From 26045f3ac998bf848a4dfaa7eb00848bda1a75b8 Mon Sep 17 00:00:00 2001 From: Aryan Bhosale <36108149+aryanbhosale@users.noreply.github.com> Date: Thu, 26 Mar 2026 16:18:48 +0530 Subject: [PATCH] =?UTF-8?q?Non-Record:=2011L=20Parallel=20Muon=20+=20LN=20?= =?UTF-8?q?Scale=20+=20LeakyReLU=C2=B2=20MLP3x=20+=20Legal=20TTT=20(val=5F?= =?UTF-8?q?bpb=201.1215,=203-seed=208xH100)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../README.md | 66 + .../submission.json | 16 + .../train_gpt.py | 1885 +++++++++++++++++ .../train_seed1337.log | 277 +++ .../train_seed2024.log | 277 +++ .../train_seed42.log | 277 +++ 6 files changed, 2798 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/README.md create mode 100644 records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/submission.json create mode 100644 records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_gpt.py create mode 100644 records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed2024.log create mode 100644 records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed42.log diff --git a/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/README.md b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/README.md new file mode 100644 index 000000000..e06917221 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/README.md @@ -0,0 +1,66 @@ +# Non-Record: 11L Parallel Muon + LN Scale + LeakyReLU² MLP3x + Legal TTT + +**3-seed mean val_bpb: 1.1215** (std=0.0002) | **~15.85 MB** | 8xH100 SXM + +## 3-Seed Results (8xH100 80GB SXM, PyTorch 2.9.1+cu128) + +| Seed | step_avg | steps | EMA bpb | Quantized bpb | **TTT bpb** | +|------|----------|-------|---------|---------------|-------------| +| 1337 | 88.8ms | 6,759 | 1.1161 | 1.1238 | **1.1217** | +| 42 | 88.8ms | 6,757 | 1.1158 | 1.1234 | **1.1213** | +| 2024 | 88.9ms | 6,752 | 1.1160 | 1.1234 | **1.1215** | +| **Mean** | **88.8ms** | **6,756** | **1.1160** | **1.1235** | **1.1215** | + +## Architecture (26.8M parameters) + +- 11 transformer layers, dim=512, 8 heads / 4 KV heads (GQA) +- **Parallel Muon** with parameter banking (4 contiguous 3D banks) + batched Newton-Schulz +- MLP 3x expansion (hidden=1536) with **LeakyReLU(0.5)²** activation +- **LN Scale** — depth-dependent normalization: 1/sqrt(layer_idx+1) +- **SmearGate** + **BigramHash(1536, dim=128)** +- **Value Residual (ResFormer)** — cache V from layer 0, blend via learned lambda +- **Gated Attention** — per-head sigmoid gate (nn.Linear, bias init 4.0) +- **XSA on last 4 layers** — exclusive self-attention +- **Partial RoPE** — 16/64 head dimensions +- Tied FP16 embeddings, U-Net skip connections, orthogonal initialization +- Flash Attention 3 for causal attention + +## Training + +- **Parallel Muon optimizer**: 3-phase async reduce-scatter -> Adam -> NS5+all-gather + - lr=0.025, momentum 0.92->0.99/1500 steps, WD=0.04 + - No DDP -- manual gradient sync for non-bank params +- Adam for embeddings (lr=0.035) and scalars (lr=0.025) +- Batch 786,432 tokens, seq_len 2048 +- EMA (decay=0.997) + SWA (every 50 steps when scale < 0.2) +- Warmdown 3500 iterations (wallclock-based) +- Late QAT via STE (final 15% of wallclock) +- Gradient clipping 0.3 +- torch.compile(fullgraph=True) + +## Quantization + +- Int6 uniform per-row with GPTQ-lite (5-percentile clip search per row) +- FP16 passthrough for tied embeddings +- zstd-22 compression +- Unbank -> quantize -> rebank for compatibility with parameter banking + +## Legal Score-First TTT (PR #461 / #549 recipe) + +Every token scored BEFORE any weight update: + +``` +for each 32K-token chunk: + Phase 1 -- SCORE: sliding window eval (inference_mode, stride=64) + Phase 2 -- TRAIN: SGD(lr=0.002, momentum=0.9), 3 epochs, all blocks unfrozen, cosine LR +``` + +TTT improves quantized BPB by ~0.002 (1.1235 -> 1.1215). + +## Credits + +- Parallel Muon / Parameter Banking: PR #399 by @abaybektursun +- LeakyReLU²: PR #493 by @parinzee, PR #518 by @sofiabod +- LN Scale: PR #315/374 by @jfprincz +- TTT recipe: PR #461 by @Christopher-Lee-McClendon (adapted: freeze=0) +- Base model stack: PR #414 by @signalrush diff --git a/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/submission.json b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/submission.json new file mode 100644 index 000000000..a9dcc7f1c --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/submission.json @@ -0,0 +1,16 @@ +{ + "author": "Aryan Bhosale", + "github_id": "aryanbhosale", + "name": "11L Parallel Muon + LN Scale + LeakyReLU² MLP3x + Legal TTT (mean val_bpb=1.1215)", + "blurb": "11-layer 512d transformer with Parallel Muon (parameter banking + batched NS5), LN Scale, LeakyReLU(0.5)² MLP 3x, SmearGate, BigramHash(1536), Value Residual, Gated Attention, XSA4, Partial RoPE(16/64), EMA(0.997)+SWA, Late QAT, GPTQ-lite int6+zstd-22, legal score-first TTT. 3-seed mean 1.1215 BPB (std=0.0002) on 8xH100 SXM.", + "date": "2026-03-26T12:00:00Z", + "val_loss": 1.8937, + "val_bpb": 1.1215, + "bytes_total": 15850000, + "bytes_code": 80000, + "seeds": { + "1337": {"val_bpb": 1.1217, "val_loss": 1.8940, "steps": 6759, "step_avg_ms": 88.8}, + "42": {"val_bpb": 1.1213, "val_loss": 1.8933, "steps": 6757, "step_avg_ms": 88.8}, + "2024": {"val_bpb": 1.1215, "val_loss": 1.8937, "steps": 6752, "step_avg_ms": 88.9} + } +} diff --git a/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_gpt.py b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_gpt.py new file mode 100644 index 000000000..4fdba1c50 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_gpt.py @@ -0,0 +1,1885 @@ +"""SOTA config: Parallel Muon + parameter banks + MLP 3x + 11L XSA + LN Scale + GPTQ-lite int6 + zstd-22.""" +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +try: + import zstandard + def _compress(data: bytes) -> bytes: return zstandard.ZstdCompressor(level=22).compress(data) + def _decompress(data: bytes) -> bytes: return zstandard.ZstdDecompressor().decompress(data) +except ImportError: + def _compress(data: bytes) -> bytes: return zlib.compress(data, level=9) + def _decompress(data: bytes) -> bytes: return zlib.decompress(data) + +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 + +try: + from flash_attn_interface import flash_attn_func as _flash_attn_3_func + _HAS_FA3 = True +except ImportError: + _HAS_FA3 = False + +# --- HYPERPARAMETERS --- +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 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)) + 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)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 1536)) + 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"))) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "0"))) + ve_dim = int(os.environ.get("VE_DIM", 32)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + + use_smeargate = bool(int(os.environ.get("USE_SMEARGATE", "1"))) + use_bigramhash = bool(int(os.environ.get("USE_BIGRAMHASH", "1"))) + use_value_residual = bool(int(os.environ.get("USE_VALUE_RESIDUAL", "1"))) + use_gated_attention = bool(int(os.environ.get("USE_GATED_ATTENTION", "1"))) + + use_ema = bool(int(os.environ.get("USE_EMA", "1"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + use_late_qat = bool(int(os.environ.get("USE_LATE_QAT", "1"))) + qat_time_frac = float(os.environ.get("QAT_TIME_FRAC", 0.15)) + + # TTT (Test-Time Training) — legal score-first approach + use_ttt = bool(int(os.environ.get("USE_TTT", "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", 0)) + 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)) + + +# --- COMPRESSION / QUANTIZATION CONSTANTS --- +INT6_RANGE = 31 +QUANT_RANGE = INT6_RANGE +_FP16_PASSTHROUGH_NAMES = ("tok_emb.weight",) + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,mlp_scale,resid_mix,q_gain,skip_weight,skip_weights," + "smear,bigram_scale,vr_lambda,attn_gate,ve_layer_scales,ve_shared.scale", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = CONTROL_TENSOR_NAME_PATTERNS +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 + + +# --- 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] + + +# --- Sliding window eval (non-TTT) --- + +def eval_val( + args: Hyperparameters, + 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, +) -> tuple[float, float]: + """Sliding window eval with stride=eval_stride, batched for throughput.""" + seq_len = args.train_seq_len + stride = args.eval_stride + windows_per_batch = 32 + total_tokens = val_tokens.numel() - 1 + + starts = list(range(0, total_tokens - seq_len + 1, stride)) + my_starts = starts[rank::world_size] + + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_start in range(0, len(my_starts), windows_per_batch): + batch_starts = my_starts[batch_start : batch_start + windows_per_batch] + x_list = [] + y_list = [] + for s in batch_starts: + chunk = val_tokens[s : s + seq_len + 1].to(dtype=torch.int64) + x_list.append(chunk[:-1]) + y_list.append(chunk[1:]) + x = torch.stack(x_list).to(device=device, non_blocking=True) + y = torch.stack(y_list).to(device=device, non_blocking=True) + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = model(x) # target_ids=None -> returns logits + + for i, s in enumerate(batch_starts): + if s == 0: + score_start = 0 + score_len = min(seq_len, stride) + else: + score_start = seq_len - stride + score_len = stride + + window_logits = logits[i, score_start : score_start + score_len] + window_targets = y[i, score_start : score_start + score_len] + loss = F.cross_entropy(window_logits.float(), window_targets, reduction="sum") + val_loss_sum += loss.to(torch.float64) + val_token_count += score_len + + prev_ids = x[i, score_start : score_start + score_len] + tgt_ids = window_targets + 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) + + +# --- TTT evaluation (legal score-first sliding window) --- + +def eval_val_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, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT: 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 + stride = args.eval_stride + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + if log_fn is None: + log_fn = lambda msg: None + + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + 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) + + log_fn(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) + + 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) + + log_fn(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)}") + + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + batch_seqs = args.ttt_batch_seqs + 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) + + # --- 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.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_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) + 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))) + 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 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + 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) + optimizer.zero_grad(set_to_none=True) + 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) + optimizer.step() + + 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 + log_fn(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() + + log_fn(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 + + +# --- POST-TRAINING QUANTIZATION (GPTQ-lite int6 + zstd) --- + +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_simple(t: Tensor) -> tuple[Tensor, Tensor]: + """Simple int6 quantization for non-2D tensors (fallback).""" + qrange = QUANT_RANGE + t32 = t.float() + clip_q = 0.9999984 + clip_abs = float(torch.quantile(t32.abs().flatten(), clip_q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / float(qrange) if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -qrange, qrange).to(torch.int8).contiguous() + return q, scale + + +def _gptq_lite_2d(t: Tensor) -> tuple[Tensor, Tensor]: + """GPTQ-lite: per-row 5-percentile clip search for 2D weight tensors.""" + qrange = QUANT_RANGE + t32 = t.float() + _CLIP_QS = [0.9990, 0.9995, 0.9999, 0.99999, 1.0] + best_q = None + best_scale = None + best_mse = None + for cq in _CLIP_QS: + clip_abs = torch.quantile(t32.abs(), cq, dim=1) if t32.numel() else torch.empty((t32.shape[0],), dtype=torch.float32) + clipped = torch.clamp(t32, -clip_abs[:, None], clip_abs[:, None]) + s = (clip_abs / float(qrange)).clamp_min(1.0 / float(qrange)) + q = torch.clamp(torch.round(clipped / s[:, None]), -qrange, qrange) + recon = q * s[:, None] + mse = (t32 - recon).square().sum(dim=1) + if best_mse is None: + best_mse, best_q, best_scale = mse, q, s + else: + improved = mse < best_mse + if improved.any(): + best_mse = torch.where(improved, mse, best_mse) + best_q = torch.where(improved[:, None], q, best_q) + best_scale = torch.where(improved, s, best_scale) + return best_q.to(torch.int8).contiguous(), best_scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + +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 or any(p in name for p in _FP16_PASSTHROUGH_NAMES): + 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 + + if t.ndim == 2: + q, s = _gptq_lite_2d(t) + else: + q, s = quantize_float_tensor_simple(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 + + +# --- Bank <-> individual weight conversion for quantization --- + +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 + 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 + + +# --- 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 # CLASS-level flag + + 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() # STE + 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) + # Layout: [1, T, 1, D//2] for B,T,H,D attention format + 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 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.""" + def __init__(self, vocab_size: int, ve_dim: int, kv_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, kv_dim, bias=False) + 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.proj(self.embed(token_ids)) + return h * self.scale.to(dtype=h.dtype) + + +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. + 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) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + 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] + + # q/k/v are [B, T, H, D] + if _HAS_FA3: + y = _flash_attn_3_func(q, k, v, causal=True) # FA3: [B,T,H,D] in/out + else: + q_t, k_t, v_t = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2) + y = F.scaled_dot_product_attention( + q_t, k_t, v_t, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + + if self.use_xsa: + y = self._xsa_efficient(y, v) # v is [B, T, Hkv, D] + + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) # [B, T, H, 1] + y = y * gate + + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + 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: float, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: 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 + + 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) + 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: float, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + use_smeargate: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + 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.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) if use_smeargate else None + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # 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, + 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) + 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 + # Value Embeddings + 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.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + 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 + 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, 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_body(self, input_ids: Tensor) -> Tensor: + """Shared forward body: input_ids -> final hidden states.""" + 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),)) + if self.smear is not None: + 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) + return self.final_norm(x) + + def forward(self, input_ids: Tensor, target_ids: Tensor | None = None) -> Tensor: + x = self._forward_body(input_ids) + x_flat = x.reshape(-1, x.size(-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) + if target_ids is None: + # Eval mode: return logits [B, T, V] + return logits.reshape(input_ids.shape[0], input_ids.shape[1], -1) + else: + # Training mode: return loss + targets = target_ids.reshape(-1) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + x = self._forward_body(input_ids) + 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) + + +# --- QAT forward for bank weights --- + +def _qat_forward_linear(x: Tensor, w: Tensor) -> Tensor: + """Apply QAT (straight-through estimator) to a bank weight during training.""" + w_cast = w.to(x.dtype) + if CastedLinear._qat_enabled and w.requires_grad: + with torch.no_grad(): + w32 = w.float() + row_max = w32.abs().amax(dim=-1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + # For 3D banks, scale has shape [B, rows]; for 2D, [rows] + if w32.ndim == 2: + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + else: + w_q = (torch.clamp(torch.round(w32 / scale[..., None]), -32, 31) * scale[..., None]).to(x.dtype) + w_cast = w_cast + (w_q - w_cast).detach() + return F.linear(x, w_cast) + + +# --- 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 + CUDA SETUP --- + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # --- TOKENIZER + VALIDATION METRIC SETUP --- + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # --- MODEL SETUP --- + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.use_gated_attention, + value_residual=args.use_value_residual, + use_smeargate=args.use_smeargate, + ).to(device).bfloat16() + + # Banks stay FP32, 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) + 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) + if base_model.smear is not None: + 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()) + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"model_params:{n_params}") + 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}") + log0( + f"features: smeargate={args.use_smeargate} bigramhash={args.use_bigramhash} " + f"value_residual={args.use_value_residual} gated_attn={args.use_gated_attention} " + f"rope_dims={args.rope_dims} xsa_last_n={args.xsa_last_n} " + f"ema={args.use_ema}(decay={args.ema_decay}) " + f"swa={args.swa_enabled}(every={args.swa_every}) " + f"late_qat={args.use_late_qat}(time_frac={args.qat_time_frac}) " + f"muon_wd={args.muon_wd} grad_clip={args.grad_clip_norm}" + ) + + # --- DATA LOADER --- + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # --- WARMUP --- + 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) + + # --- EMA + SWA STATE INIT (keep on GPU for speed) --- + ema_state: dict[str, Tensor] = {} + if args.use_ema: + for name, t in base_model.state_dict().items(): + ema_state[name] = t.detach().float().clone() # stays on GPU + + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + + # --- MAIN TRAINING LOOP --- + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, model, rank, world_size, device, + 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) + + # Late QAT activation -- time-fraction based + if args.use_late_qat and not CastedLinear._qat_enabled and max_wallclock_ms is not None: + time_frac_elapsed = elapsed_ms / max_wallclock_ms + if time_frac_elapsed >= (1.0 - args.qat_time_frac): + CastedLinear._qat_enabled = True + log0(f"step:{step} QAT activated (time_frac={time_frac_elapsed:.3f}, 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 + if args.use_ema: + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(args.ema_decay).add_(t.detach().float(), alpha=1.0 - args.ema_decay) + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + # SWA: average EMA weights periodically during warmdown + if args.swa_enabled and args.use_ema and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: ema_state[name].clone() for name in ema_state} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name in swa_state: + swa_state[name] += ema_state[name] + swa_count += 1 + + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # --- CHOOSE BEST WEIGHTS: raw vs EMA vs SWA --- + raw_sd = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + best_bpb = val_bpb + best_source = "raw" + log0(f"raw model val_bpb:{val_bpb:.4f}") + + if args.use_ema and ema_state: + log0("Evaluating EMA weights...") + ema_sd = {name: ema_state[name].to(dtype=raw_sd[name].dtype) for name in raw_sd} + base_model.load_state_dict(ema_sd, strict=True) + torch.cuda.synchronize() + t_ema = time.perf_counter() + ema_val_loss, ema_val_bpb = eval_val( + args, model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0(f"ema_eval val_bpb:{ema_val_bpb:.4f} eval_time:{1000.0 * (time.perf_counter() - t_ema):.0f}ms") + if ema_val_bpb < best_bpb: + best_bpb = ema_val_bpb + best_source = "ema" + + if args.swa_enabled and swa_state is not None and swa_count > 0: + log0(f"Evaluating SWA ({swa_count} snapshots)...") + swa_sd = {name: (swa_state[name] / swa_count).to(dtype=raw_sd[name].dtype) for name in raw_sd} + base_model.load_state_dict(swa_sd, strict=True) + torch.cuda.synchronize() + t_swa = time.perf_counter() + swa_val_loss, swa_val_bpb = eval_val( + args, model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0(f"swa_eval val_bpb:{swa_val_bpb:.4f} eval_time:{1000.0 * (time.perf_counter() - t_swa):.0f}ms") + if swa_val_bpb < best_bpb: + best_bpb = swa_val_bpb + best_source = "swa" + + log0(f"Using {best_source} weights (val_bpb={best_bpb:.4f})") + if best_source == "raw": + base_model.load_state_dict(raw_sd, strict=True) + elif best_source == "ema": + base_model.load_state_dict(ema_sd, strict=True) + # If swa, weights are already loaded + + # --- SERIALIZATION --- + export_sd = base_model.state_dict() + 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") + + # --- Quantization: unbank -> GPTQ-lite int6 -> compress with zstd --- + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_obj, quant_stats = quantize_state_dict_int8(unbanked_sd) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress(quant_raw) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int6.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int6+zstd: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} " + f"payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission int6+zstd: {quant_file_bytes + code_bytes} bytes") + + # --- Roundtrip validation: decompress -> dequant -> rebank -> eval --- + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(_decompress(quant_blob_disk)), map_location="cpu") + deq_unbanked = dequantize_state_dict_int8(quant_state) + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + + # Build fresh eval model for roundtrip + 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, + 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, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.use_gated_attention, value_residual=args.use_value_residual, + use_smeargate=args.use_smeargate, + ).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, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int6_zstd_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_zstd_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + # Alias for leaderboard compatibility + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + # --- Legal Score-First TTT on quantized model --- + if args.use_ttt: + log0(f"Starting TTT: {args.ttt_epochs} epochs, lr={args.ttt_lr}, " + f"chunk={args.ttt_chunk_tokens}, freeze_blocks={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + log_fn=log0, + ) + torch.cuda.synchronize() + log0( + f"legal_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"legal_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed1337.log b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed1337.log new file mode 100644 index 000000000..bd5b3af1b --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed1337.log @@ -0,0 +1,277 @@ +W0326 09:02:50.644000 166226 torch/distributed/run.py:803] +W0326 09:02:50.644000 166226 torch/distributed/run.py:803] ***************************************** +W0326 09:02:50.644000 166226 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. +W0326 09:02:50.644000 166226 torch/distributed/run.py:803] ***************************************** +logs/0c534995-ca17-4dd9-a293-ff6fac157841.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:26809543 +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 +features: smeargate=True bigramhash=True value_residual=True gated_attn=True rope_dims=16 xsa_last_n=4 ema=True(decay=0.997) swa=True(every=50) late_qat=True(time_frac=0.15) muon_wd=0.04 grad_clip=0.3 +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:6.9305 val_bpb:4.1046 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9314 train_time:136ms step_avg:136.28ms +step:2/20000 train_loss:8.6744 train_time:191ms step_avg:95.35ms +step:3/20000 train_loss:7.6770 train_time:278ms step_avg:92.77ms +step:4/20000 train_loss:7.2297 train_time:365ms step_avg:91.22ms +step:5/20000 train_loss:7.2048 train_time:451ms step_avg:90.27ms +step:6/20000 train_loss:7.1270 train_time:540ms step_avg:89.98ms +step:7/20000 train_loss:7.0202 train_time:626ms step_avg:89.42ms +step:8/20000 train_loss:6.9422 train_time:713ms step_avg:89.14ms +step:9/20000 train_loss:6.5932 train_time:800ms step_avg:88.90ms +step:10/20000 train_loss:6.2041 train_time:887ms step_avg:88.71ms +step:500/20000 train_loss:2.3581 train_time:44365ms step_avg:88.73ms +step:1000/20000 train_loss:2.2458 train_time:88771ms step_avg:88.77ms +step:1500/20000 train_loss:2.2021 train_time:133162ms step_avg:88.77ms +step:2000/20000 train_loss:2.0479 train_time:177588ms step_avg:88.79ms +step:2500/20000 train_loss:2.1547 train_time:222004ms step_avg:88.80ms +step:3000/20000 train_loss:2.1476 train_time:266399ms step_avg:88.80ms +step:3500/20000 train_loss:2.1648 train_time:310831ms step_avg:88.81ms +step:4000/20000 train_loss:1.9566 train_time:355305ms step_avg:88.83ms +step:4000/20000 val_loss:2.0056 val_bpb:1.1878 train_time:355339ms step_avg:88.83ms +step:4500/20000 train_loss:2.1071 train_time:399668ms step_avg:88.82ms +step:5000/20000 train_loss:2.0845 train_time:444022ms step_avg:88.80ms +step:5500/20000 train_loss:2.0028 train_time:488376ms step_avg:88.80ms +step:5744 QAT activated (time_frac=0.850, scale=0.2894) +step:6000/20000 train_loss:1.9246 train_time:532730ms step_avg:88.79ms +swa:start step:6100 +step:6500/20000 train_loss:2.0630 train_time:577072ms step_avg:88.78ms +step:6759/20000 val_loss:1.8860 val_bpb:1.1170 train_time:600062ms step_avg:88.78ms +stopping_early: wallclock_cap train_time:600062ms step:6759/20000 +peak memory allocated: 22672 MiB reserved: 22816 MiB +raw model val_bpb:1.1170 +Evaluating EMA weights... +ema_eval val_bpb:1.1161 eval_time:71149ms +Evaluating SWA (14 snapshots)... +swa_eval val_bpb:1.1184 eval_time:71127ms +Using ema weights (val_bpb=1.1161) +Serialized model: 105826834 bytes +Code size: 81870 bytes +Serialized model int6+zstd: 15892614 bytes (payload:27693850 raw_torch:27759695 payload_ratio:3.82x) +Total submission int6+zstd: 15974484 bytes +final_int6_zstd_roundtrip val_loss:1.8975 val_bpb:1.1238 eval_time:74981ms +final_int6_zstd_roundtrip_exact val_loss:1.89745547 val_bpb:1.12378353 +final_int8_zlib_roundtrip_exact val_loss:1.89745547 val_bpb:1.12378353 +Starting TTT: 3 epochs, lr=0.002, chunk=32768, freeze_blocks=0 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26809543 frozen=0 + ttt_chunk [1/1893] bpb=1.159777 time=0.5s + ttt_chunk [11/1893] bpb=1.148422 time=2.9s + ttt_chunk [21/1893] bpb=1.132815 time=5.3s + ttt_chunk [31/1893] bpb=1.132408 time=7.8s + ttt_chunk [41/1893] bpb=1.118940 time=10.2s + ttt_chunk [51/1893] bpb=1.113445 time=12.6s + ttt_chunk [61/1893] bpb=1.120308 time=15.0s + ttt_chunk [71/1893] bpb=1.118491 time=17.4s + ttt_chunk [81/1893] bpb=1.117455 time=19.8s + ttt_chunk [91/1893] bpb=1.118237 time=22.3s + ttt_chunk [101/1893] bpb=1.121650 time=24.7s + ttt_chunk [111/1893] bpb=1.124120 time=27.1s + ttt_chunk [121/1893] bpb=1.117461 time=29.5s + ttt_chunk [131/1893] bpb=1.117592 time=31.9s + ttt_chunk [141/1893] bpb=1.123137 time=34.3s + ttt_chunk [151/1893] bpb=1.125164 time=36.7s + ttt_chunk [161/1893] bpb=1.124667 time=39.0s + ttt_chunk [171/1893] bpb=1.129027 time=41.4s + ttt_chunk [181/1893] bpb=1.131332 time=43.9s + ttt_chunk [191/1893] bpb=1.138614 time=46.2s + ttt_chunk [201/1893] bpb=1.137428 time=48.6s + ttt_chunk [211/1893] bpb=1.135268 time=51.0s + ttt_chunk [221/1893] bpb=1.136965 time=53.5s + ttt_chunk [231/1893] bpb=1.135644 time=55.9s + ttt_chunk [241/1893] bpb=1.135944 time=58.3s + ttt_chunk [251/1893] bpb=1.135539 time=60.7s + ttt_chunk [261/1893] bpb=1.132637 time=63.1s + ttt_chunk [271/1893] bpb=1.131447 time=65.5s + ttt_chunk [281/1893] bpb=1.132698 time=67.9s + ttt_chunk [291/1893] bpb=1.134439 time=70.3s + ttt_chunk [301/1893] bpb=1.135147 time=72.7s + ttt_chunk [311/1893] bpb=1.137155 time=75.1s + ttt_chunk [321/1893] bpb=1.139093 time=77.5s + ttt_chunk [331/1893] bpb=1.138970 time=79.9s + ttt_chunk [341/1893] bpb=1.137975 time=82.3s + ttt_chunk [351/1893] bpb=1.140282 time=84.7s + ttt_chunk [361/1893] bpb=1.140443 time=87.1s + ttt_chunk [371/1893] bpb=1.139766 time=89.5s + ttt_chunk [381/1893] bpb=1.139908 time=91.9s + ttt_chunk [391/1893] bpb=1.139746 time=94.3s + ttt_chunk [401/1893] bpb=1.137693 time=96.7s + ttt_chunk [411/1893] bpb=1.136587 time=99.1s + ttt_chunk [421/1893] bpb=1.135673 time=101.5s + ttt_chunk [431/1893] bpb=1.135507 time=103.9s + ttt_chunk [441/1893] bpb=1.135859 time=106.3s + ttt_chunk [451/1893] bpb=1.136173 time=108.7s + ttt_chunk [461/1893] bpb=1.135075 time=111.1s + ttt_chunk [471/1893] bpb=1.135696 time=113.6s + ttt_chunk [481/1893] bpb=1.135304 time=116.0s + ttt_chunk [491/1893] bpb=1.134229 time=118.4s + ttt_chunk [501/1893] bpb=1.133727 time=120.8s + ttt_chunk [511/1893] bpb=1.133064 time=123.2s + ttt_chunk [521/1893] bpb=1.130684 time=125.6s + ttt_chunk [531/1893] bpb=1.131866 time=128.0s + ttt_chunk [541/1893] bpb=1.132206 time=130.4s + ttt_chunk [551/1893] bpb=1.131169 time=132.8s + ttt_chunk [561/1893] bpb=1.131685 time=135.2s + ttt_chunk [571/1893] bpb=1.130679 time=137.6s + ttt_chunk [581/1893] bpb=1.129860 time=140.0s + ttt_chunk [591/1893] bpb=1.129215 time=142.4s + ttt_chunk [601/1893] bpb=1.129717 time=144.8s + ttt_chunk [611/1893] bpb=1.129582 time=147.2s + ttt_chunk [621/1893] bpb=1.129441 time=149.6s + ttt_chunk [631/1893] bpb=1.130135 time=152.1s + ttt_chunk [641/1893] bpb=1.129913 time=154.5s + ttt_chunk [651/1893] bpb=1.130020 time=156.9s + ttt_chunk [661/1893] bpb=1.129463 time=159.3s + ttt_chunk [671/1893] bpb=1.129819 time=161.7s + ttt_chunk [681/1893] bpb=1.130510 time=164.1s + ttt_chunk [691/1893] bpb=1.131525 time=166.5s + ttt_chunk [701/1893] bpb=1.130955 time=168.9s + ttt_chunk [711/1893] bpb=1.130909 time=171.3s + ttt_chunk [721/1893] bpb=1.130566 time=173.7s + ttt_chunk [731/1893] bpb=1.130610 time=176.2s + ttt_chunk [741/1893] bpb=1.130714 time=178.5s + ttt_chunk [751/1893] bpb=1.130573 time=180.9s + ttt_chunk [761/1893] bpb=1.130518 time=183.3s + ttt_chunk [771/1893] bpb=1.130187 time=185.7s + ttt_chunk [781/1893] bpb=1.130940 time=188.1s + ttt_chunk [791/1893] bpb=1.130524 time=190.5s + ttt_chunk [801/1893] bpb=1.130852 time=192.9s + ttt_chunk [811/1893] bpb=1.130613 time=195.3s + ttt_chunk [821/1893] bpb=1.130397 time=197.7s + ttt_chunk [831/1893] bpb=1.130212 time=200.1s + ttt_chunk [841/1893] bpb=1.129572 time=202.5s + ttt_chunk [851/1893] bpb=1.129319 time=204.9s + ttt_chunk [861/1893] bpb=1.129075 time=207.3s + ttt_chunk [871/1893] bpb=1.129311 time=209.7s + ttt_chunk [881/1893] bpb=1.129470 time=212.1s + ttt_chunk [891/1893] bpb=1.129052 time=214.5s + ttt_chunk [901/1893] bpb=1.128776 time=216.9s + ttt_chunk [911/1893] bpb=1.128882 time=219.3s + ttt_chunk [921/1893] bpb=1.129361 time=221.7s + ttt_chunk [931/1893] bpb=1.129323 time=224.1s + ttt_chunk [941/1893] bpb=1.129013 time=226.5s + ttt_chunk [951/1893] bpb=1.129407 time=228.9s + ttt_chunk [961/1893] bpb=1.129480 time=231.3s + ttt_chunk [971/1893] bpb=1.130338 time=233.7s + ttt_chunk [981/1893] bpb=1.130427 time=236.1s + ttt_chunk [991/1893] bpb=1.130419 time=238.5s + ttt_chunk [1001/1893] bpb=1.130383 time=240.9s + ttt_chunk [1011/1893] bpb=1.130174 time=243.2s + ttt_chunk [1021/1893] bpb=1.130493 time=245.6s + ttt_chunk [1031/1893] bpb=1.130949 time=248.0s + ttt_chunk [1041/1893] bpb=1.130625 time=250.4s + ttt_chunk [1051/1893] bpb=1.130364 time=252.8s + ttt_chunk [1061/1893] bpb=1.130416 time=255.2s + ttt_chunk [1071/1893] bpb=1.131022 time=257.6s + ttt_chunk [1081/1893] bpb=1.131293 time=260.0s + ttt_chunk [1091/1893] bpb=1.132015 time=262.4s + ttt_chunk [1101/1893] bpb=1.132024 time=264.8s + ttt_chunk [1111/1893] bpb=1.131867 time=267.2s + ttt_chunk [1121/1893] bpb=1.131654 time=269.6s + ttt_chunk [1131/1893] bpb=1.131538 time=272.0s + ttt_chunk [1141/1893] bpb=1.131227 time=274.4s + ttt_chunk [1151/1893] bpb=1.131255 time=276.8s + ttt_chunk [1161/1893] bpb=1.130890 time=279.2s + ttt_chunk [1171/1893] bpb=1.131196 time=281.6s + ttt_chunk [1181/1893] bpb=1.130463 time=284.0s + ttt_chunk [1191/1893] bpb=1.130325 time=286.4s + ttt_chunk [1201/1893] bpb=1.130714 time=288.8s + ttt_chunk [1211/1893] bpb=1.130258 time=291.2s + ttt_chunk [1221/1893] bpb=1.129971 time=293.6s + ttt_chunk [1231/1893] bpb=1.129685 time=296.0s + ttt_chunk [1241/1893] bpb=1.129342 time=298.4s + ttt_chunk [1251/1893] bpb=1.128759 time=300.8s + ttt_chunk [1261/1893] bpb=1.128730 time=303.2s + ttt_chunk [1271/1893] bpb=1.128363 time=305.6s + ttt_chunk [1281/1893] bpb=1.128177 time=308.0s + ttt_chunk [1291/1893] bpb=1.127941 time=310.5s + ttt_chunk [1301/1893] bpb=1.127356 time=312.9s + ttt_chunk [1311/1893] bpb=1.126963 time=315.3s + ttt_chunk [1321/1893] bpb=1.126645 time=317.7s + ttt_chunk [1331/1893] bpb=1.126593 time=320.1s + ttt_chunk [1341/1893] bpb=1.126484 time=322.5s + ttt_chunk [1351/1893] bpb=1.126417 time=324.9s + ttt_chunk [1361/1893] bpb=1.126479 time=327.3s + ttt_chunk [1371/1893] bpb=1.126335 time=329.7s + ttt_chunk [1381/1893] bpb=1.126317 time=332.1s + ttt_chunk [1391/1893] bpb=1.125916 time=334.5s + ttt_chunk [1401/1893] bpb=1.125891 time=336.9s + ttt_chunk [1411/1893] bpb=1.126012 time=339.3s + ttt_chunk [1421/1893] bpb=1.126268 time=341.7s + ttt_chunk [1431/1893] bpb=1.125988 time=344.1s + ttt_chunk [1441/1893] bpb=1.126493 time=346.5s + ttt_chunk [1451/1893] bpb=1.126832 time=348.9s + ttt_chunk [1461/1893] bpb=1.126377 time=351.4s + ttt_chunk [1471/1893] bpb=1.127436 time=353.8s + ttt_chunk [1481/1893] bpb=1.126982 time=356.2s + ttt_chunk [1491/1893] bpb=1.126788 time=358.6s + ttt_chunk [1501/1893] bpb=1.126702 time=361.1s + ttt_chunk [1511/1893] bpb=1.126726 time=363.5s + ttt_chunk [1521/1893] bpb=1.126753 time=365.9s + ttt_chunk [1531/1893] bpb=1.126237 time=368.3s + ttt_chunk [1541/1893] bpb=1.126085 time=370.7s + ttt_chunk [1551/1893] bpb=1.126386 time=373.1s + ttt_chunk [1561/1893] bpb=1.126410 time=375.5s + ttt_chunk [1571/1893] bpb=1.126257 time=377.9s + ttt_chunk [1581/1893] bpb=1.126375 time=380.3s + ttt_chunk [1591/1893] bpb=1.126219 time=382.7s + ttt_chunk [1601/1893] bpb=1.126395 time=385.1s + ttt_chunk [1611/1893] bpb=1.126337 time=387.5s + ttt_chunk [1621/1893] bpb=1.125932 time=389.9s + ttt_chunk [1631/1893] bpb=1.126249 time=392.4s + ttt_chunk [1641/1893] bpb=1.126273 time=394.8s + ttt_chunk [1651/1893] bpb=1.126222 time=397.2s + ttt_chunk [1661/1893] bpb=1.126118 time=399.6s + ttt_chunk [1671/1893] bpb=1.126575 time=402.0s + ttt_chunk [1681/1893] bpb=1.126718 time=404.4s + ttt_chunk [1691/1893] bpb=1.126552 time=406.7s + ttt_chunk [1701/1893] bpb=1.126711 time=409.2s + ttt_chunk [1711/1893] bpb=1.126717 time=411.6s + ttt_chunk [1721/1893] bpb=1.126731 time=414.0s + ttt_chunk [1731/1893] bpb=1.126615 time=416.4s + ttt_chunk [1741/1893] bpb=1.126412 time=418.8s + ttt_chunk [1751/1893] bpb=1.126243 time=421.2s + ttt_chunk [1761/1893] bpb=1.126383 time=423.7s + ttt_chunk [1771/1893] bpb=1.126282 time=426.0s + ttt_chunk [1781/1893] bpb=1.126319 time=428.4s + ttt_chunk [1791/1893] bpb=1.125913 time=430.8s + ttt_chunk [1801/1893] bpb=1.125800 time=433.3s + ttt_chunk [1811/1893] bpb=1.125698 time=435.7s + ttt_chunk [1821/1893] bpb=1.125755 time=438.1s + ttt_chunk [1831/1893] bpb=1.125160 time=440.5s + ttt_chunk [1841/1893] bpb=1.125112 time=442.9s + ttt_chunk [1851/1893] bpb=1.124893 time=445.3s + ttt_chunk [1861/1893] bpb=1.124527 time=447.8s + ttt_chunk [1871/1893] bpb=1.124510 time=450.2s + ttt_chunk [1881/1893] bpb=1.124062 time=452.6s + ttt_chunk [1891/1893] bpb=1.123825 time=455.0s + ttt_chunk [1893/1893] bpb=1.123866 time=455.3s +ttt_sliding:done val_loss=1.893964 val_bpb=1.121715 elapsed=455.3s +legal_ttt val_loss:1.8940 val_bpb:1.1217 eval_time:455898ms +legal_ttt_exact val_loss:1.89396354 val_bpb:1.12171486 diff --git a/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed2024.log b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed2024.log new file mode 100644 index 000000000..2ec56fef4 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed2024.log @@ -0,0 +1,277 @@ +W0326 09:55:03.717000 168727 torch/distributed/run.py:803] +W0326 09:55:03.717000 168727 torch/distributed/run.py:803] ***************************************** +W0326 09:55:03.717000 168727 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. +W0326 09:55:03.717000 168727 torch/distributed/run.py:803] ***************************************** +logs/8fe5cc84-befb-433e-b39c-86c6eba69fe9.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:26809543 +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:2024 +features: smeargate=True bigramhash=True value_residual=True gated_attn=True rope_dims=16 xsa_last_n=4 ema=True(decay=0.997) swa=True(every=50) late_qat=True(time_frac=0.15) muon_wd=0.04 grad_clip=0.3 +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:6.9310 val_bpb:4.1049 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9327 train_time:136ms step_avg:136.30ms +step:2/20000 train_loss:8.6353 train_time:190ms step_avg:95.06ms +step:3/20000 train_loss:7.6655 train_time:277ms step_avg:92.22ms +step:4/20000 train_loss:7.3087 train_time:364ms step_avg:91.12ms +step:5/20000 train_loss:7.1836 train_time:451ms step_avg:90.27ms +step:6/20000 train_loss:7.0793 train_time:538ms step_avg:89.70ms +step:7/20000 train_loss:6.9748 train_time:626ms step_avg:89.47ms +step:8/20000 train_loss:6.9001 train_time:713ms step_avg:89.13ms +step:9/20000 train_loss:6.5047 train_time:800ms step_avg:88.85ms +step:10/20000 train_loss:6.1308 train_time:887ms step_avg:88.75ms +step:500/20000 train_loss:2.3619 train_time:44430ms step_avg:88.86ms +step:1000/20000 train_loss:2.2522 train_time:88873ms step_avg:88.87ms +step:1500/20000 train_loss:2.2067 train_time:133317ms step_avg:88.88ms +step:2000/20000 train_loss:2.0513 train_time:177779ms step_avg:88.89ms +step:2500/20000 train_loss:2.1544 train_time:222348ms step_avg:88.94ms +step:3000/20000 train_loss:2.1497 train_time:266811ms step_avg:88.94ms +step:3500/20000 train_loss:2.1647 train_time:311252ms step_avg:88.93ms +step:4000/20000 train_loss:1.9541 train_time:355700ms step_avg:88.93ms +step:4000/20000 val_loss:2.0058 val_bpb:1.1879 train_time:355734ms step_avg:88.93ms +step:4500/20000 train_loss:2.1054 train_time:400083ms step_avg:88.91ms +step:5000/20000 train_loss:2.0882 train_time:444487ms step_avg:88.90ms +step:5500/20000 train_loss:2.0025 train_time:488880ms step_avg:88.89ms +step:5738 QAT activated (time_frac=0.850, scale=0.2891) +step:6000/20000 train_loss:1.9255 train_time:533251ms step_avg:88.88ms +swa:start step:6100 +step:6500/20000 train_loss:2.0610 train_time:577633ms step_avg:88.87ms +step:6752/20000 val_loss:1.8858 val_bpb:1.1169 train_time:600041ms step_avg:88.87ms +stopping_early: wallclock_cap train_time:600041ms step:6752/20000 +peak memory allocated: 22672 MiB reserved: 22816 MiB +raw model val_bpb:1.1169 +Evaluating EMA weights... +ema_eval val_bpb:1.1160 eval_time:71186ms +Evaluating SWA (14 snapshots)... +swa_eval val_bpb:1.1182 eval_time:70814ms +Using ema weights (val_bpb=1.1160) +Serialized model: 105826834 bytes +Code size: 81870 bytes +Serialized model int6+zstd: 15831649 bytes (payload:27693850 raw_torch:27759695 payload_ratio:3.82x) +Total submission int6+zstd: 15913519 bytes +final_int6_zstd_roundtrip val_loss:1.8968 val_bpb:1.1234 eval_time:75072ms +final_int6_zstd_roundtrip_exact val_loss:1.89684659 val_bpb:1.12342291 +final_int8_zlib_roundtrip_exact val_loss:1.89684659 val_bpb:1.12342291 +Starting TTT: 3 epochs, lr=0.002, chunk=32768, freeze_blocks=0 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26809543 frozen=0 + ttt_chunk [1/1893] bpb=1.163633 time=0.5s + ttt_chunk [11/1893] bpb=1.148027 time=2.9s + ttt_chunk [21/1893] bpb=1.133199 time=5.3s + ttt_chunk [31/1893] bpb=1.131264 time=7.7s + ttt_chunk [41/1893] bpb=1.117920 time=10.1s + ttt_chunk [51/1893] bpb=1.112576 time=12.5s + ttt_chunk [61/1893] bpb=1.119547 time=14.9s + ttt_chunk [71/1893] bpb=1.117924 time=17.4s + ttt_chunk [81/1893] bpb=1.116987 time=19.8s + ttt_chunk [91/1893] bpb=1.117878 time=22.2s + ttt_chunk [101/1893] bpb=1.121524 time=24.6s + ttt_chunk [111/1893] bpb=1.123917 time=27.0s + ttt_chunk [121/1893] bpb=1.117096 time=29.4s + ttt_chunk [131/1893] bpb=1.117142 time=31.7s + ttt_chunk [141/1893] bpb=1.122714 time=34.1s + ttt_chunk [151/1893] bpb=1.124413 time=36.6s + ttt_chunk [161/1893] bpb=1.124059 time=39.0s + ttt_chunk [171/1893] bpb=1.128434 time=41.3s + ttt_chunk [181/1893] bpb=1.130643 time=43.7s + ttt_chunk [191/1893] bpb=1.137990 time=46.1s + ttt_chunk [201/1893] bpb=1.136888 time=48.5s + ttt_chunk [211/1893] bpb=1.134701 time=50.9s + ttt_chunk [221/1893] bpb=1.136247 time=53.3s + ttt_chunk [231/1893] bpb=1.134852 time=55.7s + ttt_chunk [241/1893] bpb=1.135133 time=58.1s + ttt_chunk [251/1893] bpb=1.134609 time=60.5s + ttt_chunk [261/1893] bpb=1.131831 time=62.9s + ttt_chunk [271/1893] bpb=1.130667 time=65.4s + ttt_chunk [281/1893] bpb=1.132021 time=67.8s + ttt_chunk [291/1893] bpb=1.133749 time=70.2s + ttt_chunk [301/1893] bpb=1.134492 time=72.6s + ttt_chunk [311/1893] bpb=1.136611 time=75.0s + ttt_chunk [321/1893] bpb=1.138586 time=77.4s + ttt_chunk [331/1893] bpb=1.138411 time=79.9s + ttt_chunk [341/1893] bpb=1.137500 time=82.3s + ttt_chunk [351/1893] bpb=1.139814 time=84.7s + ttt_chunk [361/1893] bpb=1.140041 time=87.0s + ttt_chunk [371/1893] bpb=1.139333 time=89.4s + ttt_chunk [381/1893] bpb=1.139526 time=91.8s + ttt_chunk [391/1893] bpb=1.139357 time=94.3s + ttt_chunk [401/1893] bpb=1.137320 time=96.7s + ttt_chunk [411/1893] bpb=1.136112 time=99.1s + ttt_chunk [421/1893] bpb=1.135161 time=101.5s + ttt_chunk [431/1893] bpb=1.135017 time=103.9s + ttt_chunk [441/1893] bpb=1.135430 time=106.3s + ttt_chunk [451/1893] bpb=1.135784 time=108.7s + ttt_chunk [461/1893] bpb=1.134685 time=111.1s + ttt_chunk [471/1893] bpb=1.135278 time=113.5s + ttt_chunk [481/1893] bpb=1.134903 time=115.9s + ttt_chunk [491/1893] bpb=1.133807 time=118.3s + ttt_chunk [501/1893] bpb=1.133320 time=120.7s + ttt_chunk [511/1893] bpb=1.132624 time=123.2s + ttt_chunk [521/1893] bpb=1.130273 time=125.6s + ttt_chunk [531/1893] bpb=1.131465 time=128.0s + ttt_chunk [541/1893] bpb=1.131829 time=130.4s + ttt_chunk [551/1893] bpb=1.130801 time=132.8s + ttt_chunk [561/1893] bpb=1.131342 time=135.2s + ttt_chunk [571/1893] bpb=1.130330 time=137.6s + ttt_chunk [581/1893] bpb=1.129480 time=140.0s + ttt_chunk [591/1893] bpb=1.128864 time=142.4s + ttt_chunk [601/1893] bpb=1.129371 time=144.8s + ttt_chunk [611/1893] bpb=1.129264 time=147.2s + ttt_chunk [621/1893] bpb=1.129119 time=149.6s + ttt_chunk [631/1893] bpb=1.129822 time=152.0s + ttt_chunk [641/1893] bpb=1.129614 time=154.5s + ttt_chunk [651/1893] bpb=1.129746 time=156.8s + ttt_chunk [661/1893] bpb=1.129174 time=159.3s + ttt_chunk [671/1893] bpb=1.129541 time=161.7s + ttt_chunk [681/1893] bpb=1.130271 time=164.1s + ttt_chunk [691/1893] bpb=1.131250 time=166.4s + ttt_chunk [701/1893] bpb=1.130687 time=168.8s + ttt_chunk [711/1893] bpb=1.130659 time=171.3s + ttt_chunk [721/1893] bpb=1.130321 time=173.6s + ttt_chunk [731/1893] bpb=1.130361 time=176.0s + ttt_chunk [741/1893] bpb=1.130442 time=178.4s + ttt_chunk [751/1893] bpb=1.130321 time=180.8s + ttt_chunk [761/1893] bpb=1.130265 time=183.2s + ttt_chunk [771/1893] bpb=1.129953 time=185.6s + ttt_chunk [781/1893] bpb=1.130689 time=188.0s + ttt_chunk [791/1893] bpb=1.130288 time=190.4s + ttt_chunk [801/1893] bpb=1.130598 time=192.8s + ttt_chunk [811/1893] bpb=1.130342 time=195.2s + ttt_chunk [821/1893] bpb=1.130116 time=197.6s + ttt_chunk [831/1893] bpb=1.129945 time=200.1s + ttt_chunk [841/1893] bpb=1.129291 time=202.5s + ttt_chunk [851/1893] bpb=1.129057 time=204.9s + ttt_chunk [861/1893] bpb=1.128792 time=207.3s + ttt_chunk [871/1893] bpb=1.129057 time=209.7s + ttt_chunk [881/1893] bpb=1.129201 time=212.1s + ttt_chunk [891/1893] bpb=1.128789 time=214.5s + ttt_chunk [901/1893] bpb=1.128518 time=216.9s + ttt_chunk [911/1893] bpb=1.128673 time=219.3s + ttt_chunk [921/1893] bpb=1.129145 time=221.7s + ttt_chunk [931/1893] bpb=1.129119 time=224.1s + ttt_chunk [941/1893] bpb=1.128764 time=226.5s + ttt_chunk [951/1893] bpb=1.129162 time=228.9s + ttt_chunk [961/1893] bpb=1.129236 time=231.4s + ttt_chunk [971/1893] bpb=1.130105 time=233.7s + ttt_chunk [981/1893] bpb=1.130173 time=236.2s + ttt_chunk [991/1893] bpb=1.130193 time=238.6s + ttt_chunk [1001/1893] bpb=1.130147 time=241.0s + ttt_chunk [1011/1893] bpb=1.129929 time=243.4s + ttt_chunk [1021/1893] bpb=1.130280 time=245.8s + ttt_chunk [1031/1893] bpb=1.130752 time=248.2s + ttt_chunk [1041/1893] bpb=1.130409 time=250.6s + ttt_chunk [1051/1893] bpb=1.130165 time=253.0s + ttt_chunk [1061/1893] bpb=1.130206 time=255.4s + ttt_chunk [1071/1893] bpb=1.130817 time=257.8s + ttt_chunk [1081/1893] bpb=1.131087 time=260.2s + ttt_chunk [1091/1893] bpb=1.131851 time=262.7s + ttt_chunk [1101/1893] bpb=1.131874 time=265.1s + ttt_chunk [1111/1893] bpb=1.131748 time=267.5s + ttt_chunk [1121/1893] bpb=1.131550 time=269.9s + ttt_chunk [1131/1893] bpb=1.131423 time=272.3s + ttt_chunk [1141/1893] bpb=1.131102 time=274.7s + ttt_chunk [1151/1893] bpb=1.131118 time=277.1s + ttt_chunk [1161/1893] bpb=1.130737 time=279.5s + ttt_chunk [1171/1893] bpb=1.131068 time=281.9s + ttt_chunk [1181/1893] bpb=1.130309 time=284.3s + ttt_chunk [1191/1893] bpb=1.130190 time=286.7s + ttt_chunk [1201/1893] bpb=1.130601 time=289.1s + ttt_chunk [1211/1893] bpb=1.130138 time=291.5s + ttt_chunk [1221/1893] bpb=1.129839 time=293.9s + ttt_chunk [1231/1893] bpb=1.129559 time=296.3s + ttt_chunk [1241/1893] bpb=1.129211 time=298.7s + ttt_chunk [1251/1893] bpb=1.128616 time=301.1s + ttt_chunk [1261/1893] bpb=1.128580 time=303.5s + ttt_chunk [1271/1893] bpb=1.128225 time=305.9s + ttt_chunk [1281/1893] bpb=1.128049 time=308.3s + ttt_chunk [1291/1893] bpb=1.127820 time=310.7s + ttt_chunk [1301/1893] bpb=1.127234 time=313.1s + ttt_chunk [1311/1893] bpb=1.126852 time=315.5s + ttt_chunk [1321/1893] bpb=1.126546 time=317.9s + ttt_chunk [1331/1893] bpb=1.126495 time=320.3s + ttt_chunk [1341/1893] bpb=1.126368 time=322.7s + ttt_chunk [1351/1893] bpb=1.126297 time=325.2s + ttt_chunk [1361/1893] bpb=1.126349 time=327.6s + ttt_chunk [1371/1893] bpb=1.126211 time=330.0s + ttt_chunk [1381/1893] bpb=1.126188 time=332.4s + ttt_chunk [1391/1893] bpb=1.125791 time=334.8s + ttt_chunk [1401/1893] bpb=1.125752 time=337.2s + ttt_chunk [1411/1893] bpb=1.125850 time=339.6s + ttt_chunk [1421/1893] bpb=1.126095 time=341.9s + ttt_chunk [1431/1893] bpb=1.125802 time=344.3s + ttt_chunk [1441/1893] bpb=1.126312 time=346.8s + ttt_chunk [1451/1893] bpb=1.126636 time=349.1s + ttt_chunk [1461/1893] bpb=1.126189 time=351.6s + ttt_chunk [1471/1893] bpb=1.127237 time=354.0s + ttt_chunk [1481/1893] bpb=1.126780 time=356.3s + ttt_chunk [1491/1893] bpb=1.126588 time=358.8s + ttt_chunk [1501/1893] bpb=1.126499 time=361.1s + ttt_chunk [1511/1893] bpb=1.126523 time=363.5s + ttt_chunk [1521/1893] bpb=1.126558 time=365.9s + ttt_chunk [1531/1893] bpb=1.126047 time=368.4s + ttt_chunk [1541/1893] bpb=1.125908 time=370.8s + ttt_chunk [1551/1893] bpb=1.126216 time=373.2s + ttt_chunk [1561/1893] bpb=1.126225 time=375.5s + ttt_chunk [1571/1893] bpb=1.126061 time=377.9s + ttt_chunk [1581/1893] bpb=1.126181 time=380.4s + ttt_chunk [1591/1893] bpb=1.126030 time=382.8s + ttt_chunk [1601/1893] bpb=1.126200 time=385.2s + ttt_chunk [1611/1893] bpb=1.126131 time=387.6s + ttt_chunk [1621/1893] bpb=1.125720 time=390.0s + ttt_chunk [1631/1893] bpb=1.126021 time=392.4s + ttt_chunk [1641/1893] bpb=1.126032 time=394.8s + ttt_chunk [1651/1893] bpb=1.125985 time=397.2s + ttt_chunk [1661/1893] bpb=1.125869 time=399.6s + ttt_chunk [1671/1893] bpb=1.126333 time=402.0s + ttt_chunk [1681/1893] bpb=1.126493 time=404.4s + ttt_chunk [1691/1893] bpb=1.126345 time=406.8s + ttt_chunk [1701/1893] bpb=1.126511 time=409.2s + ttt_chunk [1711/1893] bpb=1.126512 time=411.6s + ttt_chunk [1721/1893] bpb=1.126516 time=414.0s + ttt_chunk [1731/1893] bpb=1.126400 time=416.4s + ttt_chunk [1741/1893] bpb=1.126206 time=418.8s + ttt_chunk [1751/1893] bpb=1.126039 time=421.2s + ttt_chunk [1761/1893] bpb=1.126179 time=423.6s + ttt_chunk [1771/1893] bpb=1.126080 time=426.0s + ttt_chunk [1781/1893] bpb=1.126107 time=428.4s + ttt_chunk [1791/1893] bpb=1.125710 time=430.8s + ttt_chunk [1801/1893] bpb=1.125590 time=433.2s + ttt_chunk [1811/1893] bpb=1.125480 time=435.6s + ttt_chunk [1821/1893] bpb=1.125542 time=438.0s + ttt_chunk [1831/1893] bpb=1.124948 time=440.4s + ttt_chunk [1841/1893] bpb=1.124894 time=442.8s + ttt_chunk [1851/1893] bpb=1.124684 time=445.2s + ttt_chunk [1861/1893] bpb=1.124318 time=447.6s + ttt_chunk [1871/1893] bpb=1.124308 time=450.0s + ttt_chunk [1881/1893] bpb=1.123867 time=452.4s + ttt_chunk [1891/1893] bpb=1.123630 time=454.8s + ttt_chunk [1893/1893] bpb=1.123676 time=455.1s +ttt_sliding:done val_loss=1.893681 val_bpb=1.121547 elapsed=455.1s +legal_ttt val_loss:1.8937 val_bpb:1.1215 eval_time:455666ms +legal_ttt_exact val_loss:1.89368084 val_bpb:1.12154743 diff --git a/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed42.log b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed42.log new file mode 100644 index 000000000..eab6165c4 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_11L_ParallelMuon_MLP3x_TTT/train_seed42.log @@ -0,0 +1,277 @@ +W0326 09:28:55.929000 167508 torch/distributed/run.py:803] +W0326 09:28:55.929000 167508 torch/distributed/run.py:803] ***************************************** +W0326 09:28:55.929000 167508 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. +W0326 09:28:55.929000 167508 torch/distributed/run.py:803] ***************************************** +logs/3092f97c-1d60-4df9-978c-6459e739c61b.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:26809543 +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 +features: smeargate=True bigramhash=True value_residual=True gated_attn=True rope_dims=16 xsa_last_n=4 ema=True(decay=0.997) swa=True(every=50) late_qat=True(time_frac=0.15) muon_wd=0.04 grad_clip=0.3 +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:6.9289 val_bpb:4.1037 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9306 train_time:139ms step_avg:139.07ms +step:2/20000 train_loss:8.6849 train_time:195ms step_avg:97.43ms +step:3/20000 train_loss:7.6758 train_time:282ms step_avg:93.96ms +step:4/20000 train_loss:7.2509 train_time:369ms step_avg:92.21ms +step:5/20000 train_loss:7.1238 train_time:455ms step_avg:90.97ms +step:6/20000 train_loss:7.1010 train_time:543ms step_avg:90.51ms +step:7/20000 train_loss:7.1116 train_time:630ms step_avg:89.99ms +step:8/20000 train_loss:6.9847 train_time:717ms step_avg:89.64ms +step:9/20000 train_loss:6.6380 train_time:805ms step_avg:89.40ms +step:10/20000 train_loss:6.2135 train_time:891ms step_avg:89.13ms +step:500/20000 train_loss:2.3618 train_time:44354ms step_avg:88.71ms +step:1000/20000 train_loss:2.2493 train_time:88737ms step_avg:88.74ms +step:1500/20000 train_loss:2.1998 train_time:133121ms step_avg:88.75ms +step:2000/20000 train_loss:2.0511 train_time:177557ms step_avg:88.78ms +step:2500/20000 train_loss:2.1586 train_time:221976ms step_avg:88.79ms +step:3000/20000 train_loss:2.1468 train_time:266402ms step_avg:88.80ms +step:3500/20000 train_loss:2.1661 train_time:310813ms step_avg:88.80ms +step:4000/20000 train_loss:1.9557 train_time:355245ms step_avg:88.81ms +step:4000/20000 val_loss:2.0047 val_bpb:1.1873 train_time:355278ms step_avg:88.82ms +step:4500/20000 train_loss:2.1069 train_time:399650ms step_avg:88.81ms +step:5000/20000 train_loss:2.0848 train_time:444047ms step_avg:88.81ms +step:5500/20000 train_loss:2.0019 train_time:488492ms step_avg:88.82ms +step:5742 QAT activated (time_frac=0.850, scale=0.2895) +step:6000/20000 train_loss:1.9230 train_time:532857ms step_avg:88.81ms +swa:start step:6100 +step:6500/20000 train_loss:2.0613 train_time:577206ms step_avg:88.80ms +step:6757/20000 val_loss:1.8854 val_bpb:1.1167 train_time:600033ms step_avg:88.80ms +stopping_early: wallclock_cap train_time:600033ms step:6757/20000 +peak memory allocated: 22672 MiB reserved: 22816 MiB +raw model val_bpb:1.1167 +Evaluating EMA weights... +ema_eval val_bpb:1.1158 eval_time:71265ms +Evaluating SWA (14 snapshots)... +swa_eval val_bpb:1.1180 eval_time:71264ms +Using ema weights (val_bpb=1.1158) +Serialized model: 105826834 bytes +Code size: 81870 bytes +Serialized model int6+zstd: 15821513 bytes (payload:27693850 raw_torch:27759695 payload_ratio:3.82x) +Total submission int6+zstd: 15903383 bytes +final_int6_zstd_roundtrip val_loss:1.8968 val_bpb:1.1234 eval_time:75480ms +final_int6_zstd_roundtrip_exact val_loss:1.89684242 val_bpb:1.12342044 +final_int8_zlib_roundtrip_exact val_loss:1.89684242 val_bpb:1.12342044 +Starting TTT: 3 epochs, lr=0.002, chunk=32768, freeze_blocks=0 +ttt_sliding:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=26809543 frozen=0 + ttt_chunk [1/1893] bpb=1.157865 time=0.5s + ttt_chunk [11/1893] bpb=1.147669 time=2.9s + ttt_chunk [21/1893] bpb=1.133295 time=5.3s + ttt_chunk [31/1893] bpb=1.130879 time=7.7s + ttt_chunk [41/1893] bpb=1.117995 time=10.1s + ttt_chunk [51/1893] bpb=1.112241 time=12.5s + ttt_chunk [61/1893] bpb=1.119104 time=14.9s + ttt_chunk [71/1893] bpb=1.117758 time=17.3s + ttt_chunk [81/1893] bpb=1.116978 time=19.7s + ttt_chunk [91/1893] bpb=1.118012 time=22.1s + ttt_chunk [101/1893] bpb=1.121546 time=24.6s + ttt_chunk [111/1893] bpb=1.123990 time=27.0s + ttt_chunk [121/1893] bpb=1.117243 time=29.4s + ttt_chunk [131/1893] bpb=1.117180 time=31.8s + ttt_chunk [141/1893] bpb=1.122764 time=34.2s + ttt_chunk [151/1893] bpb=1.124616 time=36.7s + ttt_chunk [161/1893] bpb=1.124071 time=39.1s + ttt_chunk [171/1893] bpb=1.128419 time=41.5s + ttt_chunk [181/1893] bpb=1.130781 time=44.0s + ttt_chunk [191/1893] bpb=1.137990 time=46.4s + ttt_chunk [201/1893] bpb=1.136808 time=48.8s + ttt_chunk [211/1893] bpb=1.134556 time=51.3s + ttt_chunk [221/1893] bpb=1.136107 time=53.7s + ttt_chunk [231/1893] bpb=1.134671 time=56.1s + ttt_chunk [241/1893] bpb=1.135045 time=58.6s + ttt_chunk [251/1893] bpb=1.134571 time=61.0s + ttt_chunk [261/1893] bpb=1.131703 time=63.4s + ttt_chunk [271/1893] bpb=1.130491 time=65.8s + ttt_chunk [281/1893] bpb=1.131886 time=68.2s + ttt_chunk [291/1893] bpb=1.133676 time=70.6s + ttt_chunk [301/1893] bpb=1.134385 time=73.1s + ttt_chunk [311/1893] bpb=1.136440 time=75.5s + ttt_chunk [321/1893] bpb=1.138416 time=77.9s + ttt_chunk [331/1893] bpb=1.138342 time=80.3s + ttt_chunk [341/1893] bpb=1.137392 time=82.7s + ttt_chunk [351/1893] bpb=1.139659 time=85.1s + ttt_chunk [361/1893] bpb=1.139954 time=87.6s + ttt_chunk [371/1893] bpb=1.139254 time=90.1s + ttt_chunk [381/1893] bpb=1.139435 time=92.5s + ttt_chunk [391/1893] bpb=1.139218 time=94.9s + ttt_chunk [401/1893] bpb=1.137169 time=97.3s + ttt_chunk [411/1893] bpb=1.135969 time=99.7s + ttt_chunk [421/1893] bpb=1.135073 time=102.1s + ttt_chunk [431/1893] bpb=1.134943 time=104.5s + ttt_chunk [441/1893] bpb=1.135314 time=106.9s + ttt_chunk [451/1893] bpb=1.135647 time=109.3s + ttt_chunk [461/1893] bpb=1.134538 time=111.7s + ttt_chunk [471/1893] bpb=1.135156 time=114.1s + ttt_chunk [481/1893] bpb=1.134783 time=116.5s + ttt_chunk [491/1893] bpb=1.133662 time=118.9s + ttt_chunk [501/1893] bpb=1.133203 time=121.3s + ttt_chunk [511/1893] bpb=1.132556 time=123.7s + ttt_chunk [521/1893] bpb=1.130154 time=126.1s + ttt_chunk [531/1893] bpb=1.131357 time=128.5s + ttt_chunk [541/1893] bpb=1.131658 time=130.9s + ttt_chunk [551/1893] bpb=1.130638 time=133.3s + ttt_chunk [561/1893] bpb=1.131169 time=135.8s + ttt_chunk [571/1893] bpb=1.130169 time=138.2s + ttt_chunk [581/1893] bpb=1.129333 time=140.6s + ttt_chunk [591/1893] bpb=1.128690 time=143.0s + ttt_chunk [601/1893] bpb=1.129184 time=145.4s + ttt_chunk [611/1893] bpb=1.129115 time=147.8s + ttt_chunk [621/1893] bpb=1.128958 time=150.2s + ttt_chunk [631/1893] bpb=1.129646 time=152.6s + ttt_chunk [641/1893] bpb=1.129410 time=155.0s + ttt_chunk [651/1893] bpb=1.129527 time=157.5s + ttt_chunk [661/1893] bpb=1.128984 time=159.8s + ttt_chunk [671/1893] bpb=1.129327 time=162.4s + ttt_chunk [681/1893] bpb=1.130033 time=164.8s + ttt_chunk [691/1893] bpb=1.131002 time=167.1s + ttt_chunk [701/1893] bpb=1.130447 time=169.5s + ttt_chunk [711/1893] bpb=1.130397 time=171.9s + ttt_chunk [721/1893] bpb=1.130084 time=174.3s + ttt_chunk [731/1893] bpb=1.130121 time=176.7s + ttt_chunk [741/1893] bpb=1.130205 time=179.1s + ttt_chunk [751/1893] bpb=1.130021 time=181.5s + ttt_chunk [761/1893] bpb=1.129976 time=183.9s + ttt_chunk [771/1893] bpb=1.129660 time=186.3s + ttt_chunk [781/1893] bpb=1.130401 time=188.7s + ttt_chunk [791/1893] bpb=1.129995 time=191.1s + ttt_chunk [801/1893] bpb=1.130310 time=193.5s + ttt_chunk [811/1893] bpb=1.130075 time=195.9s + ttt_chunk [821/1893] bpb=1.129866 time=198.3s + ttt_chunk [831/1893] bpb=1.129689 time=200.7s + ttt_chunk [841/1893] bpb=1.129043 time=203.0s + ttt_chunk [851/1893] bpb=1.128756 time=205.4s + ttt_chunk [861/1893] bpb=1.128471 time=207.8s + ttt_chunk [871/1893] bpb=1.128716 time=210.3s + ttt_chunk [881/1893] bpb=1.128899 time=212.6s + ttt_chunk [891/1893] bpb=1.128492 time=215.0s + ttt_chunk [901/1893] bpb=1.128207 time=217.4s + ttt_chunk [911/1893] bpb=1.128348 time=219.8s + ttt_chunk [921/1893] bpb=1.128842 time=222.3s + ttt_chunk [931/1893] bpb=1.128834 time=224.7s + ttt_chunk [941/1893] bpb=1.128502 time=227.1s + ttt_chunk [951/1893] bpb=1.128902 time=229.4s + ttt_chunk [961/1893] bpb=1.128996 time=231.8s + ttt_chunk [971/1893] bpb=1.129849 time=234.2s + ttt_chunk [981/1893] bpb=1.129941 time=236.6s + ttt_chunk [991/1893] bpb=1.129962 time=239.0s + ttt_chunk [1001/1893] bpb=1.129922 time=241.4s + ttt_chunk [1011/1893] bpb=1.129722 time=243.8s + ttt_chunk [1021/1893] bpb=1.130061 time=246.2s + ttt_chunk [1031/1893] bpb=1.130491 time=248.6s + ttt_chunk [1041/1893] bpb=1.130130 time=251.0s + ttt_chunk [1051/1893] bpb=1.129874 time=253.4s + ttt_chunk [1061/1893] bpb=1.129945 time=255.8s + ttt_chunk [1071/1893] bpb=1.130575 time=258.2s + ttt_chunk [1081/1893] bpb=1.130844 time=260.6s + ttt_chunk [1091/1893] bpb=1.131571 time=263.0s + ttt_chunk [1101/1893] bpb=1.131606 time=265.4s + ttt_chunk [1111/1893] bpb=1.131444 time=267.8s + ttt_chunk [1121/1893] bpb=1.131232 time=270.2s + ttt_chunk [1131/1893] bpb=1.131118 time=272.7s + ttt_chunk [1141/1893] bpb=1.130829 time=275.1s + ttt_chunk [1151/1893] bpb=1.130844 time=277.5s + ttt_chunk [1161/1893] bpb=1.130462 time=279.9s + ttt_chunk [1171/1893] bpb=1.130801 time=282.3s + ttt_chunk [1181/1893] bpb=1.130035 time=284.7s + ttt_chunk [1191/1893] bpb=1.129916 time=287.1s + ttt_chunk [1201/1893] bpb=1.130333 time=289.6s + ttt_chunk [1211/1893] bpb=1.129862 time=292.0s + ttt_chunk [1221/1893] bpb=1.129581 time=294.4s + ttt_chunk [1231/1893] bpb=1.129308 time=296.8s + ttt_chunk [1241/1893] bpb=1.128965 time=299.2s + ttt_chunk [1251/1893] bpb=1.128370 time=301.6s + ttt_chunk [1261/1893] bpb=1.128336 time=304.0s + ttt_chunk [1271/1893] bpb=1.127990 time=306.4s + ttt_chunk [1281/1893] bpb=1.127813 time=308.8s + ttt_chunk [1291/1893] bpb=1.127574 time=311.2s + ttt_chunk [1301/1893] bpb=1.126977 time=313.7s + ttt_chunk [1311/1893] bpb=1.126592 time=316.1s + ttt_chunk [1321/1893] bpb=1.126260 time=318.5s + ttt_chunk [1331/1893] bpb=1.126209 time=320.9s + ttt_chunk [1341/1893] bpb=1.126101 time=323.3s + ttt_chunk [1351/1893] bpb=1.126053 time=325.7s + ttt_chunk [1361/1893] bpb=1.126107 time=328.1s + ttt_chunk [1371/1893] bpb=1.125952 time=330.5s + ttt_chunk [1381/1893] bpb=1.125941 time=332.9s + ttt_chunk [1391/1893] bpb=1.125545 time=335.3s + ttt_chunk [1401/1893] bpb=1.125520 time=337.7s + ttt_chunk [1411/1893] bpb=1.125643 time=340.1s + ttt_chunk [1421/1893] bpb=1.125899 time=342.5s + ttt_chunk [1431/1893] bpb=1.125596 time=344.9s + ttt_chunk [1441/1893] bpb=1.126083 time=347.4s + ttt_chunk [1451/1893] bpb=1.126427 time=349.7s + ttt_chunk [1461/1893] bpb=1.125976 time=352.1s + ttt_chunk [1471/1893] bpb=1.127018 time=354.5s + ttt_chunk [1481/1893] bpb=1.126562 time=356.9s + ttt_chunk [1491/1893] bpb=1.126376 time=359.3s + ttt_chunk [1501/1893] bpb=1.126288 time=361.7s + ttt_chunk [1511/1893] bpb=1.126306 time=364.1s + ttt_chunk [1521/1893] bpb=1.126312 time=366.5s + ttt_chunk [1531/1893] bpb=1.125798 time=368.9s + ttt_chunk [1541/1893] bpb=1.125645 time=371.3s + ttt_chunk [1551/1893] bpb=1.125948 time=373.7s + ttt_chunk [1561/1893] bpb=1.125959 time=376.1s + ttt_chunk [1571/1893] bpb=1.125806 time=378.5s + ttt_chunk [1581/1893] bpb=1.125943 time=380.9s + ttt_chunk [1591/1893] bpb=1.125797 time=383.3s + ttt_chunk [1601/1893] bpb=1.125971 time=385.8s + ttt_chunk [1611/1893] bpb=1.125904 time=388.2s + ttt_chunk [1621/1893] bpb=1.125484 time=390.6s + ttt_chunk [1631/1893] bpb=1.125796 time=393.0s + ttt_chunk [1641/1893] bpb=1.125802 time=395.4s + ttt_chunk [1651/1893] bpb=1.125763 time=397.8s + ttt_chunk [1661/1893] bpb=1.125649 time=400.2s + ttt_chunk [1671/1893] bpb=1.126120 time=402.6s + ttt_chunk [1681/1893] bpb=1.126268 time=405.0s + ttt_chunk [1691/1893] bpb=1.126112 time=407.4s + ttt_chunk [1701/1893] bpb=1.126277 time=409.8s + ttt_chunk [1711/1893] bpb=1.126267 time=412.2s + ttt_chunk [1721/1893] bpb=1.126268 time=414.6s + ttt_chunk [1731/1893] bpb=1.126155 time=417.0s + ttt_chunk [1741/1893] bpb=1.125959 time=419.4s + ttt_chunk [1751/1893] bpb=1.125798 time=421.8s + ttt_chunk [1761/1893] bpb=1.125950 time=424.2s + ttt_chunk [1771/1893] bpb=1.125849 time=426.6s + ttt_chunk [1781/1893] bpb=1.125885 time=428.9s + ttt_chunk [1791/1893] bpb=1.125489 time=431.3s + ttt_chunk [1801/1893] bpb=1.125372 time=433.8s + ttt_chunk [1811/1893] bpb=1.125273 time=436.2s + ttt_chunk [1821/1893] bpb=1.125336 time=438.6s + ttt_chunk [1831/1893] bpb=1.124738 time=440.9s + ttt_chunk [1841/1893] bpb=1.124685 time=443.3s + ttt_chunk [1851/1893] bpb=1.124475 time=445.7s + ttt_chunk [1861/1893] bpb=1.124112 time=448.1s + ttt_chunk [1871/1893] bpb=1.124102 time=450.5s + ttt_chunk [1881/1893] bpb=1.123649 time=452.9s + ttt_chunk [1891/1893] bpb=1.123410 time=455.3s + ttt_chunk [1893/1893] bpb=1.123454 time=455.6s +ttt_sliding:done val_loss=1.893332 val_bpb=1.121341 elapsed=455.7s +legal_ttt val_loss:1.8933 val_bpb:1.1213 eval_time:456200ms +legal_ttt_exact val_loss:1.89333202 val_bpb:1.12134084