diff --git a/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/README.md b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/README.md new file mode 100644 index 0000000000..8fe7782f27 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/README.md @@ -0,0 +1,45 @@ +# Asynchronous Prefetching — submission notes + +### 1191 (with technique) vs 1137 (default) steps in 600s on local compute + +## Key changes + +Same model, optimizer, data layout, and training math as baseline. This is **a general purpose rework that could apply to most other approaches** for slight speed boosts. Overlap CPU data prep and host→device copies with GPU work so the GPU spends less time idle. + +| Area | Original (`train_gpt_og_linux.py`) | Improved | +|------|-------------------------------------|----------| +| **Training batches** | Each step: read tokens on CPU, then H2D — all on the main thread before forward. | **Background thread** (`PrefetchingDistributedTokenLoader`) builds the **next** pinned CPU batch while the GPU runs the current step. Primary win: **CPU work overlaps GPU compute** (not GPU-side double-buffering of H2D vs forward). | +| **H2D** | Single default stream. | Optional **dedicated CUDA copy stream** (`TRAIN_COPY_STREAM`, off when timing diagnostics are on). Transfers use pinned memory; the training path **still waits for that step’s H2D** before forward (`wait_stream`). | +| **Validation** | Simple loop: slice → GPU → forward; BPB byte math on GPU. | **Prefetch thread** for pinned CPU batches; **double-buffered** H2D with copy stream + events so the **next** batch can copy while the **current** forward runs. Default **`VAL_BYTECOUNT_DEVICE=cpu`** moves BPB byte counting off the GPU vs the original (set **`cuda`** to mirror baseline GPU LUT math). | + +## Diagnostics + +To measure how much time this actually saves, I added **`TRAINING_TIMING_BREAKDOWN`** (batch CPU vs H2D vs FWD/BWD/opt vs val; adds syncs). When enabled, lines log every **`TRAINING_TIMING_EVERY`** steps (default 200) and for early steps (first 10). Extra logs: train/val I/O mode, `val_stage_time_ms`, train vs val wall time split. + +**`VAL_BYTECOUNT_DEVICE`** defaults to **`cpu`** in the improved script (not an extra flag you must set). Use **`cuda`** if you want validation byte math on the GPU like the original. + +Optional **`VAL_PROGRESS_LOG_EVERY`** (default **0**): set to a positive value to log per-batch validation progress (`val_progress:...`). + +## Defaults & toggles + +Overlap features are **on by default** (`TRAIN_PREFETCH`, `TRAIN_COPY_STREAM`, `VAL_PREFETCH`, `VAL_COPY_STREAM`, etc.) and can be turned off via env vars if needed. **`TRAINING_TIMING_BREAKDOWN`** defaults to 0 and is not displayed. Prefetch/overlap are **automatically disabled** when `TRAINING_TIMING_BREAKDOWN=1` so timings stay interpretable. + +## Idea + +**Prefetch training and validation batches asynchronously and parallelize CPU ↔ GPU transfers with compute** to minimize pipeline bubbles under a fixed wall-clock budget. +This is an intuitive idea that I came up with that could help models with real research and architectural advancements place slightly higher. + +## Why this may be unimpactful in some cases + +With **`TRAINING_TIMING_BREAKDOWN=1`**, early-step lines look like this (same hardware / config as above; `grad_accum_steps=8`, per-micro averages for batch/forward/backward): + +```text +timing_breakdown step:1 micro_steps:8 batch_cpu_ms:0.29 batch_h2d_ms:0.35 forward_ms:30.54 backward_ms:64.93 grad_clip_ms:0.00 optimizer_ms:55.37 val_ms:121092.09 explicit_sync_ms:0.16 (per_optimizer_step; forward/backward/batch averaged over micro_steps; grad_accum_steps=8) +timing_breakdown step:2 micro_steps:8 batch_cpu_ms:0.29 batch_h2d_ms:0.35 forward_ms:30.29 backward_ms:64.72 grad_clip_ms:0.00 optimizer_ms:55.08 val_ms:0.00 explicit_sync_ms:0.00 (per_optimizer_step; forward/backward/batch averaged over micro_steps; grad_accum_steps=8) +timing_breakdown step:3 micro_steps:8 batch_cpu_ms:0.28 batch_h2d_ms:0.37 forward_ms:30.66 backward_ms:65.18 grad_clip_ms:0.00 optimizer_ms:54.45 val_ms:0.00 explicit_sync_ms:0.00 (per_optimizer_step; forward/backward/batch averaged over micro_steps; grad_accum_steps=8) +timing_breakdown step:4 micro_steps:8 batch_cpu_ms:0.31 batch_h2d_ms:0.34 forward_ms:30.34 backward_ms:64.43 grad_clip_ms:0.00 optimizer_ms:55.19 val_ms:0.00 explicit_sync_ms:0.00 (per_optimizer_step; forward/backward/batch averaged over micro_steps; grad_accum_steps=8) +``` + +**How to read this:** `batch_cpu_ms` and `batch_h2d_ms` are ~0.3 ms per micro-step; `forward_ms` and `backward_ms` are ~30 ms and ~65 ms per micro-step. Scaled by 8 micro-steps, batch prep + H2D is on the order of **~5 ms per optimizer step**, while forward + backward + optimizer is on the order of **~800+ ms**. So **data movement is a tiny slice** of the step; overlapping it cannot move wall-clock much when the GPU is already busy with compute for almost the whole step. + +**Caveat:** On a **much faster GPU** (or smaller model / larger batch so steps are shorter), the same CPU+H2D work could become a **larger fraction** of the step, and prefetch or val overlap might show up more in profiles. The breakdown above is **not** universal; it only shows why the optimization can be a no-op when **compute is the bottleneck**. diff --git a/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/final_model.int8.ptz b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/final_model.int8.ptz new file mode 100644 index 0000000000..851bd85876 Binary files /dev/null and b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/final_model.int8.ptz differ diff --git a/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/final_model.pt b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/final_model.pt new file mode 100644 index 0000000000..13c0ab5321 Binary files /dev/null and b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/final_model.pt differ diff --git a/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/meta.json b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/meta.json new file mode 100644 index 0000000000..11ea28c9b1 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/meta.json @@ -0,0 +1,80 @@ +{ + "id": "f9ba5b9c-412b-48cd-b1bc-7d2c0fc65ee7", + "started_at": "2026-03-24T00:48:35.433731+00:00", + "finished_at": "2026-03-24T01:11:47.811831+00:00", + "exit_code": 0, + "script": "train_gpt_improved_linux.py", + "argv": [ + "wsl", + "-d", + "Ubuntu", + "bash", + "-lc", + "cd '/mnt/c/Users/REDACTED/parameter-golf' && export ADAM_EPS=1e-8 && export BETA1=0.9 && export BETA2=0.95 && export DATA_PATH=./data/datasets/fineweb10B_sp1024 && export EMBED_LR=0.6 && export GRAD_CLIP_NORM=0.0 && export HEAD_LR=0.008 && export ITERATIONS=20000 && export LOGIT_SOFTCAP=30.0 && export MATRIX_LR=0.04 && export MAX_WALLCLOCK_SECONDS=600.0 && export MLP_MULT=2 && export MODEL_DIM=512 && export MUON_BACKEND_STEPS=5 && export MUON_MOMENTUM=0.95 && export MUON_MOMENTUM_WARMUP_START=0.85 && export MUON_MOMENTUM_WARMUP_STEPS=500 && export NUM_HEADS=8 && export NUM_KV_HEADS=4 && export NUM_LAYERS=9 && export PYTHONUNBUFFERED=1 && export QK_GAIN_INIT=1.5 && export ROPE_BASE=10000.0 && export RUN_ID=8gb_vram && export SCALAR_LR=0.04 && export SEED=1337 && export TIED_EMBED_INIT_STD=0.005 && export TIED_EMBED_LR=0.05 && export TIE_EMBEDDINGS=1 && export TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model && export TRAIN_BATCH_TOKENS=65536 && export TRAIN_LOG_EVERY=200 && export TRAIN_SEQ_LEN=1024 && export VAL_BATCH_SIZE=524288 && export VAL_LOSS_EVERY=1000 && export VAL_PROGRESS_LOG_EVERY=0 && export VOCAB_SIZE=1024 && export WARMDOWN_ITERS=1200 && export WARMUP_STEPS=20 && ./.venv/bin/python -u train_gpt_improved_linux.py" + ], + "env_snapshot": { + "DATA_PATH": "./data/datasets/fineweb10B_sp1024", + "TOKENIZER_PATH": "./data/tokenizers/fineweb_1024_bpe.model", + "RUN_ID": "8gb_vram", + "SEED": "1337", + "VAL_BATCH_SIZE": "524288", + "VAL_LOSS_EVERY": "1000", + "VAL_PROGRESS_LOG_EVERY": "0", + "TRAIN_LOG_EVERY": "200", + "ITERATIONS": "20000", + "WARMDOWN_ITERS": "1200", + "WARMUP_STEPS": "20", + "TRAIN_BATCH_TOKENS": "65536", + "TRAIN_SEQ_LEN": "1024", + "MAX_WALLCLOCK_SECONDS": "600.0", + "QK_GAIN_INIT": "1.5", + "VOCAB_SIZE": "1024", + "NUM_LAYERS": "9", + "NUM_KV_HEADS": "4", + "MODEL_DIM": "512", + "NUM_HEADS": "8", + "MLP_MULT": "2", + "TIE_EMBEDDINGS": "1", + "ROPE_BASE": "10000.0", + "LOGIT_SOFTCAP": "30.0", + "EMBED_LR": "0.6", + "HEAD_LR": "0.008", + "TIED_EMBED_LR": "0.05", + "TIED_EMBED_INIT_STD": "0.005", + "MATRIX_LR": "0.04", + "SCALAR_LR": "0.04", + "MUON_MOMENTUM": "0.95", + "MUON_BACKEND_STEPS": "5", + "MUON_MOMENTUM_WARMUP_START": "0.85", + "MUON_MOMENTUM_WARMUP_STEPS": "500", + "BETA1": "0.9", + "BETA2": "0.95", + "ADAM_EPS": "1e-8", + "GRAD_CLIP_NORM": "0.0", + "PYTHONUNBUFFERED": "1" + }, + "command_powershell": "(WSL runtime selected; native PowerShell command disabled.)", + "command_bash": "cd '/mnt/c/Users/REDACTED/parameter-golf' && export ADAM_EPS=1e-8 && export BETA1=0.9 && export BETA2=0.95 && export DATA_PATH=./data/datasets/fineweb10B_sp1024 && export EMBED_LR=0.6 && export GRAD_CLIP_NORM=0.0 && export HEAD_LR=0.008 && export ITERATIONS=20000 && export LOGIT_SOFTCAP=30.0 && export MATRIX_LR=0.04 && export MAX_WALLCLOCK_SECONDS=600.0 && export MLP_MULT=2 && export MODEL_DIM=512 && export MUON_BACKEND_STEPS=5 && export MUON_MOMENTUM=0.95 && export MUON_MOMENTUM_WARMUP_START=0.85 && export MUON_MOMENTUM_WARMUP_STEPS=500 && export NUM_HEADS=8 && export NUM_KV_HEADS=4 && export NUM_LAYERS=9 && export PYTHONUNBUFFERED=1 && export QK_GAIN_INIT=1.5 && export ROPE_BASE=10000.0 && export RUN_ID=8gb_vram && export SCALAR_LR=0.04 && export SEED=1337 && export TIED_EMBED_INIT_STD=0.005 && export TIED_EMBED_LR=0.05 && export TIE_EMBEDDINGS=1 && export TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model && export TRAIN_BATCH_TOKENS=65536 && export TRAIN_LOG_EVERY=200 && export TRAIN_SEQ_LEN=1024 && export VAL_BATCH_SIZE=524288 && export VAL_LOSS_EVERY=1000 && export VAL_PROGRESS_LOG_EVERY=0 && export VOCAB_SIZE=1024 && export WARMDOWN_ITERS=1200 && export WARMUP_STEPS=20 && ./.venv/bin/python -u train_gpt_improved_linux.py", + "log_file": null, + "metrics": { + "timing_val_stage_total_ms": 366515, + "timing_train_loop_ms": 600433, + "timing_val_total_ms": 366515, + "peak_memory_line": "peak memory allocated: 1552 MiB reserved: 2280 MiB", + "serialized_model_line": "Serialized model: 67224983 bytes", + "code_bytes": 70245, + "serialized_int8_zlib_line": "Serialized model int8+zlib: 12907011 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x)", + "model_int8_zlib_bytes": 12907011, + "submission_size_line": "Total submission size int8+zlib: 12977256 bytes", + "submission_total_bytes_int8_zlib": 12977256, + "timing_quantize_ms": 1082, + "timing_roundtrip_dequant_ms": 407, + "timing_roundtrip_eval_ms": 125190, + "final_roundtrip_line": "final_int8_zlib_roundtrip val_loss:2.5457 val_bpb:1.5077 eval_time:125190ms", + "val_loss": 2.5457, + "val_bpb": 1.5077, + "final_roundtrip_exact_line": "final_int8_zlib_roundtrip_exact val_loss:2.54570429 val_bpb:1.50770948", + "val_loss_exact": 2.54570429, + "val_bpb_exact": 1.50770948 + } +} \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/train_gpt.py b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/train_gpt.py new file mode 100644 index 0000000000..a921529148 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/train_gpt.py @@ -0,0 +1,1638 @@ +""" +Linux-focused GPT trainer with: +- async train/val prefetch + optional CUDA copy streams +- progress/timing logging for training, validation, and export +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import queue +import random +import subprocess +import threading +import sys +import time +import uuid +import zlib +from collections.abc import Callable +from dataclasses import dataclass +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)) + # Log every N validation micro-batches during eval_val (0 = off). See module docstring. + val_progress_log_every = int(os.environ.get("VAL_PROGRESS_LOG_EVERY", "0")) + val_prefetch = bool(int(os.environ.get("VAL_PREFETCH", "1"))) + val_copy_stream = bool(int(os.environ.get("VAL_COPY_STREAM", "1"))) + val_prefetch_queue = int(os.environ.get("VAL_PREFETCH_QUEUE", "2")) + val_bytecount_device = os.environ.get("VAL_BYTECOUNT_DEVICE", "cpu").strip().lower() + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + # When 1, log per-phase timings (batch CPU vs H2D vs forward/backward/opt vs val vs explicit sync). Adds syncs; use for diagnosis only. + training_timing_breakdown = bool(int(os.environ.get("TRAINING_TIMING_BREAKDOWN", "0"))) + training_timing_every = int(os.environ.get("TRAINING_TIMING_EVERY", "200")) + # Systems: prefetch pinned CPU batches on a worker (disabled automatically if TRAINING_TIMING_BREAKDOWN=1). + train_prefetch = bool(int(os.environ.get("TRAIN_PREFETCH", "1"))) + train_copy_stream = bool(int(os.environ.get("TRAIN_COPY_STREAM", "1"))) + + # 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, + base_bytes_lut_cpu: Tensor | None = None, + has_leading_space_lut_cpu: Tensor | None = None, + is_boundary_token_lut_cpu: Tensor | None = None, + copy_stream: torch.cuda.Stream | None = None, + enable_overlap: bool = True, + log_progress: Callable[[str], None] | None = None, +) -> 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) + use_cpu_bytecount = args.val_bytecount_device == "cpu" + if use_cpu_bytecount: + if base_bytes_lut_cpu is None or has_leading_space_lut_cpu is None or is_boundary_token_lut_cpu is None: + raise ValueError("CPU bytecount mode requires CPU LUTs") + val_byte_count_host = 0.0 + else: + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + class ValidationPrefetcher: + def __init__(self, entries: list[tuple[int, int, int]], queue_size: int): + self._entries = entries + self._queue: queue.Queue[tuple[int, Tensor, Tensor] | None] = queue.Queue(maxsize=max(queue_size, 1)) + self._stop = threading.Event() + self._worker_error: BaseException | None = None + self._thread = threading.Thread(target=self._worker_loop, name="val-token-prefetch", daemon=True) + self._thread.start() + + def _worker_loop(self) -> None: + try: + for bi, raw_start, raw_end in self._entries: + if self._stop.is_set(): + break + local = val_tokens[raw_start:raw_end].to(dtype=torch.int64) + x = local[:-1].reshape(-1, args.train_seq_len).pin_memory() + y = local[1:].reshape(-1, args.train_seq_len).pin_memory() + while not self._stop.is_set(): + try: + self._queue.put((bi, x, y), timeout=0.05) + break + except queue.Full: + pass + except BaseException as e: + self._worker_error = e + finally: + while not self._stop.is_set(): + try: + self._queue.put(None, timeout=0.05) + break + except queue.Full: + pass + + def next_item(self) -> tuple[int, Tensor, Tensor] | None: + if self._worker_error is not None: + raise self._worker_error + item = self._queue.get() + if self._worker_error is not None: + raise self._worker_error + return item + + def shutdown(self) -> None: + self._stop.set() + while True: + try: + self._queue.get_nowait() + except queue.Empty: + break + if self._thread.is_alive(): + self._thread.join(timeout=10.0) + + def _enqueue_h2d_with_event( + x_cpu: Tensor, + y_cpu: Tensor, + ) -> tuple[Tensor, Tensor, torch.cuda.Event | None]: + if copy_stream is not None: + ev = torch.cuda.Event() + with torch.cuda.stream(copy_stream): + x_d = x_cpu.to(device, non_blocking=True) + y_d = y_cpu.to(device, non_blocking=True) + ev.record(copy_stream) + return x_d, y_d, ev + return x_cpu.to(device, non_blocking=True), y_cpu.to(device, non_blocking=True), None + + model.eval() + batch_starts = list(range(seq_start, seq_end, local_batch_seqs)) + num_batches = len(batch_starts) + batch_entries = [] + for bi, batch_seq_start in enumerate(batch_starts): + 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 + batch_entries.append((bi, raw_start, raw_end)) + + use_overlap_path = enable_overlap and (args.val_prefetch or copy_stream is not None) + prefetcher: ValidationPrefetcher | None = None + with torch.inference_mode(): + try: + if use_overlap_path and batch_entries: + if args.val_prefetch: + prefetcher = ValidationPrefetcher(batch_entries, queue_size=args.val_prefetch_queue) + + def next_cpu_item() -> tuple[int, Tensor, Tensor] | None: + return prefetcher.next_item() + else: + item_idx = 0 + + def next_cpu_item() -> tuple[int, Tensor, Tensor] | None: + nonlocal item_idx + if item_idx >= len(batch_entries): + return None + bi, raw_start, raw_end = batch_entries[item_idx] + item_idx += 1 + local = val_tokens[raw_start:raw_end].to(dtype=torch.int64) + x_cpu = local[:-1].reshape(-1, args.train_seq_len).pin_memory() + y_cpu = local[1:].reshape(-1, args.train_seq_len).pin_memory() + return bi, x_cpu, y_cpu + + current = next_cpu_item() + if current is not None: + cur_bi, cur_x_cpu, cur_y_cpu = current + cur_x, cur_y, cur_ready = _enqueue_h2d_with_event(cur_x_cpu, cur_y_cpu) + while True: + nxt = next_cpu_item() + if nxt is not None: + nxt_bi, nxt_x_cpu, nxt_y_cpu = nxt + nxt_x, nxt_y, nxt_ready = _enqueue_h2d_with_event(nxt_x_cpu, nxt_y_cpu) + else: + nxt_bi = -1 + nxt_x_cpu = nxt_y_cpu = None + nxt_x = nxt_y = None + nxt_ready = None + + if cur_ready is not None: + torch.cuda.current_stream().wait_event(cur_ready) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(cur_x, cur_y).detach() + batch_token_count = float(cur_y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + + if use_cpu_bytecount: + prev_ids_cpu = cur_x_cpu.reshape(-1) + tgt_ids_cpu = cur_y_cpu.reshape(-1) + token_bytes_cpu = base_bytes_lut_cpu[tgt_ids_cpu].to(dtype=torch.int16) + token_bytes_cpu += ( + has_leading_space_lut_cpu[tgt_ids_cpu] & ~is_boundary_token_lut_cpu[prev_ids_cpu] + ).to(dtype=torch.int16) + val_byte_count_host += float(token_bytes_cpu.to(torch.float64).sum().item()) + else: + prev_ids = cur_x.reshape(-1) + tgt_ids = cur_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 ( + log_progress is not None + and rank == 0 + and args.val_progress_log_every > 0 + and num_batches > 0 + and (cur_bi % args.val_progress_log_every == 0 or cur_bi == num_batches - 1) + ): + pt = val_token_count.item() + partial = (val_loss_sum / val_token_count).item() if pt > 0 else float("nan") + scope = f"rank0/{world_size}" if world_size > 1 else "all" + log_progress( + f"val_progress:{scope} batch {cur_bi + 1}/{num_batches} " + f"partial_mean_loss:{partial:.4f} tokens_seen:{int(pt)}" + ) + + if nxt is None: + break + cur_bi, cur_x_cpu, cur_y_cpu = nxt_bi, nxt_x_cpu, nxt_y_cpu + cur_x, cur_y, cur_ready = nxt_x, nxt_y, nxt_ready + else: + for bi, raw_start, raw_end in batch_entries: + local_cpu = val_tokens[raw_start:raw_end].to(dtype=torch.int64) + x_cpu = local_cpu[:-1].reshape(-1, args.train_seq_len) + y_cpu = local_cpu[1:].reshape(-1, args.train_seq_len) + x = x_cpu.to(device=device, non_blocking=True) + y = y_cpu.to(device=device, non_blocking=True) + 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 + if use_cpu_bytecount: + prev_ids_cpu = x_cpu.reshape(-1) + tgt_ids_cpu = y_cpu.reshape(-1) + token_bytes_cpu = base_bytes_lut_cpu[tgt_ids_cpu].to(dtype=torch.int16) + token_bytes_cpu += ( + has_leading_space_lut_cpu[tgt_ids_cpu] & ~is_boundary_token_lut_cpu[prev_ids_cpu] + ).to(dtype=torch.int16) + val_byte_count_host += float(token_bytes_cpu.to(torch.float64).sum().item()) + else: + 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 ( + log_progress is not None + and rank == 0 + and args.val_progress_log_every > 0 + and num_batches > 0 + and (bi % args.val_progress_log_every == 0 or bi == num_batches - 1) + ): + pt = val_token_count.item() + partial = (val_loss_sum / val_token_count).item() if pt > 0 else float("nan") + scope = f"rank0/{world_size}" if world_size > 1 else "all" + log_progress( + f"val_progress:{scope} batch {bi + 1}/{num_batches} " + f"partial_mean_loss:{partial:.4f} tokens_seen:{int(pt)}" + ) + finally: + if prefetcher is not None: + prefetcher.shutdown() + + if use_cpu_bytecount: + val_byte_count = torch.tensor(val_byte_count_host, device=device, dtype=torch.float64) + + 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) + + +def _cpu_batch_from_stream( + stream: TokenStream, + rank: int, + world_size: int, + global_tokens: int, + seq_len: int, + grad_accum_steps: int, +) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = stream.take(per_rank_span * world_size) + start = 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, y + + +def _h2d_int64_batches( + x: Tensor, + y: Tensor, + device: torch.device, + copy_stream: torch.cuda.Stream | None, +) -> tuple[Tensor, Tensor]: + if copy_stream is not None: + with torch.cuda.stream(copy_stream): + x_d = x.to(device, non_blocking=True) + y_d = y.to(device, non_blocking=True) + torch.cuda.current_stream().wait_stream(copy_stream) + return x_d, y_d + return x.to(device, non_blocking=True), y.to(device, non_blocking=True) + + +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, + copy_stream: torch.cuda.Stream | None = None, + ): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + self.copy_stream = copy_stream + + def next_batch( + self, + global_tokens: int, + seq_len: int, + grad_accum_steps: int, + timing: dict[str, float] | None = None, + ) -> tuple[Tensor, Tensor]: + if timing is not None: + t_cpu0 = time.perf_counter() + x, y = _cpu_batch_from_stream( + self.stream, self.rank, self.world_size, global_tokens, seq_len, grad_accum_steps + ) + timing["batch_cpu_ms"] = 1000.0 * (time.perf_counter() - t_cpu0) + torch.cuda.synchronize() + t_h2d0 = time.perf_counter() + x_d, y_d = _h2d_int64_batches(x, y, self.device, None) + torch.cuda.synchronize() + timing["batch_h2d_ms"] = 1000.0 * (time.perf_counter() - t_h2d0) + return x_d, y_d + x, y = _cpu_batch_from_stream( + self.stream, self.rank, self.world_size, global_tokens, seq_len, grad_accum_steps + ) + return _h2d_int64_batches(x, y, self.device, self.copy_stream) + + +class PrefetchingDistributedTokenLoader: + """Builds the next CPU (pinned) batch on a background thread while the GPU works or validates.""" + + def __init__( + self, + pattern: str, + rank: int, + world_size: int, + device: torch.device, + global_tokens: int, + seq_len: int, + grad_accum_steps: int, + copy_stream: torch.cuda.Stream | None = None, + queue_size: int = 2, + ): + self.rank = rank + self.world_size = world_size + self.device = device + self._global_tokens = global_tokens + self._seq_len = seq_len + self._grad_accum_steps = grad_accum_steps + self.copy_stream = copy_stream + self._stream = TokenStream(pattern) + self._queue: queue.Queue[tuple[Tensor, Tensor]] = queue.Queue(maxsize=queue_size) + self._stop = threading.Event() + self._worker_error: BaseException | None = None + self._thread = threading.Thread(target=self._worker_loop, name="train-token-prefetch", daemon=True) + self._thread.start() + + def _worker_loop(self) -> None: + while not self._stop.is_set(): + try: + x, y = _cpu_batch_from_stream( + self._stream, + self.rank, + self.world_size, + self._global_tokens, + self._seq_len, + self._grad_accum_steps, + ) + x = x.pin_memory() + y = y.pin_memory() + while not self._stop.is_set(): + try: + self._queue.put((x, y), timeout=0.05) + break + except queue.Full: + pass + except BaseException as e: + self._worker_error = e + break + + def next_batch( + self, + global_tokens: int, + seq_len: int, + grad_accum_steps: int, + timing: dict[str, float] | None = None, + ) -> tuple[Tensor, Tensor]: + if timing is not None: + raise RuntimeError("PrefetchingDistributedTokenLoader does not support timing dict; use DistributedTokenLoader") + if self._worker_error is not None: + raise self._worker_error + x, y = self._queue.get() + return _h2d_int64_batches(x, y, self.device, self.copy_stream) + + def shutdown(self) -> None: + self._stop.set() + while True: + try: + self._queue.get_nowait() + except queue.Empty: + break + if self._thread.is_alive(): + self._thread.join(timeout=10.0) + + +@dataclass +class TrainingTimingBreakdown: + """Accumulates per-step ms totals when TRAINING_TIMING_BREAKDOWN=1 (not micro-averaged until finalize).""" + + batch_cpu_ms: float = 0.0 + batch_h2d_ms: float = 0.0 + forward_ms: float = 0.0 + backward_ms: float = 0.0 + optimizer_ms: float = 0.0 + grad_clip_ms: float = 0.0 + val_ms: float = 0.0 + explicit_sync_ms: float = 0.0 + micro_steps: int = 0 + + def reset(self) -> None: + self.batch_cpu_ms = 0.0 + self.batch_h2d_ms = 0.0 + self.forward_ms = 0.0 + self.backward_ms = 0.0 + self.optimizer_ms = 0.0 + self.grad_clip_ms = 0.0 + self.val_ms = 0.0 + self.explicit_sync_ms = 0.0 + self.micro_steps = 0 + + def as_log_line(self, step: int, grad_accum_steps: int) -> str: + micro = max(self.micro_steps, 1) + ga = max(grad_accum_steps, 1) + return ( + f"timing_breakdown step:{step} micro_steps:{self.micro_steps} " + f"batch_cpu_ms:{self.batch_cpu_ms / micro:.2f} batch_h2d_ms:{self.batch_h2d_ms / micro:.2f} " + f"forward_ms:{self.forward_ms / micro:.2f} backward_ms:{self.backward_ms / micro:.2f} " + f"grad_clip_ms:{self.grad_clip_ms:.2f} optimizer_ms:{self.optimizer_ms:.2f} " + f"val_ms:{self.val_ms:.2f} explicit_sync_ms:{self.explicit_sync_ms:.2f} " + f"(per_optimizer_step; forward/backward/batch averaged over micro_steps; grad_accum_steps={ga})" + ) + +# ----------------------------- +# 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() + if args.val_bytecount_device not in {"cpu", "cuda"}: + raise ValueError(f"VAL_BYTECOUNT_DEVICE must be 'cpu' or 'cuda', got {args.val_bytecount_device!r}") + 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) + + def log_val_progress(msg: str) -> None: + log0(msg) + sys.stdout.flush() + + val_log_fn = log_val_progress if args.val_progress_log_every > 0 else None + + 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 + ) + base_bytes_lut_cpu = base_bytes_lut.detach().to("cpu") + has_leading_space_lut_cpu = has_leading_space_lut.detach().to("cpu") + is_boundary_token_lut_cpu = is_boundary_token_lut.detach().to("cpu") + 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("torch_compile:on") + 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}") + + do_training_timing = args.training_timing_breakdown + training_timing = TrainingTimingBreakdown() + if do_training_timing: + log0( + f"training_timing_breakdown:enabled log_every:{args.training_timing_every} " + "(adds GPU synchronizations; use only to diagnose bottlenecks)" + ) + + use_train_prefetch = args.train_prefetch and not do_training_timing + use_train_copy_stream = args.train_copy_stream and not do_training_timing + copy_stream = torch.cuda.Stream() if use_train_copy_stream else None + use_val_overlap = not do_training_timing + use_val_copy_stream = args.val_copy_stream and use_val_overlap + val_copy_stream = torch.cuda.Stream() if use_val_copy_stream else None + + def make_train_loader() -> DistributedTokenLoader | PrefetchingDistributedTokenLoader: + if use_train_prefetch: + return PrefetchingDistributedTokenLoader( + args.train_files, + rank, + world_size, + device, + args.train_batch_tokens, + args.train_seq_len, + grad_accum_steps, + copy_stream=copy_stream, + ) + return DistributedTokenLoader(args.train_files, rank, world_size, device, copy_stream=copy_stream) + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = make_train_loader() + log0( + f"train_io:prefetch:{use_train_prefetch} copy_stream:{use_train_copy_stream} " + f"(disabled when TRAINING_TIMING_BREAKDOWN=1)" + ) + log0( + f"val_io:overlap:{use_val_overlap} prefetch:{args.val_prefetch and use_val_overlap} " + f"copy_stream:{use_val_copy_stream} prefetch_queue:{args.val_prefetch_queue} " + f"bytecount_device:{args.val_bytecount_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 + if hasattr(train_loader, "shutdown"): + train_loader.shutdown() + train_loader = make_train_loader() + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + val_total_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + pending_val_ms = 0.0 + pending_sync_ms = 0.0 + 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: + if do_training_timing: + t_sync_a = time.perf_counter() + torch.cuda.synchronize() + pending_sync_ms += 1000.0 * (time.perf_counter() - t_sync_a) + training_time_ms += 1000.0 * (time.perf_counter() - t0) + t_val0 = time.perf_counter() + 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, + base_bytes_lut_cpu=base_bytes_lut_cpu, + has_leading_space_lut_cpu=has_leading_space_lut_cpu, + is_boundary_token_lut_cpu=is_boundary_token_lut_cpu, + copy_stream=val_copy_stream, + enable_overlap=use_val_overlap, + log_progress=val_log_fn, + ) + if do_training_timing: + torch.cuda.synchronize() + pending_val_ms = 1000.0 * (time.perf_counter() - t_val0) + 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" + ) + if do_training_timing: + t_sync_b = time.perf_counter() + torch.cuda.synchronize() + pending_sync_ms += 1000.0 * (time.perf_counter() - t_sync_b) + val_elapsed_ms = 1000.0 * (time.perf_counter() - t_val0) + val_total_time_ms += val_elapsed_ms + log0(f"val_stage_time_ms:{val_elapsed_ms:.0f}ms step:{step}/{args.iterations}") + 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 + + if do_training_timing: + training_timing.reset() + training_timing.val_ms = pending_val_ms + training_timing.explicit_sync_ms = pending_sync_ms + pending_val_ms = 0.0 + pending_sync_ms = 0.0 + + 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 + micro_timing: dict[str, float] | None = {} if do_training_timing else None + x, y = train_loader.next_batch( + args.train_batch_tokens, args.train_seq_len, grad_accum_steps, micro_timing + ) + if do_training_timing: + training_timing.batch_cpu_ms += micro_timing["batch_cpu_ms"] + training_timing.batch_h2d_ms += micro_timing["batch_h2d_ms"] + if do_training_timing: + torch.cuda.synchronize() + t_fwd0 = time.perf_counter() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + if do_training_timing: + torch.cuda.synchronize() + training_timing.forward_ms += 1000.0 * (time.perf_counter() - t_fwd0) + train_loss += loss.detach() + if do_training_timing: + torch.cuda.synchronize() + t_bwd0 = time.perf_counter() + (loss * grad_scale).backward() + if do_training_timing: + torch.cuda.synchronize() + training_timing.backward_ms += 1000.0 * (time.perf_counter() - t_bwd0) + train_loss /= grad_accum_steps + if do_training_timing: + training_timing.micro_steps = 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: + if do_training_timing: + torch.cuda.synchronize() + t_clip0 = time.perf_counter() + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + if do_training_timing: + training_timing.grad_clip_ms += 1000.0 * (time.perf_counter() - t_clip0) + if do_training_timing: + torch.cuda.synchronize() + t_opt0 = time.perf_counter() + for opt in optimizers: + opt.step() + if do_training_timing: + torch.cuda.synchronize() + training_timing.optimizer_ms += 1000.0 * (time.perf_counter() - t_opt0) + zero_grad_all() + + step += 1 + if do_training_timing and (step <= 10 or step % args.training_timing_every == 0): + log0(training_timing.as_log_line(step, grad_accum_steps)) + 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 + + if hasattr(train_loader, "shutdown"): + train_loader.shutdown() + + log0( + f"timing_train_loop_ms:{training_time_ms:.0f}ms timing_val_total_ms:{val_total_time_ms:.0f}ms" + ) + 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") + + t_quant0 = time.perf_counter() + 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") + log0(f"timing_quantize_ms:{1000.0 * (time.perf_counter() - t_quant0):.0f}ms") + + if distributed: + dist.barrier() + t_dequant0 = time.perf_counter() + 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() + log0(f"timing_roundtrip_dequant_ms:{1000.0 * (time.perf_counter() - t_dequant0):.0f}ms") + 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, + base_bytes_lut_cpu=base_bytes_lut_cpu, + has_leading_space_lut_cpu=has_leading_space_lut_cpu, + is_boundary_token_lut_cpu=is_boundary_token_lut_cpu, + copy_stream=val_copy_stream, + enable_overlap=use_val_overlap, + log_progress=val_log_fn, + ) + 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/2026-03-24-AsyncTrainingValidation/training.log b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/training.log new file mode 100644 index 0000000000..941af9cbd2 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/training.log @@ -0,0 +1,67 @@ +logs/8gb_vram.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:10 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059912 +world_size:1 grad_accum_steps:8 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +torch_compile:on +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:65536 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +train_io:prefetch:True copy_stream:True (disabled when TRAINING_TIMING_BREAKDOWN=1) +val_io:overlap:True prefetch:True copy_stream:True prefetch_queue:2 bytecount_device:cpu +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9357 val_bpb:4.1077 train_time:0ms step_avg:0.01ms +val_stage_time_ms:120730ms step:0/20000 +step:1/20000 train_loss:6.9370 train_time:759ms step_avg:759.04ms +step:2/20000 train_loss:16.8171 train_time:1467ms step_avg:733.60ms +step:3/20000 train_loss:10.6699 train_time:2132ms step_avg:710.67ms +step:4/20000 train_loss:7.4092 train_time:2756ms step_avg:689.04ms +step:5/20000 train_loss:6.8772 train_time:3351ms step_avg:670.24ms +step:6/20000 train_loss:6.8297 train_time:3907ms step_avg:651.18ms +step:7/20000 train_loss:6.7256 train_time:4453ms step_avg:636.11ms +step:8/20000 train_loss:6.6050 train_time:4984ms step_avg:622.98ms +step:9/20000 train_loss:6.4395 train_time:5506ms step_avg:611.79ms +step:10/20000 train_loss:6.1579 train_time:6027ms step_avg:602.69ms +step:200/20000 train_loss:3.4292 train_time:100656ms step_avg:503.28ms +step:400/20000 train_loss:2.8769 train_time:203580ms step_avg:508.95ms +step:600/20000 train_loss:2.7293 train_time:302503ms step_avg:504.17ms +step:800/20000 train_loss:2.6021 train_time:401296ms step_avg:501.62ms +step:1000/20000 train_loss:2.6315 train_time:506084ms step_avg:506.08ms +step:1000/20000 val_loss:2.5866 val_bpb:1.5319 train_time:506089ms step_avg:506.09ms +val_stage_time_ms:122685ms step:1000/20000 +step:1191/20000 val_loss:2.5442 val_bpb:1.5068 train_time:600433ms step_avg:504.14ms +val_stage_time_ms:123100ms step:1191/20000 +stopping_early: wallclock_cap train_time:600433ms step:1191/20000 +timing_train_loop_ms:600433ms timing_val_total_ms:366515ms +peak memory allocated: 1552 MiB reserved: 2280 MiB +Serialized model: 67224983 bytes +Code size: 70245 bytes +Total submission size: 67295228 bytes +Serialized model int8+zlib: 12907011 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x) +Total submission size int8+zlib: 12977256 bytes +timing_quantize_ms:1082ms +timing_roundtrip_dequant_ms:407ms +final_int8_zlib_roundtrip val_loss:2.5457 val_bpb:1.5077 eval_time:125190ms +final_int8_zlib_roundtrip_exact val_loss:2.54570429 val_bpb:1.50770948 \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/training_default.log b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/training_default.log new file mode 100644 index 0000000000..ecb4f75538 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24-AsyncTrainingValidation/training_default.log @@ -0,0 +1,64 @@ +logs/8gb_vram.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:10 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059912 +world_size:1 grad_accum_steps:8 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:65536 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9357 val_bpb:4.1077 train_time:0ms step_avg:0.02ms +val_stage_time_ms:123547ms step:0/20000 +step:1/20000 train_loss:6.9370 train_time:706ms step_avg:705.65ms +step:2/20000 train_loss:17.2611 train_time:1357ms step_avg:678.65ms +step:3/20000 train_loss:10.9883 train_time:1951ms step_avg:650.34ms +step:4/20000 train_loss:7.0795 train_time:2525ms step_avg:631.15ms +step:5/20000 train_loss:6.3779 train_time:3078ms step_avg:615.62ms +step:6/20000 train_loss:6.4241 train_time:3609ms step_avg:601.48ms +step:7/20000 train_loss:6.2784 train_time:4121ms step_avg:588.71ms +step:8/20000 train_loss:6.2256 train_time:4634ms step_avg:579.28ms +step:9/20000 train_loss:6.1498 train_time:5117ms step_avg:568.59ms +step:10/20000 train_loss:6.0087 train_time:5652ms step_avg:565.23ms +step:200/20000 train_loss:3.3391 train_time:103334ms step_avg:516.67ms +step:400/20000 train_loss:2.8629 train_time:214247ms step_avg:535.62ms +step:600/20000 train_loss:2.7177 train_time:316599ms step_avg:527.67ms +step:800/20000 train_loss:2.5858 train_time:423217ms step_avg:529.02ms +step:1000/20000 train_loss:2.6171 train_time:526172ms step_avg:526.17ms +step:1000/20000 val_loss:2.5403 val_bpb:1.5045 train_time:526178ms step_avg:526.18ms +val_stage_time_ms:128796ms step:1000/20000 +step:1137/20000 val_loss:2.5105 val_bpb:1.4869 train_time:600535ms step_avg:528.17ms +val_stage_time_ms:125301ms step:1137/20000 +stopping_early: wallclock_cap train_time:600535ms step:1137/20000 +timing_train_loop_ms:600535ms timing_val_total_ms:377644ms +peak memory allocated: 1552 MiB reserved: 2274 MiB +Serialized model: 67224983 bytes +Code size: 48334 bytes +Total submission size: 67273317 bytes +Serialized model int8+zlib: 12758736 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x) +Total submission size int8+zlib: 12807070 bytes +timing_quantize_ms:965ms +timing_roundtrip_dequant_ms:331ms +final_int8_zlib_roundtrip val_loss:2.5120 val_bpb:1.4877 eval_time:127119ms +final_int8_zlib_roundtrip_exact val_loss:2.51197724 val_bpb:1.48773442 \ No newline at end of file