From 14cdf6f7a4c84d5016dbd6adc6df6d5208b83261 Mon Sep 17 00:00:00 2001 From: Raahil Shah Date: Fri, 20 Mar 2026 09:03:19 +0530 Subject: [PATCH 1/2] Add submission: 2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA (mean val_bpb=1.1483, 3 seeds) --- .../README.md | 74 + .../submission.json | 25 + .../train_gpt.py | 1218 +++++++++++++++++ .../train_seed1337.log | 217 +++ .../train_seed42.log | 217 +++ .../train_seed7.log | 217 +++ 6 files changed, 1968 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md create mode 100644 records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json create mode 100644 records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py create mode 100644 records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log create mode 100644 records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md new file mode 100644 index 0000000000..ac9cdab8e8 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md @@ -0,0 +1,74 @@ +# Int6 MLP3x + SmearGate + BigramHash + OrthoInit + Muon WD + SWA + +## Score: mean val_bpb = 1.1483 (3 seeds: 1.1488, 1.1485, 1.1476) + +Trained on 8×H100 SXM in 600 seconds. 15.92MB artifact (int6+zstd-22). + +## Approach + +Seven techniques stacked on the baseline 9-layer, 512-dim GPT: + +### 1. Per-Row Int6 Quantization + zstd-22 Compression +MLP and attention weight matrices are quantized to int6 ([-32, 31]) with per-row scaling. Tied embeddings remain in fp16 (quantization-sensitive). The last transformer layer's key projection is also kept in fp16 to reduce the quantization penalty on late-layer attention. zstd at level 22 provides ~5% better compression than zlib-9 on int6 data, freeing additional bytes for parameters. + +### 2. 3× MLP Expansion +MLP hidden dimension increased from 1024 (2×) to 1536 (3×), enabled by the byte savings from int6 quantization. This is the single largest contributor to the improvement over int8-based submissions. + +### 3. SmearGate +A learned gate that blends each token's embedding with the previous token's embedding, providing lightweight bigram-level context at the embedding layer. Adds ~512 parameters. Helps the model capture local dependencies without increasing sequence modeling cost. + +### 4. BigramHash Embedding +A 4096-bucket hash table (dim=128, projected to 512) that maps adjacent token pairs to learned embeddings. The hash function `(prev_token * 31 + curr_token) % 4096` provides collision-resistant coverage of the 1M possible bigram pairs. Adds ~524K parameters. Complements SmearGate by providing an additive bigram signal rather than a multiplicative gate. + +### 5. Orthogonal Weight Initialization +All large weight matrices initialized with `torch.nn.init.orthogonal_(gain=1.0)`. Output projections (attention proj, MLP proj) are additionally scaled by `1/sqrt(2 * num_layers)` following muP conventions. This accelerates early convergence by starting from a well-conditioned point, giving Muon a head start. + +### 6. Muon Optimizer with Weight Decay +Muon optimizer with decoupled weight decay (WD=0.02) applied after the Newton-Schulz gradient update. Momentum warmup from 0.92 to 0.99 over 1500 steps. AdamW (WD=0.01) for embedding and scalar parameters. Weight decay regularizes weight magnitudes, directly improving int6 quantization quality. + +### 7. Stochastic Weight Averaging (SWA) +SWA enabled over the last 50% of training, averaging checkpoints every 200 steps. Produces smoother weight distributions that quantize better, reducing the int6 quantization penalty. + +## Hyperparameters + +| Parameter | Value | +|-----------|-------| +| num_layers | 9 | +| model_dim | 512 | +| mlp_mult | 3.0 (hidden=1536) | +| train_seq_len | 2048 | +| train_batch_tokens | 786,432 | +| warmdown_iters | 3000 | +| matrix_lr | 0.02 | +| scalar_lr | 0.02 | +| tied_embed_lr | 0.03 | +| muon_momentum | 0.99 (warmup from 0.92) | +| grad_clip_norm | 0.3 | +| weight_decay | 0.01 (AdamW) / 0.02 (Muon) | +| eval_stride | 64 | +| bigram_vocab_size | 4096 | +| bigram_dim | 128 | +| compressor | zstd (level 22) | + +## Key Metrics + +- **Mean val_bpb: 1.1483** (seeds 1337, 42, 7) +- Pre-quant val_bpb: 1.1640 +- Quantization penalty: 0.016 bpb (int6 vs fp16) +- Training: 7,373 steps in 600s (81.4 ms/step) +- Model params: ~22M +- Artifact size: 15.92MB (int6+zstd-22) + +## Reproducibility + +Three independent training runs with different random seeds, all other settings identical: + +| Seed | val_loss | val_bpb | +|------|----------|---------| +| 1337 | 1.93978 | 1.14885 | +| 42 | 1.93923 | 1.14852 | +| 7 | 1.93762 | 1.14757 | +| **Mean** | **1.93888** | **1.14831** | +| **Std** | **0.00111** | **0.00066** | + +Improvement over current SOTA (1.1748): **-0.0265 bpb / -0.0459 nats** (p < 0.001 by one-sample t-test against 1.1748 with 3 samples). diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json new file mode 100644 index 0000000000..62432029c1 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json @@ -0,0 +1,25 @@ +{ + "author": "Raahil Shah", + "github_id": "raahilshah", + "name": "Int6 MLP3x + SmearGate + BigramHash + OrthoInit + Muon WD + SWA", + "blurb": "Per-row int6 quantization on MLP/attention weights with zstd-22 compression, enabling 3x MLP expansion (hidden=1536). SmearGate blends adjacent token embeddings via a learned gate. BigramHash embedding (4096 buckets, dim=128) captures token-pair context. Orthogonal weight initialization with muP output scaling. Muon optimizer with decoupled weight decay (WD=0.02) and momentum warmup (0.92→0.99 over 1500 steps). Stochastic Weight Averaging over the last 50% of training. Trained at seq_len=2048 with batch=786432 tokens, grad_clip=0.3, warmdown=3000 steps. Sliding window evaluation at stride=64.", + "date": "2026-03-20T03:30:00Z", + "val_loss": 1.93887501, + "val_bpb": 1.14831403, + "val_loss_std": 0.00111, + "val_bpb_std": 0.00066, + "seeds": [1337, 42, 7], + "seed_results": { + "1337": {"val_loss": 1.93977738, "val_bpb": 1.14884847}, + "42": {"val_loss": 1.93923034, "val_bpb": 1.14852448}, + "7": {"val_loss": 1.93761732, "val_bpb": 1.14756915} + }, + "pre_quant_val_loss": 1.9654, + "pre_quant_val_bpb": 1.1640, + "step_stop": 7373, + "wallclock_seconds": 600.035, + "eval_time_seconds": 155.128, + "bytes_total": 15916456, + "bytes_model_int6_zstd": 15858116, + "bytes_code": 58340 +} diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py new file mode 100644 index 0000000000..c143ce8a5e --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py @@ -0,0 +1,1218 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 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", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = 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", 0.01)) + + 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", 4096)) + 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", 0.5)) + swa_every = int(os.environ.get("SWA_EVERY", 200)) + +# ----------------------------- +# 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 +) +FP16_KEEP_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get("FP16_KEEP_NAME_PATTERNS", "tok_emb,blocks.8.attn.c_k").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 ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1e-12).to(torch.float16) + scale = scale.clamp_min(torch.finfo(torch.float16).tiny) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / 31.0, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).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() <= 65536: + 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: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + 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): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class SmearGate(nn.Module): + """Blend each token's embedding with the previous token's embedding.""" + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + + +class BigramHashEmbedding(nn.Module): + """Hash consecutive token pairs into a learned embedding table.""" + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: float, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList( + [ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, +) -> 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 >= 1] + 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() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # MODEL + OPTIMIZER SETUP + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.weight_decay, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=0.02, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.weight_decay, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # DATA LOADER & MODEL WARMUP + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # MAIN TRAINING LOOP + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + # SWA: collect checkpoints during warmdown + if args.swa_enabled and scale < args.swa_start_frac and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # Apply SWA if collected + if args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + current_state = base_model.state_dict() + avg_state = { + name: (tensor / swa_count).to(dtype=current_state[name].dtype) + for name, tensor in swa_state.items() + } + base_model.load_state_dict(avg_state, strict=True) + + # 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") + + # 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"}) + 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") + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + if _COMPRESSOR == "zstd": + decompressed = zstandard.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = zlib.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_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() diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log new file mode 100644 index 0000000000..6cbb3f28c6 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log @@ -0,0 +1,217 @@ +logs/repro_seed1337.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:22368841 +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:786432 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:0/20000 val_loss:6.9341 val_bpb:4.1068 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9348 train_time:141ms step_avg:140.51ms +step:2/20000 train_loss:8.1566 train_time:191ms step_avg:95.75ms +step:3/20000 train_loss:7.7060 train_time:270ms step_avg:90.09ms +step:4/20000 train_loss:6.9834 train_time:350ms step_avg:87.39ms +step:5/20000 train_loss:6.7600 train_time:429ms step_avg:85.76ms +step:6/20000 train_loss:6.5912 train_time:508ms step_avg:84.65ms +step:7/20000 train_loss:6.4885 train_time:587ms step_avg:83.91ms +step:8/20000 train_loss:6.4640 train_time:667ms step_avg:83.32ms +step:9/20000 train_loss:6.3303 train_time:746ms step_avg:82.85ms +step:10/20000 train_loss:6.0745 train_time:825ms step_avg:82.47ms +step:100/20000 train_loss:3.2146 train_time:8022ms step_avg:80.22ms +step:200/20000 train_loss:2.4109 train_time:16112ms step_avg:80.56ms +step:300/20000 train_loss:2.5634 train_time:24210ms step_avg:80.70ms +step:400/20000 train_loss:2.4297 train_time:32316ms step_avg:80.79ms +step:500/20000 train_loss:2.4170 train_time:40341ms step_avg:80.68ms +step:500/20000 val_loss:2.3733 val_bpb:1.4056 train_time:40371ms step_avg:80.74ms +step:600/20000 train_loss:2.3439 train_time:48438ms step_avg:80.73ms +step:700/20000 train_loss:2.3591 train_time:56545ms step_avg:80.78ms +step:800/20000 train_loss:2.2499 train_time:64670ms step_avg:80.84ms +step:900/20000 train_loss:2.1413 train_time:72790ms step_avg:80.88ms +step:1000/20000 train_loss:2.2830 train_time:80848ms step_avg:80.85ms +step:1000/20000 val_loss:2.2369 val_bpb:1.3248 train_time:80878ms step_avg:80.88ms +step:1100/20000 train_loss:2.3324 train_time:88978ms step_avg:80.89ms +step:1200/20000 train_loss:2.3637 train_time:97117ms step_avg:80.93ms +step:1300/20000 train_loss:2.1082 train_time:105269ms step_avg:80.98ms +step:1400/20000 train_loss:2.1930 train_time:113401ms step_avg:81.00ms +step:1500/20000 train_loss:2.2327 train_time:121469ms step_avg:80.98ms +step:1500/20000 val_loss:2.1924 val_bpb:1.2985 train_time:121499ms step_avg:81.00ms +step:1600/20000 train_loss:2.0857 train_time:129621ms step_avg:81.01ms +step:1700/20000 train_loss:2.1540 train_time:137763ms step_avg:81.04ms +step:1800/20000 train_loss:2.1688 train_time:145922ms step_avg:81.07ms +step:1900/20000 train_loss:2.1377 train_time:153991ms step_avg:81.05ms +step:2000/20000 train_loss:2.0756 train_time:162135ms step_avg:81.07ms +step:2000/20000 val_loss:2.1403 val_bpb:1.2676 train_time:162165ms step_avg:81.08ms +step:2100/20000 train_loss:2.0563 train_time:170292ms step_avg:81.09ms +step:2200/20000 train_loss:2.1475 train_time:178431ms step_avg:81.10ms +step:2300/20000 train_loss:2.1151 train_time:186585ms step_avg:81.12ms +step:2400/20000 train_loss:2.0711 train_time:194664ms step_avg:81.11ms +step:2500/20000 train_loss:2.1774 train_time:202815ms step_avg:81.13ms +step:2500/20000 val_loss:2.1111 val_bpb:1.2503 train_time:202845ms step_avg:81.14ms +step:2600/20000 train_loss:2.1118 train_time:210962ms step_avg:81.14ms +step:2700/20000 train_loss:2.1027 train_time:219098ms step_avg:81.15ms +step:2800/20000 train_loss:2.1589 train_time:227248ms step_avg:81.16ms +step:2900/20000 train_loss:2.0281 train_time:235335ms step_avg:81.15ms +step:3000/20000 train_loss:2.1634 train_time:243487ms step_avg:81.16ms +step:3000/20000 val_loss:2.0930 val_bpb:1.2396 train_time:243517ms step_avg:81.17ms +step:3100/20000 train_loss:2.0356 train_time:251647ms step_avg:81.18ms +step:3200/20000 train_loss:2.1745 train_time:259802ms step_avg:81.19ms +step:3300/20000 train_loss:2.0714 train_time:267888ms step_avg:81.18ms +step:3400/20000 train_loss:2.0224 train_time:276060ms step_avg:81.19ms +step:3500/20000 train_loss:2.1772 train_time:284217ms step_avg:81.20ms +step:3500/20000 val_loss:2.0810 val_bpb:1.2325 train_time:284248ms step_avg:81.21ms +step:3600/20000 train_loss:2.0954 train_time:292370ms step_avg:81.21ms +step:3700/20000 train_loss:2.0972 train_time:300529ms step_avg:81.22ms +step:3800/20000 train_loss:2.0742 train_time:308626ms step_avg:81.22ms +step:3900/20000 train_loss:2.0819 train_time:316791ms step_avg:81.23ms +step:4000/20000 train_loss:1.9817 train_time:324945ms step_avg:81.24ms +step:4000/20000 val_loss:2.0734 val_bpb:1.2280 train_time:324977ms step_avg:81.24ms +step:4100/20000 train_loss:2.0231 train_time:333109ms step_avg:81.25ms +step:4200/20000 train_loss:2.1629 train_time:341270ms step_avg:81.25ms +step:4300/20000 train_loss:2.0704 train_time:349383ms step_avg:81.25ms +step:4400/20000 train_loss:2.0485 train_time:357541ms step_avg:81.26ms +step:4500/20000 train_loss:2.1386 train_time:365693ms step_avg:81.27ms +step:4500/20000 val_loss:2.0620 val_bpb:1.2212 train_time:365723ms step_avg:81.27ms +step:4600/20000 train_loss:1.8586 train_time:373863ms step_avg:81.27ms +step:4700/20000 train_loss:2.2563 train_time:381953ms step_avg:81.27ms +step:4800/20000 train_loss:2.4478 train_time:390118ms step_avg:81.27ms +step:4900/20000 train_loss:2.0678 train_time:398280ms step_avg:81.28ms +step:5000/20000 train_loss:2.1230 train_time:406442ms step_avg:81.29ms +step:5000/20000 val_loss:2.0443 val_bpb:1.2107 train_time:406473ms step_avg:81.29ms +step:5100/20000 train_loss:2.1459 train_time:414602ms step_avg:81.29ms +step:5200/20000 train_loss:2.0660 train_time:422683ms step_avg:81.29ms +step:5300/20000 train_loss:2.0309 train_time:430840ms step_avg:81.29ms +step:5400/20000 train_loss:2.0720 train_time:438998ms step_avg:81.30ms +step:5500/20000 train_loss:2.0455 train_time:447152ms step_avg:81.30ms +step:5500/20000 val_loss:2.0270 val_bpb:1.2005 train_time:447183ms step_avg:81.31ms +step:5600/20000 train_loss:1.9826 train_time:455313ms step_avg:81.31ms +step:5700/20000 train_loss:2.0422 train_time:463398ms step_avg:81.30ms +step:5800/20000 train_loss:2.0308 train_time:471546ms step_avg:81.30ms +step:5900/20000 train_loss:1.9384 train_time:479713ms step_avg:81.31ms +swa:start step:6000 +step:6000/20000 train_loss:1.9721 train_time:487874ms step_avg:81.31ms +step:6000/20000 val_loss:2.0113 val_bpb:1.1912 train_time:487984ms step_avg:81.33ms +step:6100/20000 train_loss:1.9475 train_time:496046ms step_avg:81.32ms +step:6200/20000 train_loss:1.9830 train_time:504226ms step_avg:81.33ms +step:6300/20000 train_loss:1.9848 train_time:512409ms step_avg:81.33ms +step:6400/20000 train_loss:2.0354 train_time:520583ms step_avg:81.34ms +step:6500/20000 train_loss:2.1173 train_time:528775ms step_avg:81.35ms +step:6500/20000 val_loss:1.9925 val_bpb:1.1801 train_time:528804ms step_avg:81.35ms +step:6600/20000 train_loss:1.8847 train_time:536863ms step_avg:81.34ms +step:6700/20000 train_loss:1.9838 train_time:545055ms step_avg:81.35ms +step:6800/20000 train_loss:2.0685 train_time:553219ms step_avg:81.36ms +step:6900/20000 train_loss:1.8691 train_time:561427ms step_avg:81.37ms +step:7000/20000 train_loss:1.8396 train_time:569588ms step_avg:81.37ms +step:7000/20000 val_loss:1.9758 val_bpb:1.1702 train_time:569660ms step_avg:81.38ms +step:7100/20000 train_loss:1.9741 train_time:577731ms step_avg:81.37ms +step:7200/20000 train_loss:1.9235 train_time:585896ms step_avg:81.37ms +step:7300/20000 train_loss:2.0490 train_time:594100ms step_avg:81.38ms +step:7373/20000 val_loss:1.9654 val_bpb:1.1640 train_time:600035ms step_avg:81.38ms +stopping_early: wallclock_cap train_time:600035ms step:7373/20000 +peak memory allocated: 16965 MiB reserved: 17074 MiB +swa:applying averaged 7 checkpoints +Serialized model: 87413467 bytes +Code size: 52244 bytes +Total submission size: 87465711 bytes +Serialized model int6+zstd: 15864212 bytes +Total submission size int8+zlib: 15916456 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.216289 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.142874 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.144875 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.139164 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.151135 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.152228 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.153641 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.149168 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.146727 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.148293 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.157066 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.155243 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.156560 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.154786 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.153248 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.153590 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.154978 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.155521 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.161662 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.159056 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.159996 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.158657 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.158042 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.157662 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.158330 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.156005 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.155025 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.155420 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.154240 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.154178 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.153449 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.154644 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.155680 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.156212 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.155649 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.156042 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.155149 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.151210 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.151337 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.152255 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.152416 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.152287 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.151078 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.150802 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.150105 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.150178 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.150150 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.150291 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.150024 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.150638 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.150928 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.150624 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.151645 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.153561 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.152863 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.153601 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.153953 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.153924 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.153522 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.153759 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.153137 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.155947 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.155939 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.155975 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.155581 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.155099 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.154363 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.154344 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.154931 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.154970 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.154948 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.155406 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.155171 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.154795 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.155118 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.155192 +final_int8_zlib_roundtrip val_loss:1.9398 val_bpb:1.1488 eval_time:155128ms +final_int8_zlib_roundtrip_exact val_loss:1.93977738 val_bpb:1.14884847 diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log new file mode 100644 index 0000000000..8267e1bc79 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log @@ -0,0 +1,217 @@ +logs/repro_seed42.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:22368841 +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:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9299 val_bpb:4.1043 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9310 train_time:135ms step_avg:135.05ms +step:2/20000 train_loss:8.0394 train_time:189ms step_avg:94.41ms +step:3/20000 train_loss:7.5863 train_time:267ms step_avg:89.14ms +step:4/20000 train_loss:6.9690 train_time:347ms step_avg:86.72ms +step:5/20000 train_loss:6.7855 train_time:426ms step_avg:85.22ms +step:6/20000 train_loss:6.7114 train_time:505ms step_avg:84.24ms +step:7/20000 train_loss:6.5691 train_time:585ms step_avg:83.59ms +step:8/20000 train_loss:6.4877 train_time:665ms step_avg:83.13ms +step:9/20000 train_loss:6.2612 train_time:746ms step_avg:82.88ms +step:10/20000 train_loss:6.0132 train_time:825ms step_avg:82.48ms +step:100/20000 train_loss:3.2142 train_time:8038ms step_avg:80.38ms +step:200/20000 train_loss:2.4214 train_time:16137ms step_avg:80.69ms +step:300/20000 train_loss:2.5678 train_time:24263ms step_avg:80.88ms +step:400/20000 train_loss:2.4258 train_time:32379ms step_avg:80.95ms +step:500/20000 train_loss:2.4101 train_time:40431ms step_avg:80.86ms +step:500/20000 val_loss:2.3702 val_bpb:1.4037 train_time:40460ms step_avg:80.92ms +step:600/20000 train_loss:2.3439 train_time:48562ms step_avg:80.94ms +step:700/20000 train_loss:2.3577 train_time:56697ms step_avg:81.00ms +step:800/20000 train_loss:2.2490 train_time:64849ms step_avg:81.06ms +step:900/20000 train_loss:2.1349 train_time:72984ms step_avg:81.09ms +step:1000/20000 train_loss:2.2804 train_time:81068ms step_avg:81.07ms +step:1000/20000 val_loss:2.2343 val_bpb:1.3233 train_time:81098ms step_avg:81.10ms +step:1100/20000 train_loss:2.3262 train_time:89210ms step_avg:81.10ms +step:1200/20000 train_loss:2.3615 train_time:97350ms step_avg:81.12ms +step:1300/20000 train_loss:2.1056 train_time:105501ms step_avg:81.15ms +step:1400/20000 train_loss:2.1901 train_time:113656ms step_avg:81.18ms +step:1500/20000 train_loss:2.2290 train_time:121736ms step_avg:81.16ms +step:1500/20000 val_loss:2.1907 val_bpb:1.2974 train_time:121767ms step_avg:81.18ms +step:1600/20000 train_loss:2.0822 train_time:129879ms step_avg:81.17ms +step:1700/20000 train_loss:2.1529 train_time:138027ms step_avg:81.19ms +step:1800/20000 train_loss:2.1694 train_time:146194ms step_avg:81.22ms +step:1900/20000 train_loss:2.1383 train_time:154276ms step_avg:81.20ms +step:2000/20000 train_loss:2.0758 train_time:162420ms step_avg:81.21ms +step:2000/20000 val_loss:2.1396 val_bpb:1.2672 train_time:162450ms step_avg:81.22ms +step:2100/20000 train_loss:2.0556 train_time:170594ms step_avg:81.24ms +step:2200/20000 train_loss:2.1453 train_time:178748ms step_avg:81.25ms +step:2300/20000 train_loss:2.1128 train_time:186919ms step_avg:81.27ms +step:2400/20000 train_loss:2.0734 train_time:195031ms step_avg:81.26ms +step:2500/20000 train_loss:2.1765 train_time:203196ms step_avg:81.28ms +step:2500/20000 val_loss:2.1098 val_bpb:1.2496 train_time:203225ms step_avg:81.29ms +step:2600/20000 train_loss:2.1107 train_time:211355ms step_avg:81.29ms +step:2700/20000 train_loss:2.1034 train_time:219505ms step_avg:81.30ms +step:2800/20000 train_loss:2.1564 train_time:227677ms step_avg:81.31ms +step:2900/20000 train_loss:2.0267 train_time:235763ms step_avg:81.30ms +step:3000/20000 train_loss:2.1625 train_time:243900ms step_avg:81.30ms +step:3000/20000 val_loss:2.0924 val_bpb:1.2392 train_time:243930ms step_avg:81.31ms +step:3100/20000 train_loss:2.0329 train_time:252051ms step_avg:81.31ms +step:3200/20000 train_loss:2.1726 train_time:260203ms step_avg:81.31ms +step:3300/20000 train_loss:2.0708 train_time:268291ms step_avg:81.30ms +step:3400/20000 train_loss:2.0207 train_time:276454ms step_avg:81.31ms +step:3500/20000 train_loss:2.1794 train_time:284610ms step_avg:81.32ms +step:3500/20000 val_loss:2.0809 val_bpb:1.2324 train_time:284640ms step_avg:81.33ms +step:3600/20000 train_loss:2.1006 train_time:292763ms step_avg:81.32ms +step:3700/20000 train_loss:2.0947 train_time:300926ms step_avg:81.33ms +step:3800/20000 train_loss:2.0759 train_time:309013ms step_avg:81.32ms +step:3900/20000 train_loss:2.0770 train_time:317184ms step_avg:81.33ms +step:4000/20000 train_loss:1.9818 train_time:325326ms step_avg:81.33ms +step:4000/20000 val_loss:2.0736 val_bpb:1.2281 train_time:325357ms step_avg:81.34ms +step:4100/20000 train_loss:2.0247 train_time:333474ms step_avg:81.34ms +step:4200/20000 train_loss:2.1603 train_time:341627ms step_avg:81.34ms +step:4300/20000 train_loss:2.0729 train_time:349735ms step_avg:81.33ms +step:4400/20000 train_loss:2.0484 train_time:357902ms step_avg:81.34ms +step:4500/20000 train_loss:2.1404 train_time:366055ms step_avg:81.35ms +step:4500/20000 val_loss:2.0620 val_bpb:1.2213 train_time:366085ms step_avg:81.35ms +step:4600/20000 train_loss:1.8604 train_time:374223ms step_avg:81.35ms +step:4700/20000 train_loss:2.2528 train_time:382306ms step_avg:81.34ms +step:4800/20000 train_loss:2.4504 train_time:390451ms step_avg:81.34ms +step:4900/20000 train_loss:2.0692 train_time:398606ms step_avg:81.35ms +step:5000/20000 train_loss:2.1254 train_time:406761ms step_avg:81.35ms +step:5000/20000 val_loss:2.0440 val_bpb:1.2106 train_time:406794ms step_avg:81.36ms +step:5100/20000 train_loss:2.1463 train_time:414937ms step_avg:81.36ms +step:5200/20000 train_loss:2.0638 train_time:423023ms step_avg:81.35ms +step:5300/20000 train_loss:2.0309 train_time:431178ms step_avg:81.35ms +step:5400/20000 train_loss:2.0713 train_time:439324ms step_avg:81.36ms +step:5500/20000 train_loss:2.0430 train_time:447480ms step_avg:81.36ms +step:5500/20000 val_loss:2.0270 val_bpb:1.2005 train_time:447511ms step_avg:81.37ms +step:5600/20000 train_loss:1.9830 train_time:455641ms step_avg:81.36ms +step:5700/20000 train_loss:2.0419 train_time:463731ms step_avg:81.36ms +step:5800/20000 train_loss:2.0311 train_time:471887ms step_avg:81.36ms +step:5900/20000 train_loss:1.9352 train_time:480056ms step_avg:81.37ms +swa:start step:6000 +step:6000/20000 train_loss:1.9705 train_time:488217ms step_avg:81.37ms +step:6000/20000 val_loss:2.0112 val_bpb:1.1912 train_time:488324ms step_avg:81.39ms +step:6100/20000 train_loss:1.9516 train_time:496385ms step_avg:81.37ms +step:6200/20000 train_loss:1.9816 train_time:504555ms step_avg:81.38ms +step:6300/20000 train_loss:1.9841 train_time:512752ms step_avg:81.39ms +step:6400/20000 train_loss:2.0356 train_time:520915ms step_avg:81.39ms +step:6500/20000 train_loss:2.1187 train_time:529113ms step_avg:81.40ms +step:6500/20000 val_loss:1.9924 val_bpb:1.1800 train_time:529143ms step_avg:81.41ms +step:6600/20000 train_loss:1.8839 train_time:537215ms step_avg:81.40ms +step:6700/20000 train_loss:1.9833 train_time:545429ms step_avg:81.41ms +step:6800/20000 train_loss:2.0671 train_time:553598ms step_avg:81.41ms +step:6900/20000 train_loss:1.8716 train_time:561814ms step_avg:81.42ms +step:7000/20000 train_loss:1.8377 train_time:569964ms step_avg:81.42ms +step:7000/20000 val_loss:1.9757 val_bpb:1.1701 train_time:570034ms step_avg:81.43ms +step:7100/20000 train_loss:1.9747 train_time:578104ms step_avg:81.42ms +step:7200/20000 train_loss:1.9239 train_time:586263ms step_avg:81.43ms +step:7300/20000 train_loss:2.0481 train_time:594456ms step_avg:81.43ms +step:7369/20000 val_loss:1.9654 val_bpb:1.1640 train_time:600072ms step_avg:81.43ms +stopping_early: wallclock_cap train_time:600072ms step:7369/20000 +peak memory allocated: 16962 MiB reserved: 17072 MiB +swa:applying averaged 7 checkpoints +Serialized model: 87413467 bytes +Code size: 52244 bytes +Total submission size: 87465711 bytes +Serialized model int6+zstd: 15829986 bytes +Total submission size int8+zlib: 15882230 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.222434 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.143585 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.145154 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.138987 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.151322 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.152404 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.153844 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.149399 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.146812 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.148581 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.157180 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.155479 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.156651 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.154927 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.153421 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.153825 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.155192 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.155753 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.161960 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.159350 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.160298 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.158928 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.158223 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.157810 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.158468 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.156061 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.155090 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.155398 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.154158 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.154013 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.153286 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.154488 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.155554 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.156077 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.155535 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.155925 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.155077 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.151225 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.151308 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.152197 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.152359 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.152186 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.150935 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.150608 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.149901 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.149974 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.149899 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.150073 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.149781 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.150386 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.150708 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.150401 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.151451 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.153371 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.152664 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.153387 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.153717 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.153688 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.153285 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.153489 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.152888 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.155711 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.155713 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.155756 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.155396 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.154881 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.154144 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.154130 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.154755 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.154788 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.154768 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.155207 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.154940 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.154550 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.154844 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.154933 +final_int8_zlib_roundtrip val_loss:1.9392 val_bpb:1.1485 eval_time:155216ms +final_int8_zlib_roundtrip_exact val_loss:1.93923034 val_bpb:1.14852448 diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log new file mode 100644 index 0000000000..9c6bc6e45a --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log @@ -0,0 +1,217 @@ +logs/repro_seed7.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:22368841 +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:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:7 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9314 val_bpb:4.1051 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9327 train_time:134ms step_avg:134.47ms +step:2/20000 train_loss:8.0943 train_time:185ms step_avg:92.54ms +step:3/20000 train_loss:7.6481 train_time:264ms step_avg:87.98ms +step:4/20000 train_loss:6.9718 train_time:344ms step_avg:85.90ms +step:5/20000 train_loss:6.8369 train_time:423ms step_avg:84.55ms +step:6/20000 train_loss:6.7023 train_time:504ms step_avg:83.96ms +step:7/20000 train_loss:6.5070 train_time:582ms step_avg:83.18ms +step:8/20000 train_loss:6.5039 train_time:662ms step_avg:82.75ms +step:9/20000 train_loss:6.3448 train_time:741ms step_avg:82.33ms +step:10/20000 train_loss:6.0817 train_time:821ms step_avg:82.10ms +step:100/20000 train_loss:3.2045 train_time:8037ms step_avg:80.37ms +step:200/20000 train_loss:2.4053 train_time:16152ms step_avg:80.76ms +step:300/20000 train_loss:2.5614 train_time:24281ms step_avg:80.94ms +step:400/20000 train_loss:2.4350 train_time:32392ms step_avg:80.98ms +step:500/20000 train_loss:2.4122 train_time:40450ms step_avg:80.90ms +step:500/20000 val_loss:2.3740 val_bpb:1.4060 train_time:40481ms step_avg:80.96ms +step:600/20000 train_loss:2.3460 train_time:48574ms step_avg:80.96ms +step:700/20000 train_loss:2.3605 train_time:56705ms step_avg:81.01ms +step:800/20000 train_loss:2.2470 train_time:64853ms step_avg:81.07ms +step:900/20000 train_loss:2.1392 train_time:72997ms step_avg:81.11ms +step:1000/20000 train_loss:2.2828 train_time:81083ms step_avg:81.08ms +step:1000/20000 val_loss:2.2377 val_bpb:1.3253 train_time:81114ms step_avg:81.11ms +step:1100/20000 train_loss:2.3363 train_time:89243ms step_avg:81.13ms +step:1200/20000 train_loss:2.3606 train_time:97390ms step_avg:81.16ms +step:1300/20000 train_loss:2.1073 train_time:105537ms step_avg:81.18ms +step:1400/20000 train_loss:2.1886 train_time:113691ms step_avg:81.21ms +step:1500/20000 train_loss:2.2287 train_time:121777ms step_avg:81.18ms +step:1500/20000 val_loss:2.1912 val_bpb:1.2978 train_time:121808ms step_avg:81.21ms +step:1600/20000 train_loss:2.0838 train_time:129936ms step_avg:81.21ms +step:1700/20000 train_loss:2.1470 train_time:138091ms step_avg:81.23ms +step:1800/20000 train_loss:2.1716 train_time:146247ms step_avg:81.25ms +step:1900/20000 train_loss:2.1330 train_time:154351ms step_avg:81.24ms +step:2000/20000 train_loss:2.0729 train_time:162512ms step_avg:81.26ms +step:2000/20000 val_loss:2.1397 val_bpb:1.2673 train_time:162542ms step_avg:81.27ms +step:2100/20000 train_loss:2.0503 train_time:170681ms step_avg:81.28ms +step:2200/20000 train_loss:2.1452 train_time:178837ms step_avg:81.29ms +step:2300/20000 train_loss:2.1125 train_time:187011ms step_avg:81.31ms +step:2400/20000 train_loss:2.0708 train_time:195104ms step_avg:81.29ms +step:2500/20000 train_loss:2.1699 train_time:203267ms step_avg:81.31ms +step:2500/20000 val_loss:2.1088 val_bpb:1.2490 train_time:203296ms step_avg:81.32ms +step:2600/20000 train_loss:2.1110 train_time:211437ms step_avg:81.32ms +step:2700/20000 train_loss:2.1028 train_time:219592ms step_avg:81.33ms +step:2800/20000 train_loss:2.1587 train_time:227765ms step_avg:81.34ms +step:2900/20000 train_loss:2.0227 train_time:235849ms step_avg:81.33ms +step:3000/20000 train_loss:2.1601 train_time:244006ms step_avg:81.34ms +step:3000/20000 val_loss:2.0908 val_bpb:1.2383 train_time:244036ms step_avg:81.35ms +step:3100/20000 train_loss:2.0338 train_time:252170ms step_avg:81.35ms +step:3200/20000 train_loss:2.1724 train_time:260324ms step_avg:81.35ms +step:3300/20000 train_loss:2.0687 train_time:268424ms step_avg:81.34ms +step:3400/20000 train_loss:2.0210 train_time:276580ms step_avg:81.35ms +step:3500/20000 train_loss:2.1823 train_time:284743ms step_avg:81.36ms +step:3500/20000 val_loss:2.0796 val_bpb:1.2316 train_time:284772ms step_avg:81.36ms +step:3600/20000 train_loss:2.0932 train_time:292896ms step_avg:81.36ms +step:3700/20000 train_loss:2.0919 train_time:301052ms step_avg:81.37ms +step:3800/20000 train_loss:2.0734 train_time:309150ms step_avg:81.36ms +step:3900/20000 train_loss:2.0769 train_time:317315ms step_avg:81.36ms +step:4000/20000 train_loss:1.9810 train_time:325466ms step_avg:81.37ms +step:4000/20000 val_loss:2.0707 val_bpb:1.2264 train_time:325497ms step_avg:81.37ms +step:4100/20000 train_loss:2.0206 train_time:333618ms step_avg:81.37ms +step:4200/20000 train_loss:2.1607 train_time:341759ms step_avg:81.37ms +step:4300/20000 train_loss:2.0721 train_time:349844ms step_avg:81.36ms +step:4400/20000 train_loss:2.0501 train_time:357997ms step_avg:81.36ms +step:4500/20000 train_loss:2.1355 train_time:366139ms step_avg:81.36ms +step:4500/20000 val_loss:2.0599 val_bpb:1.2200 train_time:366169ms step_avg:81.37ms +step:4600/20000 train_loss:1.8575 train_time:374302ms step_avg:81.37ms +step:4700/20000 train_loss:2.2476 train_time:382390ms step_avg:81.36ms +step:4800/20000 train_loss:2.4495 train_time:390547ms step_avg:81.36ms +step:4900/20000 train_loss:2.0665 train_time:398698ms step_avg:81.37ms +step:5000/20000 train_loss:2.1236 train_time:406858ms step_avg:81.37ms +step:5000/20000 val_loss:2.0421 val_bpb:1.2094 train_time:406888ms step_avg:81.38ms +step:5100/20000 train_loss:2.1414 train_time:415019ms step_avg:81.38ms +step:5200/20000 train_loss:2.0584 train_time:423112ms step_avg:81.37ms +step:5300/20000 train_loss:2.0299 train_time:431261ms step_avg:81.37ms +step:5400/20000 train_loss:2.0701 train_time:439438ms step_avg:81.38ms +step:5500/20000 train_loss:2.0398 train_time:447597ms step_avg:81.38ms +step:5500/20000 val_loss:2.0249 val_bpb:1.1992 train_time:447627ms step_avg:81.39ms +step:5600/20000 train_loss:1.9804 train_time:455754ms step_avg:81.38ms +step:5700/20000 train_loss:2.0405 train_time:463856ms step_avg:81.38ms +step:5800/20000 train_loss:2.0266 train_time:472009ms step_avg:81.38ms +step:5900/20000 train_loss:1.9338 train_time:480166ms step_avg:81.38ms +swa:start step:6000 +step:6000/20000 train_loss:1.9689 train_time:488320ms step_avg:81.39ms +step:6000/20000 val_loss:2.0091 val_bpb:1.1899 train_time:488430ms step_avg:81.40ms +step:6100/20000 train_loss:1.9461 train_time:496507ms step_avg:81.39ms +step:6200/20000 train_loss:1.9828 train_time:504682ms step_avg:81.40ms +step:6300/20000 train_loss:1.9793 train_time:512879ms step_avg:81.41ms +step:6400/20000 train_loss:2.0335 train_time:521031ms step_avg:81.41ms +step:6500/20000 train_loss:2.1165 train_time:529232ms step_avg:81.42ms +step:6500/20000 val_loss:1.9902 val_bpb:1.1787 train_time:529262ms step_avg:81.42ms +step:6600/20000 train_loss:1.8805 train_time:537324ms step_avg:81.41ms +step:6700/20000 train_loss:1.9845 train_time:545537ms step_avg:81.42ms +step:6800/20000 train_loss:2.0656 train_time:553694ms step_avg:81.43ms +step:6900/20000 train_loss:1.8698 train_time:561893ms step_avg:81.43ms +step:7000/20000 train_loss:1.8330 train_time:570047ms step_avg:81.44ms +step:7000/20000 val_loss:1.9734 val_bpb:1.1688 train_time:570117ms step_avg:81.45ms +step:7100/20000 train_loss:1.9711 train_time:578194ms step_avg:81.44ms +step:7200/20000 train_loss:1.9240 train_time:586353ms step_avg:81.44ms +step:7300/20000 train_loss:2.0450 train_time:594564ms step_avg:81.45ms +step:7368/20000 val_loss:1.9632 val_bpb:1.1627 train_time:600093ms step_avg:81.45ms +stopping_early: wallclock_cap train_time:600093ms step:7368/20000 +peak memory allocated: 16962 MiB reserved: 17072 MiB +swa:applying averaged 7 checkpoints +Serialized model: 87413467 bytes +Code size: 52244 bytes +Total submission size: 87465711 bytes +Serialized model int6+zstd: 15823762 bytes +Total submission size int8+zlib: 15876006 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.216312 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.143074 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.144524 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.138041 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.149896 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.151041 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.152472 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.148062 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.145451 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.146927 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.155633 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.153787 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.155052 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.153311 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.151851 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.152198 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.153644 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.154183 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.160298 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.157638 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.158639 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.157312 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.156654 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.156268 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.156905 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.154552 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.153516 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.153907 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.152738 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.152602 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.151837 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.153034 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.154107 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.154615 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.154091 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.154458 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.153583 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.149718 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.149822 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.150765 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.150898 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.150747 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.149526 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.149247 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.148578 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.148680 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.148661 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.148820 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.148532 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.149157 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.149455 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.149182 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.150203 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.152094 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.151393 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.152138 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.152508 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.152472 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.152045 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.152252 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.151675 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.154459 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.154453 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.154496 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.154117 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.153594 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.152849 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.152836 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.153429 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.153450 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.153436 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.153877 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.153634 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.153252 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.153561 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.153647 +final_int8_zlib_roundtrip val_loss:1.9376 val_bpb:1.1476 eval_time:155316ms +final_int8_zlib_roundtrip_exact val_loss:1.93761732 val_bpb:1.14756915 From 3c458dcdd4c32353ad714d275a653da1597a3979 Mon Sep 17 00:00:00 2001 From: Raahil Shah Date: Fri, 20 Mar 2026 11:16:18 +0530 Subject: [PATCH 2/2] =?UTF-8?q?Update=20submission:=202026-03-20=5FInt6=5F?= =?UTF-8?q?MLP3x=5FSmearGate=5FBigramHash=5FMuonWD=5FSWA=20=E2=80=94=20imp?= =?UTF-8?q?roved=20config=20(Muon=20WD=3D0.04,=20SWA=20every=2050),=20mean?= =?UTF-8?q?=20val=5Fbpb=3D1.1458?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../README.md | 47 +-- .../submission.json | 34 +- .../train_gpt.py | 4 +- .../train_seed1337.log | 366 +++++++++--------- .../train_seed42.log | 366 +++++++++--------- .../train_seed7.log | 366 +++++++++--------- 6 files changed, 593 insertions(+), 590 deletions(-) diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md index ac9cdab8e8..1f71780175 100644 --- a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/README.md @@ -1,33 +1,33 @@ # Int6 MLP3x + SmearGate + BigramHash + OrthoInit + Muon WD + SWA -## Score: mean val_bpb = 1.1483 (3 seeds: 1.1488, 1.1485, 1.1476) +## Score: mean val_bpb = 1.1458 (3 seeds: 1.1460, 1.1466, 1.1449) -Trained on 8×H100 SXM in 600 seconds. 15.92MB artifact (int6+zstd-22). +Trained on 8×H100 SXM in 600 seconds. 15.86MB artifact (int6+zstd-22). ## Approach Seven techniques stacked on the baseline 9-layer, 512-dim GPT: ### 1. Per-Row Int6 Quantization + zstd-22 Compression -MLP and attention weight matrices are quantized to int6 ([-32, 31]) with per-row scaling. Tied embeddings remain in fp16 (quantization-sensitive). The last transformer layer's key projection is also kept in fp16 to reduce the quantization penalty on late-layer attention. zstd at level 22 provides ~5% better compression than zlib-9 on int6 data, freeing additional bytes for parameters. +MLP and attention weight matrices quantized to int6 ([-32, 31]) with per-row scaling. Tied embeddings remain in fp16 (quantization-sensitive). The last transformer layer's key projection is kept in fp16 to reduce the quantization penalty on late-layer attention. zstd at level 22 provides ~5% better compression than zlib-9 on int6 data. ### 2. 3× MLP Expansion -MLP hidden dimension increased from 1024 (2×) to 1536 (3×), enabled by the byte savings from int6 quantization. This is the single largest contributor to the improvement over int8-based submissions. +MLP hidden dimension increased from 1024 (2×) to 1536 (3×), enabled by the byte savings from int6 quantization. This is the single largest contributor to the improvement. ### 3. SmearGate -A learned gate that blends each token's embedding with the previous token's embedding, providing lightweight bigram-level context at the embedding layer. Adds ~512 parameters. Helps the model capture local dependencies without increasing sequence modeling cost. +A learned gate blending each token's embedding with the previous token's embedding, providing lightweight bigram-level context at the embedding layer. Adds ~512 parameters. ### 4. BigramHash Embedding -A 4096-bucket hash table (dim=128, projected to 512) that maps adjacent token pairs to learned embeddings. The hash function `(prev_token * 31 + curr_token) % 4096` provides collision-resistant coverage of the 1M possible bigram pairs. Adds ~524K parameters. Complements SmearGate by providing an additive bigram signal rather than a multiplicative gate. +A 4096-bucket hash table (dim=128, projected to 512) mapping adjacent token pairs to learned embeddings via `(prev_token * 31 + curr_token) % 4096`. Adds ~524K parameters. Complements SmearGate with an additive bigram signal. ### 5. Orthogonal Weight Initialization -All large weight matrices initialized with `torch.nn.init.orthogonal_(gain=1.0)`. Output projections (attention proj, MLP proj) are additionally scaled by `1/sqrt(2 * num_layers)` following muP conventions. This accelerates early convergence by starting from a well-conditioned point, giving Muon a head start. +All large weight matrices initialized with `orthogonal_(gain=1.0)`. Output projections scaled by `1/sqrt(2 * num_layers)` following muP conventions. Accelerates early convergence. ### 6. Muon Optimizer with Weight Decay -Muon optimizer with decoupled weight decay (WD=0.02) applied after the Newton-Schulz gradient update. Momentum warmup from 0.92 to 0.99 over 1500 steps. AdamW (WD=0.01) for embedding and scalar parameters. Weight decay regularizes weight magnitudes, directly improving int6 quantization quality. +Muon with decoupled weight decay WD=0.04 (swept from 0.01–0.05, optimal at 0.04). Momentum warmup from 0.92 to 0.99 over 1500 steps. AdamW WD=0.01 for embedding and scalar parameters. Weight decay regularizes magnitudes, directly improving int6 quantization quality. ### 7. Stochastic Weight Averaging (SWA) -SWA enabled over the last 50% of training, averaging checkpoints every 200 steps. Produces smoother weight distributions that quantize better, reducing the int6 quantization penalty. +SWA every 50 steps over the last 50% of training (~30 checkpoints averaged). Produces smoother weight distributions that quantize better. Swept swa_every from 200 down to 25; optimal at 50. ## Hyperparameters @@ -42,33 +42,36 @@ SWA enabled over the last 50% of training, averaging checkpoints every 200 steps | matrix_lr | 0.02 | | scalar_lr | 0.02 | | tied_embed_lr | 0.03 | -| muon_momentum | 0.99 (warmup from 0.92) | +| muon_momentum | 0.99 (warmup from 0.92 over 1500 steps) | +| muon_weight_decay | 0.04 | +| adamw_weight_decay | 0.01 | | grad_clip_norm | 0.3 | -| weight_decay | 0.01 (AdamW) / 0.02 (Muon) | | eval_stride | 64 | +| swa_every | 50 | +| swa_start_frac | 0.5 | | bigram_vocab_size | 4096 | | bigram_dim | 128 | | compressor | zstd (level 22) | ## Key Metrics -- **Mean val_bpb: 1.1483** (seeds 1337, 42, 7) -- Pre-quant val_bpb: 1.1640 +- **Mean val_bpb: 1.1458** (std: 0.0008) +- Pre-quant val_bpb: 1.1616 - Quantization penalty: 0.016 bpb (int6 vs fp16) -- Training: 7,373 steps in 600s (81.4 ms/step) +- Training: 7,379 steps in 600s (81.3 ms/step) - Model params: ~22M -- Artifact size: 15.92MB (int6+zstd-22) +- Artifact size: 15.86MB (int6+zstd-22) ## Reproducibility -Three independent training runs with different random seeds, all other settings identical: +Three independent training runs with different random seeds: | Seed | val_loss | val_bpb | |------|----------|---------| -| 1337 | 1.93978 | 1.14885 | -| 42 | 1.93923 | 1.14852 | -| 7 | 1.93762 | 1.14757 | -| **Mean** | **1.93888** | **1.14831** | -| **Std** | **0.00111** | **0.00066** | +| 1337 | 1.93492 | 1.14597 | +| 42 | 1.93591 | 1.14656 | +| 7 | 1.93314 | 1.14492 | +| **Mean** | **1.93466** | **1.14582** | +| **Std** | **0.00139** | **0.00082** | -Improvement over current SOTA (1.1748): **-0.0265 bpb / -0.0459 nats** (p < 0.001 by one-sample t-test against 1.1748 with 3 samples). +Improvement over current SOTA (1.1748): **-0.0290 bpb / -0.0503 nats** (p < 0.001). diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json index 62432029c1..fc7c902599 100644 --- a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/submission.json @@ -2,24 +2,24 @@ "author": "Raahil Shah", "github_id": "raahilshah", "name": "Int6 MLP3x + SmearGate + BigramHash + OrthoInit + Muon WD + SWA", - "blurb": "Per-row int6 quantization on MLP/attention weights with zstd-22 compression, enabling 3x MLP expansion (hidden=1536). SmearGate blends adjacent token embeddings via a learned gate. BigramHash embedding (4096 buckets, dim=128) captures token-pair context. Orthogonal weight initialization with muP output scaling. Muon optimizer with decoupled weight decay (WD=0.02) and momentum warmup (0.92→0.99 over 1500 steps). Stochastic Weight Averaging over the last 50% of training. Trained at seq_len=2048 with batch=786432 tokens, grad_clip=0.3, warmdown=3000 steps. Sliding window evaluation at stride=64.", - "date": "2026-03-20T03:30:00Z", - "val_loss": 1.93887501, - "val_bpb": 1.14831403, - "val_loss_std": 0.00111, - "val_bpb_std": 0.00066, + "blurb": "Per-row int6 quantization on MLP/attention weights with zstd-22 compression, enabling 3x MLP expansion (hidden=1536). SmearGate blends adjacent token embeddings via a learned gate. BigramHash embedding (4096 buckets, dim=128) captures token-pair context. Orthogonal weight initialization with muP output scaling. Muon optimizer with decoupled weight decay (WD=0.04) and momentum warmup (0.92->0.99 over 1500 steps). Stochastic Weight Averaging every 50 steps over the last 50% of training. Trained at seq_len=2048 with batch=786432, grad_clip=0.3, warmdown=3000. Sliding window evaluation at stride=64.", + "date": "2026-03-20T05:30:00Z", + "val_loss": 1.93465876, + "val_bpb": 1.14581692, + "val_loss_std": 0.00139, + "val_bpb_std": 0.00082, "seeds": [1337, 42, 7], "seed_results": { - "1337": {"val_loss": 1.93977738, "val_bpb": 1.14884847}, - "42": {"val_loss": 1.93923034, "val_bpb": 1.14852448}, - "7": {"val_loss": 1.93761732, "val_bpb": 1.14756915} + "1337": {"val_loss": 1.93492097, "val_bpb": 1.14597222}, + "42": {"val_loss": 1.93591485, "val_bpb": 1.14656085}, + "7": {"val_loss": 1.93314046, "val_bpb": 1.14491770} }, - "pre_quant_val_loss": 1.9654, - "pre_quant_val_bpb": 1.1640, - "step_stop": 7373, - "wallclock_seconds": 600.035, - "eval_time_seconds": 155.128, - "bytes_total": 15916456, - "bytes_model_int6_zstd": 15858116, - "bytes_code": 58340 + "pre_quant_val_loss": 1.9613, + "pre_quant_val_bpb": 1.1616, + "step_stop": 7379, + "wallclock_seconds": 600.075, + "eval_time_seconds": 155.204, + "bytes_total": 15862650, + "bytes_model_int6_zstd": 15810407, + "bytes_code": 52243 } diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py index c143ce8a5e..49438aeea8 100644 --- a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_gpt.py @@ -91,7 +91,7 @@ class Hyperparameters: swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.5)) - swa_every = int(os.environ.get("SWA_EVERY", 200)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # ----------------------------- # MUON OPTIMIZER @@ -955,7 +955,7 @@ def log0(msg: str, console: bool = True) -> None: lr=args.matrix_lr, momentum=args.muon_momentum, backend_steps=args.muon_backend_steps, - weight_decay=0.02, + weight_decay=0.04, ) for group in optimizer_muon.param_groups: group["base_lr"] = args.matrix_lr diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log index 6cbb3f28c6..5cc76640d4 100644 --- a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed1337.log @@ -29,189 +29,189 @@ warmup_step:18/20 warmup_step:19/20 warmup_step:20/20 step:0/20000 val_loss:6.9341 val_bpb:4.1068 train_time:0ms step_avg:0.02ms -step:1/20000 train_loss:6.9348 train_time:141ms step_avg:140.51ms -step:2/20000 train_loss:8.1566 train_time:191ms step_avg:95.75ms -step:3/20000 train_loss:7.7060 train_time:270ms step_avg:90.09ms -step:4/20000 train_loss:6.9834 train_time:350ms step_avg:87.39ms -step:5/20000 train_loss:6.7600 train_time:429ms step_avg:85.76ms -step:6/20000 train_loss:6.5912 train_time:508ms step_avg:84.65ms -step:7/20000 train_loss:6.4885 train_time:587ms step_avg:83.91ms -step:8/20000 train_loss:6.4640 train_time:667ms step_avg:83.32ms -step:9/20000 train_loss:6.3303 train_time:746ms step_avg:82.85ms -step:10/20000 train_loss:6.0745 train_time:825ms step_avg:82.47ms -step:100/20000 train_loss:3.2146 train_time:8022ms step_avg:80.22ms -step:200/20000 train_loss:2.4109 train_time:16112ms step_avg:80.56ms -step:300/20000 train_loss:2.5634 train_time:24210ms step_avg:80.70ms -step:400/20000 train_loss:2.4297 train_time:32316ms step_avg:80.79ms -step:500/20000 train_loss:2.4170 train_time:40341ms step_avg:80.68ms -step:500/20000 val_loss:2.3733 val_bpb:1.4056 train_time:40371ms step_avg:80.74ms -step:600/20000 train_loss:2.3439 train_time:48438ms step_avg:80.73ms -step:700/20000 train_loss:2.3591 train_time:56545ms step_avg:80.78ms -step:800/20000 train_loss:2.2499 train_time:64670ms step_avg:80.84ms -step:900/20000 train_loss:2.1413 train_time:72790ms step_avg:80.88ms -step:1000/20000 train_loss:2.2830 train_time:80848ms step_avg:80.85ms -step:1000/20000 val_loss:2.2369 val_bpb:1.3248 train_time:80878ms step_avg:80.88ms -step:1100/20000 train_loss:2.3324 train_time:88978ms step_avg:80.89ms -step:1200/20000 train_loss:2.3637 train_time:97117ms step_avg:80.93ms -step:1300/20000 train_loss:2.1082 train_time:105269ms step_avg:80.98ms -step:1400/20000 train_loss:2.1930 train_time:113401ms step_avg:81.00ms -step:1500/20000 train_loss:2.2327 train_time:121469ms step_avg:80.98ms -step:1500/20000 val_loss:2.1924 val_bpb:1.2985 train_time:121499ms step_avg:81.00ms -step:1600/20000 train_loss:2.0857 train_time:129621ms step_avg:81.01ms -step:1700/20000 train_loss:2.1540 train_time:137763ms step_avg:81.04ms -step:1800/20000 train_loss:2.1688 train_time:145922ms step_avg:81.07ms -step:1900/20000 train_loss:2.1377 train_time:153991ms step_avg:81.05ms -step:2000/20000 train_loss:2.0756 train_time:162135ms step_avg:81.07ms -step:2000/20000 val_loss:2.1403 val_bpb:1.2676 train_time:162165ms step_avg:81.08ms -step:2100/20000 train_loss:2.0563 train_time:170292ms step_avg:81.09ms -step:2200/20000 train_loss:2.1475 train_time:178431ms step_avg:81.10ms -step:2300/20000 train_loss:2.1151 train_time:186585ms step_avg:81.12ms -step:2400/20000 train_loss:2.0711 train_time:194664ms step_avg:81.11ms -step:2500/20000 train_loss:2.1774 train_time:202815ms step_avg:81.13ms -step:2500/20000 val_loss:2.1111 val_bpb:1.2503 train_time:202845ms step_avg:81.14ms -step:2600/20000 train_loss:2.1118 train_time:210962ms step_avg:81.14ms -step:2700/20000 train_loss:2.1027 train_time:219098ms step_avg:81.15ms -step:2800/20000 train_loss:2.1589 train_time:227248ms step_avg:81.16ms -step:2900/20000 train_loss:2.0281 train_time:235335ms step_avg:81.15ms -step:3000/20000 train_loss:2.1634 train_time:243487ms step_avg:81.16ms -step:3000/20000 val_loss:2.0930 val_bpb:1.2396 train_time:243517ms step_avg:81.17ms -step:3100/20000 train_loss:2.0356 train_time:251647ms step_avg:81.18ms -step:3200/20000 train_loss:2.1745 train_time:259802ms step_avg:81.19ms -step:3300/20000 train_loss:2.0714 train_time:267888ms step_avg:81.18ms -step:3400/20000 train_loss:2.0224 train_time:276060ms step_avg:81.19ms -step:3500/20000 train_loss:2.1772 train_time:284217ms step_avg:81.20ms -step:3500/20000 val_loss:2.0810 val_bpb:1.2325 train_time:284248ms step_avg:81.21ms -step:3600/20000 train_loss:2.0954 train_time:292370ms step_avg:81.21ms -step:3700/20000 train_loss:2.0972 train_time:300529ms step_avg:81.22ms -step:3800/20000 train_loss:2.0742 train_time:308626ms step_avg:81.22ms -step:3900/20000 train_loss:2.0819 train_time:316791ms step_avg:81.23ms -step:4000/20000 train_loss:1.9817 train_time:324945ms step_avg:81.24ms -step:4000/20000 val_loss:2.0734 val_bpb:1.2280 train_time:324977ms step_avg:81.24ms -step:4100/20000 train_loss:2.0231 train_time:333109ms step_avg:81.25ms -step:4200/20000 train_loss:2.1629 train_time:341270ms step_avg:81.25ms -step:4300/20000 train_loss:2.0704 train_time:349383ms step_avg:81.25ms -step:4400/20000 train_loss:2.0485 train_time:357541ms step_avg:81.26ms -step:4500/20000 train_loss:2.1386 train_time:365693ms step_avg:81.27ms -step:4500/20000 val_loss:2.0620 val_bpb:1.2212 train_time:365723ms step_avg:81.27ms -step:4600/20000 train_loss:1.8586 train_time:373863ms step_avg:81.27ms -step:4700/20000 train_loss:2.2563 train_time:381953ms step_avg:81.27ms -step:4800/20000 train_loss:2.4478 train_time:390118ms step_avg:81.27ms -step:4900/20000 train_loss:2.0678 train_time:398280ms step_avg:81.28ms -step:5000/20000 train_loss:2.1230 train_time:406442ms step_avg:81.29ms -step:5000/20000 val_loss:2.0443 val_bpb:1.2107 train_time:406473ms step_avg:81.29ms -step:5100/20000 train_loss:2.1459 train_time:414602ms step_avg:81.29ms -step:5200/20000 train_loss:2.0660 train_time:422683ms step_avg:81.29ms -step:5300/20000 train_loss:2.0309 train_time:430840ms step_avg:81.29ms -step:5400/20000 train_loss:2.0720 train_time:438998ms step_avg:81.30ms -step:5500/20000 train_loss:2.0455 train_time:447152ms step_avg:81.30ms -step:5500/20000 val_loss:2.0270 val_bpb:1.2005 train_time:447183ms step_avg:81.31ms -step:5600/20000 train_loss:1.9826 train_time:455313ms step_avg:81.31ms -step:5700/20000 train_loss:2.0422 train_time:463398ms step_avg:81.30ms -step:5800/20000 train_loss:2.0308 train_time:471546ms step_avg:81.30ms -step:5900/20000 train_loss:1.9384 train_time:479713ms step_avg:81.31ms -swa:start step:6000 -step:6000/20000 train_loss:1.9721 train_time:487874ms step_avg:81.31ms -step:6000/20000 val_loss:2.0113 val_bpb:1.1912 train_time:487984ms step_avg:81.33ms -step:6100/20000 train_loss:1.9475 train_time:496046ms step_avg:81.32ms -step:6200/20000 train_loss:1.9830 train_time:504226ms step_avg:81.33ms -step:6300/20000 train_loss:1.9848 train_time:512409ms step_avg:81.33ms -step:6400/20000 train_loss:2.0354 train_time:520583ms step_avg:81.34ms -step:6500/20000 train_loss:2.1173 train_time:528775ms step_avg:81.35ms -step:6500/20000 val_loss:1.9925 val_bpb:1.1801 train_time:528804ms step_avg:81.35ms -step:6600/20000 train_loss:1.8847 train_time:536863ms step_avg:81.34ms -step:6700/20000 train_loss:1.9838 train_time:545055ms step_avg:81.35ms -step:6800/20000 train_loss:2.0685 train_time:553219ms step_avg:81.36ms -step:6900/20000 train_loss:1.8691 train_time:561427ms step_avg:81.37ms -step:7000/20000 train_loss:1.8396 train_time:569588ms step_avg:81.37ms -step:7000/20000 val_loss:1.9758 val_bpb:1.1702 train_time:569660ms step_avg:81.38ms -step:7100/20000 train_loss:1.9741 train_time:577731ms step_avg:81.37ms -step:7200/20000 train_loss:1.9235 train_time:585896ms step_avg:81.37ms -step:7300/20000 train_loss:2.0490 train_time:594100ms step_avg:81.38ms -step:7373/20000 val_loss:1.9654 val_bpb:1.1640 train_time:600035ms step_avg:81.38ms -stopping_early: wallclock_cap train_time:600035ms step:7373/20000 +step:1/20000 train_loss:6.9348 train_time:141ms step_avg:140.86ms +step:2/20000 train_loss:8.1566 train_time:191ms step_avg:95.73ms +step:3/20000 train_loss:7.7064 train_time:270ms step_avg:90.08ms +step:4/20000 train_loss:6.9837 train_time:349ms step_avg:87.32ms +step:5/20000 train_loss:6.7600 train_time:429ms step_avg:85.70ms +step:6/20000 train_loss:6.5915 train_time:508ms step_avg:84.74ms +step:7/20000 train_loss:6.4894 train_time:588ms step_avg:83.96ms +step:8/20000 train_loss:6.4654 train_time:668ms step_avg:83.45ms +step:9/20000 train_loss:6.3310 train_time:747ms step_avg:83.00ms +step:10/20000 train_loss:6.0755 train_time:827ms step_avg:82.69ms +step:100/20000 train_loss:3.2043 train_time:8017ms step_avg:80.17ms +step:200/20000 train_loss:2.4054 train_time:16091ms step_avg:80.45ms +step:300/20000 train_loss:2.5684 train_time:24169ms step_avg:80.56ms +step:400/20000 train_loss:2.4415 train_time:32265ms step_avg:80.66ms +step:500/20000 train_loss:2.4149 train_time:40280ms step_avg:80.56ms +step:500/20000 val_loss:2.3768 val_bpb:1.4077 train_time:40311ms step_avg:80.62ms +step:600/20000 train_loss:2.3469 train_time:48365ms step_avg:80.61ms +step:700/20000 train_loss:2.3682 train_time:56464ms step_avg:80.66ms +step:800/20000 train_loss:2.2582 train_time:64564ms step_avg:80.70ms +step:900/20000 train_loss:2.1457 train_time:72674ms step_avg:80.75ms +step:1000/20000 train_loss:2.2893 train_time:80724ms step_avg:80.72ms +step:1000/20000 val_loss:2.2443 val_bpb:1.3292 train_time:80754ms step_avg:80.75ms +step:1100/20000 train_loss:2.3382 train_time:88850ms step_avg:80.77ms +step:1200/20000 train_loss:2.3689 train_time:96958ms step_avg:80.80ms +step:1300/20000 train_loss:2.1175 train_time:105083ms step_avg:80.83ms +step:1400/20000 train_loss:2.1998 train_time:113218ms step_avg:80.87ms +step:1500/20000 train_loss:2.2363 train_time:121286ms step_avg:80.86ms +step:1500/20000 val_loss:2.2003 val_bpb:1.3031 train_time:121316ms step_avg:80.88ms +step:1600/20000 train_loss:2.0913 train_time:129420ms step_avg:80.89ms +step:1700/20000 train_loss:2.1554 train_time:137554ms step_avg:80.91ms +step:1800/20000 train_loss:2.1699 train_time:145705ms step_avg:80.95ms +step:1900/20000 train_loss:2.1431 train_time:153783ms step_avg:80.94ms +step:2000/20000 train_loss:2.0824 train_time:161922ms step_avg:80.96ms +step:2000/20000 val_loss:2.1458 val_bpb:1.2709 train_time:161951ms step_avg:80.98ms +step:2100/20000 train_loss:2.0603 train_time:170068ms step_avg:80.98ms +step:2200/20000 train_loss:2.1501 train_time:178204ms step_avg:81.00ms +step:2300/20000 train_loss:2.1217 train_time:186356ms step_avg:81.02ms +step:2400/20000 train_loss:2.0800 train_time:194428ms step_avg:81.01ms +step:2500/20000 train_loss:2.1844 train_time:202574ms step_avg:81.03ms +step:2500/20000 val_loss:2.1186 val_bpb:1.2548 train_time:202604ms step_avg:81.04ms +step:2600/20000 train_loss:2.1216 train_time:210724ms step_avg:81.05ms +step:2700/20000 train_loss:2.1134 train_time:218878ms step_avg:81.07ms +step:2800/20000 train_loss:2.1697 train_time:227036ms step_avg:81.08ms +step:2900/20000 train_loss:2.0371 train_time:235109ms step_avg:81.07ms +step:3000/20000 train_loss:2.1729 train_time:243248ms step_avg:81.08ms +step:3000/20000 val_loss:2.1036 val_bpb:1.2458 train_time:243278ms step_avg:81.09ms +step:3100/20000 train_loss:2.0532 train_time:251404ms step_avg:81.10ms +step:3200/20000 train_loss:2.1856 train_time:259526ms step_avg:81.10ms +step:3300/20000 train_loss:2.0843 train_time:267598ms step_avg:81.09ms +step:3400/20000 train_loss:2.0318 train_time:275748ms step_avg:81.10ms +step:3500/20000 train_loss:2.1929 train_time:283889ms step_avg:81.11ms +step:3500/20000 val_loss:2.0947 val_bpb:1.2406 train_time:283918ms step_avg:81.12ms +step:3600/20000 train_loss:2.1124 train_time:292032ms step_avg:81.12ms +step:3700/20000 train_loss:2.1085 train_time:300166ms step_avg:81.13ms +step:3800/20000 train_loss:2.0913 train_time:308237ms step_avg:81.11ms +step:3900/20000 train_loss:2.0922 train_time:316389ms step_avg:81.13ms +step:4000/20000 train_loss:1.9969 train_time:324531ms step_avg:81.13ms +step:4000/20000 val_loss:2.0882 val_bpb:1.2368 train_time:324562ms step_avg:81.14ms +step:4100/20000 train_loss:2.0421 train_time:332679ms step_avg:81.14ms +step:4200/20000 train_loss:2.1781 train_time:340832ms step_avg:81.15ms +step:4300/20000 train_loss:2.0885 train_time:348910ms step_avg:81.14ms +step:4400/20000 train_loss:2.0679 train_time:357041ms step_avg:81.15ms +step:4500/20000 train_loss:2.1553 train_time:365175ms step_avg:81.15ms +step:4500/20000 val_loss:2.0779 val_bpb:1.2307 train_time:365205ms step_avg:81.16ms +step:4600/20000 train_loss:1.8732 train_time:373321ms step_avg:81.16ms +step:4700/20000 train_loss:2.2650 train_time:381397ms step_avg:81.15ms +step:4800/20000 train_loss:2.4581 train_time:389535ms step_avg:81.15ms +step:4900/20000 train_loss:2.0833 train_time:397690ms step_avg:81.16ms +step:5000/20000 train_loss:2.1380 train_time:405816ms step_avg:81.16ms +step:5000/20000 val_loss:2.0581 val_bpb:1.2190 train_time:405846ms step_avg:81.17ms +step:5100/20000 train_loss:2.1602 train_time:413970ms step_avg:81.17ms +step:5200/20000 train_loss:2.0790 train_time:422033ms step_avg:81.16ms +step:5300/20000 train_loss:2.0418 train_time:430163ms step_avg:81.16ms +step:5400/20000 train_loss:2.0823 train_time:438300ms step_avg:81.17ms +step:5500/20000 train_loss:2.0531 train_time:446439ms step_avg:81.17ms +step:5500/20000 val_loss:2.0397 val_bpb:1.2080 train_time:446470ms step_avg:81.18ms +step:5600/20000 train_loss:1.9954 train_time:454577ms step_avg:81.17ms +step:5700/20000 train_loss:2.0530 train_time:462635ms step_avg:81.16ms +step:5800/20000 train_loss:2.0411 train_time:470766ms step_avg:81.17ms +swa:start step:5900 +step:5900/20000 train_loss:1.9503 train_time:478911ms step_avg:81.17ms +step:6000/20000 train_loss:1.9795 train_time:487162ms step_avg:81.19ms +step:6000/20000 val_loss:2.0221 val_bpb:1.1976 train_time:487222ms step_avg:81.20ms +step:6100/20000 train_loss:1.9570 train_time:495309ms step_avg:81.20ms +step:6200/20000 train_loss:1.9956 train_time:503492ms step_avg:81.21ms +step:6300/20000 train_loss:1.9887 train_time:511672ms step_avg:81.22ms +step:6400/20000 train_loss:2.0452 train_time:519869ms step_avg:81.23ms +step:6500/20000 train_loss:2.1266 train_time:528043ms step_avg:81.24ms +step:6500/20000 val_loss:1.9989 val_bpb:1.1839 train_time:528099ms step_avg:81.25ms +step:6600/20000 train_loss:1.8915 train_time:536172ms step_avg:81.24ms +step:6700/20000 train_loss:1.9858 train_time:544370ms step_avg:81.25ms +step:6800/20000 train_loss:2.0731 train_time:552561ms step_avg:81.26ms +step:6900/20000 train_loss:1.8717 train_time:560768ms step_avg:81.27ms +step:7000/20000 train_loss:1.8351 train_time:568971ms step_avg:81.28ms +step:7000/20000 val_loss:1.9765 val_bpb:1.1706 train_time:569044ms step_avg:81.29ms +step:7100/20000 train_loss:1.9766 train_time:577112ms step_avg:81.28ms +step:7200/20000 train_loss:1.9235 train_time:585354ms step_avg:81.30ms +step:7300/20000 train_loss:2.0470 train_time:593523ms step_avg:81.30ms +step:7379/20000 val_loss:1.9613 val_bpb:1.1616 train_time:600075ms step_avg:81.32ms +stopping_early: wallclock_cap train_time:600075ms step:7379/20000 peak memory allocated: 16965 MiB reserved: 17074 MiB -swa:applying averaged 7 checkpoints +swa:applying averaged 30 checkpoints Serialized model: 87413467 bytes -Code size: 52244 bytes -Total submission size: 87465711 bytes -Serialized model int6+zstd: 15864212 bytes -Total submission size int8+zlib: 15916456 bytes +Code size: 52243 bytes +Total submission size: 87465710 bytes +Serialized model int6+zstd: 15810407 bytes +Total submission size int8+zlib: 15862650 bytes final_eval_mode:sliding_window stride:64 batch_seqs:32 - sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.216289 - sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.142874 - sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.144875 - sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.139164 - sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.151135 - sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.152228 - sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.153641 - sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.149168 - sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.146727 - sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.148293 - sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.157066 - sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.155243 - sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.156560 - sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.154786 - sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.153248 - sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.153590 - sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.154978 - sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.155521 - sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.161662 - sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.159056 - sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.159996 - sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.158657 - sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.158042 - sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.157662 - sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.158330 - sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.156005 - sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.155025 - sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.155420 - sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.154240 - sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.154178 - sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.153449 - sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.154644 - sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.155680 - sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.156212 - sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.155649 - sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.156042 - sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.155149 - sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.151210 - sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.151337 - sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.152255 - sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.152416 - sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.152287 - sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.151078 - sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.150802 - sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.150105 - sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.150178 - sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.150150 - sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.150291 - sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.150024 - sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.150638 - sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.150928 - sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.150624 - sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.151645 - sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.153561 - sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.152863 - sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.153601 - sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.153953 - sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.153924 - sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.153522 - sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.153759 - sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.153137 - sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.155947 - sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.155939 - sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.155975 - sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.155581 - sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.155099 - sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.154363 - sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.154344 - sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.154931 - sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.154970 - sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.154948 - sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.155406 - sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.155171 - sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.154795 - sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.155118 - sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.155192 -final_int8_zlib_roundtrip val_loss:1.9398 val_bpb:1.1488 eval_time:155128ms -final_int8_zlib_roundtrip_exact val_loss:1.93977738 val_bpb:1.14884847 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.208722 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.140131 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.142313 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.135811 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.148080 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.149270 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.150707 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.145960 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.143566 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.145048 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.153703 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.152061 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.153431 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.151819 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.150317 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.150648 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.152022 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.152565 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.158651 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.156082 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.156986 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.155627 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.155001 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.154637 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.155395 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.153042 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.152081 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.152401 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.151179 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.151070 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.150304 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.151501 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.152604 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.153122 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.152596 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.152950 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.152072 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.148163 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.148269 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.149229 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.149413 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.149267 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.148058 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.147801 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.147115 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.147155 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.147124 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.147277 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.146987 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.147571 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.147900 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.147616 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.148655 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.150558 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.149847 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.150566 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.150913 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.150908 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.150504 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.150744 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.150167 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.152978 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.152969 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.153008 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.152635 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.152163 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.151433 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.151416 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.152036 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.152069 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.152041 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.152486 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.152226 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.151816 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.152134 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.152203 +final_int8_zlib_roundtrip val_loss:1.9349 val_bpb:1.1460 eval_time:155204ms +final_int8_zlib_roundtrip_exact val_loss:1.93492097 val_bpb:1.14597222 diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log index 8267e1bc79..0b1d423e1b 100644 --- a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed42.log @@ -29,189 +29,189 @@ warmup_step:18/20 warmup_step:19/20 warmup_step:20/20 step:0/20000 val_loss:6.9299 val_bpb:4.1043 train_time:0ms step_avg:0.02ms -step:1/20000 train_loss:6.9310 train_time:135ms step_avg:135.05ms -step:2/20000 train_loss:8.0394 train_time:189ms step_avg:94.41ms -step:3/20000 train_loss:7.5863 train_time:267ms step_avg:89.14ms -step:4/20000 train_loss:6.9690 train_time:347ms step_avg:86.72ms -step:5/20000 train_loss:6.7855 train_time:426ms step_avg:85.22ms -step:6/20000 train_loss:6.7114 train_time:505ms step_avg:84.24ms -step:7/20000 train_loss:6.5691 train_time:585ms step_avg:83.59ms -step:8/20000 train_loss:6.4877 train_time:665ms step_avg:83.13ms -step:9/20000 train_loss:6.2612 train_time:746ms step_avg:82.88ms -step:10/20000 train_loss:6.0132 train_time:825ms step_avg:82.48ms -step:100/20000 train_loss:3.2142 train_time:8038ms step_avg:80.38ms -step:200/20000 train_loss:2.4214 train_time:16137ms step_avg:80.69ms -step:300/20000 train_loss:2.5678 train_time:24263ms step_avg:80.88ms -step:400/20000 train_loss:2.4258 train_time:32379ms step_avg:80.95ms -step:500/20000 train_loss:2.4101 train_time:40431ms step_avg:80.86ms -step:500/20000 val_loss:2.3702 val_bpb:1.4037 train_time:40460ms step_avg:80.92ms -step:600/20000 train_loss:2.3439 train_time:48562ms step_avg:80.94ms -step:700/20000 train_loss:2.3577 train_time:56697ms step_avg:81.00ms -step:800/20000 train_loss:2.2490 train_time:64849ms step_avg:81.06ms -step:900/20000 train_loss:2.1349 train_time:72984ms step_avg:81.09ms -step:1000/20000 train_loss:2.2804 train_time:81068ms step_avg:81.07ms -step:1000/20000 val_loss:2.2343 val_bpb:1.3233 train_time:81098ms step_avg:81.10ms -step:1100/20000 train_loss:2.3262 train_time:89210ms step_avg:81.10ms -step:1200/20000 train_loss:2.3615 train_time:97350ms step_avg:81.12ms -step:1300/20000 train_loss:2.1056 train_time:105501ms step_avg:81.15ms -step:1400/20000 train_loss:2.1901 train_time:113656ms step_avg:81.18ms -step:1500/20000 train_loss:2.2290 train_time:121736ms step_avg:81.16ms -step:1500/20000 val_loss:2.1907 val_bpb:1.2974 train_time:121767ms step_avg:81.18ms -step:1600/20000 train_loss:2.0822 train_time:129879ms step_avg:81.17ms -step:1700/20000 train_loss:2.1529 train_time:138027ms step_avg:81.19ms -step:1800/20000 train_loss:2.1694 train_time:146194ms step_avg:81.22ms -step:1900/20000 train_loss:2.1383 train_time:154276ms step_avg:81.20ms -step:2000/20000 train_loss:2.0758 train_time:162420ms step_avg:81.21ms -step:2000/20000 val_loss:2.1396 val_bpb:1.2672 train_time:162450ms step_avg:81.22ms -step:2100/20000 train_loss:2.0556 train_time:170594ms step_avg:81.24ms -step:2200/20000 train_loss:2.1453 train_time:178748ms step_avg:81.25ms -step:2300/20000 train_loss:2.1128 train_time:186919ms step_avg:81.27ms -step:2400/20000 train_loss:2.0734 train_time:195031ms step_avg:81.26ms -step:2500/20000 train_loss:2.1765 train_time:203196ms step_avg:81.28ms -step:2500/20000 val_loss:2.1098 val_bpb:1.2496 train_time:203225ms step_avg:81.29ms -step:2600/20000 train_loss:2.1107 train_time:211355ms step_avg:81.29ms -step:2700/20000 train_loss:2.1034 train_time:219505ms step_avg:81.30ms -step:2800/20000 train_loss:2.1564 train_time:227677ms step_avg:81.31ms -step:2900/20000 train_loss:2.0267 train_time:235763ms step_avg:81.30ms -step:3000/20000 train_loss:2.1625 train_time:243900ms step_avg:81.30ms -step:3000/20000 val_loss:2.0924 val_bpb:1.2392 train_time:243930ms step_avg:81.31ms -step:3100/20000 train_loss:2.0329 train_time:252051ms step_avg:81.31ms -step:3200/20000 train_loss:2.1726 train_time:260203ms step_avg:81.31ms -step:3300/20000 train_loss:2.0708 train_time:268291ms step_avg:81.30ms -step:3400/20000 train_loss:2.0207 train_time:276454ms step_avg:81.31ms -step:3500/20000 train_loss:2.1794 train_time:284610ms step_avg:81.32ms -step:3500/20000 val_loss:2.0809 val_bpb:1.2324 train_time:284640ms step_avg:81.33ms -step:3600/20000 train_loss:2.1006 train_time:292763ms step_avg:81.32ms -step:3700/20000 train_loss:2.0947 train_time:300926ms step_avg:81.33ms -step:3800/20000 train_loss:2.0759 train_time:309013ms step_avg:81.32ms -step:3900/20000 train_loss:2.0770 train_time:317184ms step_avg:81.33ms -step:4000/20000 train_loss:1.9818 train_time:325326ms step_avg:81.33ms -step:4000/20000 val_loss:2.0736 val_bpb:1.2281 train_time:325357ms step_avg:81.34ms -step:4100/20000 train_loss:2.0247 train_time:333474ms step_avg:81.34ms -step:4200/20000 train_loss:2.1603 train_time:341627ms step_avg:81.34ms -step:4300/20000 train_loss:2.0729 train_time:349735ms step_avg:81.33ms -step:4400/20000 train_loss:2.0484 train_time:357902ms step_avg:81.34ms -step:4500/20000 train_loss:2.1404 train_time:366055ms step_avg:81.35ms -step:4500/20000 val_loss:2.0620 val_bpb:1.2213 train_time:366085ms step_avg:81.35ms -step:4600/20000 train_loss:1.8604 train_time:374223ms step_avg:81.35ms -step:4700/20000 train_loss:2.2528 train_time:382306ms step_avg:81.34ms -step:4800/20000 train_loss:2.4504 train_time:390451ms step_avg:81.34ms -step:4900/20000 train_loss:2.0692 train_time:398606ms step_avg:81.35ms -step:5000/20000 train_loss:2.1254 train_time:406761ms step_avg:81.35ms -step:5000/20000 val_loss:2.0440 val_bpb:1.2106 train_time:406794ms step_avg:81.36ms -step:5100/20000 train_loss:2.1463 train_time:414937ms step_avg:81.36ms -step:5200/20000 train_loss:2.0638 train_time:423023ms step_avg:81.35ms -step:5300/20000 train_loss:2.0309 train_time:431178ms step_avg:81.35ms -step:5400/20000 train_loss:2.0713 train_time:439324ms step_avg:81.36ms -step:5500/20000 train_loss:2.0430 train_time:447480ms step_avg:81.36ms -step:5500/20000 val_loss:2.0270 val_bpb:1.2005 train_time:447511ms step_avg:81.37ms -step:5600/20000 train_loss:1.9830 train_time:455641ms step_avg:81.36ms -step:5700/20000 train_loss:2.0419 train_time:463731ms step_avg:81.36ms -step:5800/20000 train_loss:2.0311 train_time:471887ms step_avg:81.36ms -step:5900/20000 train_loss:1.9352 train_time:480056ms step_avg:81.37ms -swa:start step:6000 -step:6000/20000 train_loss:1.9705 train_time:488217ms step_avg:81.37ms -step:6000/20000 val_loss:2.0112 val_bpb:1.1912 train_time:488324ms step_avg:81.39ms -step:6100/20000 train_loss:1.9516 train_time:496385ms step_avg:81.37ms -step:6200/20000 train_loss:1.9816 train_time:504555ms step_avg:81.38ms -step:6300/20000 train_loss:1.9841 train_time:512752ms step_avg:81.39ms -step:6400/20000 train_loss:2.0356 train_time:520915ms step_avg:81.39ms -step:6500/20000 train_loss:2.1187 train_time:529113ms step_avg:81.40ms -step:6500/20000 val_loss:1.9924 val_bpb:1.1800 train_time:529143ms step_avg:81.41ms -step:6600/20000 train_loss:1.8839 train_time:537215ms step_avg:81.40ms -step:6700/20000 train_loss:1.9833 train_time:545429ms step_avg:81.41ms -step:6800/20000 train_loss:2.0671 train_time:553598ms step_avg:81.41ms -step:6900/20000 train_loss:1.8716 train_time:561814ms step_avg:81.42ms -step:7000/20000 train_loss:1.8377 train_time:569964ms step_avg:81.42ms -step:7000/20000 val_loss:1.9757 val_bpb:1.1701 train_time:570034ms step_avg:81.43ms -step:7100/20000 train_loss:1.9747 train_time:578104ms step_avg:81.42ms -step:7200/20000 train_loss:1.9239 train_time:586263ms step_avg:81.43ms -step:7300/20000 train_loss:2.0481 train_time:594456ms step_avg:81.43ms -step:7369/20000 val_loss:1.9654 val_bpb:1.1640 train_time:600072ms step_avg:81.43ms -stopping_early: wallclock_cap train_time:600072ms step:7369/20000 +step:1/20000 train_loss:6.9310 train_time:141ms step_avg:141.09ms +step:2/20000 train_loss:8.0394 train_time:192ms step_avg:96.13ms +step:3/20000 train_loss:7.5866 train_time:271ms step_avg:90.31ms +step:4/20000 train_loss:6.9692 train_time:350ms step_avg:87.56ms +step:5/20000 train_loss:6.7855 train_time:430ms step_avg:85.92ms +step:6/20000 train_loss:6.7112 train_time:509ms step_avg:84.87ms +step:7/20000 train_loss:6.5691 train_time:589ms step_avg:84.08ms +step:8/20000 train_loss:6.4878 train_time:668ms step_avg:83.54ms +step:9/20000 train_loss:6.2613 train_time:749ms step_avg:83.17ms +step:10/20000 train_loss:6.0139 train_time:828ms step_avg:82.78ms +step:100/20000 train_loss:3.2185 train_time:8047ms step_avg:80.47ms +step:200/20000 train_loss:2.4217 train_time:16138ms step_avg:80.69ms +step:300/20000 train_loss:2.5674 train_time:24230ms step_avg:80.77ms +step:400/20000 train_loss:2.4369 train_time:32327ms step_avg:80.82ms +step:500/20000 train_loss:2.4143 train_time:40357ms step_avg:80.71ms +step:500/20000 val_loss:2.3763 val_bpb:1.4074 train_time:40389ms step_avg:80.78ms +step:600/20000 train_loss:2.3498 train_time:48470ms step_avg:80.78ms +step:700/20000 train_loss:2.3629 train_time:56586ms step_avg:80.84ms +step:800/20000 train_loss:2.2553 train_time:64717ms step_avg:80.90ms +step:900/20000 train_loss:2.1451 train_time:72831ms step_avg:80.92ms +step:1000/20000 train_loss:2.2893 train_time:80896ms step_avg:80.90ms +step:1000/20000 val_loss:2.2408 val_bpb:1.3271 train_time:80927ms step_avg:80.93ms +step:1100/20000 train_loss:2.3339 train_time:89028ms step_avg:80.93ms +step:1200/20000 train_loss:2.3653 train_time:97141ms step_avg:80.95ms +step:1300/20000 train_loss:2.1113 train_time:105265ms step_avg:80.97ms +step:1400/20000 train_loss:2.1968 train_time:113412ms step_avg:81.01ms +step:1500/20000 train_loss:2.2328 train_time:121473ms step_avg:80.98ms +step:1500/20000 val_loss:2.1971 val_bpb:1.3012 train_time:121505ms step_avg:81.00ms +step:1600/20000 train_loss:2.0913 train_time:129610ms step_avg:81.01ms +step:1700/20000 train_loss:2.1588 train_time:137750ms step_avg:81.03ms +step:1800/20000 train_loss:2.1768 train_time:145887ms step_avg:81.05ms +step:1900/20000 train_loss:2.1430 train_time:153960ms step_avg:81.03ms +step:2000/20000 train_loss:2.0824 train_time:162084ms step_avg:81.04ms +step:2000/20000 val_loss:2.1445 val_bpb:1.2701 train_time:162115ms step_avg:81.06ms +step:2100/20000 train_loss:2.0592 train_time:170212ms step_avg:81.05ms +step:2200/20000 train_loss:2.1532 train_time:178342ms step_avg:81.06ms +step:2300/20000 train_loss:2.1209 train_time:186459ms step_avg:81.07ms +step:2400/20000 train_loss:2.0806 train_time:194529ms step_avg:81.05ms +step:2500/20000 train_loss:2.1814 train_time:202658ms step_avg:81.06ms +step:2500/20000 val_loss:2.1180 val_bpb:1.2544 train_time:202687ms step_avg:81.07ms +step:2600/20000 train_loss:2.1218 train_time:210795ms step_avg:81.08ms +step:2700/20000 train_loss:2.1135 train_time:218932ms step_avg:81.09ms +step:2800/20000 train_loss:2.1658 train_time:227072ms step_avg:81.10ms +step:2900/20000 train_loss:2.0330 train_time:235138ms step_avg:81.08ms +step:3000/20000 train_loss:2.1732 train_time:243268ms step_avg:81.09ms +step:3000/20000 val_loss:2.1037 val_bpb:1.2459 train_time:243297ms step_avg:81.10ms +step:3100/20000 train_loss:2.0492 train_time:251415ms step_avg:81.10ms +step:3200/20000 train_loss:2.1852 train_time:259549ms step_avg:81.11ms +step:3300/20000 train_loss:2.0811 train_time:267631ms step_avg:81.10ms +step:3400/20000 train_loss:2.0312 train_time:275775ms step_avg:81.11ms +step:3500/20000 train_loss:2.1952 train_time:283914ms step_avg:81.12ms +step:3500/20000 val_loss:2.0950 val_bpb:1.2408 train_time:283944ms step_avg:81.13ms +step:3600/20000 train_loss:2.1113 train_time:292049ms step_avg:81.12ms +step:3700/20000 train_loss:2.1110 train_time:300182ms step_avg:81.13ms +step:3800/20000 train_loss:2.0916 train_time:308247ms step_avg:81.12ms +step:3900/20000 train_loss:2.0976 train_time:316370ms step_avg:81.12ms +step:4000/20000 train_loss:1.9964 train_time:324501ms step_avg:81.13ms +step:4000/20000 val_loss:2.0891 val_bpb:1.2373 train_time:324533ms step_avg:81.13ms +step:4100/20000 train_loss:2.0379 train_time:332622ms step_avg:81.13ms +step:4200/20000 train_loss:2.1806 train_time:340747ms step_avg:81.13ms +step:4300/20000 train_loss:2.0848 train_time:348819ms step_avg:81.12ms +step:4400/20000 train_loss:2.0663 train_time:356948ms step_avg:81.12ms +step:4500/20000 train_loss:2.1568 train_time:365076ms step_avg:81.13ms +step:4500/20000 val_loss:2.0785 val_bpb:1.2310 train_time:365107ms step_avg:81.13ms +step:4600/20000 train_loss:1.8748 train_time:373196ms step_avg:81.13ms +step:4700/20000 train_loss:2.2622 train_time:381255ms step_avg:81.12ms +step:4800/20000 train_loss:2.4646 train_time:389386ms step_avg:81.12ms +step:4900/20000 train_loss:2.0864 train_time:397537ms step_avg:81.13ms +step:5000/20000 train_loss:2.1410 train_time:405671ms step_avg:81.13ms +step:5000/20000 val_loss:2.0593 val_bpb:1.2196 train_time:405701ms step_avg:81.14ms +step:5100/20000 train_loss:2.1628 train_time:413803ms step_avg:81.14ms +step:5200/20000 train_loss:2.0759 train_time:421867ms step_avg:81.13ms +step:5300/20000 train_loss:2.0441 train_time:430005ms step_avg:81.13ms +step:5400/20000 train_loss:2.0850 train_time:438138ms step_avg:81.14ms +step:5500/20000 train_loss:2.0548 train_time:446264ms step_avg:81.14ms +step:5500/20000 val_loss:2.0408 val_bpb:1.2087 train_time:446294ms step_avg:81.14ms +step:5600/20000 train_loss:1.9948 train_time:454407ms step_avg:81.14ms +step:5700/20000 train_loss:2.0521 train_time:462470ms step_avg:81.14ms +step:5800/20000 train_loss:2.0472 train_time:470606ms step_avg:81.14ms +swa:start step:5900 +step:5900/20000 train_loss:1.9475 train_time:478734ms step_avg:81.14ms +step:6000/20000 train_loss:1.9827 train_time:486956ms step_avg:81.16ms +step:6000/20000 val_loss:2.0230 val_bpb:1.1981 train_time:487010ms step_avg:81.17ms +step:6100/20000 train_loss:1.9602 train_time:495057ms step_avg:81.16ms +step:6200/20000 train_loss:1.9951 train_time:503235ms step_avg:81.17ms +step:6300/20000 train_loss:1.9914 train_time:511399ms step_avg:81.17ms +step:6400/20000 train_loss:2.0448 train_time:519600ms step_avg:81.19ms +step:6500/20000 train_loss:2.1264 train_time:527774ms step_avg:81.20ms +step:6500/20000 val_loss:1.9999 val_bpb:1.1845 train_time:527827ms step_avg:81.20ms +step:6600/20000 train_loss:1.8904 train_time:535881ms step_avg:81.19ms +step:6700/20000 train_loss:1.9893 train_time:544058ms step_avg:81.20ms +step:6800/20000 train_loss:2.0702 train_time:552251ms step_avg:81.21ms +step:6900/20000 train_loss:1.8737 train_time:560415ms step_avg:81.22ms +step:7000/20000 train_loss:1.8403 train_time:568606ms step_avg:81.23ms +step:7000/20000 val_loss:1.9777 val_bpb:1.1713 train_time:568677ms step_avg:81.24ms +step:7100/20000 train_loss:1.9747 train_time:576746ms step_avg:81.23ms +step:7200/20000 train_loss:1.9284 train_time:584938ms step_avg:81.24ms +step:7300/20000 train_loss:2.0473 train_time:593105ms step_avg:81.25ms +step:7385/20000 val_loss:1.9624 val_bpb:1.1622 train_time:600099ms step_avg:81.26ms +stopping_early: wallclock_cap train_time:600099ms step:7385/20000 peak memory allocated: 16962 MiB reserved: 17072 MiB -swa:applying averaged 7 checkpoints +swa:applying averaged 30 checkpoints Serialized model: 87413467 bytes -Code size: 52244 bytes -Total submission size: 87465711 bytes -Serialized model int6+zstd: 15829986 bytes -Total submission size int8+zlib: 15882230 bytes +Code size: 52243 bytes +Total submission size: 87465710 bytes +Serialized model int6+zstd: 15865061 bytes +Total submission size int8+zlib: 15917304 bytes final_eval_mode:sliding_window stride:64 batch_seqs:32 - sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.222434 - sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.143585 - sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.145154 - sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.138987 - sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.151322 - sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.152404 - sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.153844 - sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.149399 - sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.146812 - sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.148581 - sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.157180 - sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.155479 - sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.156651 - sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.154927 - sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.153421 - sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.153825 - sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.155192 - sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.155753 - sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.161960 - sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.159350 - sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.160298 - sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.158928 - sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.158223 - sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.157810 - sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.158468 - sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.156061 - sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.155090 - sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.155398 - sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.154158 - sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.154013 - sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.153286 - sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.154488 - sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.155554 - sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.156077 - sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.155535 - sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.155925 - sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.155077 - sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.151225 - sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.151308 - sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.152197 - sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.152359 - sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.152186 - sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.150935 - sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.150608 - sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.149901 - sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.149974 - sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.149899 - sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.150073 - sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.149781 - sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.150386 - sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.150708 - sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.150401 - sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.151451 - sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.153371 - sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.152664 - sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.153387 - sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.153717 - sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.153688 - sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.153285 - sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.153489 - sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.152888 - sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.155711 - sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.155713 - sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.155756 - sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.155396 - sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.154881 - sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.154144 - sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.154130 - sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.154755 - sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.154788 - sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.154768 - sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.155207 - sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.154940 - sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.154550 - sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.154844 - sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.154933 -final_int8_zlib_roundtrip val_loss:1.9392 val_bpb:1.1485 eval_time:155216ms -final_int8_zlib_roundtrip_exact val_loss:1.93923034 val_bpb:1.14852448 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.217812 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.142244 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.143869 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.137510 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.149324 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.150680 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.152112 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.147263 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.144500 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.146151 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.154878 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.153234 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.154387 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.152614 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.151132 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.151450 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.152771 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.153223 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.159424 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.156860 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.157801 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.156445 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.155906 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.155529 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.156210 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.153835 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.152829 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.153183 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.151985 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.151803 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.151061 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.152265 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.153299 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.153825 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.153325 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.153701 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.152846 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.148994 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.149131 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.150059 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.150232 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.150070 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.148854 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.148560 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.147909 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.148036 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.148014 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.148155 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.147836 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.148439 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.148726 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.148454 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.149482 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.151398 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.150673 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.151403 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.151744 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.151730 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.151305 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.151514 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.150913 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.153700 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.153718 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.153765 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.153384 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.152914 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.152196 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.152180 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.152788 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.152807 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.152799 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.153251 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.153003 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.152614 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.152912 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.153003 +final_int8_zlib_roundtrip val_loss:1.9359 val_bpb:1.1466 eval_time:154937ms +final_int8_zlib_roundtrip_exact val_loss:1.93591485 val_bpb:1.14656085 diff --git a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log index 9c6bc6e45a..cea46043a1 100644 --- a/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log +++ b/records/track_10min_16mb/2026-03-20_Int6_MLP3x_SmearGate_BigramHash_MuonWD_SWA/train_seed7.log @@ -29,189 +29,189 @@ warmup_step:18/20 warmup_step:19/20 warmup_step:20/20 step:0/20000 val_loss:6.9314 val_bpb:4.1051 train_time:0ms step_avg:0.02ms -step:1/20000 train_loss:6.9327 train_time:134ms step_avg:134.47ms -step:2/20000 train_loss:8.0943 train_time:185ms step_avg:92.54ms -step:3/20000 train_loss:7.6481 train_time:264ms step_avg:87.98ms -step:4/20000 train_loss:6.9718 train_time:344ms step_avg:85.90ms -step:5/20000 train_loss:6.8369 train_time:423ms step_avg:84.55ms -step:6/20000 train_loss:6.7023 train_time:504ms step_avg:83.96ms -step:7/20000 train_loss:6.5070 train_time:582ms step_avg:83.18ms -step:8/20000 train_loss:6.5039 train_time:662ms step_avg:82.75ms -step:9/20000 train_loss:6.3448 train_time:741ms step_avg:82.33ms -step:10/20000 train_loss:6.0817 train_time:821ms step_avg:82.10ms -step:100/20000 train_loss:3.2045 train_time:8037ms step_avg:80.37ms -step:200/20000 train_loss:2.4053 train_time:16152ms step_avg:80.76ms -step:300/20000 train_loss:2.5614 train_time:24281ms step_avg:80.94ms -step:400/20000 train_loss:2.4350 train_time:32392ms step_avg:80.98ms -step:500/20000 train_loss:2.4122 train_time:40450ms step_avg:80.90ms -step:500/20000 val_loss:2.3740 val_bpb:1.4060 train_time:40481ms step_avg:80.96ms -step:600/20000 train_loss:2.3460 train_time:48574ms step_avg:80.96ms -step:700/20000 train_loss:2.3605 train_time:56705ms step_avg:81.01ms -step:800/20000 train_loss:2.2470 train_time:64853ms step_avg:81.07ms -step:900/20000 train_loss:2.1392 train_time:72997ms step_avg:81.11ms -step:1000/20000 train_loss:2.2828 train_time:81083ms step_avg:81.08ms -step:1000/20000 val_loss:2.2377 val_bpb:1.3253 train_time:81114ms step_avg:81.11ms -step:1100/20000 train_loss:2.3363 train_time:89243ms step_avg:81.13ms -step:1200/20000 train_loss:2.3606 train_time:97390ms step_avg:81.16ms -step:1300/20000 train_loss:2.1073 train_time:105537ms step_avg:81.18ms -step:1400/20000 train_loss:2.1886 train_time:113691ms step_avg:81.21ms -step:1500/20000 train_loss:2.2287 train_time:121777ms step_avg:81.18ms -step:1500/20000 val_loss:2.1912 val_bpb:1.2978 train_time:121808ms step_avg:81.21ms -step:1600/20000 train_loss:2.0838 train_time:129936ms step_avg:81.21ms -step:1700/20000 train_loss:2.1470 train_time:138091ms step_avg:81.23ms -step:1800/20000 train_loss:2.1716 train_time:146247ms step_avg:81.25ms -step:1900/20000 train_loss:2.1330 train_time:154351ms step_avg:81.24ms -step:2000/20000 train_loss:2.0729 train_time:162512ms step_avg:81.26ms -step:2000/20000 val_loss:2.1397 val_bpb:1.2673 train_time:162542ms step_avg:81.27ms -step:2100/20000 train_loss:2.0503 train_time:170681ms step_avg:81.28ms -step:2200/20000 train_loss:2.1452 train_time:178837ms step_avg:81.29ms -step:2300/20000 train_loss:2.1125 train_time:187011ms step_avg:81.31ms -step:2400/20000 train_loss:2.0708 train_time:195104ms step_avg:81.29ms -step:2500/20000 train_loss:2.1699 train_time:203267ms step_avg:81.31ms -step:2500/20000 val_loss:2.1088 val_bpb:1.2490 train_time:203296ms step_avg:81.32ms -step:2600/20000 train_loss:2.1110 train_time:211437ms step_avg:81.32ms -step:2700/20000 train_loss:2.1028 train_time:219592ms step_avg:81.33ms -step:2800/20000 train_loss:2.1587 train_time:227765ms step_avg:81.34ms -step:2900/20000 train_loss:2.0227 train_time:235849ms step_avg:81.33ms -step:3000/20000 train_loss:2.1601 train_time:244006ms step_avg:81.34ms -step:3000/20000 val_loss:2.0908 val_bpb:1.2383 train_time:244036ms step_avg:81.35ms -step:3100/20000 train_loss:2.0338 train_time:252170ms step_avg:81.35ms -step:3200/20000 train_loss:2.1724 train_time:260324ms step_avg:81.35ms -step:3300/20000 train_loss:2.0687 train_time:268424ms step_avg:81.34ms -step:3400/20000 train_loss:2.0210 train_time:276580ms step_avg:81.35ms -step:3500/20000 train_loss:2.1823 train_time:284743ms step_avg:81.36ms -step:3500/20000 val_loss:2.0796 val_bpb:1.2316 train_time:284772ms step_avg:81.36ms -step:3600/20000 train_loss:2.0932 train_time:292896ms step_avg:81.36ms -step:3700/20000 train_loss:2.0919 train_time:301052ms step_avg:81.37ms -step:3800/20000 train_loss:2.0734 train_time:309150ms step_avg:81.36ms -step:3900/20000 train_loss:2.0769 train_time:317315ms step_avg:81.36ms -step:4000/20000 train_loss:1.9810 train_time:325466ms step_avg:81.37ms -step:4000/20000 val_loss:2.0707 val_bpb:1.2264 train_time:325497ms step_avg:81.37ms -step:4100/20000 train_loss:2.0206 train_time:333618ms step_avg:81.37ms -step:4200/20000 train_loss:2.1607 train_time:341759ms step_avg:81.37ms -step:4300/20000 train_loss:2.0721 train_time:349844ms step_avg:81.36ms -step:4400/20000 train_loss:2.0501 train_time:357997ms step_avg:81.36ms -step:4500/20000 train_loss:2.1355 train_time:366139ms step_avg:81.36ms -step:4500/20000 val_loss:2.0599 val_bpb:1.2200 train_time:366169ms step_avg:81.37ms -step:4600/20000 train_loss:1.8575 train_time:374302ms step_avg:81.37ms -step:4700/20000 train_loss:2.2476 train_time:382390ms step_avg:81.36ms -step:4800/20000 train_loss:2.4495 train_time:390547ms step_avg:81.36ms -step:4900/20000 train_loss:2.0665 train_time:398698ms step_avg:81.37ms -step:5000/20000 train_loss:2.1236 train_time:406858ms step_avg:81.37ms -step:5000/20000 val_loss:2.0421 val_bpb:1.2094 train_time:406888ms step_avg:81.38ms -step:5100/20000 train_loss:2.1414 train_time:415019ms step_avg:81.38ms -step:5200/20000 train_loss:2.0584 train_time:423112ms step_avg:81.37ms -step:5300/20000 train_loss:2.0299 train_time:431261ms step_avg:81.37ms -step:5400/20000 train_loss:2.0701 train_time:439438ms step_avg:81.38ms -step:5500/20000 train_loss:2.0398 train_time:447597ms step_avg:81.38ms -step:5500/20000 val_loss:2.0249 val_bpb:1.1992 train_time:447627ms step_avg:81.39ms -step:5600/20000 train_loss:1.9804 train_time:455754ms step_avg:81.38ms -step:5700/20000 train_loss:2.0405 train_time:463856ms step_avg:81.38ms -step:5800/20000 train_loss:2.0266 train_time:472009ms step_avg:81.38ms -step:5900/20000 train_loss:1.9338 train_time:480166ms step_avg:81.38ms -swa:start step:6000 -step:6000/20000 train_loss:1.9689 train_time:488320ms step_avg:81.39ms -step:6000/20000 val_loss:2.0091 val_bpb:1.1899 train_time:488430ms step_avg:81.40ms -step:6100/20000 train_loss:1.9461 train_time:496507ms step_avg:81.39ms -step:6200/20000 train_loss:1.9828 train_time:504682ms step_avg:81.40ms -step:6300/20000 train_loss:1.9793 train_time:512879ms step_avg:81.41ms -step:6400/20000 train_loss:2.0335 train_time:521031ms step_avg:81.41ms -step:6500/20000 train_loss:2.1165 train_time:529232ms step_avg:81.42ms -step:6500/20000 val_loss:1.9902 val_bpb:1.1787 train_time:529262ms step_avg:81.42ms -step:6600/20000 train_loss:1.8805 train_time:537324ms step_avg:81.41ms -step:6700/20000 train_loss:1.9845 train_time:545537ms step_avg:81.42ms -step:6800/20000 train_loss:2.0656 train_time:553694ms step_avg:81.43ms -step:6900/20000 train_loss:1.8698 train_time:561893ms step_avg:81.43ms -step:7000/20000 train_loss:1.8330 train_time:570047ms step_avg:81.44ms -step:7000/20000 val_loss:1.9734 val_bpb:1.1688 train_time:570117ms step_avg:81.45ms -step:7100/20000 train_loss:1.9711 train_time:578194ms step_avg:81.44ms -step:7200/20000 train_loss:1.9240 train_time:586353ms step_avg:81.44ms -step:7300/20000 train_loss:2.0450 train_time:594564ms step_avg:81.45ms -step:7368/20000 val_loss:1.9632 val_bpb:1.1627 train_time:600093ms step_avg:81.45ms -stopping_early: wallclock_cap train_time:600093ms step:7368/20000 +step:1/20000 train_loss:6.9327 train_time:139ms step_avg:138.73ms +step:2/20000 train_loss:8.0944 train_time:189ms step_avg:94.62ms +step:3/20000 train_loss:7.6485 train_time:268ms step_avg:89.29ms +step:4/20000 train_loss:6.9720 train_time:348ms step_avg:87.04ms +step:5/20000 train_loss:6.8367 train_time:427ms step_avg:85.34ms +step:6/20000 train_loss:6.7016 train_time:506ms step_avg:84.36ms +step:7/20000 train_loss:6.5065 train_time:585ms step_avg:83.61ms +step:8/20000 train_loss:6.5046 train_time:665ms step_avg:83.08ms +step:9/20000 train_loss:6.3458 train_time:745ms step_avg:82.75ms +step:10/20000 train_loss:6.0822 train_time:824ms step_avg:82.41ms +step:100/20000 train_loss:3.2092 train_time:8024ms step_avg:80.24ms +step:200/20000 train_loss:2.4059 train_time:16105ms step_avg:80.53ms +step:300/20000 train_loss:2.5671 train_time:24186ms step_avg:80.62ms +step:400/20000 train_loss:2.4327 train_time:32257ms step_avg:80.64ms +step:500/20000 train_loss:2.4160 train_time:40268ms step_avg:80.54ms +step:500/20000 val_loss:2.3756 val_bpb:1.4069 train_time:40298ms step_avg:80.60ms +step:600/20000 train_loss:2.3475 train_time:48357ms step_avg:80.60ms +step:700/20000 train_loss:2.3648 train_time:56470ms step_avg:80.67ms +step:800/20000 train_loss:2.2575 train_time:64595ms step_avg:80.74ms +step:900/20000 train_loss:2.1461 train_time:72694ms step_avg:80.77ms +step:1000/20000 train_loss:2.2902 train_time:80736ms step_avg:80.74ms +step:1000/20000 val_loss:2.2431 val_bpb:1.3285 train_time:80767ms step_avg:80.77ms +step:1100/20000 train_loss:2.3372 train_time:88855ms step_avg:80.78ms +step:1200/20000 train_loss:2.3680 train_time:96961ms step_avg:80.80ms +step:1300/20000 train_loss:2.1164 train_time:105088ms step_avg:80.84ms +step:1400/20000 train_loss:2.1986 train_time:113201ms step_avg:80.86ms +step:1500/20000 train_loss:2.2387 train_time:121247ms step_avg:80.83ms +step:1500/20000 val_loss:2.1991 val_bpb:1.3024 train_time:121278ms step_avg:80.85ms +step:1600/20000 train_loss:2.0927 train_time:129348ms step_avg:80.84ms +step:1700/20000 train_loss:2.1590 train_time:137469ms step_avg:80.86ms +step:1800/20000 train_loss:2.1712 train_time:145593ms step_avg:80.89ms +step:1900/20000 train_loss:2.1385 train_time:153651ms step_avg:80.87ms +step:2000/20000 train_loss:2.0794 train_time:161771ms step_avg:80.89ms +step:2000/20000 val_loss:2.1462 val_bpb:1.2711 train_time:161800ms step_avg:80.90ms +step:2100/20000 train_loss:2.0627 train_time:169910ms step_avg:80.91ms +step:2200/20000 train_loss:2.1523 train_time:178030ms step_avg:80.92ms +step:2300/20000 train_loss:2.1207 train_time:186161ms step_avg:80.94ms +step:2400/20000 train_loss:2.0789 train_time:194213ms step_avg:80.92ms +step:2500/20000 train_loss:2.1794 train_time:202352ms step_avg:80.94ms +step:2500/20000 val_loss:2.1183 val_bpb:1.2546 train_time:202382ms step_avg:80.95ms +step:2600/20000 train_loss:2.1236 train_time:210483ms step_avg:80.96ms +step:2700/20000 train_loss:2.1158 train_time:218597ms step_avg:80.96ms +step:2800/20000 train_loss:2.1667 train_time:226733ms step_avg:80.98ms +step:2900/20000 train_loss:2.0343 train_time:234798ms step_avg:80.96ms +step:3000/20000 train_loss:2.1712 train_time:242909ms step_avg:80.97ms +step:3000/20000 val_loss:2.1029 val_bpb:1.2455 train_time:242938ms step_avg:80.98ms +step:3100/20000 train_loss:2.0485 train_time:251038ms step_avg:80.98ms +step:3200/20000 train_loss:2.1859 train_time:259153ms step_avg:80.99ms +step:3300/20000 train_loss:2.0814 train_time:267203ms step_avg:80.97ms +step:3400/20000 train_loss:2.0306 train_time:275332ms step_avg:80.98ms +step:3500/20000 train_loss:2.1933 train_time:283448ms step_avg:80.99ms +step:3500/20000 val_loss:2.0936 val_bpb:1.2399 train_time:283478ms step_avg:80.99ms +step:3600/20000 train_loss:2.1121 train_time:291568ms step_avg:80.99ms +step:3700/20000 train_loss:2.1086 train_time:299670ms step_avg:80.99ms +step:3800/20000 train_loss:2.0914 train_time:307736ms step_avg:80.98ms +step:3900/20000 train_loss:2.0955 train_time:315873ms step_avg:80.99ms +step:4000/20000 train_loss:1.9926 train_time:323983ms step_avg:81.00ms +step:4000/20000 val_loss:2.0874 val_bpb:1.2363 train_time:324012ms step_avg:81.00ms +step:4100/20000 train_loss:2.0390 train_time:332110ms step_avg:81.00ms +step:4200/20000 train_loss:2.1809 train_time:340221ms step_avg:81.01ms +step:4300/20000 train_loss:2.0871 train_time:348269ms step_avg:80.99ms +step:4400/20000 train_loss:2.0686 train_time:356381ms step_avg:81.00ms +step:4500/20000 train_loss:2.1547 train_time:364500ms step_avg:81.00ms +step:4500/20000 val_loss:2.0773 val_bpb:1.2303 train_time:364529ms step_avg:81.01ms +step:4600/20000 train_loss:1.8733 train_time:372623ms step_avg:81.01ms +step:4700/20000 train_loss:2.2637 train_time:380674ms step_avg:80.99ms +step:4800/20000 train_loss:2.4601 train_time:388792ms step_avg:81.00ms +step:4900/20000 train_loss:2.0823 train_time:396926ms step_avg:81.01ms +step:5000/20000 train_loss:2.1381 train_time:405050ms step_avg:81.01ms +step:5000/20000 val_loss:2.0581 val_bpb:1.2189 train_time:405079ms step_avg:81.02ms +step:5100/20000 train_loss:2.1576 train_time:413163ms step_avg:81.01ms +step:5200/20000 train_loss:2.0770 train_time:421218ms step_avg:81.00ms +step:5300/20000 train_loss:2.0447 train_time:429326ms step_avg:81.00ms +step:5400/20000 train_loss:2.0824 train_time:437472ms step_avg:81.01ms +step:5500/20000 train_loss:2.0543 train_time:445595ms step_avg:81.02ms +step:5500/20000 val_loss:2.0392 val_bpb:1.2078 train_time:445625ms step_avg:81.02ms +step:5600/20000 train_loss:1.9955 train_time:453729ms step_avg:81.02ms +step:5700/20000 train_loss:2.0511 train_time:461794ms step_avg:81.02ms +step:5800/20000 train_loss:2.0449 train_time:469913ms step_avg:81.02ms +step:5900/20000 train_loss:1.9491 train_time:478054ms step_avg:81.03ms +swa:start step:5950 +step:6000/20000 train_loss:1.9804 train_time:486249ms step_avg:81.04ms +step:6000/20000 val_loss:2.0216 val_bpb:1.1973 train_time:486308ms step_avg:81.05ms +step:6100/20000 train_loss:1.9605 train_time:494353ms step_avg:81.04ms +step:6200/20000 train_loss:1.9964 train_time:502535ms step_avg:81.05ms +step:6300/20000 train_loss:1.9883 train_time:510749ms step_avg:81.07ms +step:6400/20000 train_loss:2.0438 train_time:518937ms step_avg:81.08ms +step:6500/20000 train_loss:2.1242 train_time:527090ms step_avg:81.09ms +step:6500/20000 val_loss:1.9985 val_bpb:1.1836 train_time:527147ms step_avg:81.10ms +step:6600/20000 train_loss:1.8904 train_time:535211ms step_avg:81.09ms +step:6700/20000 train_loss:1.9885 train_time:543382ms step_avg:81.10ms +step:6800/20000 train_loss:2.0730 train_time:551575ms step_avg:81.11ms +step:6900/20000 train_loss:1.8732 train_time:559768ms step_avg:81.13ms +step:7000/20000 train_loss:1.8375 train_time:567934ms step_avg:81.13ms +step:7000/20000 val_loss:1.9763 val_bpb:1.1705 train_time:567990ms step_avg:81.14ms +step:7100/20000 train_loss:1.9738 train_time:576059ms step_avg:81.14ms +step:7200/20000 train_loss:1.9234 train_time:584245ms step_avg:81.15ms +step:7300/20000 train_loss:2.0473 train_time:592436ms step_avg:81.16ms +step:7393/20000 val_loss:1.9605 val_bpb:1.1611 train_time:600101ms step_avg:81.17ms +stopping_early: wallclock_cap train_time:600101ms step:7393/20000 peak memory allocated: 16962 MiB reserved: 17072 MiB -swa:applying averaged 7 checkpoints +swa:applying averaged 29 checkpoints Serialized model: 87413467 bytes -Code size: 52244 bytes -Total submission size: 87465711 bytes -Serialized model int6+zstd: 15823762 bytes -Total submission size int8+zlib: 15876006 bytes +Code size: 52243 bytes +Total submission size: 87465710 bytes +Serialized model int6+zstd: 15892664 bytes +Total submission size int8+zlib: 15944907 bytes final_eval_mode:sliding_window stride:64 batch_seqs:32 - sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.216312 - sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.143074 - sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.144524 - sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.138041 - sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.149896 - sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.151041 - sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.152472 - sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.148062 - sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.145451 - sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.146927 - sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.155633 - sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.153787 - sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.155052 - sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.153311 - sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.151851 - sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.152198 - sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.153644 - sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.154183 - sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.160298 - sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.157638 - sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.158639 - sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.157312 - sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.156654 - sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.156268 - sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.156905 - sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.154552 - sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.153516 - sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.153907 - sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.152738 - sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.152602 - sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.151837 - sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.153034 - sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.154107 - sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.154615 - sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.154091 - sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.154458 - sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.153583 - sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.149718 - sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.149822 - sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.150765 - sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.150898 - sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.150747 - sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.149526 - sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.149247 - sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.148578 - sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.148680 - sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.148661 - sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.148820 - sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.148532 - sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.149157 - sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.149455 - sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.149182 - sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.150203 - sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.152094 - sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.151393 - sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.152138 - sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.152508 - sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.152472 - sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.152045 - sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.152252 - sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.151675 - sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.154459 - sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.154453 - sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.154496 - sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.154117 - sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.153594 - sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.152849 - sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.152836 - sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.153429 - sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.153450 - sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.153436 - sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.153877 - sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.153634 - sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.153252 - sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.153561 - sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.153647 -final_int8_zlib_roundtrip val_loss:1.9376 val_bpb:1.1476 eval_time:155316ms -final_int8_zlib_roundtrip_exact val_loss:1.93761732 val_bpb:1.14756915 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.204187 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.140479 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.141456 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.134566 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.146982 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.148125 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.149646 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.145113 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.142509 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.144101 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.153069 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.151471 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.152853 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.151038 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.149419 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.149769 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.151145 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.151649 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.157772 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.155120 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.156148 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.154779 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.154098 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.153707 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.154262 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151908 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150870 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.151243 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.150119 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.150033 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.149298 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.150513 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.151542 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.152059 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.151549 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151954 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.151101 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.147247 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.147351 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.148300 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.148482 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.148331 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.147109 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.146807 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.146093 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.146142 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.146091 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.146251 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145998 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.146598 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.146887 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.146625 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.147686 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.149608 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.148926 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.149671 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.150045 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.150024 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.149607 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.149839 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.149256 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.152058 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.152077 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.152093 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.151727 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.151235 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.150486 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.150457 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.151088 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.151106 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.151078 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.151528 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.151293 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.150913 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.151248 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.151319 +final_int8_zlib_roundtrip val_loss:1.9331 val_bpb:1.1449 eval_time:155292ms +final_int8_zlib_roundtrip_exact val_loss:1.93314046 val_bpb:1.14491770