From f4fd74efd7d5daab8d005fbad24a2ef1941f5484 Mon Sep 17 00:00:00 2001 From: Brent Girolimon Date: Sun, 22 Mar 2026 03:31:46 -0400 Subject: [PATCH] Add LLMAdvisor submission: 1.14638 BPB (track_10min_16mb) 10L Int5/Int6 + BigramHash(10240) + SmearGate + SWA Boost - Mixed int5 MLP / int6 attention quantization + FP16 embeddings + zstd-22 - Reduced batch (622592 tokens) for ~7370 steps in 600s - SWA boost: every=30 steps, start_frac=0.50, 49 averaged checkpoints - Best val_bpb: 1.14638 (seed=1337) - Artifact size: 15,736,555 bytes (under 16MB limit) --- .../README.md | 100 ++ .../submission.json | 21 + .../train_gpt.py | 1423 +++++++++++++++++ .../train_seed1337.log | 220 +++ .../train_seed1337_swa_boost.log | 220 +++ .../train_seed2024.log | 220 +++ 6 files changed, 2204 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/README.md create mode 100644 records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/submission.json create mode 100644 records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_gpt.py create mode 100644 records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337_swa_boost.log create mode 100644 records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed2024.log diff --git a/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/README.md b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/README.md new file mode 100644 index 000000000..e0b65e33d --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/README.md @@ -0,0 +1,100 @@ +# 10L Int5/Int6 + BigramHash(10240) + SmearGate + SWA Boost + +**LLMAdvisor.ai — powered by HighSignal™** + +## Score + +**Best val_bpb = 1.14638** (seed=1337, SWA boost config, 8×H100 SXM, 600s wallclock) + +| Run | Seed | SWA Config | Steps | val_loss | val_bpb | +|-----|------|-----------|-------|----------|---------| +| SWA boost | 1337 | every=30, frac=0.50 (49 ckpts) | 7,372 | 1.93562 | **1.14638** | +| Standard | 1337 | every=50, frac=0.36 (21 ckpts) | 7,376 | 1.93571 | 1.14644 | +| WD=2250 | 2024 | every=50, frac=0.36 | 7,387 | 1.93680 | 1.14709 | + +Artifact size: **15,736,555 bytes** (263,445 bytes under the 16MB limit) + +## Run Command + +```bash +# Setup (once) +pip install sentencepiece zstandard huggingface_hub +python cached_challenge_fineweb.py --variant sp1024 + +# Train + evaluate (best config: SWA boost) +SEED=1337 SWA_EVERY=30 SWA_START_FRAC=0.50 \ + torchrun --nproc_per_node=8 --standalone train_gpt.py +``` + +Default env vars reproduce the standard run. Override `SEED`, `SWA_EVERY`, and `SWA_START_FRAC` for the SWA boost config above. + +## Approach + +### 1. Mixed Int5/Int6 Quantization +- **Int5 [-16,15]** for MLP weights — saves ~1.86MB vs uniform int6, funding the 10th layer +- **Int6 [-32,31]** for attention weights — precision-sensitive +- **FP16** for tied embeddings and last-layer key projections +- **zstd level 22** compression + +### 2. BigramHash(10240, dim=128) +- Hash consecutive token pairs into 10,240-bucket embedding table (dim=128) +- Projected to model dim=512 via learned linear — captures local token-pair context + +### 3. SmearGate +- Learned per-dimension gate blending current + previous token embeddings +- Initialized near-identity for stable early training + +### 4. SWA Density Sweep +- **SWA boost**: every=30 steps, start_frac=0.50 → 49 averaged checkpoints (best: 1.14638) +- **Standard**: every=50 steps, start_frac=0.36 → 21 averaged checkpoints (1.14644) +- Denser SWA collection provides marginal but consistent improvement + +### 5. Reduced Batch Size (622,592 tokens) +- 75% of the standard 786K batch → ~81ms/step (vs ~117ms at full batch) +- ~7,370 training steps in 600s wallclock (vs ~5,100) +- More steps overcomes slightly noisier gradients + +## Architecture + +| Parameter | Value | +|-----------|-------| +| Layers | 10 | +| Model dim | 512 | +| Heads | 8 (4 KV heads, GQA) | +| MLP hidden | 1536 (3× expansion) | +| Activation | relu² | +| Vocab size | 1024 (sp1024 BPE) | +| Embeddings | Tied input/output, FP16 | +| Init | Orthogonal with muP-scaled outputs | +| Skip connections | U-Net style | + +## Training Hyperparameters + +| Parameter | Value | +|-----------|-------| +| Optimizer (matrix) | Muon, lr=0.02, momentum=0.99 | +| Optimizer (embed/scalar) | AdamW, lr=0.02 | +| Weight decay | 0.04 (decoupled) | +| Batch size | 622,592 tokens | +| Sequence length | 2,048 | +| Warmup | 20 steps | +| Warmdown | 3,000 iters | +| Gradient clipping | 0.3, 3% magnitude pruning | +| SWA (boost) | start_frac=0.50, every=30 steps | +| Wall clock cap | 600 seconds | + +## Evaluation + +- Sliding-window evaluation with stride=64 +- BPB = (val_loss / ln(2)) × (tokens / bytes) +- Eval time: ~259 seconds + +## Hardware + +- 8× NVIDIA H100 SXM GPUs +- RunPod cloud instance +- DDP training with torchrun + +## Acknowledgments + +Built on the SOTA techniques from [thwu1](https://github.com/KellerJordan/parameter-golf/tree/main/records/track_10min_16mb/2026-03-20_10L_Int5MLP_MuonWD04_SWA50) (1.14276 BPB) and [Raahil Shah](https://github.com/KellerJordan/parameter-golf/tree/main/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA) (1.1458 BPB). Key adaptations: reduced batch size for faster step throughput, SWA density sweep (every=30/frac=0.50 vs every=50/frac=0.40), and PyTorch version auto-detection for GQA compatibility. diff --git a/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/submission.json b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/submission.json new file mode 100644 index 000000000..84bc772db --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/submission.json @@ -0,0 +1,21 @@ +{ + "author": "LLMAdvisor.ai", + "github_id": "harborglowvintage-oss", + "name": "10L Int5/Int6 + BigramHash(10240) + SmearGate + SWA Boost", + "blurb": "10L GQA (8h/4kv), MLP 3x relu^2, BigramHash(10240, dim=128), SmearGate, orthogonal init, U-Net skips. Mixed int5 MLP / int6 attention quantization + FP16 embeddings + zstd-22. Reduced batch (622592 tokens) for faster step throughput (~81ms, ~7370 steps in 600s). SWA boost: every=30 steps, start_frac=0.50, averaging 49 checkpoints. Sliding-window eval stride=64.", + "date": "2026-03-22", + "val_loss": 1.93561706, + "val_bpb": 1.14638448, + "seeds": [1337], + "seed_results": { + "1337_swa_boost": {"val_loss": 1.93561706, "val_bpb": 1.14638448, "note": "SWA_EVERY=30, SWA_START_FRAC=0.50, 49 ckpts"}, + "1337_standard": {"val_loss": 1.93570710, "val_bpb": 1.14643781, "note": "SWA_EVERY=50, SWA_START_FRAC=0.36, 21 ckpts"}, + "2024_wd2250": {"val_loss": 1.93680082, "val_bpb": 1.14708558, "note": "WARMDOWN=2250"} + }, + "step_stop": 7372, + "wallclock_seconds": 599.989, + "eval_time_seconds": 258.721, + "bytes_total": 15736555, + "bytes_model_int6_zstd": 15673630, + "bytes_code": 62925 +} diff --git a/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_gpt.py b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_gpt.py new file mode 100644 index 000000000..4be7c518f --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_gpt.py @@ -0,0 +1,1423 @@ +""" +train_gpt_custom.py — Competition submission for Parameter Golf (track_10min_16mb). + +Targets SOTA 1.14276 BPB (thwu1) with identical architecture + tuned defaults: + - VOCAB_SIZE=1024 (sp1024, available on HuggingFace via competition runner) + - BigramHash(10240, dim=128), SmearGate, OrthoInit + - 10L, 512 dim, 8 heads, 4 KV heads (GQA), MLP 3x, relu^2 + - Int5-MLP / Int6-Attn mixed quantisation + zstd-22 compression + - SWA start_frac=0.50, every=30 steps + - MAX_WALLCLOCK_SECONDS=600, WARMDOWN_ITERS=3000, TRAIN_BATCH_TOKENS=622592 + - SMOKE_TEST=1: 500 iters, no SWA, no wallclock cap — for single-GPU validation + - Use DATA_PATH/TOKENIZER_PATH/VOCAB_SIZE env vars to override for sp3072 experiments +""" + +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 + +def _rms_norm(input: Tensor, normalized_shape, weight: "Tensor | None" = None, eps: float = 1e-5) -> Tensor: + return F.rms_norm(input, normalized_shape, weight, eps) + +_pytorch_version = tuple(int(x) for x in torch.__version__.split(".")[:2]) +_has_enable_gqa = _pytorch_version >= (2, 5) + +# ----------------------------- +# 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", 1337)) + + 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", 622_592)) + 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)) + + 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", 0.50)) + swa_every = int(os.environ.get("SWA_EVERY", 30)) + + # Experiment flags + stochastic_depth = float(os.environ.get("STOCHASTIC_DEPTH", 0.0)) + lr_schedule = os.environ.get("LR_SCHEDULE", "linear") + progressive_seq = bool(int(os.environ.get("PROGRESSIVE_SEQ", "0"))) + encoder_layers_override = int(os.environ.get("ENCODER_LAYERS", 0)) + wd_warmup_frac = float(os.environ.get("WD_WARMUP_FRAC", 0.0)) + embed_dropout = float(os.environ.get("EMBED_DROPOUT", 0.0)) + reverse_shards = bool(int(os.environ.get("REVERSE_SHARDS", "0"))) + +# ----------------------------- +# 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 +) +# At vocab >= 2048 the embedding table is large enough that keeping it in FP16 +# would cost ~2 MB (compressed) and push the submission over 16 MB. Drop it +# from the FP16 keep-list so it falls through to int8 quantisation instead. +# Override with FP16_KEEP_NAME_PATTERNS env-var to restore FP16 for small-vocab runs. +_default_fp16_keep = ( + "tok_emb,blocks.8.attn.c_k" + if int(os.environ.get("VOCAB_SIZE", 3072)) <= 1024 + else "blocks.8.attn.c_k" +) +FP16_KEEP_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get("FP16_KEEP_NAME_PATTERNS", _default_fp16_keep).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 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) + scale = scale.clamp_min(torch.finfo(torch.float16).tiny).to(torch.float16) + 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 mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 8192: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if any(pattern in name for pattern in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if cat in int6_cats and t.ndim >= 1: + clip = 15 if cat == "mlp" else 31 # int5 for MLP, int6 for attention + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{5 if cat == 'mlp' else 6}"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(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, reverse: bool = False): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern, reverse=reverse) + + 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 _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 = _rms_norm(q, (q.size(-1),)) + k = _rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + if _has_enable_gqa: + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + else: + if self.num_kv_heads != self.num_heads: + r = self.num_heads // self.num_kv_heads + k = k.repeat_interleave(r, dim=1) + v = v.repeat_interleave(r, dim=1) + y = F.scaled_dot_product_attention(q, k, v, attn_mask=None, is_causal=True) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + 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, + stochastic_depth: float = 0.0, + encoder_layers_override: int = 0, + embed_dropout: float = 0.0, + ): + 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 + # Bake these as Python bools at construction time so torch.compile + # fullgraph sees no dynamic branches — zero cost when disabled. + self._use_stochastic_depth = stochastic_depth > 0.0 + self.stochastic_depth = stochastic_depth + self._use_embed_dropout = embed_dropout > 0.0 + self.embed_dropout_p = embed_dropout + 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 = encoder_layers_override if encoder_layers_override > 0 else 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 = _rms_norm(x, (x.size(-1),)) + if self._use_embed_dropout and self.training: + x = F.dropout(x, p=self.embed_dropout_p, training=True) + x = self.smear(x) + x0 = x + num_total = len(self.blocks) + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + if self._use_stochastic_depth and self.training: + drop_prob = (i / max(num_total - 1, 1)) * self.stochastic_depth + if random.random() < drop_prob: + skips.append(x) + continue + 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() + block_idx = self.num_encoder_layers + i + if self._use_stochastic_depth and self.training: + drop_prob = (block_idx / max(num_total - 1, 1)) * self.stochastic_depth + if random.random() < drop_prob: + continue + x = self.blocks[block_idx](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 = _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, +) -> 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] + 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] + + 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 (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 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) + + 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 + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + smoke_test = bool(int(os.environ.get("SMOKE_TEST", "0"))) + + # SMOKE TEST MODE: fast single-GPU pipeline validation (e.g. on Jetson AGX Orin). + # Set SMOKE_TEST=1 to activate. Individual overrides still respect ITERATIONS etc. + if smoke_test: + args.iterations = int(os.environ.get("ITERATIONS", "2")) + args.warmdown_iters = min(args.warmdown_iters, 2) + args.warmup_steps = min(args.warmup_steps, 3) + args.val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", "0")) + args.train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", "1")) + args.train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", min(args.train_seq_len, 1024))) + args.train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", min(args.train_batch_tokens, 65_536))) + args.val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", min(args.val_batch_size, 131_072))) + args.eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", min(args.eval_batch_seqs, 8))) + args.swa_enabled = False + args.max_wallclock_seconds = 0.0 # no wallclock cap during smoke test + + # torch.compile: use inductor backend for full speed on H100. + # Disable with TORCHDYNAMO_DISABLE=1 for debugging or older PyTorch. + _pytorch_version = tuple(int(x) for x in torch.__version__.split('.')[:2]) + _use_compile = _pytorch_version >= (2, 2) and not bool(int(os.environ.get("TORCHDYNAMO_DISABLE", "0"))) + if _use_compile: + 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_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + try: + from torch.backends.cuda import enable_cudnn_sdp + enable_cudnn_sdp(False) + except ImportError: + pass # enable_cudnn_sdp added in PyTorch 2.2; absent in JetPack 5.1.2's 2.1.0a + # Test whether flash SDPA actually works at runtime (it may not be compiled for this arch). + _flash_works = False + try: + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + _q = torch.randn(1, 1, 4, 32, device=device, dtype=torch.bfloat16) + torch.nn.functional.scaled_dot_product_attention(_q, _q, _q) + _flash_works = True + del _q + except RuntimeError: + pass + if not _flash_works: + enable_flash_sdp(False) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + if master_process: + print("[COMPAT] Flash SDPA not available — using mem_efficient + math backends") + + 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) + try: + smi = subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout + except FileNotFoundError: + smi = "(nvidia-smi not available — Jetson uses tegrastats)" + log0(smi, 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) + if smoke_test: + smoke_val_tokens = int(os.environ.get("SMOKE_VAL_TOKENS", str(args.train_seq_len * 64))) + val_tokens = val_tokens[: min(val_tokens.numel(), smoke_val_tokens + 1)] + max_val_tokens = int(os.environ.get("MAX_VAL_TOKENS", "0")) + if max_val_tokens > 0 and not smoke_test: + val_tokens = val_tokens[: min(val_tokens.numel(), max_val_tokens + 1)] + log0(f"MAX_VAL_TOKENS={max_val_tokens} — capping validation to {val_tokens.numel() - 1} tokens") + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # 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, + stochastic_depth=args.stochastic_depth, + encoder_layers_override=args.encoder_layers_override, + embed_dropout=args.embed_dropout, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if _use_compile: + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + else: + compiled_model = base_model + if master_process: + print("[INFO] torch.compile disabled (TORCHDYNAMO_DISABLE=1 or PyTorch < 2.2) — running eager mode") + 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, reverse=args.reverse_shards) + + 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.lr_schedule == "cosine_restart": + T_0 = max(args.iterations // 3, 1) + t = step % T_0 + base = 0.5 * (1 + math.cos(math.pi * t / T_0)) + if args.warmdown_iters > 0: + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + if warmdown_start <= step < args.iterations: + base *= max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) + else: + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + if remaining_ms <= warmdown_ms: + base *= remaining_ms / max(warmdown_ms, 1e-9) + return base + if args.lr_schedule == "inv_sqrt": + base = 1.0 / math.sqrt(max(step, 1)) + if args.warmdown_iters > 0: + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + if warmdown_start <= step < args.iterations: + base *= max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) + else: + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + if remaining_ms <= warmdown_ms: + base *= remaining_ms / max(warmdown_ms, 1e-9) + return base + # Default: linear warmdown + 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) + if len(optimizers) != len(initial_optimizer_states): + raise RuntimeError("optimizer warmup state mismatch") + for opt, state in zip(optimizers, initial_optimizer_states): + 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, reverse=args.reverse_shards) + + # 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 > 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) + + # Progressive sequence length: ramp from 64 to target + current_seq_len = args.train_seq_len + if args.progressive_seq: + max_pre_warmdown = max(args.iterations - args.warmdown_iters, 1) + frac = min(step / max_pre_warmdown, 1.0) + targets = [s for s in [64, 128, 256, 512, 1024, 2048] if s <= args.train_seq_len] + if not targets: + targets = [args.train_seq_len] + idx = min(int(frac * len(targets)), len(targets) - 1) + current_seq_len = targets[idx] + + # Weight decay warmup + if args.wd_warmup_frac > 0: + wd_ramp = min(step / max(args.iterations * args.wd_warmup_frac, 1), 1.0) + current_wd = args.weight_decay * wd_ramp + for opt in optimizers: + for group in opt.param_groups: + if group.get("weight_decay", 0) > 0 or wd_ramp < 1.0: + group["weight_decay"] = current_wd + + 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, current_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" + ) + + if smoke_test: + log0("smoke_test: training path validated; skipping export, quantization, and final eval") + if distributed: + dist.destroy_process_group() + return + + # 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) + + # SERIALIZATION + ROUNDTRIP VALIDATION + 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") + + # 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) + + # INT6 mixed quantization + zstd/zlib export + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn", "bigram"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = zlib.compress(quant_raw, 9) + 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")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + total_submission_bytes = quant_file_bytes + code_bytes + _limit = 16_000_000 # decimal 16 MB per contest rules (NOT 16 MiB) + if total_submission_bytes > _limit: + log0( + f"WARNING: submission too large! " + f"{total_submission_bytes:,} bytes > {_limit:,} bytes limit — " + f"reduce bigram_vocab_size, num_layers, or model_dim." + ) + else: + log0( + f"Total submission size: {total_submission_bytes:,} bytes " + f"({_limit - total_submission_bytes:,} bytes under the 16 MB limit)" + ) + + if distributed: + dist.barrier() + 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", weights_only=True) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + base_model.load_state_dict(deq_state, strict=True) + + # Sliding window eval on int6-roundtripped 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} batch_seqs:{args.eval_batch_seqs}") + 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, + ) + 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.destroy_process_group() + + +if __name__ == "__main__": + main() +# fixes applied +# tuned diff --git a/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337.log b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337.log new file mode 100644 index 000000000..49decd147 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337.log @@ -0,0 +1,220 @@ +W0322 06:26:49.977000 138264933192320 torch/distributed/run.py:779] +W0322 06:26:49.977000 138264933192320 torch/distributed/run.py:779] ***************************************** +W0322 06:26:49.977000 138264933192320 torch/distributed/run.py:779] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0322 06:26:49.977000 138264933192320 torch/distributed/run.py:779] ***************************************** +logs/71eed499-68c2-48d9-987c-bf2d711ddf70.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:622592 train_seq_len:2048 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:1/20000 train_loss:6.9301 train_time:164ms step_avg:163.80ms +step:2/20000 train_loss:7.9505 train_time:230ms step_avg:115.24ms +step:3/20000 train_loss:7.5314 train_time:316ms step_avg:105.49ms +step:4/20000 train_loss:6.9175 train_time:403ms step_avg:100.70ms +step:5/20000 train_loss:6.7833 train_time:490ms step_avg:98.05ms +step:6/20000 train_loss:6.6808 train_time:577ms step_avg:96.24ms +step:7/20000 train_loss:6.5274 train_time:664ms step_avg:94.81ms +step:8/20000 train_loss:6.3556 train_time:749ms step_avg:93.61ms +step:9/20000 train_loss:6.2316 train_time:835ms step_avg:92.80ms +step:10/20000 train_loss:6.1070 train_time:922ms step_avg:92.20ms +step:100/20000 train_loss:3.2005 train_time:8129ms step_avg:81.29ms +step:200/20000 train_loss:2.6128 train_time:16277ms step_avg:81.39ms +step:300/20000 train_loss:2.5887 train_time:24327ms step_avg:81.09ms +step:400/20000 train_loss:2.4322 train_time:32468ms step_avg:81.17ms +step:500/20000 train_loss:2.4257 train_time:40620ms step_avg:81.24ms +step:500/20000 val_loss:2.3749 val_bpb:1.4066 train_time:40650ms step_avg:81.30ms +step:600/20000 train_loss:2.2457 train_time:48646ms step_avg:81.08ms +step:700/20000 train_loss:2.3116 train_time:56792ms step_avg:81.13ms +step:800/20000 train_loss:2.2702 train_time:64835ms step_avg:81.04ms +step:900/20000 train_loss:2.2569 train_time:72986ms step_avg:81.10ms +step:1000/20000 train_loss:2.1637 train_time:81139ms step_avg:81.14ms +step:1000/20000 val_loss:2.2551 val_bpb:1.3356 train_time:81169ms step_avg:81.17ms +step:1100/20000 train_loss:2.1843 train_time:89175ms step_avg:81.07ms +step:1200/20000 train_loss:2.2120 train_time:97331ms step_avg:81.11ms +step:1300/20000 train_loss:2.2095 train_time:105491ms step_avg:81.15ms +step:1400/20000 train_loss:2.0905 train_time:113560ms step_avg:81.11ms +step:1500/20000 train_loss:2.1510 train_time:121735ms step_avg:81.16ms +step:1500/20000 val_loss:2.2083 val_bpb:1.3079 train_time:121763ms step_avg:81.18ms +step:1600/20000 train_loss:2.1692 train_time:129802ms step_avg:81.13ms +step:1700/20000 train_loss:2.2663 train_time:137985ms step_avg:81.17ms +step:1800/20000 train_loss:2.2443 train_time:146172ms step_avg:81.21ms +step:1900/20000 train_loss:2.1946 train_time:154250ms step_avg:81.18ms +step:2000/20000 train_loss:2.0001 train_time:162440ms step_avg:81.22ms +step:2000/20000 val_loss:2.1576 val_bpb:1.2778 train_time:162470ms step_avg:81.23ms +step:2100/20000 train_loss:2.0391 train_time:170604ms step_avg:81.24ms +step:2200/20000 train_loss:2.0356 train_time:178685ms step_avg:81.22ms +step:2300/20000 train_loss:2.1122 train_time:186874ms step_avg:81.25ms +step:2400/20000 train_loss:2.1327 train_time:194957ms step_avg:81.23ms +step:2500/20000 train_loss:2.0504 train_time:203145ms step_avg:81.26ms +step:2500/20000 val_loss:2.1308 val_bpb:1.2620 train_time:203177ms step_avg:81.27ms +step:2600/20000 train_loss:2.1861 train_time:211333ms step_avg:81.28ms +step:2700/20000 train_loss:2.3065 train_time:219425ms step_avg:81.27ms +step:2800/20000 train_loss:2.1466 train_time:227615ms step_avg:81.29ms +step:2900/20000 train_loss:2.1965 train_time:235800ms step_avg:81.31ms +step:3000/20000 train_loss:2.0078 train_time:243895ms step_avg:81.30ms +step:3000/20000 val_loss:2.1162 val_bpb:1.2533 train_time:243925ms step_avg:81.31ms +step:3100/20000 train_loss:2.0092 train_time:252067ms step_avg:81.31ms +step:3200/20000 train_loss:2.1343 train_time:260135ms step_avg:81.29ms +step:3300/20000 train_loss:2.0824 train_time:268300ms step_avg:81.30ms +step:3400/20000 train_loss:2.0699 train_time:276459ms step_avg:81.31ms +step:3500/20000 train_loss:2.2127 train_time:284519ms step_avg:81.29ms +step:3500/20000 val_loss:2.1012 val_bpb:1.2444 train_time:284550ms step_avg:81.30ms +step:3600/20000 train_loss:1.9929 train_time:292693ms step_avg:81.30ms +step:3700/20000 train_loss:2.1146 train_time:300861ms step_avg:81.31ms +step:3800/20000 train_loss:2.1147 train_time:308933ms step_avg:81.30ms +step:3900/20000 train_loss:1.9960 train_time:317097ms step_avg:81.31ms +step:4000/20000 train_loss:1.9919 train_time:325161ms step_avg:81.29ms +step:4000/20000 val_loss:2.0996 val_bpb:1.2435 train_time:325191ms step_avg:81.30ms +step:4100/20000 train_loss:2.0132 train_time:333322ms step_avg:81.30ms +step:4200/20000 train_loss:2.3421 train_time:341486ms step_avg:81.31ms +step:4300/20000 train_loss:2.1821 train_time:349553ms step_avg:81.29ms +step:4400/20000 train_loss:2.1703 train_time:357716ms step_avg:81.30ms +step:4500/20000 train_loss:2.0917 train_time:365881ms step_avg:81.31ms +step:4500/20000 val_loss:2.0905 val_bpb:1.2381 train_time:365910ms step_avg:81.31ms +step:4600/20000 train_loss:2.1420 train_time:373922ms step_avg:81.29ms +step:4700/20000 train_loss:2.1092 train_time:382058ms step_avg:81.29ms +step:4800/20000 train_loss:1.9908 train_time:390116ms step_avg:81.27ms +step:4900/20000 train_loss:2.0166 train_time:398266ms step_avg:81.28ms +step:5000/20000 train_loss:2.1695 train_time:406421ms step_avg:81.28ms +step:5000/20000 val_loss:2.0740 val_bpb:1.2283 train_time:406451ms step_avg:81.29ms +step:5100/20000 train_loss:2.0517 train_time:414485ms step_avg:81.27ms +step:5200/20000 train_loss:2.0826 train_time:422646ms step_avg:81.28ms +step:5300/20000 train_loss:1.9705 train_time:430707ms step_avg:81.27ms +step:5400/20000 train_loss:2.0683 train_time:438854ms step_avg:81.27ms +step:5500/20000 train_loss:2.1024 train_time:447003ms step_avg:81.27ms +step:5500/20000 val_loss:2.0478 val_bpb:1.2128 train_time:447033ms step_avg:81.28ms +step:5600/20000 train_loss:2.0085 train_time:455043ms step_avg:81.26ms +step:5700/20000 train_loss:2.0026 train_time:463184ms step_avg:81.26ms +step:5800/20000 train_loss:2.5597 train_time:471334ms step_avg:81.26ms +step:5900/20000 train_loss:2.0118 train_time:479373ms step_avg:81.25ms +step:6000/20000 train_loss:2.0583 train_time:487513ms step_avg:81.25ms +step:6000/20000 val_loss:2.0241 val_bpb:1.1988 train_time:487544ms step_avg:81.26ms +step:6100/20000 train_loss:2.0268 train_time:495560ms step_avg:81.24ms +step:6200/20000 train_loss:1.9324 train_time:503710ms step_avg:81.24ms +step:6300/20000 train_loss:2.0849 train_time:511882ms step_avg:81.25ms +swa:start step:6350 +step:6400/20000 train_loss:1.8707 train_time:520067ms step_avg:81.26ms +step:6500/20000 train_loss:1.9212 train_time:528268ms step_avg:81.27ms +step:6500/20000 val_loss:1.9985 val_bpb:1.1836 train_time:528330ms step_avg:81.28ms +step:6600/20000 train_loss:2.0824 train_time:536496ms step_avg:81.29ms +step:6700/20000 train_loss:1.9673 train_time:544593ms step_avg:81.28ms +step:6800/20000 train_loss:1.9258 train_time:552795ms step_avg:81.29ms +step:6900/20000 train_loss:1.9872 train_time:560899ms step_avg:81.29ms +step:7000/20000 train_loss:1.9384 train_time:569116ms step_avg:81.30ms +step:7000/20000 val_loss:1.9706 val_bpb:1.1671 train_time:569185ms step_avg:81.31ms +step:7100/20000 train_loss:2.0271 train_time:577345ms step_avg:81.32ms +step:7200/20000 train_loss:2.0117 train_time:585467ms step_avg:81.31ms +step:7300/20000 train_loss:1.9759 train_time:593713ms step_avg:81.33ms +step:7376/20000 val_loss:1.9540 val_bpb:1.1573 train_time:599935ms step_avg:81.34ms +stopping_early: wallclock_cap train_time:599935ms step:7376/20000 +peak memory allocated: 18844 MiB reserved: 18918 MiB +swa:applying averaged 21 checkpoints +Serialized model: 98437014 bytes +Code size: 62925 bytes +Total submission size: 98499939 bytes +Serialized model int6+zstd: 15789200 bytes +Total submission size: 15,852,125 bytes (147,875 bytes under the 16 MB limit) +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.211444 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.141330 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.142880 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.136481 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.148720 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.149808 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.151103 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.146336 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.143560 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.145279 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.154039 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.152509 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.153746 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.152100 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.150665 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.151044 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.152476 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.152875 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.158980 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.156416 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.157399 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.156048 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.155404 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.154897 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.155577 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.153208 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.152181 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.152514 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.151301 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.151143 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.150407 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.151602 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.152656 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.153174 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.152676 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.153056 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.152133 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.148218 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.148312 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.149234 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.149397 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.149282 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.148033 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.147770 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.147035 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.147100 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.147065 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.147241 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.146970 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.147530 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.147822 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.147537 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.148564 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.150459 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.149755 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.150473 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.150854 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.150861 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.150456 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.150649 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.150049 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.152856 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.152862 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.152906 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.152559 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.152038 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.151281 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.151227 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.151846 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.151867 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.151868 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.152315 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.152058 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.151677 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.151958 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.152046 +final_int8_zlib_roundtrip val_loss:1.9357 val_bpb:1.1464 eval_time:258650ms +final_int8_zlib_roundtrip_exact val_loss:1.93570710 val_bpb:1.14643781 diff --git a/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337_swa_boost.log b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337_swa_boost.log new file mode 100644 index 000000000..12c4ae1e5 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed1337_swa_boost.log @@ -0,0 +1,220 @@ +W0322 07:04:32.776000 132578243961472 torch/distributed/run.py:779] +W0322 07:04:32.776000 132578243961472 torch/distributed/run.py:779] ***************************************** +W0322 07:04:32.776000 132578243961472 torch/distributed/run.py:779] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0322 07:04:32.776000 132578243961472 torch/distributed/run.py:779] ***************************************** +logs/c4112fb9-8542-4758-ac8e-12ee2597cefb.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:622592 train_seq_len:2048 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:1/20000 train_loss:6.9301 train_time:212ms step_avg:211.55ms +step:2/20000 train_loss:7.9505 train_time:276ms step_avg:138.25ms +step:3/20000 train_loss:7.5434 train_time:363ms step_avg:120.97ms +step:4/20000 train_loss:6.8867 train_time:449ms step_avg:112.28ms +step:5/20000 train_loss:6.7795 train_time:537ms step_avg:107.31ms +step:6/20000 train_loss:6.7055 train_time:624ms step_avg:104.00ms +step:7/20000 train_loss:6.5380 train_time:711ms step_avg:101.54ms +step:8/20000 train_loss:6.3444 train_time:798ms step_avg:99.69ms +step:9/20000 train_loss:6.2448 train_time:884ms step_avg:98.24ms +step:10/20000 train_loss:6.1196 train_time:970ms step_avg:96.99ms +step:100/20000 train_loss:3.1995 train_time:8185ms step_avg:81.85ms +step:200/20000 train_loss:2.6124 train_time:16325ms step_avg:81.62ms +step:300/20000 train_loss:2.5884 train_time:24360ms step_avg:81.20ms +step:400/20000 train_loss:2.4356 train_time:32512ms step_avg:81.28ms +step:500/20000 train_loss:2.4287 train_time:40676ms step_avg:81.35ms +step:500/20000 val_loss:2.3748 val_bpb:1.4065 train_time:40705ms step_avg:81.41ms +step:600/20000 train_loss:2.2481 train_time:48678ms step_avg:81.13ms +step:700/20000 train_loss:2.3140 train_time:56820ms step_avg:81.17ms +step:800/20000 train_loss:2.2707 train_time:64859ms step_avg:81.07ms +step:900/20000 train_loss:2.2576 train_time:72983ms step_avg:81.09ms +step:1000/20000 train_loss:2.1669 train_time:81119ms step_avg:81.12ms +step:1000/20000 val_loss:2.2554 val_bpb:1.3358 train_time:81149ms step_avg:81.15ms +step:1100/20000 train_loss:2.1834 train_time:89169ms step_avg:81.06ms +step:1200/20000 train_loss:2.2120 train_time:97320ms step_avg:81.10ms +step:1300/20000 train_loss:2.2060 train_time:105471ms step_avg:81.13ms +step:1400/20000 train_loss:2.0872 train_time:113515ms step_avg:81.08ms +step:1500/20000 train_loss:2.1528 train_time:121661ms step_avg:81.11ms +step:1500/20000 val_loss:2.2076 val_bpb:1.3074 train_time:121691ms step_avg:81.13ms +step:1600/20000 train_loss:2.1646 train_time:129717ms step_avg:81.07ms +step:1700/20000 train_loss:2.2615 train_time:137861ms step_avg:81.09ms +step:1800/20000 train_loss:2.2430 train_time:146021ms step_avg:81.12ms +step:1900/20000 train_loss:2.1938 train_time:154081ms step_avg:81.10ms +step:2000/20000 train_loss:1.9978 train_time:162224ms step_avg:81.11ms +step:2000/20000 val_loss:2.1572 val_bpb:1.2776 train_time:162254ms step_avg:81.13ms +step:2100/20000 train_loss:2.0348 train_time:170357ms step_avg:81.12ms +step:2200/20000 train_loss:2.0354 train_time:178406ms step_avg:81.09ms +step:2300/20000 train_loss:2.1107 train_time:186553ms step_avg:81.11ms +step:2400/20000 train_loss:2.1261 train_time:194591ms step_avg:81.08ms +step:2500/20000 train_loss:2.0523 train_time:202729ms step_avg:81.09ms +step:2500/20000 val_loss:2.1280 val_bpb:1.2603 train_time:202758ms step_avg:81.10ms +step:2600/20000 train_loss:2.1872 train_time:210875ms step_avg:81.11ms +step:2700/20000 train_loss:2.3111 train_time:218917ms step_avg:81.08ms +step:2800/20000 train_loss:2.1436 train_time:227058ms step_avg:81.09ms +step:2900/20000 train_loss:2.1934 train_time:235179ms step_avg:81.10ms +step:3000/20000 train_loss:2.0093 train_time:243233ms step_avg:81.08ms +step:3000/20000 val_loss:2.1161 val_bpb:1.2533 train_time:243264ms step_avg:81.09ms +step:3100/20000 train_loss:2.0080 train_time:251400ms step_avg:81.10ms +step:3200/20000 train_loss:2.1320 train_time:259449ms step_avg:81.08ms +step:3300/20000 train_loss:2.0743 train_time:267602ms step_avg:81.09ms +step:3400/20000 train_loss:2.0718 train_time:275756ms step_avg:81.10ms +step:3500/20000 train_loss:2.2123 train_time:283803ms step_avg:81.09ms +step:3500/20000 val_loss:2.1008 val_bpb:1.2442 train_time:283834ms step_avg:81.10ms +step:3600/20000 train_loss:1.9894 train_time:291958ms step_avg:81.10ms +step:3700/20000 train_loss:2.1112 train_time:300119ms step_avg:81.11ms +step:3800/20000 train_loss:2.1172 train_time:308178ms step_avg:81.10ms +step:3900/20000 train_loss:1.9921 train_time:316350ms step_avg:81.12ms +step:4000/20000 train_loss:1.9912 train_time:324420ms step_avg:81.10ms +step:4000/20000 val_loss:2.0989 val_bpb:1.2431 train_time:324450ms step_avg:81.11ms +step:4100/20000 train_loss:2.0137 train_time:332601ms step_avg:81.12ms +step:4200/20000 train_loss:2.3431 train_time:340747ms step_avg:81.13ms +step:4300/20000 train_loss:2.1806 train_time:348807ms step_avg:81.12ms +step:4400/20000 train_loss:2.1689 train_time:356963ms step_avg:81.13ms +step:4500/20000 train_loss:2.0927 train_time:365115ms step_avg:81.14ms +step:4500/20000 val_loss:2.0915 val_bpb:1.2387 train_time:365145ms step_avg:81.14ms +step:4600/20000 train_loss:2.1402 train_time:373174ms step_avg:81.12ms +step:4700/20000 train_loss:2.1116 train_time:381337ms step_avg:81.14ms +step:4800/20000 train_loss:1.9868 train_time:389407ms step_avg:81.13ms +step:4900/20000 train_loss:2.0189 train_time:397581ms step_avg:81.14ms +step:5000/20000 train_loss:2.1679 train_time:405748ms step_avg:81.15ms +step:5000/20000 val_loss:2.0735 val_bpb:1.2280 train_time:405779ms step_avg:81.16ms +step:5100/20000 train_loss:2.0522 train_time:413816ms step_avg:81.14ms +step:5200/20000 train_loss:2.0806 train_time:421978ms step_avg:81.15ms +step:5300/20000 train_loss:1.9688 train_time:430040ms step_avg:81.14ms +step:5400/20000 train_loss:2.0698 train_time:438197ms step_avg:81.15ms +step:5500/20000 train_loss:2.1000 train_time:446364ms step_avg:81.16ms +step:5500/20000 val_loss:2.0473 val_bpb:1.2126 train_time:446394ms step_avg:81.16ms +step:5600/20000 train_loss:2.0088 train_time:454431ms step_avg:81.15ms +step:5700/20000 train_loss:2.0027 train_time:462605ms step_avg:81.16ms +step:5800/20000 train_loss:2.5674 train_time:470782ms step_avg:81.17ms +step:5900/20000 train_loss:2.0129 train_time:478862ms step_avg:81.16ms +swa:start step:5910 +step:6000/20000 train_loss:2.0580 train_time:487216ms step_avg:81.20ms +step:6000/20000 val_loss:2.0233 val_bpb:1.1983 train_time:487276ms step_avg:81.21ms +step:6100/20000 train_loss:2.0298 train_time:495370ms step_avg:81.21ms +step:6200/20000 train_loss:1.9317 train_time:503603ms step_avg:81.23ms +step:6300/20000 train_loss:2.0821 train_time:511831ms step_avg:81.24ms +step:6400/20000 train_loss:1.8733 train_time:519993ms step_avg:81.25ms +step:6500/20000 train_loss:1.9204 train_time:528224ms step_avg:81.27ms +step:6500/20000 val_loss:1.9977 val_bpb:1.1832 train_time:528254ms step_avg:81.27ms +step:6600/20000 train_loss:2.0755 train_time:536481ms step_avg:81.29ms +step:6700/20000 train_loss:1.9668 train_time:544653ms step_avg:81.29ms +step:6800/20000 train_loss:1.9274 train_time:552930ms step_avg:81.31ms +step:6900/20000 train_loss:1.9860 train_time:561096ms step_avg:81.32ms +step:7000/20000 train_loss:1.9374 train_time:569373ms step_avg:81.34ms +step:7000/20000 val_loss:1.9695 val_bpb:1.1665 train_time:569404ms step_avg:81.34ms +step:7100/20000 train_loss:2.0215 train_time:577678ms step_avg:81.36ms +step:7200/20000 train_loss:2.0132 train_time:585839ms step_avg:81.37ms +step:7300/20000 train_loss:1.9765 train_time:594111ms step_avg:81.39ms +step:7372/20000 val_loss:1.9534 val_bpb:1.1569 train_time:599989ms step_avg:81.39ms +stopping_early: wallclock_cap train_time:599989ms step:7372/20000 +peak memory allocated: 18844 MiB reserved: 18918 MiB +swa:applying averaged 49 checkpoints +Serialized model: 98437014 bytes +Code size: 62925 bytes +Total submission size: 98499939 bytes +Serialized model int6+zstd: 15673630 bytes +Total submission size: 15,736,555 bytes (263,445 bytes under the 16 MB limit) +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.212045 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.141736 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.143057 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.135712 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.147314 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.148881 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.150366 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.145702 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.143224 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.144971 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.153583 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.152185 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.153525 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.151807 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.150389 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.150698 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.152067 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.152545 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.158688 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.156167 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.157046 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.155778 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.155086 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.154659 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.155310 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.152911 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.151877 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.152222 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.151004 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.150932 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.150220 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.151446 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.152519 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.152993 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.152498 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.152882 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.151973 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.148111 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.148262 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.149214 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.149370 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.149207 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.147970 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.147690 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.146958 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.147042 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.146993 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.147162 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.146883 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.147466 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.147786 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.147498 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.148551 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.150454 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.149763 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.150479 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.150811 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.150811 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.150377 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.150582 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.149978 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.152783 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.152777 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.152808 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.152441 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.151935 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.151174 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.151164 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.151786 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.151831 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.151812 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.152244 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.151998 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.151617 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.151893 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.151961 +final_int8_zlib_roundtrip val_loss:1.9356 val_bpb:1.1464 eval_time:258721ms +final_int8_zlib_roundtrip_exact val_loss:1.93561706 val_bpb:1.14638448 diff --git a/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed2024.log b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed2024.log new file mode 100644 index 000000000..42a5b6af6 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_Int5Int6_BigramHash_SmearGate_SWA_LLMAdvisor/train_seed2024.log @@ -0,0 +1,220 @@ +W0322 06:44:07.835000 134389206397568 torch/distributed/run.py:779] +W0322 06:44:07.835000 134389206397568 torch/distributed/run.py:779] ***************************************** +W0322 06:44:07.835000 134389206397568 torch/distributed/run.py:779] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0322 06:44:07.835000 134389206397568 torch/distributed/run.py:779] ***************************************** +logs/d4746345-a8bf-453e-a53b-c7b0fc2e7031.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:622592 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:2024 +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:1/20000 train_loss:6.9341 train_time:218ms step_avg:217.98ms +step:2/20000 train_loss:8.1463 train_time:284ms step_avg:142.03ms +step:3/20000 train_loss:7.6968 train_time:369ms step_avg:123.08ms +step:4/20000 train_loss:6.9720 train_time:456ms step_avg:114.09ms +step:5/20000 train_loss:6.8212 train_time:544ms step_avg:108.75ms +step:6/20000 train_loss:6.6291 train_time:631ms step_avg:105.12ms +step:7/20000 train_loss:6.4682 train_time:716ms step_avg:102.35ms +step:8/20000 train_loss:6.3516 train_time:802ms step_avg:100.30ms +step:9/20000 train_loss:6.2828 train_time:889ms step_avg:98.78ms +step:10/20000 train_loss:6.1234 train_time:976ms step_avg:97.56ms +step:100/20000 train_loss:3.1903 train_time:8188ms step_avg:81.88ms +step:200/20000 train_loss:2.6104 train_time:16313ms step_avg:81.56ms +step:300/20000 train_loss:2.5828 train_time:24331ms step_avg:81.10ms +step:400/20000 train_loss:2.4344 train_time:32480ms step_avg:81.20ms +step:500/20000 train_loss:2.4257 train_time:40651ms step_avg:81.30ms +step:500/20000 val_loss:2.3748 val_bpb:1.4065 train_time:40681ms step_avg:81.36ms +step:600/20000 train_loss:2.2428 train_time:48669ms step_avg:81.11ms +step:700/20000 train_loss:2.3064 train_time:56823ms step_avg:81.18ms +step:800/20000 train_loss:2.2663 train_time:64870ms step_avg:81.09ms +step:900/20000 train_loss:2.2530 train_time:73050ms step_avg:81.17ms +step:1000/20000 train_loss:2.1613 train_time:81241ms step_avg:81.24ms +step:1000/20000 val_loss:2.2506 val_bpb:1.3329 train_time:81272ms step_avg:81.27ms +step:1100/20000 train_loss:2.1770 train_time:89317ms step_avg:81.20ms +step:1200/20000 train_loss:2.2067 train_time:97476ms step_avg:81.23ms +step:1300/20000 train_loss:2.2008 train_time:105620ms step_avg:81.25ms +step:1400/20000 train_loss:2.0889 train_time:113663ms step_avg:81.19ms +step:1500/20000 train_loss:2.1448 train_time:121818ms step_avg:81.21ms +step:1500/20000 val_loss:2.2042 val_bpb:1.3055 train_time:121849ms step_avg:81.23ms +step:1600/20000 train_loss:2.1582 train_time:129854ms step_avg:81.16ms +step:1700/20000 train_loss:2.2619 train_time:137997ms step_avg:81.17ms +step:1800/20000 train_loss:2.2417 train_time:146143ms step_avg:81.19ms +step:1900/20000 train_loss:2.1915 train_time:154186ms step_avg:81.15ms +step:2000/20000 train_loss:1.9943 train_time:162362ms step_avg:81.18ms +step:2000/20000 val_loss:2.1548 val_bpb:1.2762 train_time:162392ms step_avg:81.20ms +step:2100/20000 train_loss:2.0353 train_time:170523ms step_avg:81.20ms +step:2200/20000 train_loss:2.0290 train_time:178613ms step_avg:81.19ms +step:2300/20000 train_loss:2.1099 train_time:186798ms step_avg:81.22ms +step:2400/20000 train_loss:2.1264 train_time:194868ms step_avg:81.19ms +step:2500/20000 train_loss:2.0498 train_time:203035ms step_avg:81.21ms +step:2500/20000 val_loss:2.1256 val_bpb:1.2589 train_time:203064ms step_avg:81.23ms +step:2600/20000 train_loss:2.1846 train_time:211200ms step_avg:81.23ms +step:2700/20000 train_loss:2.3041 train_time:219269ms step_avg:81.21ms +step:2800/20000 train_loss:2.1396 train_time:227435ms step_avg:81.23ms +step:2900/20000 train_loss:2.1926 train_time:235575ms step_avg:81.23ms +step:3000/20000 train_loss:2.0062 train_time:243640ms step_avg:81.21ms +step:3000/20000 val_loss:2.1144 val_bpb:1.2522 train_time:243670ms step_avg:81.22ms +step:3100/20000 train_loss:2.0072 train_time:251784ms step_avg:81.22ms +step:3200/20000 train_loss:2.1301 train_time:259831ms step_avg:81.20ms +step:3300/20000 train_loss:2.0815 train_time:267995ms step_avg:81.21ms +step:3400/20000 train_loss:2.0695 train_time:276145ms step_avg:81.22ms +step:3500/20000 train_loss:2.2123 train_time:284195ms step_avg:81.20ms +step:3500/20000 val_loss:2.0991 val_bpb:1.2432 train_time:284226ms step_avg:81.21ms +step:3600/20000 train_loss:1.9904 train_time:292326ms step_avg:81.20ms +step:3700/20000 train_loss:2.1079 train_time:300480ms step_avg:81.21ms +step:3800/20000 train_loss:2.1123 train_time:308547ms step_avg:81.20ms +step:3900/20000 train_loss:1.9899 train_time:316701ms step_avg:81.21ms +step:4000/20000 train_loss:1.9874 train_time:324756ms step_avg:81.19ms +step:4000/20000 val_loss:2.0964 val_bpb:1.2416 train_time:324785ms step_avg:81.20ms +step:4100/20000 train_loss:2.0104 train_time:332911ms step_avg:81.20ms +step:4200/20000 train_loss:2.3410 train_time:341072ms step_avg:81.21ms +step:4300/20000 train_loss:2.1805 train_time:349140ms step_avg:81.20ms +step:4400/20000 train_loss:2.1641 train_time:357309ms step_avg:81.21ms +step:4500/20000 train_loss:2.0886 train_time:365473ms step_avg:81.22ms +step:4500/20000 val_loss:2.0919 val_bpb:1.2389 train_time:365502ms step_avg:81.22ms +step:4600/20000 train_loss:2.1463 train_time:373534ms step_avg:81.20ms +step:4700/20000 train_loss:2.1208 train_time:381686ms step_avg:81.21ms +step:4800/20000 train_loss:1.9998 train_time:389733ms step_avg:81.19ms +step:4900/20000 train_loss:2.0300 train_time:397887ms step_avg:81.20ms +step:5000/20000 train_loss:2.1903 train_time:406031ms step_avg:81.21ms +step:5000/20000 val_loss:2.0919 val_bpb:1.2389 train_time:406062ms step_avg:81.21ms +step:5100/20000 train_loss:2.0732 train_time:414065ms step_avg:81.19ms +step:5200/20000 train_loss:2.1009 train_time:422189ms step_avg:81.19ms +step:5300/20000 train_loss:1.9902 train_time:430225ms step_avg:81.17ms +step:5400/20000 train_loss:2.0898 train_time:438358ms step_avg:81.18ms +step:5500/20000 train_loss:2.1221 train_time:446485ms step_avg:81.18ms +step:5500/20000 val_loss:2.0680 val_bpb:1.2248 train_time:446514ms step_avg:81.18ms +step:5600/20000 train_loss:2.0242 train_time:454525ms step_avg:81.17ms +step:5700/20000 train_loss:2.0237 train_time:462657ms step_avg:81.17ms +step:5800/20000 train_loss:2.6022 train_time:470794ms step_avg:81.17ms +step:5900/20000 train_loss:2.0317 train_time:478858ms step_avg:81.16ms +step:6000/20000 train_loss:2.0735 train_time:487023ms step_avg:81.17ms +step:6000/20000 val_loss:2.0398 val_bpb:1.2081 train_time:487055ms step_avg:81.18ms +step:6100/20000 train_loss:2.0417 train_time:495090ms step_avg:81.16ms +step:6200/20000 train_loss:1.9450 train_time:503252ms step_avg:81.17ms +step:6300/20000 train_loss:2.0977 train_time:511415ms step_avg:81.18ms +step:6400/20000 train_loss:1.8816 train_time:519458ms step_avg:81.17ms +step:6500/20000 train_loss:1.9352 train_time:527592ms step_avg:81.17ms +step:6500/20000 val_loss:2.0094 val_bpb:1.1901 train_time:527623ms step_avg:81.17ms +swa:start step:6600 +step:6600/20000 train_loss:2.0895 train_time:535743ms step_avg:81.17ms +step:6700/20000 train_loss:1.9735 train_time:543890ms step_avg:81.18ms +step:6800/20000 train_loss:1.9380 train_time:552088ms step_avg:81.19ms +step:6900/20000 train_loss:1.9905 train_time:560186ms step_avg:81.19ms +step:7000/20000 train_loss:1.9437 train_time:568374ms step_avg:81.20ms +step:7000/20000 val_loss:1.9768 val_bpb:1.1708 train_time:568434ms step_avg:81.20ms +step:7100/20000 train_loss:2.0303 train_time:576564ms step_avg:81.21ms +step:7200/20000 train_loss:2.0171 train_time:584664ms step_avg:81.20ms +step:7300/20000 train_loss:1.9781 train_time:592856ms step_avg:81.21ms +step:7387/20000 val_loss:1.9554 val_bpb:1.1581 train_time:599943ms step_avg:81.22ms +stopping_early: wallclock_cap train_time:599943ms step:7387/20000 +peak memory allocated: 18843 MiB reserved: 18932 MiB +swa:applying averaged 16 checkpoints +Serialized model: 98437014 bytes +Code size: 62925 bytes +Total submission size: 98499939 bytes +Serialized model int6+zstd: 15571435 bytes +Total submission size: 15,634,360 bytes (365,640 bytes under the 16 MB limit) +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.211705 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.139700 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.142098 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.135403 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.147827 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.149213 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.150452 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.146094 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.143701 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.145515 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.154293 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.152880 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.154222 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.152338 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.150865 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.151230 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.152571 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.153012 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.159092 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.156571 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.157558 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.156209 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.155573 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.155184 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.155786 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.153368 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.152368 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.152774 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.151598 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.151430 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.150680 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.151899 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.152997 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.153470 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.152960 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.153344 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.152467 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.148587 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.148698 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.149625 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.149806 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.149695 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.148441 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.148141 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.147456 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.147551 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.147547 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.147725 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.147457 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.148072 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.148401 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.148098 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.149151 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.151074 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.150351 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.151093 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.151424 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.151417 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.150981 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.151204 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.150605 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.153415 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.153419 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.153440 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.153070 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.152587 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.151836 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.151817 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.152437 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.152467 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.152470 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.152926 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.152699 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.152327 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.152645 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.152731 +final_int8_zlib_roundtrip val_loss:1.9368 val_bpb:1.1471 eval_time:258830ms +final_int8_zlib_roundtrip_exact val_loss:1.93680082 val_bpb:1.14708558