From 180e457d886c4abbc56d8414846a4d9c42ee4f10 Mon Sep 17 00:00:00 2001 From: Tim Pietrusky Date: Sat, 28 Mar 2026 21:59:00 +0100 Subject: [PATCH] =?UTF-8?q?Record:=20Muon=20TTT=20+=20Entropy-Adaptive=20E?= =?UTF-8?q?pochs=20=E2=80=94=20val=5Fbpb=201.1179=20(3-seed=20mean)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../README.md | 40 + .../submission.json | 25 + .../train_gpt.py | 1954 +++++++++++++++++ .../train_seed1337.log | 275 +++ .../train_seed2025.log | 275 +++ .../train_seed42.log | 275 +++ 6 files changed, 2844 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/README.md create mode 100644 records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/submission.json create mode 100755 records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_gpt.py create mode 100644 records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed2025.log create mode 100644 records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed42.log diff --git a/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/README.md b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/README.md new file mode 100644 index 000000000..a1e4dba38 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/README.md @@ -0,0 +1,40 @@ +# Record: Muon TTT + Entropy-Adaptive Epochs — val_bpb 1.1179 (3-seed mean) + +## 3-Seed Results + +| Seed | val_bpb | Eval time | Artifact | +|------|---------|-----------|----------| +| 1337 | **1.1173** | 594s | 15.95MB | +| 42 | **1.1181** | 598s | 16.06MB | +| 2025 | **1.1183** | 603s | 15.94MB | +| **Mean** | **1.1179** | | | +| **Std** | **0.0005** | | | + +## Method + +### Architecture +- 11 layers, 512 dim, 8 heads, 4 KV heads (GQA) +- MLP 3.0x with LeakyReLU(0.5)^2 +- XSA on last 4 layers, partial RoPE (16 dims) +- BigramHash embeddings, SmearGate +- Value embeddings on layers 9-10 +- Tied embeddings, logit softcap=30 + +### Training +- Muon optimizer (lr=0.025, momentum 0.99 warmed from 0.92) +- 786K batch tokens, 2048 seq len, 600s wallclock +- EMA (0.997) + SWA every 50 steps +- CROWN-Q QAT during warmdown +- Int6 GPTQ + LZMA compression, 4% pruning + +### Test-Time Training (Legal, Score-First) +- Muon-style Newton-Schulz orthogonalized TTT updates +- Entropy-adaptive epoch selection (harder chunks get more epochs) +- 32K token chunks, stride 64 +- All blocks unfrozen during TTT +- Score-first: tokens scored BEFORE any weight update + +## Credits +- Based on PR #999 by @contributor (Muon TTT + entropy-adaptive epochs) +- CROWN-Q from PR #995 architecture stack +- Built and validated by @TimPietrusky with Claude Code diff --git a/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/submission.json b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/submission.json new file mode 100644 index 000000000..b60449c62 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/submission.json @@ -0,0 +1,25 @@ +{ + "record_name": "Muon TTT + Entropy-Adaptive Epochs + GQA + CROWN-Q", + "val_bpb_mean": 0.1179, + "val_bpb_std": 0.0005, + "seeds": [1337, 42, 2025], + "seed_results": { + "1337": 1.1173, + "42": 1.1181, + "2025": 1.1183 + }, + "hardware": "8xH100 SXM 80GB", + "train_time_seconds": 600, + "eval_time_seconds": 595, + "artifact_bytes": 15950000, + "techniques": [ + "Muon TTT (Newton-Schulz orthogonalized updates)", + "Entropy-adaptive epoch selection", + "GQA (8 heads, 4 KV heads)", + "CROWN-Q soft-round QAT", + "LeakyReLU(0.5)^2 activation", + "EMA + SWA", + "Int6 GPTQ + LZMA compression", + "Score-first legal TTT" + ] +} diff --git a/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_gpt.py b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_gpt.py new file mode 100755 index 000000000..60f976337 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_gpt.py @@ -0,0 +1,1954 @@ +from __future__ import annotations +import copy +import glob +import io +import lzma +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 +from flash_attn_interface import flash_attn_func as flash_attn_3_func +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", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + 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)) + eval_seq_len = int(os.environ.get("EVAL_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", 11)) + 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", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + 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)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 2)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_muon = bool(int(os.environ.get("TTT_MUON", "1"))) # use Muon-style NS update for TTT + ttt_ns_steps = int(os.environ.get("TTT_NS_STEPS", "3")) # Newton-Schulz steps for TTT Muon + ttt_entropy_adapt = bool(int(os.environ.get("TTT_ENTROPY_ADAPT", "1"))) # entropy-adaptive epochs + ttt_entropy_high = float(os.environ.get("TTT_ENTROPY_HIGH", "2.1")) # nats – hard chunk threshold + ttt_entropy_low = float(os.environ.get("TTT_ENTROPY_LOW", "1.75")) # nats – easy chunk threshold + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + 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), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + 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: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +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, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < 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}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // 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.no_grad(): + 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 * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, 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) + +# --- Quantization helpers --- + +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,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + 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 quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +# --- Data loading --- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + 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): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + 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, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + 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 + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + 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, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + 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, + gated_attention: bool = False, + value_residual: bool = False, + ): + 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") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + 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, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + 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): + 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 ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + 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()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + 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.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + 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.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + 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.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = 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) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + 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) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + 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 + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, 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_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + 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 + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + 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) + +# --- Sliding window evaluation --- + +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, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or 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() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.no_grad(): + 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 = compiled_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 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 + + +def eval_val_sliding_ttt( + 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, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + 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) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + # Muon TTT: no external optimizer; we apply Newton-Schulz orthogonalized updates + # For non-matrix params (1-D) we use plain LR-scaled gradient updates (like AdamW scalar track) + use_muon_ttt = args.ttt_muon + if not use_muon_ttt: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + else: + optimizer = None # manual update below + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # Track per-chunk NLL for entropy-adaptive epoch selection + chunk_loss_sum = 0.0 + chunk_token_count = 0 + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.no_grad(): + 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_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[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) + chunk_loss_sum += scored_nll.sum().item() + chunk_token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], 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() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if not use_muon_ttt: + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + # --- Entropy-adaptive epoch count (synchronized across ranks) --- + # Sync chunk NLL across all ranks so every rank uses the same effective_epochs + chunk_nll = float('inf') + if args.ttt_entropy_adapt: + cls_t = torch.tensor(chunk_loss_sum, device=device, dtype=torch.float64) + ctc_t = torch.tensor(chunk_token_count, device=device, dtype=torch.float64) + if world_size > 1: + dist.all_reduce(cls_t, op=dist.ReduceOp.SUM) + dist.all_reduce(ctc_t, op=dist.ReduceOp.SUM) + if ctc_t.item() > 0: + chunk_nll = (cls_t / ctc_t).item() + effective_epochs = args.ttt_epochs + if args.ttt_entropy_adapt: + if chunk_nll > args.ttt_entropy_high: + effective_epochs = args.ttt_epochs + 1 # hard chunk → extra epoch + elif chunk_nll < args.ttt_entropy_low: + effective_epochs = max(args.ttt_epochs - 1, 1) # easy chunk → save time + # Wall-clock guard: if we're past 85% of soft eval cap, cap at baseline epochs + elapsed_now = time.perf_counter() - t0 + eval_soft_cap = float(os.environ.get("EVAL_TIME_SOFT_CAP_SECONDS", "570")) + if elapsed_now > eval_soft_cap * 0.85: + effective_epochs = min(effective_epochs, args.ttt_epochs) + + for _ep in range(effective_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + if not use_muon_ttt: + optimizer.zero_grad(set_to_none=True) + else: + for p in ttt_params: + p.grad = None + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + if not use_muon_ttt: + optimizer.step() + else: + # Muon-style update: Newton-Schulz orthogonalization for matrix params, + # plain LR-scaled gradient for vector params (norms, biases, scalars). + with torch.no_grad(): + for p in ttt_params: + if p.grad is None: + continue + g = p.grad.detach().float() + if g.ndim >= 2: + g = zeropower_via_newtonschulz5(g, steps=args.ttt_ns_steps) + p.data.add_(g.to(p.dtype), alpha=-cos_lr) + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + 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() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +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, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + 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 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.get(name) + if info is None: + continue + 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 + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + 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"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_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}") + CastedLinear._qat_enabled = args.qat_enabled + 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, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + 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: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + 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.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + 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, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + 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): + 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() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + 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() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + 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) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 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 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + 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 weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_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"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "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") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=7) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_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, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_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_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed1337.log b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed1337.log new file mode 100644 index 000000000..7059d8a3b --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed1337.log @@ -0,0 +1,275 @@ +W0328 16:40:38.067000 3358211 torch/distributed/run.py:803] +W0328 16:40:38.067000 3358211 torch/distributed/run.py:803] ***************************************** +W0328 16:40:38.067000 3358211 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0328 16:40:38.067000 3358211 torch/distributed/run.py:803] ***************************************** +logs/c5b8380c-bf4c-4584-8c0b-15071c56991c.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:26993756 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +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:6.9309 val_bpb:4.1049 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9317 train_time:134ms step_avg:133.87ms +step:2/20000 train_loss:8.6535 train_time:164ms step_avg:82.11ms +step:3/20000 train_loss:7.6847 train_time:244ms step_avg:81.24ms +step:4/20000 train_loss:7.2552 train_time:323ms step_avg:80.80ms +step:5/20000 train_loss:7.1506 train_time:404ms step_avg:80.85ms +step:6/20000 train_loss:7.1066 train_time:485ms step_avg:80.86ms +step:7/20000 train_loss:6.9991 train_time:567ms step_avg:80.95ms +step:8/20000 train_loss:6.9259 train_time:648ms step_avg:81.03ms +step:9/20000 train_loss:6.5602 train_time:731ms step_avg:81.19ms +step:10/20000 train_loss:6.1613 train_time:813ms step_avg:81.32ms +step:500/20000 train_loss:2.3866 train_time:41702ms step_avg:83.40ms +step:1000/20000 train_loss:2.2663 train_time:83662ms step_avg:83.66ms +step:1500/20000 train_loss:2.2105 train_time:125710ms step_avg:83.81ms +step:2000/20000 train_loss:2.0516 train_time:167798ms step_avg:83.90ms +step:2500/20000 train_loss:2.1573 train_time:209966ms step_avg:83.99ms +step:3000/20000 train_loss:2.1488 train_time:252119ms step_avg:84.04ms +step:3500/20000 train_loss:2.1688 train_time:294215ms step_avg:84.06ms +step:4000/20000 train_loss:1.9646 train_time:336318ms step_avg:84.08ms +step:4000/20000 val_loss:2.0545 val_bpb:1.2168 train_time:336375ms step_avg:84.09ms +step:4500/20000 train_loss:2.1103 train_time:378475ms step_avg:84.11ms +step:5000/20000 train_loss:2.0954 train_time:420651ms step_avg:84.13ms +step:5500/20000 train_loss:2.0133 train_time:462786ms step_avg:84.14ms +step:6000/20000 train_loss:1.9341 train_time:504957ms step_avg:84.16ms +swa:start step:6450 +step:6500/20000 train_loss:2.0771 train_time:547245ms step_avg:84.19ms +late_qat:enabled step:6599 scale:0.1499 +step:7000/20000 train_loss:1.7844 train_time:590200ms step_avg:84.31ms +step:7115/20000 val_loss:1.9203 val_bpb:1.1373 train_time:600122ms step_avg:84.35ms +stopping_early: wallclock_cap train_time:600122ms step:7115/20000 +peak memory allocated: 21472 MiB reserved: 22004 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9185 val_bpb:1.1363 eval_time:2012ms +Serialized model: 106158518 bytes +Code size: 93017 bytes +Serialized model int6+lzma: 15854080 bytes +Total submission size int6+lzma: 15947097 bytes +final_int6_roundtrip val_loss:1.9325 val_bpb:1.1446 eval_time:4182ms +final_int6_roundtrip_exact val_loss:1.93254463 val_bpb:1.14456178 +final_int6_sliding_window val_loss:1.8928 val_bpb:1.1210 stride:64 eval_time:76585ms +final_int6_sliding_window_exact val_loss:1.89277953 val_bpb:1.12101362 +final_int8_zlib_roundtrip_exact val_loss:1.89277953 val_bpb:1.12101362 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=2 +ttt_sliding:params unfrozen=26989644 frozen=4112 + ttt_chunk [1/1893] bpb=1.154444 time=0.6s + ttt_chunk [11/1893] bpb=1.142298 time=3.3s + ttt_chunk [21/1893] bpb=1.129419 time=6.0s + ttt_chunk [31/1893] bpb=1.127100 time=8.9s + ttt_chunk [41/1893] bpb=1.113281 time=11.6s + ttt_chunk [51/1893] bpb=1.107757 time=14.3s + ttt_chunk [61/1893] bpb=1.114474 time=17.2s + ttt_chunk [71/1893] bpb=1.113251 time=20.2s + ttt_chunk [81/1893] bpb=1.112575 time=23.0s + ttt_chunk [91/1893] bpb=1.113679 time=25.7s + ttt_chunk [101/1893] bpb=1.117318 time=28.6s + ttt_chunk [111/1893] bpb=1.119911 time=31.3s + ttt_chunk [121/1893] bpb=1.113054 time=34.0s + ttt_chunk [131/1893] bpb=1.113091 time=37.0s + ttt_chunk [141/1893] bpb=1.118843 time=40.0s + ttt_chunk [151/1893] bpb=1.120559 time=43.1s + ttt_chunk [161/1893] bpb=1.119972 time=46.2s + ttt_chunk [171/1893] bpb=1.124444 time=49.2s + ttt_chunk [181/1893] bpb=1.126657 time=52.2s + ttt_chunk [191/1893] bpb=1.133905 time=55.3s + ttt_chunk [201/1893] bpb=1.132634 time=58.3s + ttt_chunk [211/1893] bpb=1.130499 time=61.3s + ttt_chunk [221/1893] bpb=1.132024 time=64.2s + ttt_chunk [231/1893] bpb=1.130757 time=67.2s + ttt_chunk [241/1893] bpb=1.131066 time=70.2s + ttt_chunk [251/1893] bpb=1.130639 time=73.2s + ttt_chunk [261/1893] bpb=1.127807 time=76.1s + ttt_chunk [271/1893] bpb=1.126724 time=79.0s + ttt_chunk [281/1893] bpb=1.128104 time=81.9s + ttt_chunk [291/1893] bpb=1.129955 time=84.7s + ttt_chunk [301/1893] bpb=1.130669 time=87.5s + ttt_chunk [311/1893] bpb=1.132755 time=90.2s + ttt_chunk [321/1893] bpb=1.134801 time=93.0s + ttt_chunk [331/1893] bpb=1.134670 time=95.8s + ttt_chunk [341/1893] bpb=1.133678 time=98.8s + ttt_chunk [351/1893] bpb=1.136000 time=102.0s + ttt_chunk [361/1893] bpb=1.136075 time=105.0s + ttt_chunk [371/1893] bpb=1.135377 time=108.2s + ttt_chunk [381/1893] bpb=1.135587 time=111.5s + ttt_chunk [391/1893] bpb=1.135460 time=114.7s + ttt_chunk [401/1893] bpb=1.133381 time=117.9s + ttt_chunk [411/1893] bpb=1.132219 time=121.1s + ttt_chunk [421/1893] bpb=1.131358 time=124.2s + ttt_chunk [431/1893] bpb=1.131278 time=127.5s + ttt_chunk [441/1893] bpb=1.131646 time=130.7s + ttt_chunk [451/1893] bpb=1.131938 time=133.9s + ttt_chunk [461/1893] bpb=1.130850 time=137.2s + ttt_chunk [471/1893] bpb=1.131547 time=140.5s + ttt_chunk [481/1893] bpb=1.131175 time=143.7s + ttt_chunk [491/1893] bpb=1.130080 time=146.8s + ttt_chunk [501/1893] bpb=1.129648 time=150.1s + ttt_chunk [511/1893] bpb=1.128980 time=153.3s + ttt_chunk [521/1893] bpb=1.126561 time=156.5s + ttt_chunk [531/1893] bpb=1.127723 time=159.8s + ttt_chunk [541/1893] bpb=1.128051 time=163.0s + ttt_chunk [551/1893] bpb=1.127000 time=166.3s + ttt_chunk [561/1893] bpb=1.127611 time=169.5s + ttt_chunk [571/1893] bpb=1.126597 time=172.7s + ttt_chunk [581/1893] bpb=1.125781 time=175.8s + ttt_chunk [591/1893] bpb=1.125144 time=178.8s + ttt_chunk [601/1893] bpb=1.125657 time=182.0s + ttt_chunk [611/1893] bpb=1.125568 time=185.2s + ttt_chunk [621/1893] bpb=1.125443 time=188.4s + ttt_chunk [631/1893] bpb=1.126152 time=191.5s + ttt_chunk [641/1893] bpb=1.125918 time=194.6s + ttt_chunk [651/1893] bpb=1.126053 time=197.7s + ttt_chunk [661/1893] bpb=1.125531 time=200.8s + ttt_chunk [671/1893] bpb=1.125908 time=204.0s + ttt_chunk [681/1893] bpb=1.126637 time=207.1s + ttt_chunk [691/1893] bpb=1.127606 time=210.3s + ttt_chunk [701/1893] bpb=1.127044 time=213.4s + ttt_chunk [711/1893] bpb=1.126984 time=216.6s + ttt_chunk [721/1893] bpb=1.126639 time=219.7s + ttt_chunk [731/1893] bpb=1.126706 time=222.9s + ttt_chunk [741/1893] bpb=1.126830 time=226.0s + ttt_chunk [751/1893] bpb=1.126674 time=229.2s + ttt_chunk [761/1893] bpb=1.126626 time=232.4s + ttt_chunk [771/1893] bpb=1.126303 time=235.6s + ttt_chunk [781/1893] bpb=1.127054 time=238.8s + ttt_chunk [791/1893] bpb=1.126619 time=241.9s + ttt_chunk [801/1893] bpb=1.126915 time=245.1s + ttt_chunk [811/1893] bpb=1.126683 time=248.3s + ttt_chunk [821/1893] bpb=1.126449 time=251.5s + ttt_chunk [831/1893] bpb=1.126266 time=254.7s + ttt_chunk [841/1893] bpb=1.125624 time=257.9s + ttt_chunk [851/1893] bpb=1.125343 time=261.1s + ttt_chunk [861/1893] bpb=1.125081 time=264.3s + ttt_chunk [871/1893] bpb=1.125332 time=267.6s + ttt_chunk [881/1893] bpb=1.125502 time=270.5s + ttt_chunk [891/1893] bpb=1.125061 time=273.5s + ttt_chunk [901/1893] bpb=1.124799 time=276.6s + ttt_chunk [911/1893] bpb=1.124875 time=279.7s + ttt_chunk [921/1893] bpb=1.125331 time=283.0s + ttt_chunk [931/1893] bpb=1.125299 time=286.2s + ttt_chunk [941/1893] bpb=1.124950 time=289.5s + ttt_chunk [951/1893] bpb=1.125346 time=292.6s + ttt_chunk [961/1893] bpb=1.125423 time=295.9s + ttt_chunk [971/1893] bpb=1.126285 time=299.1s + ttt_chunk [981/1893] bpb=1.126344 time=302.4s + ttt_chunk [991/1893] bpb=1.126347 time=305.6s + ttt_chunk [1001/1893] bpb=1.126319 time=308.8s + ttt_chunk [1011/1893] bpb=1.126085 time=312.1s + ttt_chunk [1021/1893] bpb=1.126433 time=315.3s + ttt_chunk [1031/1893] bpb=1.126880 time=318.6s + ttt_chunk [1041/1893] bpb=1.126535 time=322.0s + ttt_chunk [1051/1893] bpb=1.126292 time=325.3s + ttt_chunk [1061/1893] bpb=1.126333 time=328.5s + ttt_chunk [1071/1893] bpb=1.126919 time=331.7s + ttt_chunk [1081/1893] bpb=1.127188 time=335.0s + ttt_chunk [1091/1893] bpb=1.127911 time=338.3s + ttt_chunk [1101/1893] bpb=1.127938 time=341.5s + ttt_chunk [1111/1893] bpb=1.127795 time=344.7s + ttt_chunk [1121/1893] bpb=1.127599 time=347.9s + ttt_chunk [1131/1893] bpb=1.127464 time=351.0s + ttt_chunk [1141/1893] bpb=1.127178 time=354.1s + ttt_chunk [1151/1893] bpb=1.127179 time=357.3s + ttt_chunk [1161/1893] bpb=1.126796 time=360.5s + ttt_chunk [1171/1893] bpb=1.127118 time=363.7s + ttt_chunk [1181/1893] bpb=1.126359 time=366.8s + ttt_chunk [1191/1893] bpb=1.126218 time=370.1s + ttt_chunk [1201/1893] bpb=1.126638 time=373.4s + ttt_chunk [1211/1893] bpb=1.126155 time=376.4s + ttt_chunk [1221/1893] bpb=1.125858 time=379.6s + ttt_chunk [1231/1893] bpb=1.125576 time=382.9s + ttt_chunk [1241/1893] bpb=1.125226 time=386.1s + ttt_chunk [1251/1893] bpb=1.124636 time=389.0s + ttt_chunk [1261/1893] bpb=1.124618 time=392.2s + ttt_chunk [1271/1893] bpb=1.124207 time=395.3s + ttt_chunk [1281/1893] bpb=1.124012 time=398.7s + ttt_chunk [1291/1893] bpb=1.123777 time=401.8s + ttt_chunk [1301/1893] bpb=1.123178 time=405.1s + ttt_chunk [1311/1893] bpb=1.122789 time=408.2s + ttt_chunk [1321/1893] bpb=1.122455 time=411.4s + ttt_chunk [1331/1893] bpb=1.122409 time=414.6s + ttt_chunk [1341/1893] bpb=1.122288 time=417.8s + ttt_chunk [1351/1893] bpb=1.122218 time=421.0s + ttt_chunk [1361/1893] bpb=1.122262 time=424.2s + ttt_chunk [1371/1893] bpb=1.122128 time=427.4s + ttt_chunk [1381/1893] bpb=1.122118 time=430.6s + ttt_chunk [1391/1893] bpb=1.121701 time=433.7s + ttt_chunk [1401/1893] bpb=1.121664 time=436.8s + ttt_chunk [1411/1893] bpb=1.121757 time=440.1s + ttt_chunk [1421/1893] bpb=1.121997 time=443.2s + ttt_chunk [1431/1893] bpb=1.121684 time=446.4s + ttt_chunk [1441/1893] bpb=1.122192 time=449.7s + ttt_chunk [1451/1893] bpb=1.122526 time=452.9s + ttt_chunk [1461/1893] bpb=1.122057 time=456.0s + ttt_chunk [1471/1893] bpb=1.123083 time=459.3s + ttt_chunk [1481/1893] bpb=1.122619 time=462.6s + ttt_chunk [1491/1893] bpb=1.122427 time=465.8s + ttt_chunk [1501/1893] bpb=1.122311 time=468.9s + ttt_chunk [1511/1893] bpb=1.122336 time=472.1s + ttt_chunk [1521/1893] bpb=1.122348 time=475.3s + ttt_chunk [1531/1893] bpb=1.121827 time=478.6s + ttt_chunk [1541/1893] bpb=1.121680 time=481.7s + ttt_chunk [1551/1893] bpb=1.121985 time=484.9s + ttt_chunk [1561/1893] bpb=1.121976 time=488.1s + ttt_chunk [1571/1893] bpb=1.121806 time=491.3s + ttt_chunk [1581/1893] bpb=1.121901 time=494.5s + ttt_chunk [1591/1893] bpb=1.121738 time=497.7s + ttt_chunk [1601/1893] bpb=1.121901 time=501.0s + ttt_chunk [1611/1893] bpb=1.121830 time=504.3s + ttt_chunk [1621/1893] bpb=1.121423 time=507.3s + ttt_chunk [1631/1893] bpb=1.121723 time=510.5s + ttt_chunk [1641/1893] bpb=1.121736 time=513.7s + ttt_chunk [1651/1893] bpb=1.121680 time=516.9s + ttt_chunk [1661/1893] bpb=1.121554 time=520.1s + ttt_chunk [1671/1893] bpb=1.122027 time=523.3s + ttt_chunk [1681/1893] bpb=1.122173 time=526.5s + ttt_chunk [1691/1893] bpb=1.122001 time=529.7s + ttt_chunk [1701/1893] bpb=1.122150 time=532.9s + ttt_chunk [1711/1893] bpb=1.122156 time=536.1s + ttt_chunk [1721/1893] bpb=1.122155 time=539.2s + ttt_chunk [1731/1893] bpb=1.122032 time=542.4s + ttt_chunk [1741/1893] bpb=1.121832 time=545.7s + ttt_chunk [1751/1893] bpb=1.121659 time=548.9s + ttt_chunk [1761/1893] bpb=1.121807 time=552.0s + ttt_chunk [1771/1893] bpb=1.121701 time=555.2s + ttt_chunk [1781/1893] bpb=1.121713 time=558.4s + ttt_chunk [1791/1893] bpb=1.121316 time=561.5s + ttt_chunk [1801/1893] bpb=1.121178 time=564.7s + ttt_chunk [1811/1893] bpb=1.121070 time=567.9s + ttt_chunk [1821/1893] bpb=1.121122 time=571.1s + ttt_chunk [1831/1893] bpb=1.120511 time=574.4s + ttt_chunk [1841/1893] bpb=1.120442 time=577.5s + ttt_chunk [1851/1893] bpb=1.120230 time=580.6s + ttt_chunk [1861/1893] bpb=1.119865 time=583.7s + ttt_chunk [1871/1893] bpb=1.119864 time=587.0s + ttt_chunk [1881/1893] bpb=1.119402 time=590.0s + ttt_chunk [1891/1893] bpb=1.119160 time=593.2s + ttt_chunk [1893/1893] bpb=1.119204 time=593.6s +ttt_sliding:done val_loss=1.886513 val_bpb=1.117302 elapsed=593.6s +legal_ttt val_loss:1.8865 val_bpb:1.1173 eval_time:594055ms +legal_ttt_exact val_loss:1.88651332 val_bpb:1.11730240 diff --git a/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed2025.log b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed2025.log new file mode 100644 index 000000000..22e8478f1 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed2025.log @@ -0,0 +1,275 @@ +W0328 17:48:43.737000 3360929 torch/distributed/run.py:803] +W0328 17:48:43.737000 3360929 torch/distributed/run.py:803] ***************************************** +W0328 17:48:43.737000 3360929 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0328 17:48:43.737000 3360929 torch/distributed/run.py:803] ***************************************** +logs/b68abfb9-3ea0-4bd0-9c46-c54ec3f12f9e.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:26993756 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:2025 +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:6.9277 val_bpb:4.1030 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9281 train_time:131ms step_avg:130.59ms +step:2/20000 train_loss:8.6097 train_time:159ms step_avg:79.40ms +step:3/20000 train_loss:7.6978 train_time:240ms step_avg:79.86ms +step:4/20000 train_loss:7.2941 train_time:322ms step_avg:80.41ms +step:5/20000 train_loss:7.1366 train_time:404ms step_avg:80.77ms +step:6/20000 train_loss:7.0599 train_time:486ms step_avg:80.97ms +step:7/20000 train_loss:6.9509 train_time:567ms step_avg:81.02ms +step:8/20000 train_loss:6.9313 train_time:649ms step_avg:81.07ms +step:9/20000 train_loss:6.6158 train_time:730ms step_avg:81.17ms +step:10/20000 train_loss:6.2143 train_time:813ms step_avg:81.31ms +step:500/20000 train_loss:2.3996 train_time:41867ms step_avg:83.73ms +step:1000/20000 train_loss:2.2648 train_time:83940ms step_avg:83.94ms +step:1500/20000 train_loss:2.2109 train_time:126045ms step_avg:84.03ms +step:2000/20000 train_loss:2.0574 train_time:168289ms step_avg:84.14ms +step:2500/20000 train_loss:2.1617 train_time:210549ms step_avg:84.22ms +step:3000/20000 train_loss:2.1502 train_time:252824ms step_avg:84.27ms +step:3500/20000 train_loss:2.1686 train_time:295031ms step_avg:84.29ms +step:4000/20000 train_loss:1.9655 train_time:337271ms step_avg:84.32ms +step:4000/20000 val_loss:2.0567 val_bpb:1.2181 train_time:337322ms step_avg:84.33ms +step:4500/20000 train_loss:2.1183 train_time:379619ms step_avg:84.36ms +step:5000/20000 train_loss:2.0960 train_time:421970ms step_avg:84.39ms +step:5500/20000 train_loss:2.0135 train_time:464254ms step_avg:84.41ms +step:6000/20000 train_loss:1.9354 train_time:506500ms step_avg:84.42ms +swa:start step:6450 +step:6500/20000 train_loss:2.0778 train_time:548917ms step_avg:84.45ms +late_qat:enabled step:6578 scale:0.1498 +step:7000/20000 train_loss:1.7862 train_time:592009ms step_avg:84.57ms +step:7093/20000 val_loss:1.9219 val_bpb:1.1383 train_time:600052ms step_avg:84.60ms +stopping_early: wallclock_cap train_time:600052ms step:7093/20000 +peak memory allocated: 21472 MiB reserved: 22004 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9202 val_bpb:1.1372 eval_time:2015ms +Serialized model: 106158518 bytes +Code size: 93017 bytes +Serialized model int6+lzma: 15848608 bytes +Total submission size int6+lzma: 15941625 bytes +final_int6_roundtrip val_loss:1.9341 val_bpb:1.1455 eval_time:3928ms +final_int6_roundtrip_exact val_loss:1.93408738 val_bpb:1.14547549 +final_int6_sliding_window val_loss:1.8945 val_bpb:1.1220 stride:64 eval_time:76068ms +final_int6_sliding_window_exact val_loss:1.89450684 val_bpb:1.12203663 +final_int8_zlib_roundtrip_exact val_loss:1.89450684 val_bpb:1.12203663 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=2 +ttt_sliding:params unfrozen=26989644 frozen=4112 + ttt_chunk [1/1893] bpb=1.159020 time=0.6s + ttt_chunk [11/1893] bpb=1.146969 time=3.4s + ttt_chunk [21/1893] bpb=1.131451 time=6.3s + ttt_chunk [31/1893] bpb=1.129174 time=9.6s + ttt_chunk [41/1893] bpb=1.115529 time=12.8s + ttt_chunk [51/1893] bpb=1.109809 time=15.8s + ttt_chunk [61/1893] bpb=1.116315 time=19.0s + ttt_chunk [71/1893] bpb=1.114715 time=22.2s + ttt_chunk [81/1893] bpb=1.113503 time=25.2s + ttt_chunk [91/1893] bpb=1.114492 time=28.4s + ttt_chunk [101/1893] bpb=1.118159 time=31.6s + ttt_chunk [111/1893] bpb=1.120714 time=34.7s + ttt_chunk [121/1893] bpb=1.113941 time=37.7s + ttt_chunk [131/1893] bpb=1.114232 time=40.9s + ttt_chunk [141/1893] bpb=1.119929 time=44.1s + ttt_chunk [151/1893] bpb=1.121732 time=47.3s + ttt_chunk [161/1893] bpb=1.121016 time=50.6s + ttt_chunk [171/1893] bpb=1.125521 time=53.9s + ttt_chunk [181/1893] bpb=1.127866 time=57.1s + ttt_chunk [191/1893] bpb=1.135137 time=60.3s + ttt_chunk [201/1893] bpb=1.134009 time=63.5s + ttt_chunk [211/1893] bpb=1.131706 time=66.7s + ttt_chunk [221/1893] bpb=1.133223 time=69.8s + ttt_chunk [231/1893] bpb=1.131894 time=73.0s + ttt_chunk [241/1893] bpb=1.132245 time=76.2s + ttt_chunk [251/1893] bpb=1.131804 time=79.4s + ttt_chunk [261/1893] bpb=1.128973 time=82.5s + ttt_chunk [271/1893] bpb=1.127789 time=85.5s + ttt_chunk [281/1893] bpb=1.129251 time=88.7s + ttt_chunk [291/1893] bpb=1.131072 time=91.9s + ttt_chunk [301/1893] bpb=1.131843 time=95.2s + ttt_chunk [311/1893] bpb=1.133852 time=98.3s + ttt_chunk [321/1893] bpb=1.135816 time=101.6s + ttt_chunk [331/1893] bpb=1.135699 time=104.7s + ttt_chunk [341/1893] bpb=1.134687 time=108.0s + ttt_chunk [351/1893] bpb=1.136994 time=111.2s + ttt_chunk [361/1893] bpb=1.137119 time=114.2s + ttt_chunk [371/1893] bpb=1.136371 time=117.5s + ttt_chunk [381/1893] bpb=1.136624 time=120.7s + ttt_chunk [391/1893] bpb=1.136528 time=124.0s + ttt_chunk [401/1893] bpb=1.134474 time=127.2s + ttt_chunk [411/1893] bpb=1.133295 time=130.4s + ttt_chunk [421/1893] bpb=1.132379 time=133.6s + ttt_chunk [431/1893] bpb=1.132276 time=137.0s + ttt_chunk [441/1893] bpb=1.132605 time=140.2s + ttt_chunk [451/1893] bpb=1.132948 time=143.4s + ttt_chunk [461/1893] bpb=1.131836 time=146.7s + ttt_chunk [471/1893] bpb=1.132486 time=150.0s + ttt_chunk [481/1893] bpb=1.132134 time=153.2s + ttt_chunk [491/1893] bpb=1.131034 time=156.4s + ttt_chunk [501/1893] bpb=1.130551 time=159.7s + ttt_chunk [511/1893] bpb=1.129900 time=162.9s + ttt_chunk [521/1893] bpb=1.127511 time=166.2s + ttt_chunk [531/1893] bpb=1.128701 time=169.6s + ttt_chunk [541/1893] bpb=1.129045 time=172.8s + ttt_chunk [551/1893] bpb=1.128027 time=176.0s + ttt_chunk [561/1893] bpb=1.128548 time=179.3s + ttt_chunk [571/1893] bpb=1.127492 time=182.5s + ttt_chunk [581/1893] bpb=1.126685 time=185.6s + ttt_chunk [591/1893] bpb=1.126051 time=188.7s + ttt_chunk [601/1893] bpb=1.126530 time=191.9s + ttt_chunk [611/1893] bpb=1.126435 time=195.1s + ttt_chunk [621/1893] bpb=1.126324 time=198.4s + ttt_chunk [631/1893] bpb=1.127054 time=201.5s + ttt_chunk [641/1893] bpb=1.126790 time=204.6s + ttt_chunk [651/1893] bpb=1.126891 time=207.8s + ttt_chunk [661/1893] bpb=1.126347 time=210.9s + ttt_chunk [671/1893] bpb=1.126721 time=214.1s + ttt_chunk [681/1893] bpb=1.127470 time=217.2s + ttt_chunk [691/1893] bpb=1.128447 time=220.5s + ttt_chunk [701/1893] bpb=1.127885 time=223.6s + ttt_chunk [711/1893] bpb=1.127848 time=226.8s + ttt_chunk [721/1893] bpb=1.127512 time=229.9s + ttt_chunk [731/1893] bpb=1.127558 time=233.2s + ttt_chunk [741/1893] bpb=1.127669 time=236.3s + ttt_chunk [751/1893] bpb=1.127496 time=239.5s + ttt_chunk [761/1893] bpb=1.127465 time=242.7s + ttt_chunk [771/1893] bpb=1.127142 time=246.0s + ttt_chunk [781/1893] bpb=1.127882 time=249.2s + ttt_chunk [791/1893] bpb=1.127446 time=252.4s + ttt_chunk [801/1893] bpb=1.127761 time=255.6s + ttt_chunk [811/1893] bpb=1.127502 time=258.9s + ttt_chunk [821/1893] bpb=1.127270 time=262.1s + ttt_chunk [831/1893] bpb=1.127081 time=265.2s + ttt_chunk [841/1893] bpb=1.126414 time=268.4s + ttt_chunk [851/1893] bpb=1.126138 time=271.6s + ttt_chunk [861/1893] bpb=1.125884 time=274.7s + ttt_chunk [871/1893] bpb=1.126137 time=278.0s + ttt_chunk [881/1893] bpb=1.126280 time=280.9s + ttt_chunk [891/1893] bpb=1.125839 time=283.9s + ttt_chunk [901/1893] bpb=1.125560 time=287.0s + ttt_chunk [911/1893] bpb=1.125664 time=290.1s + ttt_chunk [921/1893] bpb=1.126142 time=293.3s + ttt_chunk [931/1893] bpb=1.126114 time=296.6s + ttt_chunk [941/1893] bpb=1.125795 time=299.8s + ttt_chunk [951/1893] bpb=1.126164 time=303.0s + ttt_chunk [961/1893] bpb=1.126264 time=306.3s + ttt_chunk [971/1893] bpb=1.127132 time=309.5s + ttt_chunk [981/1893] bpb=1.127209 time=312.7s + ttt_chunk [991/1893] bpb=1.127233 time=316.0s + ttt_chunk [1001/1893] bpb=1.127211 time=319.3s + ttt_chunk [1011/1893] bpb=1.126966 time=322.6s + ttt_chunk [1021/1893] bpb=1.127299 time=325.8s + ttt_chunk [1031/1893] bpb=1.127744 time=329.1s + ttt_chunk [1041/1893] bpb=1.127424 time=332.4s + ttt_chunk [1051/1893] bpb=1.127163 time=335.7s + ttt_chunk [1061/1893] bpb=1.127206 time=338.9s + ttt_chunk [1071/1893] bpb=1.127819 time=342.0s + ttt_chunk [1081/1893] bpb=1.128083 time=345.4s + ttt_chunk [1091/1893] bpb=1.128799 time=348.6s + ttt_chunk [1101/1893] bpb=1.128800 time=351.9s + ttt_chunk [1111/1893] bpb=1.128661 time=355.0s + ttt_chunk [1121/1893] bpb=1.128439 time=358.2s + ttt_chunk [1131/1893] bpb=1.128307 time=361.3s + ttt_chunk [1141/1893] bpb=1.128024 time=364.4s + ttt_chunk [1151/1893] bpb=1.128028 time=367.6s + ttt_chunk [1161/1893] bpb=1.127625 time=370.9s + ttt_chunk [1171/1893] bpb=1.127947 time=374.2s + ttt_chunk [1181/1893] bpb=1.127180 time=377.4s + ttt_chunk [1191/1893] bpb=1.127024 time=380.6s + ttt_chunk [1201/1893] bpb=1.127421 time=383.9s + ttt_chunk [1211/1893] bpb=1.126947 time=386.9s + ttt_chunk [1221/1893] bpb=1.126660 time=390.2s + ttt_chunk [1231/1893] bpb=1.126382 time=393.4s + ttt_chunk [1241/1893] bpb=1.126058 time=396.6s + ttt_chunk [1251/1893] bpb=1.125474 time=399.5s + ttt_chunk [1261/1893] bpb=1.125455 time=402.6s + ttt_chunk [1271/1893] bpb=1.125062 time=405.8s + ttt_chunk [1281/1893] bpb=1.124869 time=409.1s + ttt_chunk [1291/1893] bpb=1.124622 time=412.2s + ttt_chunk [1301/1893] bpb=1.124045 time=415.4s + ttt_chunk [1311/1893] bpb=1.123639 time=418.5s + ttt_chunk [1321/1893] bpb=1.123315 time=421.7s + ttt_chunk [1331/1893] bpb=1.123251 time=424.9s + ttt_chunk [1341/1893] bpb=1.123128 time=428.1s + ttt_chunk [1351/1893] bpb=1.123064 time=431.2s + ttt_chunk [1361/1893] bpb=1.123116 time=434.5s + ttt_chunk [1371/1893] bpb=1.122965 time=437.7s + ttt_chunk [1381/1893] bpb=1.122964 time=440.8s + ttt_chunk [1391/1893] bpb=1.122553 time=443.9s + ttt_chunk [1401/1893] bpb=1.122520 time=447.1s + ttt_chunk [1411/1893] bpb=1.122637 time=450.3s + ttt_chunk [1421/1893] bpb=1.122891 time=453.4s + ttt_chunk [1431/1893] bpb=1.122593 time=456.6s + ttt_chunk [1441/1893] bpb=1.123087 time=459.9s + ttt_chunk [1451/1893] bpb=1.123418 time=463.1s + ttt_chunk [1461/1893] bpb=1.122975 time=466.2s + ttt_chunk [1471/1893] bpb=1.124000 time=469.5s + ttt_chunk [1481/1893] bpb=1.123532 time=472.7s + ttt_chunk [1491/1893] bpb=1.123352 time=475.9s + ttt_chunk [1501/1893] bpb=1.123251 time=479.0s + ttt_chunk [1511/1893] bpb=1.123253 time=482.2s + ttt_chunk [1521/1893] bpb=1.123277 time=485.4s + ttt_chunk [1531/1893] bpb=1.122763 time=488.5s + ttt_chunk [1541/1893] bpb=1.122599 time=491.7s + ttt_chunk [1551/1893] bpb=1.122899 time=494.9s + ttt_chunk [1561/1893] bpb=1.122888 time=498.2s + ttt_chunk [1571/1893] bpb=1.122728 time=501.3s + ttt_chunk [1581/1893] bpb=1.122833 time=504.5s + ttt_chunk [1591/1893] bpb=1.122675 time=507.7s + ttt_chunk [1601/1893] bpb=1.122834 time=510.9s + ttt_chunk [1611/1893] bpb=1.122771 time=514.1s + ttt_chunk [1621/1893] bpb=1.122375 time=517.1s + ttt_chunk [1631/1893] bpb=1.122674 time=520.3s + ttt_chunk [1641/1893] bpb=1.122680 time=523.5s + ttt_chunk [1651/1893] bpb=1.122633 time=526.7s + ttt_chunk [1661/1893] bpb=1.122527 time=529.8s + ttt_chunk [1671/1893] bpb=1.122994 time=533.0s + ttt_chunk [1681/1893] bpb=1.123129 time=536.2s + ttt_chunk [1691/1893] bpb=1.122962 time=539.4s + ttt_chunk [1701/1893] bpb=1.123125 time=542.5s + ttt_chunk [1711/1893] bpb=1.123126 time=545.7s + ttt_chunk [1721/1893] bpb=1.123135 time=548.7s + ttt_chunk [1731/1893] bpb=1.123011 time=551.9s + ttt_chunk [1741/1893] bpb=1.122801 time=555.1s + ttt_chunk [1751/1893] bpb=1.122628 time=558.3s + ttt_chunk [1761/1893] bpb=1.122763 time=561.5s + ttt_chunk [1771/1893] bpb=1.122668 time=564.6s + ttt_chunk [1781/1893] bpb=1.122702 time=567.8s + ttt_chunk [1791/1893] bpb=1.122299 time=570.9s + ttt_chunk [1801/1893] bpb=1.122178 time=574.0s + ttt_chunk [1811/1893] bpb=1.122077 time=577.2s + ttt_chunk [1821/1893] bpb=1.122142 time=580.4s + ttt_chunk [1831/1893] bpb=1.121533 time=583.6s + ttt_chunk [1841/1893] bpb=1.121476 time=586.7s + ttt_chunk [1851/1893] bpb=1.121258 time=589.8s + ttt_chunk [1861/1893] bpb=1.120892 time=593.0s + ttt_chunk [1871/1893] bpb=1.120864 time=596.2s + ttt_chunk [1881/1893] bpb=1.120413 time=599.2s + ttt_chunk [1891/1893] bpb=1.120188 time=602.3s + ttt_chunk [1893/1893] bpb=1.120233 time=602.8s +ttt_sliding:done val_loss=1.888212 val_bpb=1.118308 elapsed=602.8s +legal_ttt val_loss:1.8882 val_bpb:1.1183 eval_time:603230ms +legal_ttt_exact val_loss:1.88821177 val_bpb:1.11830833 diff --git a/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed42.log b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed42.log new file mode 100644 index 000000000..76915faa0 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_MuonTTT_EntropyAdaptive_1.1179/train_seed42.log @@ -0,0 +1,275 @@ +W0328 17:25:49.794000 3359713 torch/distributed/run.py:803] +W0328 17:25:49.794000 3359713 torch/distributed/run.py:803] ***************************************** +W0328 17:25:49.794000 3359713 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0328 17:25:49.794000 3359713 torch/distributed/run.py:803] ***************************************** +logs/21a7b63f-64c3-4dc4-9c4e-667fa039c19e.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:26993756 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +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:6.9297 val_bpb:4.1042 train_time:0ms step_avg:0.03ms +step:1/20000 train_loss:6.9319 train_time:157ms step_avg:157.29ms +step:2/20000 train_loss:8.6254 train_time:192ms step_avg:95.86ms +step:3/20000 train_loss:7.7123 train_time:272ms step_avg:90.56ms +step:4/20000 train_loss:7.2839 train_time:353ms step_avg:88.27ms +step:5/20000 train_loss:7.1730 train_time:434ms step_avg:86.82ms +step:6/20000 train_loss:7.0089 train_time:515ms step_avg:85.87ms +step:7/20000 train_loss:6.9175 train_time:596ms step_avg:85.19ms +step:8/20000 train_loss:6.8685 train_time:678ms step_avg:84.75ms +step:9/20000 train_loss:6.5568 train_time:760ms step_avg:84.48ms +step:10/20000 train_loss:6.2118 train_time:842ms step_avg:84.17ms +step:500/20000 train_loss:2.3958 train_time:41845ms step_avg:83.69ms +step:1000/20000 train_loss:2.2631 train_time:83895ms step_avg:83.89ms +step:1500/20000 train_loss:2.2109 train_time:126002ms step_avg:84.00ms +step:2000/20000 train_loss:2.0535 train_time:168153ms step_avg:84.08ms +step:2500/20000 train_loss:2.1619 train_time:210322ms step_avg:84.13ms +step:3000/20000 train_loss:2.1512 train_time:252545ms step_avg:84.18ms +step:3500/20000 train_loss:2.1691 train_time:294807ms step_avg:84.23ms +step:4000/20000 train_loss:1.9593 train_time:337059ms step_avg:84.26ms +step:4000/20000 val_loss:2.0550 val_bpb:1.2171 train_time:337117ms step_avg:84.28ms +step:4500/20000 train_loss:2.1151 train_time:379304ms step_avg:84.29ms +step:5000/20000 train_loss:2.0983 train_time:421513ms step_avg:84.30ms +step:5500/20000 train_loss:2.0105 train_time:463742ms step_avg:84.32ms +step:6000/20000 train_loss:1.9353 train_time:505954ms step_avg:84.33ms +swa:start step:6450 +step:6500/20000 train_loss:2.0784 train_time:548304ms step_avg:84.35ms +late_qat:enabled step:6586 scale:0.1498 +step:7000/20000 train_loss:1.7866 train_time:591384ms step_avg:84.48ms +step:7101/20000 val_loss:1.9214 val_bpb:1.1380 train_time:600151ms step_avg:84.52ms +stopping_early: wallclock_cap train_time:600151ms step:7101/20000 +peak memory allocated: 21472 MiB reserved: 22004 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9196 val_bpb:1.1369 eval_time:2018ms +Serialized model: 106158518 bytes +Code size: 93017 bytes +Serialized model int6+lzma: 15968952 bytes +Total submission size int6+lzma: 16061969 bytes +final_int6_roundtrip val_loss:1.9338 val_bpb:1.1453 eval_time:3977ms +final_int6_roundtrip_exact val_loss:1.93382619 val_bpb:1.14532079 +final_int6_sliding_window val_loss:1.8941 val_bpb:1.1218 stride:64 eval_time:76671ms +final_int6_sliding_window_exact val_loss:1.89407841 val_bpb:1.12178289 +final_int8_zlib_roundtrip_exact val_loss:1.89407841 val_bpb:1.12178289 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=2 +ttt_sliding:params unfrozen=26989644 frozen=4112 + ttt_chunk [1/1893] bpb=1.161872 time=0.6s + ttt_chunk [11/1893] bpb=1.147042 time=3.3s + ttt_chunk [21/1893] bpb=1.132523 time=6.0s + ttt_chunk [31/1893] bpb=1.129538 time=8.9s + ttt_chunk [41/1893] bpb=1.116299 time=11.7s + ttt_chunk [51/1893] bpb=1.110537 time=14.3s + ttt_chunk [61/1893] bpb=1.116769 time=17.2s + ttt_chunk [71/1893] bpb=1.115344 time=20.0s + ttt_chunk [81/1893] bpb=1.114533 time=22.7s + ttt_chunk [91/1893] bpb=1.115600 time=25.4s + ttt_chunk [101/1893] bpb=1.119092 time=28.2s + ttt_chunk [111/1893] bpb=1.121485 time=31.0s + ttt_chunk [121/1893] bpb=1.114764 time=33.6s + ttt_chunk [131/1893] bpb=1.114914 time=36.7s + ttt_chunk [141/1893] bpb=1.120504 time=39.9s + ttt_chunk [151/1893] bpb=1.122167 time=43.0s + ttt_chunk [161/1893] bpb=1.121553 time=46.2s + ttt_chunk [171/1893] bpb=1.125924 time=49.5s + ttt_chunk [181/1893] bpb=1.128065 time=52.6s + ttt_chunk [191/1893] bpb=1.135321 time=55.8s + ttt_chunk [201/1893] bpb=1.134126 time=59.0s + ttt_chunk [211/1893] bpb=1.131959 time=62.1s + ttt_chunk [221/1893] bpb=1.133563 time=65.2s + ttt_chunk [231/1893] bpb=1.132277 time=68.3s + ttt_chunk [241/1893] bpb=1.132584 time=71.5s + ttt_chunk [251/1893] bpb=1.132111 time=74.6s + ttt_chunk [261/1893] bpb=1.129187 time=77.7s + ttt_chunk [271/1893] bpb=1.127955 time=80.8s + ttt_chunk [281/1893] bpb=1.129401 time=84.0s + ttt_chunk [291/1893] bpb=1.131279 time=87.2s + ttt_chunk [301/1893] bpb=1.131907 time=90.4s + ttt_chunk [311/1893] bpb=1.133895 time=93.5s + ttt_chunk [321/1893] bpb=1.135819 time=96.7s + ttt_chunk [331/1893] bpb=1.135615 time=99.9s + ttt_chunk [341/1893] bpb=1.134611 time=103.1s + ttt_chunk [351/1893] bpb=1.136911 time=106.3s + ttt_chunk [361/1893] bpb=1.137029 time=109.3s + ttt_chunk [371/1893] bpb=1.136287 time=112.5s + ttt_chunk [381/1893] bpb=1.136510 time=115.8s + ttt_chunk [391/1893] bpb=1.136389 time=119.0s + ttt_chunk [401/1893] bpb=1.134240 time=122.2s + ttt_chunk [411/1893] bpb=1.133065 time=125.4s + ttt_chunk [421/1893] bpb=1.132133 time=128.5s + ttt_chunk [431/1893] bpb=1.131952 time=131.8s + ttt_chunk [441/1893] bpb=1.132325 time=135.0s + ttt_chunk [451/1893] bpb=1.132671 time=138.3s + ttt_chunk [461/1893] bpb=1.131574 time=141.5s + ttt_chunk [471/1893] bpb=1.132207 time=144.8s + ttt_chunk [481/1893] bpb=1.131884 time=148.0s + ttt_chunk [491/1893] bpb=1.130797 time=151.1s + ttt_chunk [501/1893] bpb=1.130308 time=154.4s + ttt_chunk [511/1893] bpb=1.129614 time=157.7s + ttt_chunk [521/1893] bpb=1.127167 time=160.9s + ttt_chunk [531/1893] bpb=1.128359 time=164.3s + ttt_chunk [541/1893] bpb=1.128639 time=167.5s + ttt_chunk [551/1893] bpb=1.127623 time=170.7s + ttt_chunk [561/1893] bpb=1.128140 time=174.1s + ttt_chunk [571/1893] bpb=1.127094 time=177.2s + ttt_chunk [581/1893] bpb=1.126280 time=180.3s + ttt_chunk [591/1893] bpb=1.125648 time=183.4s + ttt_chunk [601/1893] bpb=1.126146 time=186.5s + ttt_chunk [611/1893] bpb=1.126086 time=189.7s + ttt_chunk [621/1893] bpb=1.125948 time=193.0s + ttt_chunk [631/1893] bpb=1.126679 time=196.1s + ttt_chunk [641/1893] bpb=1.126471 time=199.2s + ttt_chunk [651/1893] bpb=1.126615 time=202.3s + ttt_chunk [661/1893] bpb=1.126089 time=205.4s + ttt_chunk [671/1893] bpb=1.126466 time=208.6s + ttt_chunk [681/1893] bpb=1.127154 time=211.7s + ttt_chunk [691/1893] bpb=1.128113 time=215.0s + ttt_chunk [701/1893] bpb=1.127555 time=218.1s + ttt_chunk [711/1893] bpb=1.127585 time=221.2s + ttt_chunk [721/1893] bpb=1.127231 time=224.3s + ttt_chunk [731/1893] bpb=1.127280 time=227.5s + ttt_chunk [741/1893] bpb=1.127395 time=230.6s + ttt_chunk [751/1893] bpb=1.127233 time=233.8s + ttt_chunk [761/1893] bpb=1.127141 time=237.0s + ttt_chunk [771/1893] bpb=1.126849 time=240.2s + ttt_chunk [781/1893] bpb=1.127587 time=243.4s + ttt_chunk [791/1893] bpb=1.127149 time=246.5s + ttt_chunk [801/1893] bpb=1.127458 time=249.7s + ttt_chunk [811/1893] bpb=1.127198 time=252.9s + ttt_chunk [821/1893] bpb=1.126962 time=256.1s + ttt_chunk [831/1893] bpb=1.126782 time=259.3s + ttt_chunk [841/1893] bpb=1.126128 time=262.4s + ttt_chunk [851/1893] bpb=1.125878 time=265.6s + ttt_chunk [861/1893] bpb=1.125630 time=268.7s + ttt_chunk [871/1893] bpb=1.125896 time=272.0s + ttt_chunk [881/1893] bpb=1.126070 time=274.9s + ttt_chunk [891/1893] bpb=1.125635 time=277.9s + ttt_chunk [901/1893] bpb=1.125364 time=280.9s + ttt_chunk [911/1893] bpb=1.125484 time=284.1s + ttt_chunk [921/1893] bpb=1.125967 time=287.3s + ttt_chunk [931/1893] bpb=1.125931 time=290.5s + ttt_chunk [941/1893] bpb=1.125624 time=293.7s + ttt_chunk [951/1893] bpb=1.125995 time=296.9s + ttt_chunk [961/1893] bpb=1.126078 time=300.1s + ttt_chunk [971/1893] bpb=1.126929 time=303.3s + ttt_chunk [981/1893] bpb=1.127021 time=306.5s + ttt_chunk [991/1893] bpb=1.127049 time=309.8s + ttt_chunk [1001/1893] bpb=1.127009 time=313.0s + ttt_chunk [1011/1893] bpb=1.126798 time=316.3s + ttt_chunk [1021/1893] bpb=1.127136 time=319.5s + ttt_chunk [1031/1893] bpb=1.127594 time=322.8s + ttt_chunk [1041/1893] bpb=1.127258 time=326.0s + ttt_chunk [1051/1893] bpb=1.127011 time=329.3s + ttt_chunk [1061/1893] bpb=1.127045 time=332.5s + ttt_chunk [1071/1893] bpb=1.127648 time=335.7s + ttt_chunk [1081/1893] bpb=1.127926 time=339.0s + ttt_chunk [1091/1893] bpb=1.128648 time=342.3s + ttt_chunk [1101/1893] bpb=1.128661 time=345.5s + ttt_chunk [1111/1893] bpb=1.128524 time=348.7s + ttt_chunk [1121/1893] bpb=1.128316 time=351.9s + ttt_chunk [1131/1893] bpb=1.128203 time=355.0s + ttt_chunk [1141/1893] bpb=1.127907 time=358.1s + ttt_chunk [1151/1893] bpb=1.127938 time=361.4s + ttt_chunk [1161/1893] bpb=1.127557 time=364.6s + ttt_chunk [1171/1893] bpb=1.127846 time=367.8s + ttt_chunk [1181/1893] bpb=1.127078 time=370.9s + ttt_chunk [1191/1893] bpb=1.126955 time=374.1s + ttt_chunk [1201/1893] bpb=1.127377 time=377.4s + ttt_chunk [1211/1893] bpb=1.126901 time=380.5s + ttt_chunk [1221/1893] bpb=1.126607 time=383.7s + ttt_chunk [1231/1893] bpb=1.126308 time=387.0s + ttt_chunk [1241/1893] bpb=1.125951 time=390.1s + ttt_chunk [1251/1893] bpb=1.125361 time=393.1s + ttt_chunk [1261/1893] bpb=1.125339 time=396.3s + ttt_chunk [1271/1893] bpb=1.124967 time=399.4s + ttt_chunk [1281/1893] bpb=1.124774 time=402.8s + ttt_chunk [1291/1893] bpb=1.124551 time=405.9s + ttt_chunk [1301/1893] bpb=1.123948 time=409.1s + ttt_chunk [1311/1893] bpb=1.123567 time=412.3s + ttt_chunk [1321/1893] bpb=1.123257 time=415.5s + ttt_chunk [1331/1893] bpb=1.123186 time=418.8s + ttt_chunk [1341/1893] bpb=1.123052 time=422.0s + ttt_chunk [1351/1893] bpb=1.122984 time=425.3s + ttt_chunk [1361/1893] bpb=1.123010 time=428.5s + ttt_chunk [1371/1893] bpb=1.122880 time=431.7s + ttt_chunk [1381/1893] bpb=1.122861 time=434.8s + ttt_chunk [1391/1893] bpb=1.122464 time=437.9s + ttt_chunk [1401/1893] bpb=1.122409 time=441.1s + ttt_chunk [1411/1893] bpb=1.122520 time=444.3s + ttt_chunk [1421/1893] bpb=1.122762 time=447.5s + ttt_chunk [1431/1893] bpb=1.122451 time=450.7s + ttt_chunk [1441/1893] bpb=1.122928 time=454.0s + ttt_chunk [1451/1893] bpb=1.123246 time=457.2s + ttt_chunk [1461/1893] bpb=1.122773 time=460.4s + ttt_chunk [1471/1893] bpb=1.123806 time=463.7s + ttt_chunk [1481/1893] bpb=1.123364 time=466.9s + ttt_chunk [1491/1893] bpb=1.123159 time=470.1s + ttt_chunk [1501/1893] bpb=1.123059 time=473.3s + ttt_chunk [1511/1893] bpb=1.123052 time=476.5s + ttt_chunk [1521/1893] bpb=1.123064 time=479.8s + ttt_chunk [1531/1893] bpb=1.122530 time=483.0s + ttt_chunk [1541/1893] bpb=1.122353 time=486.2s + ttt_chunk [1551/1893] bpb=1.122678 time=489.4s + ttt_chunk [1561/1893] bpb=1.122649 time=492.6s + ttt_chunk [1571/1893] bpb=1.122486 time=495.8s + ttt_chunk [1581/1893] bpb=1.122591 time=499.0s + ttt_chunk [1591/1893] bpb=1.122444 time=502.2s + ttt_chunk [1601/1893] bpb=1.122611 time=505.4s + ttt_chunk [1611/1893] bpb=1.122542 time=508.7s + ttt_chunk [1621/1893] bpb=1.122134 time=511.7s + ttt_chunk [1631/1893] bpb=1.122436 time=514.9s + ttt_chunk [1641/1893] bpb=1.122450 time=518.0s + ttt_chunk [1651/1893] bpb=1.122395 time=521.3s + ttt_chunk [1661/1893] bpb=1.122276 time=524.5s + ttt_chunk [1671/1893] bpb=1.122734 time=527.7s + ttt_chunk [1681/1893] bpb=1.122882 time=530.9s + ttt_chunk [1691/1893] bpb=1.122714 time=534.2s + ttt_chunk [1701/1893] bpb=1.122862 time=537.4s + ttt_chunk [1711/1893] bpb=1.122862 time=540.6s + ttt_chunk [1721/1893] bpb=1.122867 time=543.7s + ttt_chunk [1731/1893] bpb=1.122731 time=546.9s + ttt_chunk [1741/1893] bpb=1.122511 time=550.2s + ttt_chunk [1751/1893] bpb=1.122339 time=553.4s + ttt_chunk [1761/1893] bpb=1.122485 time=556.6s + ttt_chunk [1771/1893] bpb=1.122385 time=559.8s + ttt_chunk [1781/1893] bpb=1.122423 time=563.1s + ttt_chunk [1791/1893] bpb=1.122011 time=566.1s + ttt_chunk [1801/1893] bpb=1.121883 time=569.4s + ttt_chunk [1811/1893] bpb=1.121777 time=572.6s + ttt_chunk [1821/1893] bpb=1.121829 time=575.8s + ttt_chunk [1831/1893] bpb=1.121230 time=579.0s + ttt_chunk [1841/1893] bpb=1.121149 time=582.1s + ttt_chunk [1851/1893] bpb=1.120941 time=585.2s + ttt_chunk [1861/1893] bpb=1.120582 time=588.4s + ttt_chunk [1871/1893] bpb=1.120572 time=591.6s + ttt_chunk [1881/1893] bpb=1.120119 time=594.6s + ttt_chunk [1891/1893] bpb=1.119890 time=597.9s + ttt_chunk [1893/1893] bpb=1.119935 time=598.3s +ttt_sliding:done val_loss=1.887786 val_bpb=1.118056 elapsed=598.3s +legal_ttt val_loss:1.8878 val_bpb:1.1181 eval_time:598743ms +legal_ttt_exact val_loss:1.88778631 val_bpb:1.11805635