diff --git a/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/README.md b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/README.md new file mode 100644 index 000000000..fd6370e29 --- /dev/null +++ b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/README.md @@ -0,0 +1,29 @@ +# 11L BackoffNgramMixer + +## Results + +| Seed | val_bpb | Eval time | +|------|---------|-----------| +| 42 | 0.6672 | 512s | +| 1337 | 0.6673 | ~512s | +| 2024 | 0.6667 | ~512s | +| **Mean** | **0.6671** | | +| Std | 0.0003 | | + +Artifact: ~16.0 MB. Train: 600s on 8xH100 SXM. Eval: ~512s. + +## Architecture + +- 11 layers, 512 dim, 8/8 full MHA heads +- XSA-all, LeakyReLU(0.5)^2, 3.5x MLP +- BigramHash, SmearGate, Value Residual, Gated Attention +- int5 quantization + zstd compression +- EMA, Tight SWA, Soft-Round QAT + +## BackoffNgramMixer + +GPU-vectorized multi-order n-gram backoff (orders 2-7) with entropy-adaptive alpha mixing. Score-first backward-looking cache with per-token entropy gating. + +## Acknowledgments + +Architecture and mixer based on community techniques. diff --git a/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/submission.json b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/submission.json new file mode 100644 index 000000000..6df21b4f7 --- /dev/null +++ b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/submission.json @@ -0,0 +1,9 @@ +{ + "name": "11L BackoffNgramMixer", + "val_loss": 0.6671, + "bytes_total": 15995959, + "blurb": "11L XSA-all 8/8 MHA with BackoffNgramMixer (entropy-adaptive, orders 2-7). Mean 0.6671 (std 0.0003) across 3 seeds.", + "author": "hypery11", + "github_id": "hypery11", + "date": "2026-03-26" +} diff --git a/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_gpt.py b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_gpt.py new file mode 100644 index 000000000..17f716703 --- /dev/null +++ b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_gpt.py @@ -0,0 +1,1757 @@ +"""V27: CROWN-Q training + stride=64 + 4 TTT epochs.""" +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True +except ImportError: + try: + from flash_attn import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True + except ImportError: + _HAS_FA3 = False + flash_attn_3_func = None + +class BackoffNgramMixer: + """Multi-order n-gram backoff with entropy-adaptive alpha.""" + + def __init__(self, vocab_size: int = 1024, device: str = 'cuda', eta: float = 0.1): + self.V = vocab_size + self.device = device + self.eta = eta + self.total_tokens = 0 + self.max_order = 7 + self.min_order = 2 + import numpy as _np + self._np = _np + self.BUCKETS = 4_194_304 + self.primes = [_np.uint64(p) for p in [36313, 27191, 51647, 81929, 131071, 174763, 233017]] + self.ctx_counts = [_np.zeros(self.BUCKETS, dtype=_np.uint32) for _ in range(6)] + self.full_counts = [_np.zeros(self.BUCKETS, dtype=_np.uint32) for _ in range(6)] + + def update(self, tokens): + np = self._np + if hasattr(tokens, 'cpu'): + t = tokens.cpu().numpy().astype(np.int64) + else: + t = np.array(tokens, dtype=np.int64) + n = len(t) + if n == 0: + return + self.total_tokens += n + mask = np.uint64(self.BUCKETS - 1) + for oi, order in enumerate(range(self.min_order, self.max_order + 1)): + if n < order: + continue + cw = order - 1 + ctx_hash = np.zeros(n - order + 1, dtype=np.uint64) + for k in range(cw): + ctx_hash ^= t[k:n - order + 1 + k].astype(np.uint64) * self.primes[k] + ctx_key = (ctx_hash & mask).astype(np.int64) + tgt = t[order - 1:].astype(np.uint64) + full_key = ((ctx_hash ^ (tgt * self.primes[cw])) & mask).astype(np.int64) + np.add.at(self.ctx_counts[oi], ctx_key, 1) + np.add.at(self.full_counts[oi], full_key, 1) + + def mix_and_score(self, neural_logits, x_batch, y_batch, wlens): + np = self._np + bsz, slen, V = neural_logits.shape + device = neural_logits.device + neural_lp = F.log_softmax(neural_logits, dim=-1) + neural_nll = -neural_lp.gather(2, y_batch.unsqueeze(2)).squeeze(2) + if self.total_tokens < 100: + return neural_nll, None + with torch.no_grad(): + probs = neural_lp.exp() + entropy = -(probs * neural_lp).sum(dim=-1) + alpha = 0.05 + 0.55 * torch.sigmoid(2.0 * (entropy - 4.0)) + neural_p = neural_lp.gather(2, y_batch.unsqueeze(2)).squeeze(2).exp() + x_np = x_batch.cpu().numpy().astype(np.int64) + y_np = y_batch.cpu().numpy().astype(np.int64) + mask = np.uint64(self.BUCKETS - 1) + uniform_nll = math.log(self.V) + ngram_p = np.zeros((bsz, slen), dtype=np.float64) + ngram_hit = np.zeros((bsz, slen), dtype=np.bool_) + for oi_rev in range(5, -1, -1): + order = oi_rev + 2 + cw = order - 1 + if slen < cw: + continue + ctx_hash = np.zeros((bsz, slen), dtype=np.uint64) + for k in range(cw): + shift = cw - 1 - k + shifted = np.zeros_like(x_np, dtype=np.uint64) + if shift > 0 and shift < slen: + shifted[:, shift:] = x_np[:, :slen - shift].astype(np.uint64) + elif shift == 0: + shifted = x_np.astype(np.uint64) + ctx_hash ^= shifted * self.primes[k] + ctx_key = (ctx_hash & mask).astype(np.int64) + full_key = ((ctx_hash ^ (y_np.astype(np.uint64) * self.primes[cw])) & mask).astype(np.int64) + ctx_c = self.ctx_counts[oi_rev][ctx_key.reshape(-1)].astype(np.float64).reshape(bsz, slen) + full_c = self.full_counts[oi_rev][full_key.reshape(-1)].astype(np.float64).reshape(bsz, slen) + valid = (ctx_c >= 2) & (~ngram_hit) + if cw > 0: + valid[:, :cw] = False + p = np.minimum(full_c, ctx_c) / np.maximum(ctx_c, 1.0) + p = np.clip(p, 0.0, 1.0) + ngram_p[valid] = p[valid] + ngram_hit[valid] = True + ngram_p[~ngram_hit] = 1.0 / self.V + ngram_p_t = torch.tensor(ngram_p, device=device, dtype=torch.float32) + mixed_p = (1.0 - alpha) * neural_p + alpha * ngram_p_t + mixed_nll = -torch.log(mixed_p.clamp(min=1e-12)) + return mixed_nll, None + + def update_weights(self, expert_nll, wlens): + pass + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 8)) + 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", 3.5)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + int6_last_n = int(os.environ.get("INT6_LAST_N", 0)) # all int5 (saves ~300KB vs int6 for last 2 blocks) + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) # post-TTT temperature calibration + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 6144)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.5)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + prune_pct = float(os.environ.get("PRUNE_PCT", 0.03)) + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + +def eval_val(args: Hyperparameters, model: nn.Module, rank: int, world_size: int, + device: torch.device, grad_accum_steps: int, val_tokens: Tensor, + base_bytes_lut: Tensor, has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, eval_seq_len: int | None = None) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_Q = 0.9999984 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def 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) + +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 + _soft_round_alpha: float = 1.0 # temperature for soft-round (annealed during training) + _use_soft_round: bool = False # enable soft-round QAT instead of STE + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self._clip_range = 15 # default int5, set to 31 for int6 layers + + @staticmethod + def soft_round(y: Tensor, alpha: float) -> Tensor: + """Differentiable approximation to round() from Agustsson & Theis (NeurIPS 2020). + s_alpha(y) = floor(y) + 0.5 * tanh(alpha * r) / tanh(alpha/2) + 0.5 + where r = y - floor(y) - 0.5 (centered fractional part) + """ + fl = torch.floor(y) + r = y - fl - 0.5 + return fl + 0.5 * torch.tanh(alpha * r) / (math.tanh(alpha / 2) + 1e-10) + 0.5 + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + cr = self._clip_range + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._use_soft_round: + # Soft-Round QAT: differentiable rounding with temperature annealing + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_scaled = w32 / scale[:, None] + w_rounded = CastedLinear.soft_round(w_scaled, CastedLinear._soft_round_alpha) + w_q = (torch.clamp(w_rounded, -(cr+1), cr) * scale[:, None]).to(x.dtype) + w = w_q # fully differentiable path + else: + # Original STE QAT + with torch.no_grad(): + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -(cr+1), cr) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, + rope_base: float, qk_gain_init: float): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + y_g = y.reshape(B, T, Hkv, H // Hkv, 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, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + 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] + if _HAS_FA3: + y = flash_attn_3_func(q, k, v, causal=True).contiguous() + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + 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) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) + +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): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + return self.proj(F.leaky_relu(self.fc(x), negative_slope=0.5).square()) + +class Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: int, + rope_base: float, qk_gain_init: float, layer_idx: int = 0, + ln_scale: bool = False, dtg: 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) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) + 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) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out + +class GPT(nn.Module): + def __init__(self, vocab_size: int, num_layers: int, model_dim: int, num_heads: int, + num_kv_heads: int, mlp_mult: int, tie_embeddings: bool, tied_embed_init_std: float, + logit_softcap: float, rope_base: float, qk_gain_init: float, + bigram_vocab_size: int = 0, bigram_dim: int = 128, xsa_last_n: int = 0, + rope_dims: int = 0, ln_scale: bool = False, dtg: bool = False, + ve_enabled: bool = False, ve_dim: int = 128, ve_layers: str = "9,10"): + 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.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, + qk_gain_init, layer_idx=i, ln_scale=ln_scale, dtg=dtg) + for i in range(num_layers) + ]) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + 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 + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + 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, v_embed=ve) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + 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, v_embed=ve) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +def eval_val_sliding(args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, eval_seq_len: int | None = None) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + last_full_start = max(total_tokens - seq_len, 0) + window_starts = list(range(0, last_full_start + 1, stride)) + if not window_starts or window_starts[-1] != last_full_start: + window_starts.append(last_full_start) + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + + # Pre-compile: dummy forward+backward with TTT shapes to warm the compile cache + if rank == 0: + print(" ttt: pre-compiling forward+backward kernels...", flush=True) + _dummy_x = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + _dummy_y = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + _dummy_logits = base_model.forward_logits(_dummy_x) + _dummy_loss = F.cross_entropy(_dummy_logits.reshape(-1, _dummy_logits.size(-1)), _dummy_y.reshape(-1)) + _dummy_loss.backward() + base_model.zero_grad(set_to_none=True) + if rank == 0: + print(" ttt: pre-compile done", flush=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, ttt_epochs: int = 3, ttt_lr: float = 0.001, + ttt_momentum: float = 0.9, ttt_freeze_blocks: int = 2, + batch_seqs: int = 32, eval_seq_len: int | None = None, + ttt_chunk_tokens: int = 32768, ttt_optimizer: str = "adamw", + ttt_temp: float = 1.0, + ppm_alpha: float = 0.85, + byte_weighted_ttt: bool = True, + use_cache: bool = True, + cache_alpha: float = 0.3, + adaptive_lr: bool = True, + adaptive_lr_max_mult: float = 3.0, +) -> tuple[float, float]: + """Legal score-first TTT: score each chunk, then train on it. + Every token scored BEFORE any update that could use it.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Initialize GPU-vectorized logistic context mixer + use_mixer = os.environ.get("USE_MIXER", "1") == "1" + mixer = BackoffNgramMixer( + vocab_size=val_tokens.to(torch.int32).max().item() + 1, + device=device, + eta=float(os.environ.get("MIXER_ETA", "0.1")), + ) if use_mixer else None + if use_mixer and rank == 0: + print(f" Logistic context mixer enabled: eta={mixer.eta}") + if adaptive_lr and rank == 0: + print(f" Adaptive LR enabled: max_mult={adaptive_lr_max_mult}") + + # Pre-compute all window starts + last_full_start = max(total_tokens - seq_len, 0) + window_starts = list(range(0, last_full_start + 1, stride)) + if not window_starts or window_starts[-1] != last_full_start: + window_starts.append(last_full_start) + + # Assign each window to a chunk based on scored token position + num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens + 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_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + if rank == 0: + print(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " + f"windows={len(window_starts)} stride={stride} " + f"lr={ttt_lr} epochs={ttt_epochs} opt={ttt_optimizer} " + f"freeze_first={ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze everything, then selectively unfreeze for TTT + num_blocks = len(base_model.blocks) + for p in base_model.parameters(): + p.requires_grad_(False) + ttt_params = [] + ttt_param_ids = set() + use_qttt = os.environ.get("QTTT", "0") == "1" + if use_qttt: + # qTTT: only unfreeze Q projections in last N blocks + norms + head + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for name, p in base_model.blocks[i].named_parameters(): + if "c_q" in name: + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + else: + # Standard: unfreeze all params in last N blocks + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for p in base_model.blocks[i].parameters(): + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + # Unfreeze norms, scales, lm_head + for name, p in base_model.named_parameters(): + if "norm" in name or "scale" in name or "lm_head" in name: + p.requires_grad_(True) + if id(p) not in ttt_param_ids: + ttt_params.append(p) + ttt_param_ids.add(id(p)) + + if rank == 0: + n_unfrozen = sum(p.numel() for p in ttt_params) + n_frozen = sum(p.numel() for p in base_model.parameters() if not p.requires_grad) + print(f"ttt:params unfrozen={n_unfrozen} frozen={n_frozen}") + + if ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=ttt_lr, momentum=ttt_momentum) + + # Polyak averaging (TTT weight EMA) for smoother scoring + use_polyak = os.environ.get("USE_POLYAK", "1") == "1" + polyak_decay = float(os.environ.get("POLYAK_DECAY", "0.998")) + if use_polyak: + polyak_state = {id(p): p.data.clone() for p in ttt_params} + if rank == 0: + print(f" Polyak averaging enabled: decay={polyak_decay}") + + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + + # --- Phase 1: SCORE this chunk (inference_mode, no grad) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # Swap in Polyak-averaged weights for scoring + if use_polyak and ci > 0: + _saved_weights = {} + for p in ttt_params: + _saved_weights[id(p)] = p.data.clone() + p.data.copy_(polyak_state[id(p)]) + + 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) + logits_scaled = logits.float() / ttt_temp + + # Adaptive temperature: sharpen confident predictions more + if ttt_temp != 1.0: + with torch.no_grad(): + probs_for_entropy = F.softmax(logits.float(), dim=-1) + token_entropy = -(probs_for_entropy * (probs_for_entropy + 1e-10).log()).sum(-1) + max_ent = math.log(logits.size(-1)) + # Confident tokens (low entropy) get more sharpening + adaptive_temp = 1.0 - (1.0 - ttt_temp) * (1.0 - token_entropy / max_ent) + adaptive_temp = adaptive_temp.clamp(min=0.9, max=1.05) + logits_scaled = logits.float() / adaptive_temp.unsqueeze(-1) + + # Logistic context mixing (GPU-vectorized) or plain CE + if mixer is not None: + nll, expert_nll = mixer.mix_and_score(logits_scaled, x_batch, y_batch, wlens) + mixer.update_weights(expert_nll, wlens) + else: + nll = F.cross_entropy( + logits_scaled.reshape(-1, logits_scaled.size(-1)), + 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() + + # --- Update context mixer with scored chunk tokens (GPU-vectorized) --- + chunk_start_tok = ci * ttt_chunk_tokens + chunk_end_tok = min((ci + 1) * ttt_chunk_tokens, total_tokens) + if mixer is not None: + mixer.update(val_tokens[chunk_start_tok:chunk_end_tok + 1]) + + # Swap back training weights after scoring + if use_polyak and ci > 0: + for p in ttt_params: + p.data.copy_(_saved_weights[id(p)]) + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and ttt_epochs > 0: + chunk_start = ci * ttt_chunk_tokens + chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) + chunk_seqs = (chunk_end - chunk_start) // seq_len + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] seqs={chunk_seqs} start_train...", flush=True) + if chunk_seqs > 0: + # Cosine LR across chunks + adaptive scaling + cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if adaptive_lr: + # Increase LR as we've seen more data (more confident adaptation) + progress = min(ci / max(num_chunks * 0.3, 1), 1.0) # ramp over first 30% of chunks + lr_mult = 1.0 + (adaptive_lr_max_mult - 1.0) * progress + cos_lr = cos_lr * lr_mult + 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(ttt_epochs): + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] epoch={_ep+1}/{ttt_epochs} batches={my_chunk_seqs} ...", flush=True) + 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): + if byte_weighted_ttt: + # Byte-weighted loss: tokens covering more bytes matter more + ttt_logits = base_model.forward_logits(x) + per_token_loss = F.cross_entropy( + ttt_logits.reshape(-1, ttt_logits.size(-1)), + y.reshape(-1), reduction='none' + ).reshape(y.shape) + byte_weights = base_bytes_lut[y].float() + byte_weights = byte_weights + (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).float() + ttt_loss = (per_token_loss * byte_weights).sum() / byte_weights.sum() + else: + ttt_loss = base_model(x, y) + ttt_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, 1.0) + optimizer.step() + # Update Polyak EMA after each step + if use_polyak: + for p in ttt_params: + polyak_state[id(p)].lerp_(p.data, 1.0 - polyak_decay) + if rank == 0 and ci < 3: + print(f" step done ep={_ep+1} bs={bs} loss={ttt_loss.item():.4f}", flush=True) + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 5): + 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 + print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s", flush=True) + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + 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() + + if rank == 0: + print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + + +def _get_layer_clip_range(name: str, num_layers: int, int6_last_n: int) -> int: + """Return clip_range based on which layer the param belongs to.""" + import re + m = re.search(r'blocks\.(\d+)\.', name) + if m: + layer_idx = int(m.group(1)) + if layer_idx >= num_layers - int6_last_n: + return 31 # int6 + return 15 # int5 + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0(f"Python {sys.version} PyTorch {torch.__version__}", console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + 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: + matrix_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, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + # Set int6 clip_range for last N layers (mixed precision) + int6_start = args.num_layers - args.int6_last_n + for i, block in enumerate(base_model.blocks): + if i >= int6_start: + for m in block.modules(): + if isinstance(m, CastedLinear): + m._clip_range = 31 # int6 + if master_process: + int5_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 15) + int6_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 31) + log0(f"mixed_precision: {int5_count} int5 layers, {int6_count} int6 layers (last {args.int6_last_n} blocks)") + log0(f"model_params:{n_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:{xsa_layers} ws:{world_size} gqa:{args.num_heads}/{args.num_kv_heads}") + log0(f"lr:embed={token_lr} matrix={args.matrix_lr} scalar={args.scalar_lr} batch:{args.train_batch_tokens} wall:{args.max_wallclock_seconds:.0f}s seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + train_reserve_ms = 18000 + effective_train_ms = (max_wallclock_ms - train_reserve_ms) if max_wallclock_ms is not None else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if effective_train_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(effective_train_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + # TTT_ONLY mode: skip training, load saved model, run TTT eval + if os.environ.get("TTT_ONLY", "0") == "1": + log0("TTT_ONLY mode: skipping training, loading saved model...") + sd_cpu = {k: v.cpu() for k, v in torch.load("final_model.pt", map_location="cpu").items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + 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) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + log0(f"TTT_ONLY: model loaded, starting TTT eval...") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = float(os.environ.get("EMA_DECAY", "0.997")) + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + # Anneal soft-round alpha based on QAT progress + if CastedLinear._use_soft_round and CastedLinear._qat_enabled: + qat_progress = max(0.0, 1.0 - scale / max(args.late_qat_threshold, 0.01)) + CastedLinear._soft_round_alpha = 1.0 + 15.0 * qat_progress + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + CastedLinear._use_soft_round = os.environ.get("SOFT_ROUND_QAT", "0") == "1" + if CastedLinear._use_soft_round and master_process: + log0(f"soft_round_qat:enabled initial_alpha=1.0") + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + # CROWN-Q: penalize quantization-sensitive weights during warmdown + crownq_lambda = float(os.environ.get("CROWN_Q_LAMBDA", "0.01")) + if CastedLinear._qat_enabled and crownq_lambda > 0: + cq_loss = torch.zeros((), device=device) + for m in base_model.modules(): + if isinstance(m, CastedLinear) and m.weight.ndim == 2: + w = m.weight.float() + cr = float(m._clip_range) + row_max = w.detach().abs().amax(dim=1) + delta = row_max / cr # quantization step size + cq_loss = cq_loss + (w.pow(2) * delta.pow(2).unsqueeze(1)).mean() + loss = loss + crownq_lambda * cq_loss / 12.0 + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + 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 = effective_train_ms is not None and approx_training_time_ms >= effective_train_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply EMA weights directly (skip diagnostic evals to save ~5s of reserve) + log0("ema:applying EMA weights (skipping diagnostic evals)") + current_state = base_model.state_dict() + ema_sd = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(ema_sd, strict=True) + 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") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + if master_process: + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + 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) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + if sw_seq_len != effective_eval_seq_len and rank == 0: + log0(f"Eval seq_len override: {effective_eval_seq_len} -> {sw_seq_len}") + if args.eval_stride > 0 and args.eval_stride < sw_seq_len and not os.environ.get("SKIP_SLIDING"): + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ppm_alpha_val = float(os.environ.get("PPM_ALPHA", "0.85")) + bw_ttt = os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1" + log0(f"PPM alpha: {ppm_alpha_val}, Byte-weighted TTT: {bw_ttt}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_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/2026-03-26_BackoffNgramMixer/train_seed1337.log b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed1337.log new file mode 100644 index 000000000..4eb76f85b --- /dev/null +++ b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed1337.log @@ -0,0 +1,297 @@ +W0326 04:26:06.154000 324829 torch/distributed/run.py:803] +W0326 04:26:06.154000 324829 torch/distributed/run.py:803] ***************************************** +W0326 04:26:06.154000 324829 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 04:26:06.154000 324829 torch/distributed/run.py:803] ***************************************** +logs/10ce789a-0272-41ab-8a69-43b98ec2a33c.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 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9285 val_bpb:4.1034 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9305 train_time:145ms step_avg:144.90ms +step:2/20000 train_loss:8.6408 train_time:242ms step_avg:120.78ms +step:3/20000 train_loss:7.7277 train_time:342ms step_avg:114.06ms +step:4/20000 train_loss:7.2808 train_time:442ms step_avg:110.56ms +step:5/20000 train_loss:7.0685 train_time:542ms step_avg:108.47ms +step:6/20000 train_loss:6.9656 train_time:642ms step_avg:107.07ms +step:7/20000 train_loss:6.8519 train_time:743ms step_avg:106.09ms +step:8/20000 train_loss:6.7077 train_time:843ms step_avg:105.36ms +step:9/20000 train_loss:6.3629 train_time:943ms step_avg:104.75ms +step:10/20000 train_loss:6.0310 train_time:1043ms step_avg:104.33ms +step:500/20000 train_loss:2.3622 train_time:48826ms step_avg:97.65ms +step:1000/20000 train_loss:2.2347 train_time:97720ms step_avg:97.72ms +step:1500/20000 train_loss:2.1856 train_time:146660ms step_avg:97.77ms +step:2000/20000 train_loss:2.0300 train_time:195655ms step_avg:97.83ms +step:2500/20000 train_loss:2.1340 train_time:244676ms step_avg:97.87ms +step:3000/20000 train_loss:2.1162 train_time:293701ms step_avg:97.90ms +step:3500/20000 train_loss:2.1181 train_time:342745ms step_avg:97.93ms +step:4000/20000 train_loss:1.9070 train_time:391751ms step_avg:97.94ms +step:4000/20000 val_loss:1.9981 val_bpb:1.1834 train_time:391763ms step_avg:97.94ms +late_qat:enabled step:4193 scale:0.4999 +step:4500/20000 train_loss:2.0558 train_time:441694ms step_avg:98.15ms +step:5000/20000 train_loss:2.0307 train_time:492117ms step_avg:98.42ms +swa:start step:5250 +step:5500/20000 train_loss:1.9355 train_time:542870ms step_avg:98.70ms +step:5885/20000 val_loss:1.9032 val_bpb:1.1272 train_time:582051ms step_avg:98.90ms +stopping_early: wallclock_cap train_time:582051ms step:5885/20000 +peak memory allocated: 26197 MiB reserved: 26810 MiB +ema:applying EMA weights (skipping diagnostic evals) +Serialized model: 130432585 bytes +Code size: 87336 bytes +pruning:3.0% magnitude pruning applied +Serialized model int6+zlib: 16865151 bytes +Total submission size int6+zlib: 16952487 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.9184 val_bpb:1.1362 stride:64 eval_time:84815ms +final_int6_sliding_window_exact val_loss:1.91842159 val_bpb:1.13619732 +TTT: epochs=3 lr=0.0005 freeze_first=2 chunk=32768 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=1893 chunk_tokens=32768 windows=969057 stride=64 lr=0.0005 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=16 start_train... + ttt_train [1] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=2.3214 + ttt_train [1] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=2.2462 + ttt_train [1] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=2.2185 + ttt_chunk [1/1893] bpb=1.168290 time=0.3s + ttt_train [2] seqs=16 start_train... + ttt_train [2] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=2.7399 + ttt_train [2] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=2.7281 + ttt_train [2] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=2.7114 + ttt_chunk [2/1893] bpb=1.265442 time=0.6s + ttt_train [3] seqs=16 start_train... + ttt_train [3] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=1.8777 + ttt_train [3] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=1.8687 + ttt_train [3] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=1.8534 + ttt_chunk [3/1893] bpb=1.221009 time=0.9s + ttt_chunk [4/1893] bpb=1.220198 time=1.2s + ttt_chunk [5/1893] bpb=1.212990 time=1.4s + ttt_chunk [11/1893] bpb=1.190102 time=3.1s + ttt_chunk [21/1893] bpb=1.170965 time=5.9s + ttt_chunk [31/1893] bpb=1.161922 time=8.8s + ttt_chunk [41/1893] bpb=1.140608 time=11.7s + ttt_chunk [51/1893] bpb=1.127624 time=14.4s + ttt_chunk [61/1893] bpb=1.124732 time=17.3s + ttt_chunk [71/1893] bpb=1.114150 time=20.0s + ttt_chunk [81/1893] bpb=1.104690 time=22.8s + ttt_chunk [91/1893] bpb=1.097244 time=25.6s + ttt_chunk [101/1893] bpb=1.090816 time=28.4s + ttt_chunk [111/1893] bpb=1.083942 time=31.2s + ttt_chunk [121/1893] bpb=1.069580 time=33.9s + ttt_chunk [131/1893] bpb=1.060707 time=36.7s + ttt_chunk [141/1893] bpb=1.057328 time=39.5s + ttt_chunk [151/1893] bpb=1.049838 time=42.3s + ttt_chunk [161/1893] bpb=1.041519 time=45.1s + ttt_chunk [171/1893] bpb=1.035623 time=47.8s + ttt_chunk [181/1893] bpb=1.029285 time=50.6s + ttt_chunk [191/1893] bpb=1.025552 time=53.4s + ttt_chunk [201/1893] bpb=1.017070 time=56.1s + ttt_chunk [211/1893] bpb=1.007699 time=58.9s + ttt_chunk [221/1893] bpb=1.000488 time=61.7s + ttt_chunk [231/1893] bpb=0.991797 time=64.4s + ttt_chunk [241/1893] bpb=0.984456 time=67.2s + ttt_chunk [251/1893] bpb=0.977353 time=70.0s + ttt_chunk [261/1893] bpb=0.967998 time=72.8s + ttt_chunk [271/1893] bpb=0.960129 time=75.6s + ttt_chunk [281/1893] bpb=0.953803 time=78.3s + ttt_chunk [291/1893] bpb=0.948064 time=81.1s + ttt_chunk [301/1893] bpb=0.941508 time=83.9s + ttt_chunk [311/1893] bpb=0.936138 time=86.7s + ttt_chunk [321/1893] bpb=0.930703 time=89.4s + ttt_chunk [331/1893] bpb=0.924779 time=92.2s + ttt_chunk [341/1893] bpb=0.917923 time=94.9s + ttt_chunk [351/1893] bpb=0.913039 time=97.7s + ttt_chunk [361/1893] bpb=0.907925 time=100.5s + ttt_chunk [371/1893] bpb=0.901862 time=103.2s + ttt_chunk [381/1893] bpb=0.896817 time=106.0s + ttt_chunk [391/1893] bpb=0.891652 time=108.9s + ttt_chunk [401/1893] bpb=0.885752 time=111.8s + ttt_chunk [411/1893] bpb=0.880612 time=114.7s + ttt_chunk [421/1893] bpb=0.875347 time=117.5s + ttt_chunk [431/1893] bpb=0.870496 time=120.3s + ttt_chunk [441/1893] bpb=0.866159 time=123.0s + ttt_chunk [451/1893] bpb=0.861574 time=125.8s + ttt_chunk [461/1893] bpb=0.856699 time=128.6s + ttt_chunk [471/1893] bpb=0.852287 time=131.4s + ttt_chunk [481/1893] bpb=0.848275 time=134.1s + ttt_chunk [491/1893] bpb=0.843626 time=136.9s + ttt_chunk [501/1893] bpb=0.839784 time=139.6s + ttt_chunk [511/1893] bpb=0.836040 time=142.3s + ttt_chunk [521/1893] bpb=0.831557 time=145.1s + ttt_chunk [531/1893] bpb=0.828283 time=147.8s + ttt_chunk [541/1893] bpb=0.825049 time=150.5s + ttt_chunk [551/1893] bpb=0.821294 time=153.3s + ttt_chunk [561/1893] bpb=0.818098 time=156.0s + ttt_chunk [571/1893] bpb=0.814563 time=158.7s + ttt_chunk [581/1893] bpb=0.811189 time=161.4s + ttt_chunk [591/1893] bpb=0.807961 time=164.2s + ttt_chunk [601/1893] bpb=0.805166 time=166.9s + ttt_chunk [611/1893] bpb=0.802336 time=169.7s + ttt_chunk [621/1893] bpb=0.799557 time=172.4s + ttt_chunk [631/1893] bpb=0.797040 time=175.2s + ttt_chunk [641/1893] bpb=0.794492 time=177.9s + ttt_chunk [651/1893] bpb=0.791729 time=180.7s + ttt_chunk [661/1893] bpb=0.789151 time=183.5s + ttt_chunk [671/1893] bpb=0.786874 time=186.2s + ttt_chunk [681/1893] bpb=0.784405 time=189.0s + ttt_chunk [691/1893] bpb=0.782543 time=191.7s + ttt_chunk [701/1893] bpb=0.780100 time=194.4s + ttt_chunk [711/1893] bpb=0.778052 time=197.1s + ttt_chunk [721/1893] bpb=0.775891 time=199.7s + ttt_chunk [731/1893] bpb=0.773908 time=202.4s + ttt_chunk [741/1893] bpb=0.771872 time=205.1s + ttt_chunk [751/1893] bpb=0.769838 time=207.8s + ttt_chunk [761/1893] bpb=0.767966 time=210.5s + ttt_chunk [771/1893] bpb=0.766114 time=213.2s + ttt_chunk [781/1893] bpb=0.764751 time=215.9s + ttt_chunk [791/1893] bpb=0.762825 time=218.7s + ttt_chunk [801/1893] bpb=0.760982 time=221.5s + ttt_chunk [811/1893] bpb=0.759267 time=224.2s + ttt_chunk [821/1893] bpb=0.757481 time=226.9s + ttt_chunk [831/1893] bpb=0.755965 time=229.7s + ttt_chunk [841/1893] bpb=0.754139 time=232.4s + ttt_chunk [851/1893] bpb=0.752564 time=235.1s + ttt_chunk [861/1893] bpb=0.750949 time=237.8s + ttt_chunk [871/1893] bpb=0.749485 time=240.5s + ttt_chunk [881/1893] bpb=0.748130 time=243.2s + ttt_chunk [891/1893] bpb=0.746736 time=245.9s + ttt_chunk [901/1893] bpb=0.745493 time=248.7s + ttt_chunk [911/1893] bpb=0.744213 time=251.4s + ttt_chunk [921/1893] bpb=0.742991 time=254.1s + ttt_chunk [931/1893] bpb=0.741673 time=256.8s + ttt_chunk [941/1893] bpb=0.740295 time=259.5s + ttt_chunk [951/1893] bpb=0.739129 time=262.2s + ttt_chunk [961/1893] bpb=0.737786 time=264.9s + ttt_chunk [971/1893] bpb=0.736849 time=267.7s + ttt_chunk [981/1893] bpb=0.735666 time=270.4s + ttt_chunk [991/1893] bpb=0.734575 time=273.2s + ttt_chunk [1001/1893] bpb=0.733301 time=275.9s + ttt_chunk [1011/1893] bpb=0.731996 time=278.6s + ttt_chunk [1021/1893] bpb=0.730982 time=281.4s + ttt_chunk [1031/1893] bpb=0.729859 time=284.1s + ttt_chunk [1041/1893] bpb=0.728538 time=286.8s + ttt_chunk [1051/1893] bpb=0.727294 time=289.5s + ttt_chunk [1061/1893] bpb=0.726139 time=292.2s + ttt_chunk [1071/1893] bpb=0.725332 time=295.0s + ttt_chunk [1081/1893] bpb=0.724320 time=297.7s + ttt_chunk [1091/1893] bpb=0.723233 time=300.4s + ttt_chunk [1101/1893] bpb=0.722134 time=303.1s + ttt_chunk [1111/1893] bpb=0.720949 time=305.9s + ttt_chunk [1121/1893] bpb=0.719874 time=308.6s + ttt_chunk [1131/1893] bpb=0.718765 time=311.4s + ttt_chunk [1141/1893] bpb=0.717727 time=314.1s + ttt_chunk [1151/1893] bpb=0.716676 time=316.8s + ttt_chunk [1161/1893] bpb=0.715557 time=319.6s + ttt_chunk [1171/1893] bpb=0.714605 time=322.3s + ttt_chunk [1181/1893] bpb=0.713369 time=325.1s + ttt_chunk [1191/1893] bpb=0.712449 time=327.8s + ttt_chunk [1201/1893] bpb=0.711493 time=330.5s + ttt_chunk [1211/1893] bpb=0.710415 time=333.2s + ttt_chunk [1221/1893] bpb=0.709443 time=335.9s + ttt_chunk [1231/1893] bpb=0.708347 time=338.6s + ttt_chunk [1241/1893] bpb=0.707272 time=341.3s + ttt_chunk [1251/1893] bpb=0.706232 time=344.0s + ttt_chunk [1261/1893] bpb=0.705405 time=346.7s + ttt_chunk [1271/1893] bpb=0.704491 time=349.4s + ttt_chunk [1281/1893] bpb=0.703530 time=352.1s + ttt_chunk [1291/1893] bpb=0.702692 time=354.9s + ttt_chunk [1301/1893] bpb=0.701693 time=357.6s + ttt_chunk [1311/1893] bpb=0.700737 time=360.3s + ttt_chunk [1321/1893] bpb=0.699852 time=363.0s + ttt_chunk [1331/1893] bpb=0.699057 time=365.7s + ttt_chunk [1341/1893] bpb=0.698271 time=368.5s + ttt_chunk [1351/1893] bpb=0.697549 time=371.2s + ttt_chunk [1361/1893] bpb=0.696907 time=374.0s + ttt_chunk [1371/1893] bpb=0.696203 time=376.7s + ttt_chunk [1381/1893] bpb=0.695600 time=379.4s + ttt_chunk [1391/1893] bpb=0.694765 time=382.1s + ttt_chunk [1401/1893] bpb=0.694144 time=384.8s + ttt_chunk [1411/1893] bpb=0.693639 time=387.6s + ttt_chunk [1421/1893] bpb=0.693082 time=390.2s + ttt_chunk [1431/1893] bpb=0.692385 time=392.9s + ttt_chunk [1441/1893] bpb=0.691919 time=395.6s + ttt_chunk [1451/1893] bpb=0.691463 time=398.3s + ttt_chunk [1461/1893] bpb=0.690770 time=401.0s + ttt_chunk [1471/1893] bpb=0.690396 time=403.7s + ttt_chunk [1481/1893] bpb=0.689644 time=406.5s + ttt_chunk [1491/1893] bpb=0.689002 time=409.2s + ttt_chunk [1501/1893] bpb=0.688478 time=411.9s + ttt_chunk [1511/1893] bpb=0.687889 time=414.6s + ttt_chunk [1521/1893] bpb=0.687313 time=417.4s + ttt_chunk [1531/1893] bpb=0.686674 time=420.1s + ttt_chunk [1541/1893] bpb=0.686027 time=422.9s + ttt_chunk [1551/1893] bpb=0.685582 time=425.6s + ttt_chunk [1561/1893] bpb=0.685085 time=428.4s + ttt_chunk [1571/1893] bpb=0.684495 time=431.2s + ttt_chunk [1581/1893] bpb=0.684014 time=433.9s + ttt_chunk [1591/1893] bpb=0.683448 time=436.6s + ttt_chunk [1601/1893] bpb=0.683014 time=439.3s + ttt_chunk [1611/1893] bpb=0.682495 time=442.0s + ttt_chunk [1621/1893] bpb=0.681863 time=444.7s + ttt_chunk [1631/1893] bpb=0.681385 time=447.4s + ttt_chunk [1641/1893] bpb=0.680878 time=450.1s + ttt_chunk [1651/1893] bpb=0.680342 time=452.8s + ttt_chunk [1661/1893] bpb=0.679802 time=455.6s + ttt_chunk [1671/1893] bpb=0.679448 time=458.3s + ttt_chunk [1681/1893] bpb=0.678965 time=461.0s + ttt_chunk [1691/1893] bpb=0.678353 time=463.8s + ttt_chunk [1701/1893] bpb=0.677853 time=466.5s + ttt_chunk [1711/1893] bpb=0.677330 time=469.2s + ttt_chunk [1721/1893] bpb=0.676821 time=471.9s + ttt_chunk [1731/1893] bpb=0.676327 time=474.7s + ttt_chunk [1741/1893] bpb=0.675840 time=477.5s + ttt_chunk [1751/1893] bpb=0.675261 time=480.2s + ttt_chunk [1761/1893] bpb=0.674879 time=482.9s + ttt_chunk [1771/1893] bpb=0.674395 time=485.6s + ttt_chunk [1781/1893] bpb=0.674019 time=488.4s + ttt_chunk [1791/1893] bpb=0.673388 time=491.1s + ttt_chunk [1801/1893] bpb=0.672912 time=493.9s + ttt_chunk [1811/1893] bpb=0.672451 time=496.6s + ttt_chunk [1821/1893] bpb=0.671982 time=499.3s + ttt_chunk [1831/1893] bpb=0.671364 time=502.0s + ttt_chunk [1841/1893] bpb=0.670810 time=504.7s + ttt_chunk [1851/1893] bpb=0.670319 time=507.4s + ttt_chunk [1861/1893] bpb=0.669743 time=510.1s + ttt_chunk [1871/1893] bpb=0.669323 time=512.9s + ttt_chunk [1881/1893] bpb=0.668792 time=515.6s + ttt_chunk [1891/1893] bpb=0.668302 time=518.4s + ttt_chunk [1893/1893] bpb=0.668258 time=518.8s +ttt:done val_loss=1.126632 val_bpb=0.667255 elapsed=518.8s +final_int6_ttt val_loss:1.1266 val_bpb:0.6673 stride:64 eval_time:519433ms +final_int6_ttt_exact val_loss:1.12663163 val_bpb:0.66725471 diff --git a/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed2024.log b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed2024.log new file mode 100644 index 000000000..18ed097a1 --- /dev/null +++ b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed2024.log @@ -0,0 +1,297 @@ +W0326 04:47:03.829000 325693 torch/distributed/run.py:803] +W0326 04:47:03.829000 325693 torch/distributed/run.py:803] ***************************************** +W0326 04:47:03.829000 325693 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 04:47:03.829000 325693 torch/distributed/run.py:803] ***************************************** +logs/43c75892-bc70-4bb2-8a05-4b03e2930df5.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 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:2024 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9299 val_bpb:4.1043 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9321 train_time:149ms step_avg:148.70ms +step:2/20000 train_loss:8.7256 train_time:243ms step_avg:121.70ms +step:3/20000 train_loss:7.7945 train_time:343ms step_avg:114.50ms +step:4/20000 train_loss:7.3072 train_time:443ms step_avg:110.87ms +step:5/20000 train_loss:7.0587 train_time:544ms step_avg:108.77ms +step:6/20000 train_loss:6.8772 train_time:644ms step_avg:107.37ms +step:7/20000 train_loss:6.8213 train_time:745ms step_avg:106.36ms +step:8/20000 train_loss:6.6930 train_time:845ms step_avg:105.58ms +step:9/20000 train_loss:6.3713 train_time:945ms step_avg:104.99ms +step:10/20000 train_loss:5.9868 train_time:1046ms step_avg:104.55ms +step:500/20000 train_loss:2.3567 train_time:48852ms step_avg:97.70ms +step:1000/20000 train_loss:2.2412 train_time:97779ms step_avg:97.78ms +step:1500/20000 train_loss:2.1855 train_time:146756ms step_avg:97.84ms +step:2000/20000 train_loss:2.0300 train_time:195812ms step_avg:97.91ms +step:2500/20000 train_loss:2.1327 train_time:244875ms step_avg:97.95ms +step:3000/20000 train_loss:2.1105 train_time:293945ms step_avg:97.98ms +step:3500/20000 train_loss:2.1201 train_time:343019ms step_avg:98.01ms +step:4000/20000 train_loss:1.9097 train_time:392074ms step_avg:98.02ms +step:4000/20000 val_loss:1.9984 val_bpb:1.1836 train_time:392086ms step_avg:98.02ms +late_qat:enabled step:4188 scale:0.4999 +step:4500/20000 train_loss:2.0556 train_time:442133ms step_avg:98.25ms +step:5000/20000 train_loss:2.0280 train_time:492663ms step_avg:98.53ms +swa:start step:5250 +step:5500/20000 train_loss:1.9376 train_time:543449ms step_avg:98.81ms +step:5878/20000 val_loss:1.9034 val_bpb:1.1273 train_time:581994ms step_avg:99.01ms +stopping_early: wallclock_cap train_time:581994ms step:5878/20000 +peak memory allocated: 26197 MiB reserved: 26810 MiB +ema:applying EMA weights (skipping diagnostic evals) +Serialized model: 130432585 bytes +Code size: 87336 bytes +pruning:3.0% magnitude pruning applied +Serialized model int6+zlib: 16871584 bytes +Total submission size int6+zlib: 16958920 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.9231 val_bpb:1.1390 stride:64 eval_time:84852ms +final_int6_sliding_window_exact val_loss:1.92308795 val_bpb:1.13896101 +TTT: epochs=3 lr=0.0005 freeze_first=2 chunk=32768 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=1893 chunk_tokens=32768 windows=969057 stride=64 lr=0.0005 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=16 start_train... + ttt_train [1] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=2.3353 + ttt_train [1] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=2.2590 + ttt_train [1] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=2.2303 + ttt_chunk [1/1893] bpb=1.173197 time=0.4s + ttt_train [2] seqs=16 start_train... + ttt_train [2] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=2.7237 + ttt_train [2] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=2.7083 + ttt_train [2] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=2.6898 + ttt_chunk [2/1893] bpb=1.267828 time=0.6s + ttt_train [3] seqs=16 start_train... + ttt_train [3] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=1.8670 + ttt_train [3] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=1.8575 + ttt_train [3] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=1.8440 + ttt_chunk [3/1893] bpb=1.223267 time=0.9s + ttt_chunk [4/1893] bpb=1.223600 time=1.2s + ttt_chunk [5/1893] bpb=1.216356 time=1.5s + ttt_chunk [11/1893] bpb=1.197371 time=3.1s + ttt_chunk [21/1893] bpb=1.177215 time=5.9s + ttt_chunk [31/1893] bpb=1.167302 time=8.7s + ttt_chunk [41/1893] bpb=1.145790 time=11.6s + ttt_chunk [51/1893] bpb=1.132462 time=14.3s + ttt_chunk [61/1893] bpb=1.129251 time=17.1s + ttt_chunk [71/1893] bpb=1.118372 time=19.8s + ttt_chunk [81/1893] bpb=1.108461 time=22.6s + ttt_chunk [91/1893] bpb=1.100569 time=25.3s + ttt_chunk [101/1893] bpb=1.094122 time=28.0s + ttt_chunk [111/1893] bpb=1.087097 time=30.8s + ttt_chunk [121/1893] bpb=1.072642 time=33.5s + ttt_chunk [131/1893] bpb=1.063388 time=36.1s + ttt_chunk [141/1893] bpb=1.059660 time=38.8s + ttt_chunk [151/1893] bpb=1.051953 time=41.5s + ttt_chunk [161/1893] bpb=1.043562 time=44.2s + ttt_chunk [171/1893] bpb=1.037549 time=47.0s + ttt_chunk [181/1893] bpb=1.031082 time=49.8s + ttt_chunk [191/1893] bpb=1.027287 time=52.6s + ttt_chunk [201/1893] bpb=1.018693 time=55.3s + ttt_chunk [211/1893] bpb=1.009177 time=58.0s + ttt_chunk [221/1893] bpb=1.001921 time=60.6s + ttt_chunk [231/1893] bpb=0.993041 time=63.3s + ttt_chunk [241/1893] bpb=0.985591 time=66.0s + ttt_chunk [251/1893] bpb=0.978421 time=68.7s + ttt_chunk [261/1893] bpb=0.968962 time=71.4s + ttt_chunk [271/1893] bpb=0.960982 time=74.2s + ttt_chunk [281/1893] bpb=0.954606 time=76.9s + ttt_chunk [291/1893] bpb=0.948776 time=79.6s + ttt_chunk [301/1893] bpb=0.942204 time=82.3s + ttt_chunk [311/1893] bpb=0.936766 time=85.0s + ttt_chunk [321/1893] bpb=0.931269 time=87.7s + ttt_chunk [331/1893] bpb=0.925284 time=90.4s + ttt_chunk [341/1893] bpb=0.918380 time=93.1s + ttt_chunk [351/1893] bpb=0.913464 time=95.9s + ttt_chunk [361/1893] bpb=0.908245 time=98.6s + ttt_chunk [371/1893] bpb=0.902112 time=101.3s + ttt_chunk [381/1893] bpb=0.897032 time=104.0s + ttt_chunk [391/1893] bpb=0.891834 time=106.6s + ttt_chunk [401/1893] bpb=0.885877 time=109.3s + ttt_chunk [411/1893] bpb=0.880714 time=112.1s + ttt_chunk [421/1893] bpb=0.875391 time=114.8s + ttt_chunk [431/1893] bpb=0.870510 time=117.5s + ttt_chunk [441/1893] bpb=0.866121 time=120.2s + ttt_chunk [451/1893] bpb=0.861497 time=122.9s + ttt_chunk [461/1893] bpb=0.856588 time=125.6s + ttt_chunk [471/1893] bpb=0.852102 time=128.3s + ttt_chunk [481/1893] bpb=0.848040 time=131.0s + ttt_chunk [491/1893] bpb=0.843400 time=133.7s + ttt_chunk [501/1893] bpb=0.839506 time=136.3s + ttt_chunk [511/1893] bpb=0.835765 time=139.0s + ttt_chunk [521/1893] bpb=0.831284 time=141.7s + ttt_chunk [531/1893] bpb=0.827981 time=144.3s + ttt_chunk [541/1893] bpb=0.824743 time=147.0s + ttt_chunk [551/1893] bpb=0.820973 time=149.6s + ttt_chunk [561/1893] bpb=0.817762 time=152.3s + ttt_chunk [571/1893] bpb=0.814217 time=155.0s + ttt_chunk [581/1893] bpb=0.810845 time=157.7s + ttt_chunk [591/1893] bpb=0.807622 time=160.5s + ttt_chunk [601/1893] bpb=0.804788 time=163.2s + ttt_chunk [611/1893] bpb=0.801936 time=165.9s + ttt_chunk [621/1893] bpb=0.799140 time=168.6s + ttt_chunk [631/1893] bpb=0.796641 time=171.4s + ttt_chunk [641/1893] bpb=0.794047 time=174.1s + ttt_chunk [651/1893] bpb=0.791279 time=176.7s + ttt_chunk [661/1893] bpb=0.788684 time=179.4s + ttt_chunk [671/1893] bpb=0.786389 time=182.0s + ttt_chunk [681/1893] bpb=0.783930 time=184.7s + ttt_chunk [691/1893] bpb=0.782045 time=187.3s + ttt_chunk [701/1893] bpb=0.779580 time=190.0s + ttt_chunk [711/1893] bpb=0.777541 time=192.7s + ttt_chunk [721/1893] bpb=0.775382 time=195.4s + ttt_chunk [731/1893] bpb=0.773392 time=198.0s + ttt_chunk [741/1893] bpb=0.771354 time=200.7s + ttt_chunk [751/1893] bpb=0.769304 time=203.3s + ttt_chunk [761/1893] bpb=0.767413 time=206.0s + ttt_chunk [771/1893] bpb=0.765537 time=208.7s + ttt_chunk [781/1893] bpb=0.764172 time=211.5s + ttt_chunk [791/1893] bpb=0.762210 time=214.2s + ttt_chunk [801/1893] bpb=0.760368 time=216.9s + ttt_chunk [811/1893] bpb=0.758646 time=219.6s + ttt_chunk [821/1893] bpb=0.756864 time=222.3s + ttt_chunk [831/1893] bpb=0.755341 time=225.0s + ttt_chunk [841/1893] bpb=0.753518 time=227.7s + ttt_chunk [851/1893] bpb=0.751938 time=230.4s + ttt_chunk [861/1893] bpb=0.750328 time=233.0s + ttt_chunk [871/1893] bpb=0.748877 time=235.7s + ttt_chunk [881/1893] bpb=0.747490 time=238.3s + ttt_chunk [891/1893] bpb=0.746076 time=241.0s + ttt_chunk [901/1893] bpb=0.744829 time=243.6s + ttt_chunk [911/1893] bpb=0.743554 time=246.3s + ttt_chunk [921/1893] bpb=0.742329 time=249.0s + ttt_chunk [931/1893] bpb=0.741010 time=251.6s + ttt_chunk [941/1893] bpb=0.739638 time=254.3s + ttt_chunk [951/1893] bpb=0.738476 time=257.0s + ttt_chunk [961/1893] bpb=0.737134 time=259.8s + ttt_chunk [971/1893] bpb=0.736163 time=262.5s + ttt_chunk [981/1893] bpb=0.734976 time=265.2s + ttt_chunk [991/1893] bpb=0.733880 time=267.9s + ttt_chunk [1001/1893] bpb=0.732616 time=270.6s + ttt_chunk [1011/1893] bpb=0.731309 time=273.3s + ttt_chunk [1021/1893] bpb=0.730303 time=276.0s + ttt_chunk [1031/1893] bpb=0.729176 time=278.7s + ttt_chunk [1041/1893] bpb=0.727856 time=281.4s + ttt_chunk [1051/1893] bpb=0.726619 time=284.0s + ttt_chunk [1061/1893] bpb=0.725480 time=286.7s + ttt_chunk [1071/1893] bpb=0.724679 time=289.4s + ttt_chunk [1081/1893] bpb=0.723670 time=292.1s + ttt_chunk [1091/1893] bpb=0.722575 time=294.7s + ttt_chunk [1101/1893] bpb=0.721465 time=297.4s + ttt_chunk [1111/1893] bpb=0.720289 time=300.0s + ttt_chunk [1121/1893] bpb=0.719199 time=302.7s + ttt_chunk [1131/1893] bpb=0.718100 time=305.4s + ttt_chunk [1141/1893] bpb=0.717065 time=308.1s + ttt_chunk [1151/1893] bpb=0.716020 time=310.8s + ttt_chunk [1161/1893] bpb=0.714907 time=313.5s + ttt_chunk [1171/1893] bpb=0.713952 time=316.2s + ttt_chunk [1181/1893] bpb=0.712714 time=318.9s + ttt_chunk [1191/1893] bpb=0.711794 time=321.5s + ttt_chunk [1201/1893] bpb=0.710833 time=324.3s + ttt_chunk [1211/1893] bpb=0.709751 time=326.9s + ttt_chunk [1221/1893] bpb=0.708798 time=329.6s + ttt_chunk [1231/1893] bpb=0.707700 time=332.2s + ttt_chunk [1241/1893] bpb=0.706626 time=334.9s + ttt_chunk [1251/1893] bpb=0.705577 time=337.5s + ttt_chunk [1261/1893] bpb=0.704749 time=340.2s + ttt_chunk [1271/1893] bpb=0.703824 time=342.9s + ttt_chunk [1281/1893] bpb=0.702862 time=345.6s + ttt_chunk [1291/1893] bpb=0.702034 time=348.3s + ttt_chunk [1301/1893] bpb=0.701027 time=351.0s + ttt_chunk [1311/1893] bpb=0.700064 time=353.7s + ttt_chunk [1321/1893] bpb=0.699170 time=356.4s + ttt_chunk [1331/1893] bpb=0.698374 time=359.1s + ttt_chunk [1341/1893] bpb=0.697572 time=361.7s + ttt_chunk [1351/1893] bpb=0.696849 time=364.4s + ttt_chunk [1361/1893] bpb=0.696209 time=367.0s + ttt_chunk [1371/1893] bpb=0.695508 time=369.7s + ttt_chunk [1381/1893] bpb=0.694896 time=372.4s + ttt_chunk [1391/1893] bpb=0.694051 time=375.0s + ttt_chunk [1401/1893] bpb=0.693445 time=377.7s + ttt_chunk [1411/1893] bpb=0.692957 time=380.3s + ttt_chunk [1421/1893] bpb=0.692404 time=383.0s + ttt_chunk [1431/1893] bpb=0.691717 time=385.6s + ttt_chunk [1441/1893] bpb=0.691262 time=388.3s + ttt_chunk [1451/1893] bpb=0.690808 time=390.9s + ttt_chunk [1461/1893] bpb=0.690113 time=393.6s + ttt_chunk [1471/1893] bpb=0.689758 time=396.3s + ttt_chunk [1481/1893] bpb=0.689024 time=399.0s + ttt_chunk [1491/1893] bpb=0.688382 time=401.7s + ttt_chunk [1501/1893] bpb=0.687854 time=404.4s + ttt_chunk [1511/1893] bpb=0.687261 time=407.1s + ttt_chunk [1521/1893] bpb=0.686680 time=409.7s + ttt_chunk [1531/1893] bpb=0.686042 time=412.4s + ttt_chunk [1541/1893] bpb=0.685407 time=415.0s + ttt_chunk [1551/1893] bpb=0.684959 time=417.7s + ttt_chunk [1561/1893] bpb=0.684463 time=420.3s + ttt_chunk [1571/1893] bpb=0.683874 time=423.0s + ttt_chunk [1581/1893] bpb=0.683403 time=425.6s + ttt_chunk [1591/1893] bpb=0.682836 time=428.3s + ttt_chunk [1601/1893] bpb=0.682401 time=430.9s + ttt_chunk [1611/1893] bpb=0.681889 time=433.6s + ttt_chunk [1621/1893] bpb=0.681263 time=436.2s + ttt_chunk [1631/1893] bpb=0.680787 time=438.9s + ttt_chunk [1641/1893] bpb=0.680284 time=441.6s + ttt_chunk [1651/1893] bpb=0.679750 time=444.3s + ttt_chunk [1661/1893] bpb=0.679214 time=447.0s + ttt_chunk [1671/1893] bpb=0.678853 time=449.7s + ttt_chunk [1681/1893] bpb=0.678353 time=452.4s + ttt_chunk [1691/1893] bpb=0.677733 time=455.1s + ttt_chunk [1701/1893] bpb=0.677248 time=457.7s + ttt_chunk [1711/1893] bpb=0.676727 time=460.4s + ttt_chunk [1721/1893] bpb=0.676221 time=463.0s + ttt_chunk [1731/1893] bpb=0.675736 time=465.7s + ttt_chunk [1741/1893] bpb=0.675252 time=468.4s + ttt_chunk [1751/1893] bpb=0.674672 time=471.0s + ttt_chunk [1761/1893] bpb=0.674285 time=473.7s + ttt_chunk [1771/1893] bpb=0.673795 time=476.3s + ttt_chunk [1781/1893] bpb=0.673424 time=479.0s + ttt_chunk [1791/1893] bpb=0.672801 time=481.6s + ttt_chunk [1801/1893] bpb=0.672334 time=484.3s + ttt_chunk [1811/1893] bpb=0.671871 time=487.0s + ttt_chunk [1821/1893] bpb=0.671401 time=489.7s + ttt_chunk [1831/1893] bpb=0.670787 time=492.4s + ttt_chunk [1841/1893] bpb=0.670238 time=495.1s + ttt_chunk [1851/1893] bpb=0.669737 time=497.7s + ttt_chunk [1861/1893] bpb=0.669159 time=500.4s + ttt_chunk [1871/1893] bpb=0.668745 time=503.1s + ttt_chunk [1881/1893] bpb=0.668216 time=505.8s + ttt_chunk [1891/1893] bpb=0.667724 time=508.4s + ttt_chunk [1893/1893] bpb=0.667681 time=508.8s +ttt:done val_loss=1.125764 val_bpb=0.666741 elapsed=508.8s +final_int6_ttt val_loss:1.1258 val_bpb:0.6667 stride:64 eval_time:509364ms +final_int6_ttt_exact val_loss:1.12576387 val_bpb:0.66674078 diff --git a/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed42.log b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed42.log new file mode 100644 index 000000000..f6f9f0519 --- /dev/null +++ b/records/track_10min_16mb/2026-03-26_BackoffNgramMixer/train_seed42.log @@ -0,0 +1,297 @@ +W0326 04:05:22.064000 323983 torch/distributed/run.py:803] +W0326 04:05:22.064000 323983 torch/distributed/run.py:803] ***************************************** +W0326 04:05:22.064000 323983 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 04:05:22.064000 323983 torch/distributed/run.py:803] ***************************************** +logs/856a360b-07d8-4ec8-b9e1-c8ae88bd6ee7.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 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33317980 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9309 train_time:143ms step_avg:142.88ms +step:2/20000 train_loss:8.7072 train_time:240ms step_avg:120.10ms +step:3/20000 train_loss:7.7701 train_time:340ms step_avg:113.47ms +step:4/20000 train_loss:7.2998 train_time:440ms step_avg:110.08ms +step:5/20000 train_loss:7.0351 train_time:540ms step_avg:108.07ms +step:6/20000 train_loss:6.9496 train_time:640ms step_avg:106.75ms +step:7/20000 train_loss:6.8531 train_time:741ms step_avg:105.80ms +step:8/20000 train_loss:6.7345 train_time:840ms step_avg:105.03ms +step:9/20000 train_loss:6.3556 train_time:940ms step_avg:104.46ms +step:10/20000 train_loss:6.0343 train_time:1041ms step_avg:104.06ms +step:500/20000 train_loss:2.3589 train_time:48779ms step_avg:97.56ms +step:1000/20000 train_loss:2.2410 train_time:97660ms step_avg:97.66ms +step:1500/20000 train_loss:2.1871 train_time:146627ms step_avg:97.75ms +step:2000/20000 train_loss:2.0315 train_time:195624ms step_avg:97.81ms +step:2500/20000 train_loss:2.1331 train_time:244634ms step_avg:97.85ms +step:3000/20000 train_loss:2.1146 train_time:293654ms step_avg:97.88ms +step:3500/20000 train_loss:2.1201 train_time:342666ms step_avg:97.90ms +step:4000/20000 train_loss:1.9082 train_time:391666ms step_avg:97.92ms +step:4000/20000 val_loss:1.9995 val_bpb:1.1842 train_time:391678ms step_avg:97.92ms +late_qat:enabled step:4195 scale:0.4995 +step:4500/20000 train_loss:2.0528 train_time:441662ms step_avg:98.15ms +step:5000/20000 train_loss:2.0284 train_time:492078ms step_avg:98.42ms +swa:start step:5250 +step:5500/20000 train_loss:1.9370 train_time:542803ms step_avg:98.69ms +step:5886/20000 val_loss:1.9040 val_bpb:1.1276 train_time:582087ms step_avg:98.89ms +stopping_early: wallclock_cap train_time:582087ms step:5886/20000 +peak memory allocated: 26197 MiB reserved: 26810 MiB +ema:applying EMA weights (skipping diagnostic evals) +Serialized model: 130432585 bytes +Code size: 87336 bytes +pruning:3.0% magnitude pruning applied +Serialized model int6+zlib: 16869201 bytes +Total submission size int6+zlib: 16956537 bytes + ttt: pre-compiling forward+backward kernels... + ttt: pre-compile done +final_int6_sliding_window val_loss:1.9188 val_bpb:1.1364 stride:64 eval_time:84606ms +final_int6_sliding_window_exact val_loss:1.91880344 val_bpb:1.13642348 +TTT: epochs=3 lr=0.0005 freeze_first=2 chunk=32768 opt=adamw +TTT temperature: 0.98 +PPM alpha: 0.85, Byte-weighted TTT: True + Logistic context mixer enabled: eta=0.1 + Adaptive LR enabled: max_mult=3.0 +ttt:start chunks=1893 chunk_tokens=32768 windows=969057 stride=64 lr=0.0005 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27537480 + Polyak averaging enabled: decay=0.998 + ttt_train [1] seqs=16 start_train... + ttt_train [1] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=2.3501 + ttt_train [1] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=2.2730 + ttt_train [1] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=2.2465 + ttt_chunk [1/1893] bpb=1.179795 time=0.3s + ttt_train [2] seqs=16 start_train... + ttt_train [2] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=2.7310 + ttt_train [2] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=2.7180 + ttt_train [2] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=2.7016 + ttt_chunk [2/1893] bpb=1.270454 time=0.6s + ttt_train [3] seqs=16 start_train... + ttt_train [3] epoch=1/3 batches=2 ... + step done ep=1 bs=0 loss=1.8837 + ttt_train [3] epoch=2/3 batches=2 ... + step done ep=2 bs=0 loss=1.8739 + ttt_train [3] epoch=3/3 batches=2 ... + step done ep=3 bs=0 loss=1.8579 + ttt_chunk [3/1893] bpb=1.224386 time=0.9s + ttt_chunk [4/1893] bpb=1.224589 time=1.1s + ttt_chunk [5/1893] bpb=1.217872 time=1.4s + ttt_chunk [11/1893] bpb=1.191312 time=3.0s + ttt_chunk [21/1893] bpb=1.170163 time=5.7s + ttt_chunk [31/1893] bpb=1.161655 time=8.5s + ttt_chunk [41/1893] bpb=1.140674 time=11.2s + ttt_chunk [51/1893] bpb=1.127605 time=14.0s + ttt_chunk [61/1893] bpb=1.124751 time=16.7s + ttt_chunk [71/1893] bpb=1.114339 time=19.3s + ttt_chunk [81/1893] bpb=1.104786 time=22.0s + ttt_chunk [91/1893] bpb=1.097411 time=24.7s + ttt_chunk [101/1893] bpb=1.091109 time=27.5s + ttt_chunk [111/1893] bpb=1.084362 time=30.2s + ttt_chunk [121/1893] bpb=1.070085 time=32.9s + ttt_chunk [131/1893] bpb=1.060968 time=35.6s + ttt_chunk [141/1893] bpb=1.057379 time=38.3s + ttt_chunk [151/1893] bpb=1.049839 time=41.0s + ttt_chunk [161/1893] bpb=1.041642 time=43.6s + ttt_chunk [171/1893] bpb=1.035775 time=46.3s + ttt_chunk [181/1893] bpb=1.029386 time=49.1s + ttt_chunk [191/1893] bpb=1.025760 time=51.8s + ttt_chunk [201/1893] bpb=1.017185 time=54.5s + ttt_chunk [211/1893] bpb=1.007817 time=57.1s + ttt_chunk [221/1893] bpb=1.000739 time=59.8s + ttt_chunk [231/1893] bpb=0.991998 time=62.5s + ttt_chunk [241/1893] bpb=0.984681 time=65.3s + ttt_chunk [251/1893] bpb=0.977611 time=67.9s + ttt_chunk [261/1893] bpb=0.968294 time=70.7s + ttt_chunk [271/1893] bpb=0.960490 time=73.3s + ttt_chunk [281/1893] bpb=0.954080 time=76.0s + ttt_chunk [291/1893] bpb=0.948315 time=78.7s + ttt_chunk [301/1893] bpb=0.941725 time=81.4s + ttt_chunk [311/1893] bpb=0.936387 time=84.1s + ttt_chunk [321/1893] bpb=0.930945 time=86.8s + ttt_chunk [331/1893] bpb=0.925014 time=89.4s + ttt_chunk [341/1893] bpb=0.918131 time=92.1s + ttt_chunk [351/1893] bpb=0.913238 time=94.8s + ttt_chunk [361/1893] bpb=0.908072 time=97.5s + ttt_chunk [371/1893] bpb=0.901982 time=100.2s + ttt_chunk [381/1893] bpb=0.896930 time=102.8s + ttt_chunk [391/1893] bpb=0.891759 time=105.5s + ttt_chunk [401/1893] bpb=0.885872 time=108.2s + ttt_chunk [411/1893] bpb=0.880690 time=110.9s + ttt_chunk [421/1893] bpb=0.875423 time=113.6s + ttt_chunk [431/1893] bpb=0.870574 time=116.3s + ttt_chunk [441/1893] bpb=0.866261 time=119.0s + ttt_chunk [451/1893] bpb=0.861674 time=121.7s + ttt_chunk [461/1893] bpb=0.856794 time=124.4s + ttt_chunk [471/1893] bpb=0.852342 time=127.1s + ttt_chunk [481/1893] bpb=0.848318 time=129.8s + ttt_chunk [491/1893] bpb=0.843687 time=132.4s + ttt_chunk [501/1893] bpb=0.839866 time=135.1s + ttt_chunk [511/1893] bpb=0.836115 time=137.7s + ttt_chunk [521/1893] bpb=0.831606 time=140.4s + ttt_chunk [531/1893] bpb=0.828333 time=143.0s + ttt_chunk [541/1893] bpb=0.825114 time=145.7s + ttt_chunk [551/1893] bpb=0.821330 time=148.3s + ttt_chunk [561/1893] bpb=0.818133 time=151.0s + ttt_chunk [571/1893] bpb=0.814590 time=153.7s + ttt_chunk [581/1893] bpb=0.811219 time=156.3s + ttt_chunk [591/1893] bpb=0.807998 time=159.0s + ttt_chunk [601/1893] bpb=0.805170 time=161.7s + ttt_chunk [611/1893] bpb=0.802330 time=164.4s + ttt_chunk [621/1893] bpb=0.799537 time=167.1s + ttt_chunk [631/1893] bpb=0.797006 time=169.8s + ttt_chunk [641/1893] bpb=0.794436 time=172.5s + ttt_chunk [651/1893] bpb=0.791656 time=175.2s + ttt_chunk [661/1893] bpb=0.789072 time=177.9s + ttt_chunk [671/1893] bpb=0.786799 time=180.5s + ttt_chunk [681/1893] bpb=0.784302 time=183.2s + ttt_chunk [691/1893] bpb=0.782427 time=185.8s + ttt_chunk [701/1893] bpb=0.779977 time=188.4s + ttt_chunk [711/1893] bpb=0.777932 time=191.0s + ttt_chunk [721/1893] bpb=0.775760 time=193.7s + ttt_chunk [731/1893] bpb=0.773794 time=196.3s + ttt_chunk [741/1893] bpb=0.771760 time=199.0s + ttt_chunk [751/1893] bpb=0.769709 time=201.6s + ttt_chunk [761/1893] bpb=0.767815 time=204.2s + ttt_chunk [771/1893] bpb=0.765959 time=206.9s + ttt_chunk [781/1893] bpb=0.764579 time=209.5s + ttt_chunk [791/1893] bpb=0.762659 time=212.2s + ttt_chunk [801/1893] bpb=0.760842 time=214.9s + ttt_chunk [811/1893] bpb=0.759123 time=217.6s + ttt_chunk [821/1893] bpb=0.757338 time=220.2s + ttt_chunk [831/1893] bpb=0.755830 time=222.9s + ttt_chunk [841/1893] bpb=0.754009 time=225.6s + ttt_chunk [851/1893] bpb=0.752428 time=228.3s + ttt_chunk [861/1893] bpb=0.750819 time=230.9s + ttt_chunk [871/1893] bpb=0.749363 time=233.6s + ttt_chunk [881/1893] bpb=0.747982 time=236.3s + ttt_chunk [891/1893] bpb=0.746565 time=238.9s + ttt_chunk [901/1893] bpb=0.745314 time=241.5s + ttt_chunk [911/1893] bpb=0.744031 time=244.2s + ttt_chunk [921/1893] bpb=0.742792 time=246.8s + ttt_chunk [931/1893] bpb=0.741465 time=249.4s + ttt_chunk [941/1893] bpb=0.740081 time=252.1s + ttt_chunk [951/1893] bpb=0.738928 time=254.7s + ttt_chunk [961/1893] bpb=0.737583 time=257.4s + ttt_chunk [971/1893] bpb=0.736655 time=260.1s + ttt_chunk [981/1893] bpb=0.735471 time=262.8s + ttt_chunk [991/1893] bpb=0.734383 time=265.4s + ttt_chunk [1001/1893] bpb=0.733113 time=268.1s + ttt_chunk [1011/1893] bpb=0.731816 time=270.8s + ttt_chunk [1021/1893] bpb=0.730803 time=273.5s + ttt_chunk [1031/1893] bpb=0.729684 time=276.1s + ttt_chunk [1041/1893] bpb=0.728376 time=278.8s + ttt_chunk [1051/1893] bpb=0.727136 time=281.4s + ttt_chunk [1061/1893] bpb=0.725994 time=284.1s + ttt_chunk [1071/1893] bpb=0.725199 time=286.7s + ttt_chunk [1081/1893] bpb=0.724195 time=289.4s + ttt_chunk [1091/1893] bpb=0.723097 time=292.1s + ttt_chunk [1101/1893] bpb=0.721992 time=294.7s + ttt_chunk [1111/1893] bpb=0.720811 time=297.4s + ttt_chunk [1121/1893] bpb=0.719745 time=300.0s + ttt_chunk [1131/1893] bpb=0.718663 time=302.7s + ttt_chunk [1141/1893] bpb=0.717635 time=305.4s + ttt_chunk [1151/1893] bpb=0.716588 time=308.1s + ttt_chunk [1161/1893] bpb=0.715479 time=310.7s + ttt_chunk [1171/1893] bpb=0.714525 time=313.4s + ttt_chunk [1181/1893] bpb=0.713301 time=316.1s + ttt_chunk [1191/1893] bpb=0.712387 time=318.8s + ttt_chunk [1201/1893] bpb=0.711440 time=321.4s + ttt_chunk [1211/1893] bpb=0.710371 time=324.0s + ttt_chunk [1221/1893] bpb=0.709408 time=326.7s + ttt_chunk [1231/1893] bpb=0.708305 time=329.4s + ttt_chunk [1241/1893] bpb=0.707231 time=332.0s + ttt_chunk [1251/1893] bpb=0.706190 time=334.7s + ttt_chunk [1261/1893] bpb=0.705366 time=337.3s + ttt_chunk [1271/1893] bpb=0.704445 time=340.0s + ttt_chunk [1281/1893] bpb=0.703466 time=342.6s + ttt_chunk [1291/1893] bpb=0.702620 time=345.3s + ttt_chunk [1301/1893] bpb=0.701629 time=347.9s + ttt_chunk [1311/1893] bpb=0.700671 time=350.6s + ttt_chunk [1321/1893] bpb=0.699786 time=353.3s + ttt_chunk [1331/1893] bpb=0.698992 time=355.9s + ttt_chunk [1341/1893] bpb=0.698200 time=358.7s + ttt_chunk [1351/1893] bpb=0.697486 time=361.3s + ttt_chunk [1361/1893] bpb=0.696845 time=364.0s + ttt_chunk [1371/1893] bpb=0.696142 time=366.7s + ttt_chunk [1381/1893] bpb=0.695540 time=369.4s + ttt_chunk [1391/1893] bpb=0.694691 time=372.0s + ttt_chunk [1401/1893] bpb=0.694078 time=374.7s + ttt_chunk [1411/1893] bpb=0.693581 time=377.3s + ttt_chunk [1421/1893] bpb=0.693024 time=379.9s + ttt_chunk [1431/1893] bpb=0.692322 time=382.6s + ttt_chunk [1441/1893] bpb=0.691856 time=385.3s + ttt_chunk [1451/1893] bpb=0.691394 time=387.9s + ttt_chunk [1461/1893] bpb=0.690705 time=390.6s + ttt_chunk [1471/1893] bpb=0.690329 time=393.2s + ttt_chunk [1481/1893] bpb=0.689576 time=395.9s + ttt_chunk [1491/1893] bpb=0.688936 time=398.5s + ttt_chunk [1501/1893] bpb=0.688410 time=401.2s + ttt_chunk [1511/1893] bpb=0.687828 time=403.9s + ttt_chunk [1521/1893] bpb=0.687248 time=406.6s + ttt_chunk [1531/1893] bpb=0.686622 time=409.3s + ttt_chunk [1541/1893] bpb=0.685983 time=412.0s + ttt_chunk [1551/1893] bpb=0.685540 time=414.7s + ttt_chunk [1561/1893] bpb=0.685041 time=417.4s + ttt_chunk [1571/1893] bpb=0.684455 time=420.0s + ttt_chunk [1581/1893] bpb=0.683989 time=422.6s + ttt_chunk [1591/1893] bpb=0.683423 time=425.3s + ttt_chunk [1601/1893] bpb=0.682989 time=428.0s + ttt_chunk [1611/1893] bpb=0.682472 time=430.6s + ttt_chunk [1621/1893] bpb=0.681849 time=433.3s + ttt_chunk [1631/1893] bpb=0.681381 time=435.9s + ttt_chunk [1641/1893] bpb=0.680882 time=438.6s + ttt_chunk [1651/1893] bpb=0.680350 time=441.2s + ttt_chunk [1661/1893] bpb=0.679806 time=443.9s + ttt_chunk [1671/1893] bpb=0.679447 time=446.6s + ttt_chunk [1681/1893] bpb=0.678956 time=449.2s + ttt_chunk [1691/1893] bpb=0.678339 time=451.9s + ttt_chunk [1701/1893] bpb=0.677844 time=454.6s + ttt_chunk [1711/1893] bpb=0.677318 time=457.3s + ttt_chunk [1721/1893] bpb=0.676801 time=460.0s + ttt_chunk [1731/1893] bpb=0.676305 time=462.6s + ttt_chunk [1741/1893] bpb=0.675831 time=465.3s + ttt_chunk [1751/1893] bpb=0.675249 time=468.0s + ttt_chunk [1761/1893] bpb=0.674868 time=470.7s + ttt_chunk [1771/1893] bpb=0.674382 time=473.4s + ttt_chunk [1781/1893] bpb=0.673999 time=476.0s + ttt_chunk [1791/1893] bpb=0.673374 time=478.6s + ttt_chunk [1801/1893] bpb=0.672899 time=481.3s + ttt_chunk [1811/1893] bpb=0.672434 time=484.0s + ttt_chunk [1821/1893] bpb=0.671960 time=486.6s + ttt_chunk [1831/1893] bpb=0.671348 time=489.3s + ttt_chunk [1841/1893] bpb=0.670790 time=492.0s + ttt_chunk [1851/1893] bpb=0.670286 time=494.6s + ttt_chunk [1861/1893] bpb=0.669710 time=497.3s + ttt_chunk [1871/1893] bpb=0.669287 time=500.0s + ttt_chunk [1881/1893] bpb=0.668761 time=502.6s + ttt_chunk [1891/1893] bpb=0.668266 time=505.3s + ttt_chunk [1893/1893] bpb=0.668225 time=505.7s +ttt:done val_loss=1.126612 val_bpb=0.667243 elapsed=505.7s +final_int6_ttt val_loss:1.1266 val_bpb:0.6672 stride:64 eval_time:506274ms +final_int6_ttt_exact val_loss:1.12661235 val_bpb:0.66724329