scripts/modal_repro_longcontext.py new file mode 100644 index 000000000..f08231c87 --- /dev/null +++ b/scripts/modal_repro_longcontext.py @@ -0,0 +1,77 @@ +import os +import re +import subprocess +import modal + +APP_NAME = "parameter-golf-repro-longcontext-8h100" +REPO_REMOTE_PATH = "/workspace/parameter-golf" +TARGET_SCRIPT = "records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_gpt.py" + +app = modal.App(APP_NAME) +image = ( + modal.Image.from_registry("pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime") + .apt_install("build-essential") + .pip_install( + "numpy", + "tqdm", + "huggingface-hub", + "kernels", + "setuptools", + "typing-extensions==4.15.0", + "datasets", + "tiktoken", + "sentencepiece", + ) + .add_local_dir(".", remote_path=REPO_REMOTE_PATH) +) + +@app.function(image=image, gpu="H100:8", timeout=90 * 60, cpu=32, memory=196608) +def run(): + os.chdir(REPO_REMOTE_PATH) + + subprocess.run(["python", "data/cached_challenge_fineweb.py", "--variant", "sp1024"], check=True) + + env = os.environ.copy() + env.update( + { + "RUN_ID": "modal_longcontext_8h100_repro", + "DATA_PATH": "./data/datasets/fineweb10B_sp1024", + "TOKENIZER_PATH": "./data/tokenizers/fineweb_1024_bpe.model", + "VOCAB_SIZE": "1024", + "MAX_WALLCLOCK_SECONDS": "600", + "NCCL_IB_DISABLE": "1", + "CC": "gcc", + "CXX": "g++", + } + ) + + cmd = ["torchrun", "--standalone", "--nproc_per_node=8", TARGET_SCRIPT] + proc = subprocess.Popen(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1) + lines = [] + assert proc.stdout is not None + for line in proc.stdout: + print(line, end="") + lines.append(line) + rc = proc.wait() + if rc != 0: + raise RuntimeError(f"Training failed with exit code {rc}") + + log = "".join(lines) + exact = re.search(r"final_int8_zlib_roundtrip_exact\\s+val_loss:([0-9.]+)\\s+val_bpb:([0-9.]+)", log) + size = re.search(r"Total submission size int8\\+zlib:\\s*([0-9]+)\\s*bytes", log) + step = re.search(r"stopping_early: wallclock_cap .* step:([0-9]+)/([0-9]+)", log) + + out = { + "val_loss": float(exact.group(1)) if exact else None, + "val_bpb": float(exact.group(2)) if exact else None, + "bytes_total_int8_zlib": int(size.group(1)) if size else None, + "steps_done": int(step.group(1)) if step else None, + "steps_target": int(step.group(2)) if step else None, + } + print("\\n=== REPRO SUMMARY ===") + print(out) + return out + +@app.local_entrypoint() +def main(): + print(run.remote()) diff --git a/scripts/modal_repro_top_gated.py b/scripts/modal_repro_top_gated.py new file mode 100644 index 000000000..d4931ea4e --- /dev/null +++ b/scripts/modal_repro_top_gated.py @@ -0,0 +1,145 @@ +from __future__ import annotations + +import os +import re +import subprocess + +import modal + +APP_NAME = "parameter-golf-repro-top-gated-8h100" +REPO_REMOTE_PATH = "/workspace/parameter-golf" +TARGET_SCRIPT = "records/track_10min_16mb/2026-03-19_SlidingWindow_FP16Emb_10L_MuonWD_OvertoneInit/train_gpt.py" + +STEP_RE = re.compile(r"step:(\d+)/(\d+).*step_avg:([0-9.]+)ms") +FINAL_RE = re.compile(r"final_int8_zlib_roundtrip_exact\s+val_loss:([0-9.]+)\s+val_bpb:([0-9.]+)") +SIZE_RE = re.compile(r"Total submission size int8\+zlib:\s*([0-9]+)\s*bytes") +STOP_RE = re.compile(r"stopping_early: wallclock_cap .* step:([0-9]+)/([0-9]+)") + +app = modal.App(APP_NAME) +image = ( + modal.Image.from_registry("pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime") + .apt_install("build-essential") + .pip_install( + "numpy", + "tqdm", + "huggingface-hub", + "kernels", + "setuptools", + "typing-extensions==4.15.0", + "datasets", + "tiktoken", + "sentencepiece", + ) + .add_local_dir(".", remote_path=REPO_REMOTE_PATH) +) + + +@app.function(image=image, gpu="H100:8", timeout=90 * 60, cpu=32, memory=196608) +def run_top( + run_id: str, + max_wallclock_seconds: int = 120, + enable_throughput_gate: bool = True, + gate_step_avg_ms: float = 60.0, + gate_check_step: int = 1000, +) -> dict[str, object]: + os.chdir(REPO_REMOTE_PATH) + + subprocess.run(["python", "data/cached_challenge_fineweb.py", "--variant", "sp1024"], check=True) + + env = os.environ.copy() + env.update( + { + "RUN_ID": run_id, + "DATA_PATH": "./data/datasets/fineweb10B_sp1024", + "TOKENIZER_PATH": "./data/tokenizers/fineweb_1024_bpe.model", + "VOCAB_SIZE": "1024", + "MAX_WALLCLOCK_SECONDS": str(max_wallclock_seconds), + "NCCL_IB_DISABLE": "1", + "CC": "gcc", + "CXX": "g++", + } + ) + + cmd = ["torchrun", "--standalone", "--nproc_per_node=8", TARGET_SCRIPT] + proc = subprocess.Popen(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1) + + lines: list[str] = [] + last_step = 0 + last_step_avg_ms = 0.0 + gate_checked = False + + assert proc.stdout is not None + for line in proc.stdout: + print(line, end="") + lines.append(line) + m = STEP_RE.search(line) + if m: + last_step = int(m.group(1)) + last_step_avg_ms = float(m.group(3)) + if enable_throughput_gate and (not gate_checked) and last_step >= gate_check_step: + gate_checked = True + if last_step_avg_ms > gate_step_avg_ms: + print( + f"THROUGHPUT_GATE_FAIL step={last_step} step_avg_ms={last_step_avg_ms:.2f} " + f"threshold_ms={gate_step_avg_ms:.2f}" + ) + proc.terminate() + try: + proc.wait(timeout=30) + except subprocess.TimeoutExpired: + proc.kill() + raise RuntimeError( + f"THROUGHPUT_GATE_FAIL step={last_step} step_avg_ms={last_step_avg_ms:.2f} " + f"threshold_ms={gate_step_avg_ms:.2f}" + ) + + rc = proc.wait() + if rc != 0: + raise RuntimeError(f"Training failed with exit code {rc}") + + if enable_throughput_gate and not gate_checked: + raise RuntimeError( + f"THROUGHPUT_GATE_INCONCLUSIVE did not reach step {gate_check_step} (last_step={last_step})" + ) + + full_log = "".join(lines) + exact = FINAL_RE.search(full_log) + size = SIZE_RE.search(full_log) + stop = STOP_RE.search(full_log) + + out = { + "run_id": run_id, + "steps_seen": last_step, + "step_avg_ms": last_step_avg_ms, + "gate_enabled": enable_throughput_gate, + "gate_step_avg_ms": gate_step_avg_ms, + "gate_check_step": gate_check_step, + "val_loss": float(exact.group(1)) if exact else None, + "val_bpb": float(exact.group(2)) if exact else None, + "bytes_total_int8_zlib": int(size.group(1)) if size else None, + "steps_done": int(stop.group(1)) if stop else None, + "steps_target": int(stop.group(2)) if stop else None, + } + + print("\n=== REPRO SUMMARY ===") + print(out) + return out + + +@app.local_entrypoint() +def main( + run_id: str = "modal_top_preflight", + max_wallclock_seconds: int = 120, + enable_throughput_gate: bool = True, + gate_step_avg_ms: float = 60.0, + gate_check_step: int = 1000, +): + print( + run_top.remote( + run_id=run_id, + max_wallclock_seconds=max_wallclock_seconds, + enable_throughput_gate=enable_throughput_gate, + gate_step_avg_ms=gate_step_avg_ms, + gate_check_step=gate_check_step, + ) + ) diff --git a/scripts/modal_run_baseline_1h100.py b/scripts/modal_run_baseline_1h100.py new file mode 100644 index 000000000..e6ee4cd5b --- /dev/null +++ b/scripts/modal_run_baseline_1h100.py @@ -0,0 +1,159 @@ +from __future__ import annotations + +import os +import re +import subprocess +from pathlib import Path + +import modal + +APP_NAME = "parameter-golf-baseline-1h100" +REPO_REMOTE_PATH = "/workspace/parameter-golf" +RECORD_DIR = "records/track_10min_16mb/2026-03-19_WIP_PLACEHOLDER" +TRAIN_SCRIPT = f"{RECORD_DIR}/train_gpt.py" +TRAIN_LOG = f"{RECORD_DIR}/train.log" + +app = modal.App(APP_NAME) + +image = ( + modal.Image.from_registry("pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime") + .apt_install("build-essential") + .pip_install( + "numpy", + "tqdm", + "huggingface-hub", + "kernels", + "setuptools", + "typing-extensions==4.15.0", + "datasets", + "tiktoken", + "sentencepiece", + ) + .add_local_dir(".", remote_path=REPO_REMOTE_PATH) +) + + +def _extract_metrics(log_text: str) -> dict[str, float | int | None]: + bpb_match = re.search(r"final_int8_zlib_roundtrip_exact\\s+val_loss:([0-9.]+)\\s+val_bpb:([0-9.]+)", log_text) + size_match = re.search(r"Total submission size int8\\+zlib:\\s*([0-9]+)\\s*bytes", log_text) + steps_match = re.search(r"stopping_early: wallclock_cap .* step:([0-9]+)/([0-9]+)", log_text) + + out: dict[str, float | int | None] = { + "val_loss": None, + "val_bpb": None, + "bytes_total_int8_zlib": None, + "steps_done": None, + "steps_target": None, + } + if bpb_match: + out["val_loss"] = float(bpb_match.group(1)) + out["val_bpb"] = float(bpb_match.group(2)) + if size_match: + out["bytes_total_int8_zlib"] = int(size_match.group(1)) + if steps_match: + out["steps_done"] = int(steps_match.group(1)) + out["steps_target"] = int(steps_match.group(2)) + return out + + +@app.function( + image=image, + gpu="H100", + timeout=60 * 60, + cpu=8, + memory=64 * 1024, +) +def run_baseline_one_h100( + train_shards: int = 1, + max_wallclock_seconds: int = 180, + iterations: int = 600, + val_loss_every: int = 200, + train_log_every: int = 100, + run_id: str = "modal_1h100_baseline_smoke", +) -> dict[str, object]: + os.chdir(REPO_REMOTE_PATH) + Path(RECORD_DIR).mkdir(parents=True, exist_ok=True) + + # Download challenge data/tokenizer cache (small shard count for smoke by default). + subprocess.run( + [ + "python", + "data/cached_challenge_fineweb.py", + "--variant", + "sp1024", + "--train-shards", + str(train_shards), + ], + check=True, + ) + + env = os.environ.copy() + env.update( + { + "RUN_ID": run_id, + "DATA_PATH": "./data/datasets/fineweb10B_sp1024", + "TOKENIZER_PATH": "./data/tokenizers/fineweb_1024_bpe.model", + "VOCAB_SIZE": "1024", + "MAX_WALLCLOCK_SECONDS": str(max_wallclock_seconds), + "ITERATIONS": str(iterations), + "VAL_LOSS_EVERY": str(val_loss_every), + "TRAIN_LOG_EVERY": str(train_log_every), + "CC": "gcc", + "CXX": "g++", + } + ) + + cmd = ["torchrun", "--standalone", "--nproc_per_node=1", TRAIN_SCRIPT] + + # Stream logs to Modal output and save a local copy in the record folder. + lines: list[str] = [] + with open(TRAIN_LOG, "w", encoding="utf-8") as fout: + proc = subprocess.Popen( + cmd, + env=env, + stdout=subprocess.PIPE, + stderr=subprocess.STDOUT, + text=True, + bufsize=1, + ) + assert proc.stdout is not None + for line in proc.stdout: + print(line, end="") + fout.write(line) + lines.append(line) + rc = proc.wait() + if rc != 0: + raise RuntimeError(f"Training command failed with exit code {rc}") + + full_log = "".join(lines) + metrics = _extract_metrics(full_log) + return { + "run_id": run_id, + "train_log_path": TRAIN_LOG, + "train_shards": train_shards, + "max_wallclock_seconds": max_wallclock_seconds, + "iterations": iterations, + "metrics": metrics, + "log_tail": lines[-40:], + } + + +@app.local_entrypoint() +def main( + train_shards: int = 1, + max_wallclock_seconds: int = 180, + iterations: int = 600, + val_loss_every: int = 200, + train_log_every: int = 100, + run_id: str = "modal_1h100_baseline_smoke", +): + result = run_baseline_one_h100.remote( + train_shards=train_shards, + max_wallclock_seconds=max_wallclock_seconds, + iterations=iterations, + val_loss_every=val_loss_every, + train_log_every=train_log_every, + run_id=run_id, + ) + print("\\n=== Modal run summary ===") + print(result)