diff --git a/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/README.md b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/README.md new file mode 100644 index 0000000000..470899fee3 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/README.md @@ -0,0 +1,99 @@ +# Record: 1.0400 BPB — Hedge Mixer + VRL + AdamW TTT + Polyak EMA + +**3-seed mean: pending (seed 42 = 1.0400, seeds 1337 and 2024 in progress)** + +## Summary + +11-layer transformer (26.99M params) with Value Residual Learning (VRL), combined with a 5-expert Hedge Mixer during evaluation. The Hedge Mixer blends neural model predictions with online n-gram statistics (unigram, bigram, trigram) using the multiplicative-weights algorithm, achieving a 0.078 BPB improvement over the merged SOTA (1.1194). + +## Key Techniques + +### Architecture (Training) +- 11 layers, model_dim=512, GQA (8 query / 4 KV heads), MLP 3x expansion +- **Value Residual Learning (VRL):** Residual connection from layer 0's value output to all subsequent layers (~22 extra params, -0.01 BPB) +- **LeakyReLU(0.5)^2** activation +- **XSA-4:** Cross-Token Self-Attention on last 4 layers +- Tied embeddings, BigramHash (2048 buckets), SmearGate +- Soft-Round QAT (late activation at scale < 0.15) +- EMA weight averaging (decay=0.997) +- Muon optimizer (matrix params) + AdamW (embeddings/scalars) +- Int6 mixed quantization + lzma compression + +### Evaluation (Legal Score-First) + +**TTT (Test-Time Training):** +- AdamW optimizer (lr=0.0005, no weight decay) +- Polyak EMA (decay=0.998): train on raw weights, score with smoothed EMA weights +- Byte-weighted loss: tokens weighted by byte coverage +- Adaptive cosine LR: ramps from 1x to 3x over first 30% of chunks, then cosine decay +- Temperature scaling: T=0.98 on eval logits +- Freeze first 9 blocks, unfreeze last 2 + norms/scales/embeddings (~5.26M of 27M params) +- 3 epochs per chunk, 32K tokens per chunk + +**Hedge Mixer (5-expert online ensemble):** +- Expert 0: Neural model (temperature-scaled logits) +- Expert 1: Unigram frequency counts (Laplace-smoothed) +- Expert 2: Bigram P(next|prev) counts +- Expert 3: Trigram P(next|hash(prev2,prev1)) with 64K hashed buckets +- Expert 4: Neural entropy as confidence meta-signal +- Hedge algorithm (eta=0.1) learns expert weights online +- **Deferred weight updates:** Expert weights for chunk N computed from chunks 0..N-1 only. All windows in a chunk scored with identical weights. Updates happen AFTER chunk scoring is complete. +- N-gram tables updated AFTER chunk scoring (score-first, legal) +- Mixer activates after 10,000 tokens accumulated + +### Legality + +Every token's probability is committed using ONLY information from previously-scored chunks: +1. Hedge weights for chunk N reflect performance on chunks 0..N-1 (deferred update) +2. N-gram tables contain only tokens from chunks 0..N-1 (updated after scoring) +3. Polyak EMA uses fixed decay (0.998), no selection of best snapshot +4. TTT trains only on already-scored chunks (score-first) +5. No validation data accessed during training; no training data accessed during evaluation + +## Results + +| Metric | Value | +|--------|-------| +| **TTT + Hedge BPB** | **1.0400** | +| Sliding window BPB (no TTT) | 1.1263 | +| Int6 roundtrip BPB | 1.1499 | +| Post-EMA BPB | 1.1425 | +| Training steps | 6104 | +| Step avg | 98.36 ms | +| Artifact size | 15,999,919 bytes | +| Eval time (TTT+Hedge) | 486s / 600s budget | + +## Ablation (approximate, from development testing) + +| Component | BPB Impact | +|-----------|-----------| +| VRL | -0.010 | +| LeakyReLU(0.5)^2 | -0.001 | +| XSA-4 | -0.005 | +| TTT (SGD baseline) | -0.001 | +| AdamW + Polyak TTT | ~-0.003 vs SGD TTT | +| Hedge Mixer | ~-0.080 | + +## Running + +```bash +# Full training + eval (single seed): +VRL_ENABLED=1 LEAKY_RELU=1 GATED_ATTENTION=0 TTT_ENABLED=1 \ +TTT_OPTIMIZER=adamw TTT_LR=0.0005 TTT_EPOCHS=3 TTT_FREEZE_BLOCKS=2 \ +TTT_TEMPERATURE=0.98 USE_HEDGE_MIXER=1 SEED=42 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py + +# Eval-only (requires saved checkpoint): +EVAL_ONLY=1 CHECKPOINT_PATH=final_model.pt \ +VRL_ENABLED=1 LEAKY_RELU=1 GATED_ATTENTION=0 TTT_ENABLED=1 \ +TTT_OPTIMIZER=adamw TTT_LR=0.0005 TTT_EPOCHS=3 TTT_FREEZE_BLOCKS=2 \ +TTT_TEMPERATURE=0.98 USE_HEDGE_MIXER=1 SEED=42 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## Acknowledgments + +- Hedge Mixer concept inspired by PR #700 (RoyiRa) and PR #702 (lukacf) +- TTT recipe based on PR #606 (EthanYangTW) — AdamW + Polyak EMA + byte-weighted loss +- VRL technique from the Value Residual literature +- Base architecture from PR #198/#287/#549 lineage diff --git a/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/eval_seed42.log b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/eval_seed42.log new file mode 100644 index 0000000000..856a404e9c --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/eval_seed42.log @@ -0,0 +1,2210 @@ +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 lzma +from pathlib import Path +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 + +_FA_VERSION = 0 +_fa_func = None +try: + from flash_attn_interface import flash_attn_func as _fa_func + _FA_VERSION = 3 +except ImportError: + try: + from flash_attn import flash_attn_func as _fa_func + _FA_VERSION = 2 + except ImportError: + _FA_VERSION = 0 + _fa_func = None + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + soft_round_qat = bool(int(os.environ.get("SOFT_ROUND_QAT", "1"))) + soft_round_temp_start = float(os.environ.get("SOFT_ROUND_TEMP_START", 1.0)) + soft_round_temp_end = float(os.environ.get("SOFT_ROUND_TEMP_END", 0.05)) + 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") + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + leaky_relu = bool(int(os.environ.get("LEAKY_RELU", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adamw") + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + polyak_decay = float(os.environ.get("POLYAK_DECAY", 0.998)) + use_polyak = bool(int(os.environ.get("USE_POLYAK", "1"))) + byte_weighted_ttt = bool(int(os.environ.get("BYTE_WEIGHTED_TTT", "1"))) + adaptive_lr = bool(int(os.environ.get("ADAPTIVE_LR", "1"))) + adaptive_lr_max = float(os.environ.get("ADAPTIVE_LR_MAX", 3.0)) + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + checkpoint_path = os.environ.get("CHECKPOINT_PATH", "final_model.pt") + ttt_max_chunks = int(os.environ.get("TTT_MAX_CHUNKS", 0)) + skip_sliding_window = bool(int(os.environ.get("SKIP_SLIDING_WINDOW", "0"))) + use_hedge_mixer = bool(int(os.environ.get("USE_HEDGE_MIXER", "1"))) + mixer_eta = float(os.environ.get("MIXER_ETA", 0.1)) + mixer_min_tokens = int(os.environ.get("MIXER_MIN_TOKENS", 10000)) + +class HedgeMixer: + """5-expert online mixer: neural, unigram, bigram, trigram(hashed), entropy.""" + def __init__(self, vocab_size: int, device: torch.device, eta: float = 0.1, + tri_buckets: int = 65536, min_tokens: int = 10000): + self.V = vocab_size + self.K = 5 + self.eta = eta + self.min_tokens = min_tokens + self.device = device + self.log_weights = torch.zeros(self.K, device=device) + self.log_weights[0] = 2.0 + self.tokens_seen = 0 + self.uni_counts = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.uni_total = 0.0 + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_row_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.tri_buckets = tri_buckets + self.tri_counts = torch.zeros(tri_buckets, vocab_size, device=device, dtype=torch.float32) + self.tri_row_totals = torch.zeros(tri_buckets, device=device, dtype=torch.float32) + self._smooth = 0.1 + def update(self, tokens: Tensor): + t = tokens.to(self.device).long() + n = t.numel() + self.tokens_seen += n + ones = torch.ones(n, device=self.device, dtype=torch.float32) + self.uni_counts.scatter_add_(0, t, ones) + self.uni_total += n + if n >= 2: + ctx = t[:-1]; nxt = t[1:] + bi_idx = ctx * self.V + nxt + self.bi_counts.reshape(-1).scatter_add_(0, bi_idx, ones[:n-1]) + self.bi_row_totals.scatter_add_(0, ctx, ones[:n-1]) + if n >= 3: + prev2 = t[:-2]; prev1 = t[1:-1]; nxt3 = t[2:] + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.tri_buckets + tri_idx = tri_ctx * self.V + nxt3 + self.tri_counts.reshape(-1).scatter_add_(0, tri_idx, ones[:n-2]) + self.tri_row_totals.scatter_add_(0, tri_ctx, ones[:n-2]) + def get_expert_nll(self, logits: Tensor, x_batch: Tensor, y_batch: Tensor, + wlens: list[int], temperature: float = 1.0) -> Tensor: + bsz, slen, V = logits.shape + if temperature != 1.0: + logits = logits / temperature + log_probs_neural = F.log_softmax(logits.float(), dim=-1) + neural_nll = -log_probs_neural.gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + s = self._smooth + uni_probs = (self.uni_counts + s) / (self.uni_total + s * self.V) + uni_nll = -uni_probs.log()[y_batch] + prev_tokens = x_batch + bi_row = self.bi_counts[prev_tokens] + bi_row_sum = self.bi_row_totals[prev_tokens].unsqueeze(-1) + bi_probs = (bi_row + s) / (bi_row_sum + s * self.V) + bi_nll = -bi_probs.log().gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + tri_nll = torch.full((bsz, slen), 10.0, device=self.device) + if slen >= 2: + prev2 = torch.zeros_like(x_batch) + prev2[:, 1:] = x_batch[:, :-1] + prev1 = x_batch + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.tri_buckets + tri_row = self.tri_counts[tri_ctx] + tri_row_sum = self.tri_row_totals[tri_ctx].unsqueeze(-1) + tri_probs = (tri_row + s) / (tri_row_sum + s * self.V) + tri_nll = -tri_probs.log().gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + probs_neural = log_probs_neural.exp() + entropy = -(probs_neural * log_probs_neural).sum(dim=-1) + return torch.stack([neural_nll, uni_nll, bi_nll, tri_nll, entropy], dim=-1) + def mix_and_score(self, expert_nll: Tensor) -> Tensor: + log_w = self.log_weights - self.log_weights.logsumexp(0) + mixed_lp = (-expert_nll + log_w).logsumexp(dim=-1) + return -mixed_lp + def update_weights(self, expert_nll: Tensor, mask: Tensor): + masked_nll = expert_nll * mask.unsqueeze(-1).float() + expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) + self.log_weights -= self.eta * expert_mean_loss + +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("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.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,vrl_scales", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +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_qat: bool = True + _soft_round_temp: float = 1.0 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._soft_round_qat: + w32 = self.weight.float() + row_max = w32.detach().abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_s = w32 / scale[:, None] + residual = w_s - w_s.detach().round() + temp = CastedLinear._soft_round_temp + w_soft = w_s.detach().round() + 0.5 * torch.tanh(residual / temp) + w = (w_soft.clamp(-32, 31) * scale[:, None]).to(x.dtype) + else: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + ): + 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 + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, 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 _FA_VERSION == 3: + y = _fa_func(q, k, v, causal=True) + elif _FA_VERSION == 2: + y = _fa_func(q.bfloat16(), k.bfloat16(), v.bfloat16(), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, enable_gqa=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + 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, leaky: bool = False): + 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 + self._neg_slope = 0.5 if leaky else 0.0 + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), self._neg_slope) + return self.proj(x.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, + **kwargs, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=kwargs.get("gated_attention", False)) + self.mlp = MLP(dim, mlp_mult, leaky=kwargs.get("leaky", False)) + 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, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + vrl_enabled: bool = False, + leaky_relu: bool = False, + gated_attention: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + 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, + leaky=leaky_relu, + gated_attention=gated_attention, + ) + 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() + if self.vrl_enabled: + self.vrl_scales = nn.ParameterList( + [nn.Parameter(torch.zeros(1, dtype=torch.float32)) for _ in range(num_layers - 1)] + ) + else: + self.vrl_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + 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 = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + 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) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + + def forward_logits(self, input_ids: Tensor) -> Tensor: + 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 = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + 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) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + 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_ttt( + args, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """#606-style TTT + Hedge Mixer (safe: deferred weight updates between chunks).""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + if args.ttt_max_chunks > 0: + num_chunks = min(num_chunks, args.ttt_max_chunks) + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + if ci < num_chunks: + chunk_windows[ci].append(ws) + log0(f"ttt606:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={sum(len(w) for w in chunk_windows)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks} optimizer={args.ttt_optimizer} " + f"polyak={args.use_polyak}({args.polyak_decay}) byte_weighted={args.byte_weighted_ttt} " + f"adaptive_lr={args.adaptive_lr}(max={args.adaptive_lr_max}) temp={args.ttt_temperature}") + 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) + num_blocks = len(base_model.blocks) + unfrozen_block_start = max(0, num_blocks - args.ttt_freeze_blocks) if args.ttt_freeze_blocks > 0 else 0 + ttt_params = [] + for name, p in base_model.named_parameters(): + unfreeze = False + if args.ttt_freeze_blocks <= 0: + unfreeze = True + elif "norm" in name or "scale" in name or "lm_head" in name or "tok_emb" in name: + unfreeze = True + else: + for bi in range(unfrozen_block_start, num_blocks): + if f"blocks.{bi}." in name: + unfreeze = True + break + if unfreeze: + p.requires_grad_(True) + ttt_params.append(p) + else: + p.requires_grad_(False) + log0(f"ttt606:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + polyak_state: dict[str, Tensor] | None = None + if args.use_polyak: + polyak_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + mixer: HedgeMixer | None = None + if args.use_hedge_mixer: + mixer = HedgeMixer(args.vocab_size, device, eta=args.mixer_eta, min_tokens=args.mixer_min_tokens) + log0(f"hedge_mixer:enabled K={mixer.K} eta={mixer.eta} min_tokens={mixer.min_tokens}") + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + raw_state: dict[str, Tensor] | None = None + if polyak_state is not None: + raw_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + for n, p in base_model.named_parameters(): + if n in polyak_state: + p.data.copy_(polyak_state[n]) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + if mixer is not None and mixer.tokens_seen >= mixer.min_tokens: + expert_nll = mixer.get_expert_nll(logits, x_batch, y_batch, wlens, args.ttt_temperature) + nll = mixer.mix_and_score(expert_nll) + scored_mask = torch.zeros(bsz, seq_len, device=device, dtype=torch.bool) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_mask[i, s:wlen] = True + if not hasattr(mixer, '_chunk_expert_nll'): + mixer._chunk_expert_nll = [] + mixer._chunk_masks = [] + mixer._chunk_expert_nll.append(expert_nll.detach()) + mixer._chunk_masks.append(scored_mask) + else: + if args.ttt_temperature != 1.0: + logits = logits / args.ttt_temperature + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if mixer is not None: + if hasattr(mixer, '_chunk_expert_nll') and mixer._chunk_expert_nll: + all_enll = torch.cat(mixer._chunk_expert_nll, dim=0) + all_masks = torch.cat(mixer._chunk_masks, dim=0) + mixer.update_weights(all_enll, all_masks) + mixer._chunk_expert_nll = [] + mixer._chunk_masks = [] + chunk_tokens = val_tokens[chunk_start:chunk_end].to(device) + mixer.update(chunk_tokens) + if raw_state is not None: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if args.adaptive_lr: + progress = min(ci / (num_chunks * 0.3), 1.0) + lr_mult = 1.0 + (args.adaptive_lr_max - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + distributed = dist.is_available() and dist.is_initialized() + my_seq_s = (chunk_seqs * rank) // world_size if distributed else 0 + my_seq_e = (chunk_seqs * (rank + 1)) // world_size if distributed else chunk_seqs + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_t = base_model.forward_logits(x) + if args.byte_weighted_ttt: + per_tok_nll = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), reduction="none", + ) + byte_weights = base_bytes_lut[y.reshape(-1)].float() + byte_weights = byte_weights / byte_weights.mean().clamp(min=1e-6) + loss = (per_tok_nll * byte_weights).mean() + else: + loss = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), + ) + loss.backward() + if distributed and world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if polyak_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in polyak_state: + polyak_state[n].lerp_(p.data, 1.0 - args.polyak_decay) + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +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 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.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 _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +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 + + _gpu_name = torch.cuda.get_device_name(0) + _is_high_end = "H100" in _gpu_name or "A100" in _gpu_name + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + if _is_high_end: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(True) + enable_flash_sdp(True) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + log0(f"flash_attn_version:{_FA_VERSION} gpu:{_gpu_name} high_end:{_is_high_end}") + 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 + CastedLinear._soft_round_qat = args.soft_round_qat + CastedLinear._soft_round_temp = args.soft_round_temp_start + qat_start_step = 0 if args.qat_enabled else -1 + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, + leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if distributed: + torch._dynamo.config.optimize_ddp = False + 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) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + 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) + if base_model.vrl_enabled: + for s in base_model.vrl_scales: + scalar_params.append(s) + 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()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"sdp_backends:high_end={_is_high_end}") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0(f"vrl_enabled:{args.vrl_enabled} leaky_relu:{args.leaky_relu} ttt_enabled:{args.ttt_enabled}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0 and not args.eval_only: + 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) + if args.eval_only: + log0(f"eval_only:loading checkpoint from {args.checkpoint_path}") + ckpt_state = torch.load(args.checkpoint_path, map_location="cpu") + base_model.load_state_dict(ckpt_state, strict=True) + log0(f"eval_only:checkpoint loaded, {sum(p.numel() for p in base_model.parameters())} params") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + log0(f"eval_only:quantized model size: {len(quant_blob)} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_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, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, gated_attention=args.gated_attention, + ).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 = effective_eval_seq_len + if not args.skip_sliding_window and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0(f"eval_only:sliding_window val_bpb:{sw_val_bpb:.4f} stride:{args.eval_stride} " + f"eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms") + elif args.skip_sliding_window: + log0("eval_only:skipping sliding window eval") + if args.ttt_enabled: + log0(f"eval_only:ttt lr={args.ttt_lr} epochs={args.ttt_epochs} " + f"chunk={args.ttt_chunk_tokens} freeze_blocks={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_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, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"eval_only:ttt val_bpb:{ttt_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + if distributed: + dist.destroy_process_group() + return + + 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 = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + qat_start_step = step + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + if CastedLinear._qat_enabled and CastedLinear._soft_round_qat and qat_start_step >= 0: + qat_total = max(args.iterations - qat_start_step, 1) + qat_progress = min((step - qat_start_step) / qat_total, 1.0) + log_start = math.log(args.soft_round_temp_start) + log_end = math.log(args.soft_round_temp_end) + CastedLinear._soft_round_temp = math.exp(log_start + qat_progress * (log_end - log_start)) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + 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 = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_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, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, gated_attention=args.gated_attention, + ).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) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + + if args.ttt_enabled: + log0("ttt:starting test-time training evaluation") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + 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, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0( + f"ttt_eval val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"ttt_eval_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.11.10 (main, Sep 7 2024, 18:35:41) [GCC 11.4.0] +Running PyTorch 2.11.0+cu128 +Wed Mar 25 14:57:03 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 565.57.01 Driver Version: 565.57.01 CUDA Version: 12.7 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 34C P0 149W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:2A:00.0 Off | 0 | +| N/A 35C P0 150W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3A:00.0 Off | 0 | +| N/A 37C P0 152W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 35C P0 153W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9A:00.0 Off | 0 | +| N/A 34C P0 148W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:AB:00.0 Off | 0 | +| N/A 36C P0 146W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BA:00.0 Off | 0 | +| N/A 34C P0 146W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 35C P0 152W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +flash_attn_version:0 gpu:NVIDIA H100 80GB HBM3 high_end:True +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26993766 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:high_end=True +attention_mode:gqa num_heads:8 num_kv_heads:4 +vrl_enabled:True leaky_relu:True ttt_enabled:True +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:42 +eval_only:loading checkpoint from final_model.pt +eval_only:checkpoint loaded, 26993766 params +eval_only:quantized model size: 15905672 bytes +eval_only:sliding_window val_bpb:1.1263 stride:64 eval_time:92931ms +eval_only:ttt lr=0.0005 epochs=3 chunk=32768 freeze_blocks=2 +ttt606:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.0005 ttt_epochs=3 freeze_blocks=2 optimizer=adamw polyak=True(0.998) byte_weighted=True adaptive_lr=True(max=3.0) temp=0.98 +ttt606:params unfrozen=5256222 frozen=21737544 +hedge_mixer:enabled K=5 eta=0.1 min_tokens=10000 + ttt_chunk [1/1893] bpb=1.168129 time=1.9s + ttt_chunk [11/1893] bpb=1.103862 time=4.0s + ttt_chunk [21/1893] bpb=1.063417 time=6.1s + ttt_chunk [31/1893] bpb=1.050750 time=8.2s + ttt_chunk [41/1893] bpb=1.033851 time=10.3s + ttt_chunk [51/1893] bpb=1.025208 time=12.4s + ttt_chunk [61/1893] bpb=1.028682 time=14.5s + ttt_chunk [71/1893] bpb=1.025204 time=16.7s + ttt_chunk [81/1893] bpb=1.022983 time=18.8s + ttt_chunk [91/1893] bpb=1.022750 time=20.9s + ttt_chunk [101/1893] bpb=1.024831 time=23.0s + ttt_chunk [111/1893] bpb=1.026340 time=25.1s + ttt_chunk [121/1893] bpb=1.019328 time=27.2s + ttt_chunk [131/1893] bpb=1.018845 time=29.3s + ttt_chunk [141/1893] bpb=1.023513 time=31.4s + ttt_chunk [151/1893] bpb=1.024634 time=33.5s + ttt_chunk [161/1893] bpb=1.023497 time=35.6s + ttt_chunk [171/1893] bpb=1.026772 time=38.0s + ttt_chunk [181/1893] bpb=1.028517 time=40.4s + ttt_chunk [191/1893] bpb=1.034545 time=42.9s + ttt_chunk [201/1893] bpb=1.032907 time=45.3s + ttt_chunk [211/1893] bpb=1.030549 time=47.7s + ttt_chunk [221/1893] bpb=1.031586 time=49.9s + ttt_chunk [231/1893] bpb=1.030037 time=52.0s + ttt_chunk [241/1893] bpb=1.030045 time=54.1s + ttt_chunk [251/1893] bpb=1.029018 time=56.3s + ttt_chunk [261/1893] bpb=1.025817 time=58.7s + ttt_chunk [271/1893] bpb=1.024470 time=61.1s + ttt_chunk [281/1893] bpb=1.025279 time=63.5s + ttt_chunk [291/1893] bpb=1.026463 time=65.8s + ttt_chunk [301/1893] bpb=1.026718 time=68.1s + ttt_chunk [311/1893] bpb=1.028256 time=70.2s + ttt_chunk [321/1893] bpb=1.029890 time=72.3s + ttt_chunk [331/1893] bpb=1.029878 time=74.4s + ttt_chunk [341/1893] bpb=1.028695 time=76.5s + ttt_chunk [351/1893] bpb=1.030751 time=78.6s + ttt_chunk [361/1893] bpb=1.031224 time=80.7s + ttt_chunk [371/1893] bpb=1.030747 time=82.8s + ttt_chunk [381/1893] bpb=1.031392 time=84.9s + ttt_chunk [391/1893] bpb=1.031487 time=87.0s + ttt_chunk [401/1893] bpb=1.030018 time=89.1s + ttt_chunk [411/1893] bpb=1.029425 time=91.2s + ttt_chunk [421/1893] bpb=1.029253 time=93.3s + ttt_chunk [431/1893] bpb=1.029785 time=95.4s + ttt_chunk [441/1893] bpb=1.030680 time=97.5s + ttt_chunk [451/1893] bpb=1.031984 time=99.6s + ttt_chunk [461/1893] bpb=1.031785 time=101.7s + ttt_chunk [471/1893] bpb=1.033176 time=103.8s + ttt_chunk [481/1893] bpb=1.033719 time=105.9s + ttt_chunk [491/1893] bpb=1.033610 time=108.0s + ttt_chunk [501/1893] bpb=1.033876 time=110.1s + ttt_chunk [511/1893] bpb=1.034118 time=112.2s + ttt_chunk [521/1893] bpb=1.032858 time=114.3s + ttt_chunk [531/1893] bpb=1.034885 time=116.5s + ttt_chunk [541/1893] bpb=1.036009 time=118.6s + ttt_chunk [551/1893] bpb=1.035926 time=120.6s + ttt_chunk [561/1893] bpb=1.037414 time=122.7s + ttt_chunk [571/1893] bpb=1.037349 time=124.8s + ttt_chunk [581/1893] bpb=1.037408 time=126.9s + ttt_chunk [591/1893] bpb=1.037810 time=129.0s + ttt_chunk [601/1893] bpb=1.038941 time=131.1s + ttt_chunk [611/1893] bpb=1.039654 time=133.2s + ttt_chunk [621/1893] bpb=1.040274 time=135.3s + ttt_chunk [631/1893] bpb=1.041696 time=137.4s + ttt_chunk [641/1893] bpb=1.042092 time=139.5s + ttt_chunk [651/1893] bpb=1.042813 time=141.7s + ttt_chunk [661/1893] bpb=1.042890 time=143.8s + ttt_chunk [671/1893] bpb=1.043734 time=145.9s + ttt_chunk [681/1893] bpb=1.044852 time=147.9s + ttt_chunk [691/1893] bpb=1.046334 time=150.0s + ttt_chunk [701/1893] bpb=1.046421 time=152.1s + ttt_chunk [711/1893] bpb=1.047166 time=154.2s + ttt_chunk [721/1893] bpb=1.047165 time=156.3s + ttt_chunk [731/1893] bpb=1.047743 time=158.5s + ttt_chunk [741/1893] bpb=1.048238 time=160.6s + ttt_chunk [751/1893] bpb=1.048414 time=162.7s + ttt_chunk [761/1893] bpb=1.048670 time=164.8s + ttt_chunk [771/1893] bpb=1.048892 time=166.9s + ttt_chunk [781/1893] bpb=1.049969 time=169.0s + ttt_chunk [791/1893] bpb=1.049816 time=171.1s + ttt_chunk [801/1893] bpb=1.050357 time=173.2s + ttt_chunk [811/1893] bpb=1.050515 time=175.5s + ttt_chunk [821/1893] bpb=1.050631 time=177.6s + ttt_chunk [831/1893] bpb=1.050757 time=179.7s + ttt_chunk [841/1893] bpb=1.050412 time=181.8s + ttt_chunk [851/1893] bpb=1.050491 time=183.9s + ttt_chunk [861/1893] bpb=1.050407 time=186.0s + ttt_chunk [871/1893] bpb=1.050921 time=188.1s + ttt_chunk [881/1893] bpb=1.051364 time=190.2s + ttt_chunk [891/1893] bpb=1.051220 time=192.3s + ttt_chunk [901/1893] bpb=1.051198 time=194.4s + ttt_chunk [911/1893] bpb=1.051529 time=196.5s + ttt_chunk [921/1893] bpb=1.052094 time=198.7s + ttt_chunk [931/1893] bpb=1.052178 time=200.8s + ttt_chunk [941/1893] bpb=1.052022 time=202.9s + ttt_chunk [951/1893] bpb=1.052566 time=205.0s + ttt_chunk [961/1893] bpb=1.052724 time=207.1s + ttt_chunk [971/1893] bpb=1.053501 time=209.2s + ttt_chunk [981/1893] bpb=1.053640 time=211.3s + ttt_chunk [991/1893] bpb=1.053746 time=213.4s + ttt_chunk [1001/1893] bpb=1.053705 time=215.5s + ttt_chunk [1011/1893] bpb=1.053625 time=217.6s + ttt_chunk [1021/1893] bpb=1.053909 time=219.7s + ttt_chunk [1031/1893] bpb=1.054360 time=221.8s + ttt_chunk [1041/1893] bpb=1.054030 time=223.9s + ttt_chunk [1051/1893] bpb=1.053748 time=226.0s + ttt_chunk [1061/1893] bpb=1.053822 time=228.1s + ttt_chunk [1071/1893] bpb=1.054410 time=230.2s + ttt_chunk [1081/1893] bpb=1.054643 time=232.3s + ttt_chunk [1091/1893] bpb=1.055296 time=234.4s + ttt_chunk [1101/1893] bpb=1.055272 time=236.6s + ttt_chunk [1111/1893] bpb=1.055038 time=239.0s + ttt_chunk [1121/1893] bpb=1.054788 time=241.3s + ttt_chunk [1131/1893] bpb=1.054682 time=243.4s + ttt_chunk [1141/1893] bpb=1.054406 time=245.5s + ttt_chunk [1151/1893] bpb=1.054329 time=247.6s + ttt_chunk [1161/1893] bpb=1.053974 time=249.7s + ttt_chunk [1171/1893] bpb=1.054125 time=251.8s + ttt_chunk [1181/1893] bpb=1.053374 time=253.9s + ttt_chunk [1191/1893] bpb=1.053159 time=256.0s + ttt_chunk [1201/1893] bpb=1.053413 time=258.1s + ttt_chunk [1211/1893] bpb=1.052940 time=260.2s + ttt_chunk [1221/1893] bpb=1.052575 time=262.3s + ttt_chunk [1231/1893] bpb=1.052262 time=264.4s + ttt_chunk [1241/1893] bpb=1.051800 time=266.5s + ttt_chunk [1251/1893] bpb=1.051166 time=268.6s + ttt_chunk [1261/1893] bpb=1.051000 time=270.7s + ttt_chunk [1271/1893] bpb=1.050608 time=272.8s + ttt_chunk [1281/1893] bpb=1.050388 time=274.9s + ttt_chunk [1291/1893] bpb=1.050155 time=277.0s + ttt_chunk [1301/1893] bpb=1.049488 time=279.3s + ttt_chunk [1311/1893] bpb=1.049096 time=281.4s + ttt_chunk [1321/1893] bpb=1.048729 time=283.5s + ttt_chunk [1331/1893] bpb=1.048614 time=285.6s + ttt_chunk [1341/1893] bpb=1.048429 time=287.7s + ttt_chunk [1351/1893] bpb=1.048277 time=289.9s + ttt_chunk [1361/1893] bpb=1.048280 time=292.0s + ttt_chunk [1371/1893] bpb=1.048062 time=294.1s + ttt_chunk [1381/1893] bpb=1.047966 time=296.2s + ttt_chunk [1391/1893] bpb=1.047553 time=298.3s + ttt_chunk [1401/1893] bpb=1.047479 time=300.4s + ttt_chunk [1411/1893] bpb=1.047526 time=302.5s + ttt_chunk [1421/1893] bpb=1.047709 time=304.6s + ttt_chunk [1431/1893] bpb=1.047339 time=306.8s + ttt_chunk [1441/1893] bpb=1.047660 time=308.9s + ttt_chunk [1451/1893] bpb=1.047970 time=311.0s + ttt_chunk [1461/1893] bpb=1.047573 time=313.1s + ttt_chunk [1471/1893] bpb=1.048419 time=315.2s + ttt_chunk [1481/1893] bpb=1.047938 time=317.3s + ttt_chunk [1491/1893] bpb=1.047732 time=319.4s + ttt_chunk [1501/1893] bpb=1.047607 time=321.5s + ttt_chunk [1511/1893] bpb=1.047468 time=323.6s + ttt_chunk [1521/1893] bpb=1.047434 time=325.7s + ttt_chunk [1531/1893] bpb=1.046910 time=327.8s + ttt_chunk [1541/1893] bpb=1.046651 time=329.9s + ttt_chunk [1551/1893] bpb=1.046896 time=332.0s + ttt_chunk [1561/1893] bpb=1.046834 time=334.1s + ttt_chunk [1571/1893] bpb=1.046636 time=336.2s + ttt_chunk [1581/1893] bpb=1.046698 time=338.3s + ttt_chunk [1591/1893] bpb=1.046543 time=340.4s + ttt_chunk [1601/1893] bpb=1.046661 time=342.5s + ttt_chunk [1611/1893] bpb=1.046533 time=344.6s + ttt_chunk [1621/1893] bpb=1.046076 time=346.7s + ttt_chunk [1631/1893] bpb=1.046255 time=348.8s + ttt_chunk [1641/1893] bpb=1.046174 time=350.9s + ttt_chunk [1651/1893] bpb=1.046006 time=353.0s + ttt_chunk [1661/1893] bpb=1.045789 time=355.2s + ttt_chunk [1671/1893] bpb=1.046114 time=357.3s + ttt_chunk [1681/1893] bpb=1.046188 time=359.4s + ttt_chunk [1691/1893] bpb=1.045958 time=361.5s + ttt_chunk [1701/1893] bpb=1.046061 time=363.6s + ttt_chunk [1711/1893] bpb=1.046011 time=365.7s + ttt_chunk [1721/1893] bpb=1.045867 time=367.8s + ttt_chunk [1731/1893] bpb=1.045693 time=369.9s + ttt_chunk [1741/1893] bpb=1.045501 time=372.0s + ttt_chunk [1751/1893] bpb=1.045236 time=374.1s + ttt_chunk [1761/1893] bpb=1.045274 time=376.2s + ttt_chunk [1771/1893] bpb=1.045104 time=378.3s + ttt_chunk [1781/1893] bpb=1.045073 time=380.4s + ttt_chunk [1791/1893] bpb=1.044574 time=382.5s + ttt_chunk [1801/1893] bpb=1.044396 time=384.6s + ttt_chunk [1811/1893] bpb=1.044230 time=386.7s + ttt_chunk [1821/1893] bpb=1.044213 time=388.8s + ttt_chunk [1831/1893] bpb=1.043569 time=390.9s + ttt_chunk [1841/1893] bpb=1.043480 time=393.0s + ttt_chunk [1851/1893] bpb=1.043210 time=395.1s + ttt_chunk [1861/1893] bpb=1.042787 time=397.2s + ttt_chunk [1871/1893] bpb=1.042656 time=399.3s + ttt_chunk [1881/1893] bpb=1.042145 time=401.4s + ttt_chunk [1891/1893] bpb=1.041844 time=403.5s + ttt_chunk [1893/1893] bpb=1.041876 time=404.1s +ttt_sliding:done val_loss=1.755931 val_bpb=1.039964 elapsed=404.1s +eval_only:ttt val_bpb:1.0400 eval_time:404494ms diff --git a/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/submission.json b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/submission.json new file mode 100644 index 0000000000..bf8946eac9 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/submission.json @@ -0,0 +1,30 @@ +{ + "submission_name": "HedgeMixer + VRL + AdamW TTT + Polyak EMA", + "description": "11L transformer with Value Residual, LeakyReLU(0.5)^2, XSA-4, legal score-first TTT with AdamW + Polyak EMA + byte-weighted loss + adaptive cosine LR, and a 5-expert Hedge Mixer (neural + unigram + bigram + trigram + entropy) with deferred weight updates.", + "bpb": 1.0400, + "val_loss": 1.7559, + "seeds": [42], + "num_params": 26993766, + "training_time_seconds": 600, + "steps": 6104, + "artifact_bytes": 15999919, + "code_bytes": 94305, + "model_bytes": 15905672, + "compression": "int6+lzma", + "gpu": "8xH100 SXM 80GB", + "pytorch_version": "2.11.0+cu128", + "techniques": [ + "11 layers, 512 dim, GQA 8/4 heads", + "Value Residual (VRL) from layer 0 to all subsequent layers", + "LeakyReLU(0.5)^2 activation", + "XSA-4 (Cross-Token Self-Attention on last 4 layers)", + "Soft-Round QAT with late activation", + "EMA weight averaging (decay=0.997)", + "Muon optimizer + AdamW for embeddings", + "Int6 mixed quantization + lzma compression", + "Sliding window eval (stride=64)", + "Legal score-first TTT: AdamW (lr=0.0005), Polyak EMA (0.998), byte-weighted loss, adaptive cosine LR (1x->3x ramp), temperature=0.98, freeze first 9 of 11 blocks", + "5-expert Hedge Mixer: neural model + unigram + bigram + trigram(64K hashed) + entropy, with deferred between-chunk weight updates (eta=0.1)" + ], + "legality_notes": "All eval-time adaptations are strictly score-first: each chunk is scored before any adaptation. Hedge mixer weights update ONLY between chunks (after all windows scored). N-gram tables update ONLY after chunk scoring. Polyak EMA uses fixed decay, no selection. No validation data accessed during training." +} diff --git a/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/train_gpt.py b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/train_gpt.py new file mode 100644 index 0000000000..66c7e3313f --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/train_gpt.py @@ -0,0 +1,1942 @@ +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 lzma +from pathlib import Path +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 + +_FA_VERSION = 0 +_fa_func = None +try: + from flash_attn_interface import flash_attn_func as _fa_func + _FA_VERSION = 3 +except ImportError: + try: + from flash_attn import flash_attn_func as _fa_func + _FA_VERSION = 2 + except ImportError: + _FA_VERSION = 0 + _fa_func = None + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + soft_round_qat = bool(int(os.environ.get("SOFT_ROUND_QAT", "1"))) + soft_round_temp_start = float(os.environ.get("SOFT_ROUND_TEMP_START", 1.0)) + soft_round_temp_end = float(os.environ.get("SOFT_ROUND_TEMP_END", 0.05)) + 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") + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + leaky_relu = bool(int(os.environ.get("LEAKY_RELU", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adamw") + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + polyak_decay = float(os.environ.get("POLYAK_DECAY", 0.998)) + use_polyak = bool(int(os.environ.get("USE_POLYAK", "1"))) + byte_weighted_ttt = bool(int(os.environ.get("BYTE_WEIGHTED_TTT", "1"))) + adaptive_lr = bool(int(os.environ.get("ADAPTIVE_LR", "1"))) + adaptive_lr_max = float(os.environ.get("ADAPTIVE_LR_MAX", 3.0)) + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + checkpoint_path = os.environ.get("CHECKPOINT_PATH", "final_model.pt") + ttt_max_chunks = int(os.environ.get("TTT_MAX_CHUNKS", 0)) + skip_sliding_window = bool(int(os.environ.get("SKIP_SLIDING_WINDOW", "0"))) + use_hedge_mixer = bool(int(os.environ.get("USE_HEDGE_MIXER", "1"))) + mixer_eta = float(os.environ.get("MIXER_ETA", 0.1)) + mixer_min_tokens = int(os.environ.get("MIXER_MIN_TOKENS", 10000)) + +class HedgeMixer: + """5-expert online mixer: neural, unigram, bigram, trigram(hashed), entropy.""" + def __init__(self, vocab_size: int, device: torch.device, eta: float = 0.1, + tri_buckets: int = 65536, min_tokens: int = 10000): + self.V = vocab_size + self.K = 5 + self.eta = eta + self.min_tokens = min_tokens + self.device = device + self.log_weights = torch.zeros(self.K, device=device) + self.log_weights[0] = 2.0 + self.tokens_seen = 0 + self.uni_counts = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.uni_total = 0.0 + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_row_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.tri_buckets = tri_buckets + self.tri_counts = torch.zeros(tri_buckets, vocab_size, device=device, dtype=torch.float32) + self.tri_row_totals = torch.zeros(tri_buckets, device=device, dtype=torch.float32) + self._smooth = 0.1 + def update(self, tokens: Tensor): + t = tokens.to(self.device).long() + n = t.numel() + self.tokens_seen += n + ones = torch.ones(n, device=self.device, dtype=torch.float32) + self.uni_counts.scatter_add_(0, t, ones) + self.uni_total += n + if n >= 2: + ctx = t[:-1]; nxt = t[1:] + bi_idx = ctx * self.V + nxt + self.bi_counts.reshape(-1).scatter_add_(0, bi_idx, ones[:n-1]) + self.bi_row_totals.scatter_add_(0, ctx, ones[:n-1]) + if n >= 3: + prev2 = t[:-2]; prev1 = t[1:-1]; nxt3 = t[2:] + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.tri_buckets + tri_idx = tri_ctx * self.V + nxt3 + self.tri_counts.reshape(-1).scatter_add_(0, tri_idx, ones[:n-2]) + self.tri_row_totals.scatter_add_(0, tri_ctx, ones[:n-2]) + def get_expert_nll(self, logits: Tensor, x_batch: Tensor, y_batch: Tensor, + wlens: list[int], temperature: float = 1.0) -> Tensor: + bsz, slen, V = logits.shape + if temperature != 1.0: + logits = logits / temperature + log_probs_neural = F.log_softmax(logits.float(), dim=-1) + neural_nll = -log_probs_neural.gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + s = self._smooth + uni_probs = (self.uni_counts + s) / (self.uni_total + s * self.V) + uni_nll = -uni_probs.log()[y_batch] + prev_tokens = x_batch + bi_row = self.bi_counts[prev_tokens] + bi_row_sum = self.bi_row_totals[prev_tokens].unsqueeze(-1) + bi_probs = (bi_row + s) / (bi_row_sum + s * self.V) + bi_nll = -bi_probs.log().gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + tri_nll = torch.full((bsz, slen), 10.0, device=self.device) + if slen >= 2: + prev2 = torch.zeros_like(x_batch) + prev2[:, 1:] = x_batch[:, :-1] + prev1 = x_batch + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.tri_buckets + tri_row = self.tri_counts[tri_ctx] + tri_row_sum = self.tri_row_totals[tri_ctx].unsqueeze(-1) + tri_probs = (tri_row + s) / (tri_row_sum + s * self.V) + tri_nll = -tri_probs.log().gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + probs_neural = log_probs_neural.exp() + entropy = -(probs_neural * log_probs_neural).sum(dim=-1) + return torch.stack([neural_nll, uni_nll, bi_nll, tri_nll, entropy], dim=-1) + def mix_and_score(self, expert_nll: Tensor) -> Tensor: + log_w = self.log_weights - self.log_weights.logsumexp(0) + mixed_lp = (-expert_nll + log_w).logsumexp(dim=-1) + return -mixed_lp + def update_weights(self, expert_nll: Tensor, mask: Tensor): + masked_nll = expert_nll * mask.unsqueeze(-1).float() + expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) + self.log_weights -= self.eta * expert_mean_loss + +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("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.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,vrl_scales", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +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_qat: bool = True + _soft_round_temp: float = 1.0 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._soft_round_qat: + w32 = self.weight.float() + row_max = w32.detach().abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_s = w32 / scale[:, None] + residual = w_s - w_s.detach().round() + temp = CastedLinear._soft_round_temp + w_soft = w_s.detach().round() + 0.5 * torch.tanh(residual / temp) + w = (w_soft.clamp(-32, 31) * scale[:, None]).to(x.dtype) + else: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + ): + 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 + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, 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 _FA_VERSION == 3: + y = _fa_func(q, k, v, causal=True) + elif _FA_VERSION == 2: + y = _fa_func(q.bfloat16(), k.bfloat16(), v.bfloat16(), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, enable_gqa=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + 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, leaky: bool = False): + 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 + self._neg_slope = 0.5 if leaky else 0.0 + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), self._neg_slope) + return self.proj(x.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, + **kwargs, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=kwargs.get("gated_attention", False)) + self.mlp = MLP(dim, mlp_mult, leaky=kwargs.get("leaky", False)) + 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, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + vrl_enabled: bool = False, + leaky_relu: bool = False, + gated_attention: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + 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, + leaky=leaky_relu, + gated_attention=gated_attention, + ) + 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() + if self.vrl_enabled: + self.vrl_scales = nn.ParameterList( + [nn.Parameter(torch.zeros(1, dtype=torch.float32)) for _ in range(num_layers - 1)] + ) + else: + self.vrl_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + 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 = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + 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) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + + def forward_logits(self, input_ids: Tensor) -> Tensor: + 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 = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + 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) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + 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_ttt( + args, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """#606-style TTT + Hedge Mixer (safe: deferred weight updates between chunks).""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + if args.ttt_max_chunks > 0: + num_chunks = min(num_chunks, args.ttt_max_chunks) + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + if ci < num_chunks: + chunk_windows[ci].append(ws) + log0(f"ttt606:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={sum(len(w) for w in chunk_windows)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks} optimizer={args.ttt_optimizer} " + f"polyak={args.use_polyak}({args.polyak_decay}) byte_weighted={args.byte_weighted_ttt} " + f"adaptive_lr={args.adaptive_lr}(max={args.adaptive_lr_max}) temp={args.ttt_temperature}") + 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) + num_blocks = len(base_model.blocks) + unfrozen_block_start = max(0, num_blocks - args.ttt_freeze_blocks) if args.ttt_freeze_blocks > 0 else 0 + ttt_params = [] + for name, p in base_model.named_parameters(): + unfreeze = False + if args.ttt_freeze_blocks <= 0: + unfreeze = True + elif "norm" in name or "scale" in name or "lm_head" in name or "tok_emb" in name: + unfreeze = True + else: + for bi in range(unfrozen_block_start, num_blocks): + if f"blocks.{bi}." in name: + unfreeze = True + break + if unfreeze: + p.requires_grad_(True) + ttt_params.append(p) + else: + p.requires_grad_(False) + log0(f"ttt606:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + polyak_state: dict[str, Tensor] | None = None + if args.use_polyak: + polyak_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + mixer: HedgeMixer | None = None + if args.use_hedge_mixer: + mixer = HedgeMixer(args.vocab_size, device, eta=args.mixer_eta, min_tokens=args.mixer_min_tokens) + log0(f"hedge_mixer:enabled K={mixer.K} eta={mixer.eta} min_tokens={mixer.min_tokens}") + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + raw_state: dict[str, Tensor] | None = None + if polyak_state is not None: + raw_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + for n, p in base_model.named_parameters(): + if n in polyak_state: + p.data.copy_(polyak_state[n]) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + if mixer is not None and mixer.tokens_seen >= mixer.min_tokens: + expert_nll = mixer.get_expert_nll(logits, x_batch, y_batch, wlens, args.ttt_temperature) + nll = mixer.mix_and_score(expert_nll) + scored_mask = torch.zeros(bsz, seq_len, device=device, dtype=torch.bool) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_mask[i, s:wlen] = True + if not hasattr(mixer, '_chunk_expert_nll'): + mixer._chunk_expert_nll = [] + mixer._chunk_masks = [] + mixer._chunk_expert_nll.append(expert_nll.detach()) + mixer._chunk_masks.append(scored_mask) + else: + if args.ttt_temperature != 1.0: + logits = logits / args.ttt_temperature + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if mixer is not None: + if hasattr(mixer, '_chunk_expert_nll') and mixer._chunk_expert_nll: + all_enll = torch.cat(mixer._chunk_expert_nll, dim=0) + all_masks = torch.cat(mixer._chunk_masks, dim=0) + mixer.update_weights(all_enll, all_masks) + mixer._chunk_expert_nll = [] + mixer._chunk_masks = [] + chunk_tokens = val_tokens[chunk_start:chunk_end].to(device) + mixer.update(chunk_tokens) + if raw_state is not None: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if args.adaptive_lr: + progress = min(ci / (num_chunks * 0.3), 1.0) + lr_mult = 1.0 + (args.adaptive_lr_max - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + distributed = dist.is_available() and dist.is_initialized() + my_seq_s = (chunk_seqs * rank) // world_size if distributed else 0 + my_seq_e = (chunk_seqs * (rank + 1)) // world_size if distributed else chunk_seqs + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_t = base_model.forward_logits(x) + if args.byte_weighted_ttt: + per_tok_nll = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), reduction="none", + ) + byte_weights = base_bytes_lut[y.reshape(-1)].float() + byte_weights = byte_weights / byte_weights.mean().clamp(min=1e-6) + loss = (per_tok_nll * byte_weights).mean() + else: + loss = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), + ) + loss.backward() + if distributed and world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if polyak_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in polyak_state: + polyak_state[n].lerp_(p.data, 1.0 - args.polyak_decay) + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +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 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.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 _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +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 + + _gpu_name = torch.cuda.get_device_name(0) + _is_high_end = "H100" in _gpu_name or "A100" in _gpu_name + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + if _is_high_end: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(True) + enable_flash_sdp(True) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + log0(f"flash_attn_version:{_FA_VERSION} gpu:{_gpu_name} high_end:{_is_high_end}") + 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 + CastedLinear._soft_round_qat = args.soft_round_qat + CastedLinear._soft_round_temp = args.soft_round_temp_start + qat_start_step = 0 if args.qat_enabled else -1 + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, + leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if distributed: + torch._dynamo.config.optimize_ddp = False + 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) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + 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) + if base_model.vrl_enabled: + for s in base_model.vrl_scales: + scalar_params.append(s) + 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()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"sdp_backends:high_end={_is_high_end}") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0(f"vrl_enabled:{args.vrl_enabled} leaky_relu:{args.leaky_relu} ttt_enabled:{args.ttt_enabled}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0 and not args.eval_only: + 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) + if args.eval_only: + log0(f"eval_only:loading checkpoint from {args.checkpoint_path}") + ckpt_state = torch.load(args.checkpoint_path, map_location="cpu") + base_model.load_state_dict(ckpt_state, strict=True) + log0(f"eval_only:checkpoint loaded, {sum(p.numel() for p in base_model.parameters())} params") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + log0(f"eval_only:quantized model size: {len(quant_blob)} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_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, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, gated_attention=args.gated_attention, + ).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 = effective_eval_seq_len + if not args.skip_sliding_window and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0(f"eval_only:sliding_window val_bpb:{sw_val_bpb:.4f} stride:{args.eval_stride} " + f"eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms") + elif args.skip_sliding_window: + log0("eval_only:skipping sliding window eval") + if args.ttt_enabled: + log0(f"eval_only:ttt lr={args.ttt_lr} epochs={args.ttt_epochs} " + f"chunk={args.ttt_chunk_tokens} freeze_blocks={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_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, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"eval_only:ttt val_bpb:{ttt_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + if distributed: + dist.destroy_process_group() + return + + 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 = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + qat_start_step = step + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + if CastedLinear._qat_enabled and CastedLinear._soft_round_qat and qat_start_step >= 0: + qat_total = max(args.iterations - qat_start_step, 1) + qat_progress = min((step - qat_start_step) / qat_total, 1.0) + log_start = math.log(args.soft_round_temp_start) + log_end = math.log(args.soft_round_temp_end) + CastedLinear._soft_round_temp = math.exp(log_start + qat_progress * (log_end - log_start)) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + 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 = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_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, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, gated_attention=args.gated_attention, + ).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) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + + if args.ttt_enabled: + log0("ttt:starting test-time training evaluation") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + 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, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0( + f"ttt_eval val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"ttt_eval_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/train_seed42.log b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/train_seed42.log new file mode 100644 index 0000000000..2ff02b2f32 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/train_seed42.log @@ -0,0 +1,2266 @@ +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 lzma +from pathlib import Path +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 + +_FA_VERSION = 0 +_fa_func = None +try: + from flash_attn_interface import flash_attn_func as _fa_func + _FA_VERSION = 3 +except ImportError: + try: + from flash_attn import flash_attn_func as _fa_func + _FA_VERSION = 2 + except ImportError: + _FA_VERSION = 0 + _fa_func = None + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + soft_round_qat = bool(int(os.environ.get("SOFT_ROUND_QAT", "1"))) + soft_round_temp_start = float(os.environ.get("SOFT_ROUND_TEMP_START", 1.0)) + soft_round_temp_end = float(os.environ.get("SOFT_ROUND_TEMP_END", 0.05)) + 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") + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + leaky_relu = bool(int(os.environ.get("LEAKY_RELU", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adamw") + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + polyak_decay = float(os.environ.get("POLYAK_DECAY", 0.998)) + use_polyak = bool(int(os.environ.get("USE_POLYAK", "1"))) + byte_weighted_ttt = bool(int(os.environ.get("BYTE_WEIGHTED_TTT", "1"))) + adaptive_lr = bool(int(os.environ.get("ADAPTIVE_LR", "1"))) + adaptive_lr_max = float(os.environ.get("ADAPTIVE_LR_MAX", 3.0)) + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + checkpoint_path = os.environ.get("CHECKPOINT_PATH", "final_model.pt") + ttt_max_chunks = int(os.environ.get("TTT_MAX_CHUNKS", 0)) + skip_sliding_window = bool(int(os.environ.get("SKIP_SLIDING_WINDOW", "0"))) + use_hedge_mixer = bool(int(os.environ.get("USE_HEDGE_MIXER", "1"))) + mixer_eta = float(os.environ.get("MIXER_ETA", 0.1)) + mixer_min_tokens = int(os.environ.get("MIXER_MIN_TOKENS", 10000)) + +class HedgeMixer: + """5-expert online mixer: neural, unigram, bigram, trigram(hashed), entropy.""" + def __init__(self, vocab_size: int, device: torch.device, eta: float = 0.1, + tri_buckets: int = 65536, min_tokens: int = 10000): + self.V = vocab_size + self.K = 5 + self.eta = eta + self.min_tokens = min_tokens + self.device = device + self.log_weights = torch.zeros(self.K, device=device) + self.log_weights[0] = 2.0 + self.tokens_seen = 0 + self.uni_counts = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.uni_total = 0.0 + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_row_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.tri_buckets = tri_buckets + self.tri_counts = torch.zeros(tri_buckets, vocab_size, device=device, dtype=torch.float32) + self.tri_row_totals = torch.zeros(tri_buckets, device=device, dtype=torch.float32) + self._smooth = 0.1 + def update(self, tokens: Tensor): + t = tokens.to(self.device).long() + n = t.numel() + self.tokens_seen += n + ones = torch.ones(n, device=self.device, dtype=torch.float32) + self.uni_counts.scatter_add_(0, t, ones) + self.uni_total += n + if n >= 2: + ctx = t[:-1]; nxt = t[1:] + bi_idx = ctx * self.V + nxt + self.bi_counts.reshape(-1).scatter_add_(0, bi_idx, ones[:n-1]) + self.bi_row_totals.scatter_add_(0, ctx, ones[:n-1]) + if n >= 3: + prev2 = t[:-2]; prev1 = t[1:-1]; nxt3 = t[2:] + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.tri_buckets + tri_idx = tri_ctx * self.V + nxt3 + self.tri_counts.reshape(-1).scatter_add_(0, tri_idx, ones[:n-2]) + self.tri_row_totals.scatter_add_(0, tri_ctx, ones[:n-2]) + def get_expert_nll(self, logits: Tensor, x_batch: Tensor, y_batch: Tensor, + wlens: list[int], temperature: float = 1.0) -> Tensor: + bsz, slen, V = logits.shape + if temperature != 1.0: + logits = logits / temperature + log_probs_neural = F.log_softmax(logits.float(), dim=-1) + neural_nll = -log_probs_neural.gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + s = self._smooth + uni_probs = (self.uni_counts + s) / (self.uni_total + s * self.V) + uni_nll = -uni_probs.log()[y_batch] + prev_tokens = x_batch + bi_row = self.bi_counts[prev_tokens] + bi_row_sum = self.bi_row_totals[prev_tokens].unsqueeze(-1) + bi_probs = (bi_row + s) / (bi_row_sum + s * self.V) + bi_nll = -bi_probs.log().gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + tri_nll = torch.full((bsz, slen), 10.0, device=self.device) + if slen >= 2: + prev2 = torch.zeros_like(x_batch) + prev2[:, 1:] = x_batch[:, :-1] + prev1 = x_batch + tri_ctx = ((prev2 * 36313) ^ (prev1 * 27191)) % self.tri_buckets + tri_row = self.tri_counts[tri_ctx] + tri_row_sum = self.tri_row_totals[tri_ctx].unsqueeze(-1) + tri_probs = (tri_row + s) / (tri_row_sum + s * self.V) + tri_nll = -tri_probs.log().gather(-1, y_batch.unsqueeze(-1)).squeeze(-1) + probs_neural = log_probs_neural.exp() + entropy = -(probs_neural * log_probs_neural).sum(dim=-1) + return torch.stack([neural_nll, uni_nll, bi_nll, tri_nll, entropy], dim=-1) + def mix_and_score(self, expert_nll: Tensor) -> Tensor: + log_w = self.log_weights - self.log_weights.logsumexp(0) + mixed_lp = (-expert_nll + log_w).logsumexp(dim=-1) + return -mixed_lp + def update_weights(self, expert_nll: Tensor, mask: Tensor): + masked_nll = expert_nll * mask.unsqueeze(-1).float() + expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) + self.log_weights -= self.eta * expert_mean_loss + +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("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.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,vrl_scales", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +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_qat: bool = True + _soft_round_temp: float = 1.0 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._soft_round_qat: + w32 = self.weight.float() + row_max = w32.detach().abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_s = w32 / scale[:, None] + residual = w_s - w_s.detach().round() + temp = CastedLinear._soft_round_temp + w_soft = w_s.detach().round() + 0.5 * torch.tanh(residual / temp) + w = (w_soft.clamp(-32, 31) * scale[:, None]).to(x.dtype) + else: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + ): + 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 + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, 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 _FA_VERSION == 3: + y = _fa_func(q, k, v, causal=True) + elif _FA_VERSION == 2: + y = _fa_func(q.bfloat16(), k.bfloat16(), v.bfloat16(), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, enable_gqa=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + 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, leaky: bool = False): + 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 + self._neg_slope = 0.5 if leaky else 0.0 + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), self._neg_slope) + return self.proj(x.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, + **kwargs, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=kwargs.get("gated_attention", False)) + self.mlp = MLP(dim, mlp_mult, leaky=kwargs.get("leaky", False)) + 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, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + vrl_enabled: bool = False, + leaky_relu: bool = False, + gated_attention: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + 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, + leaky=leaky_relu, + gated_attention=gated_attention, + ) + 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() + if self.vrl_enabled: + self.vrl_scales = nn.ParameterList( + [nn.Parameter(torch.zeros(1, dtype=torch.float32)) for _ in range(num_layers - 1)] + ) + else: + self.vrl_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + 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 = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + 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) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + + def forward_logits(self, input_ids: Tensor) -> Tensor: + 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 = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + 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) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + 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_ttt( + args, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """#606-style TTT + Hedge Mixer (safe: deferred weight updates between chunks).""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + if args.ttt_max_chunks > 0: + num_chunks = min(num_chunks, args.ttt_max_chunks) + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + if ci < num_chunks: + chunk_windows[ci].append(ws) + log0(f"ttt606:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={sum(len(w) for w in chunk_windows)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks} optimizer={args.ttt_optimizer} " + f"polyak={args.use_polyak}({args.polyak_decay}) byte_weighted={args.byte_weighted_ttt} " + f"adaptive_lr={args.adaptive_lr}(max={args.adaptive_lr_max}) temp={args.ttt_temperature}") + 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) + num_blocks = len(base_model.blocks) + unfrozen_block_start = max(0, num_blocks - args.ttt_freeze_blocks) if args.ttt_freeze_blocks > 0 else 0 + ttt_params = [] + for name, p in base_model.named_parameters(): + unfreeze = False + if args.ttt_freeze_blocks <= 0: + unfreeze = True + elif "norm" in name or "scale" in name or "lm_head" in name or "tok_emb" in name: + unfreeze = True + else: + for bi in range(unfrozen_block_start, num_blocks): + if f"blocks.{bi}." in name: + unfreeze = True + break + if unfreeze: + p.requires_grad_(True) + ttt_params.append(p) + else: + p.requires_grad_(False) + log0(f"ttt606:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + polyak_state: dict[str, Tensor] | None = None + if args.use_polyak: + polyak_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + mixer: HedgeMixer | None = None + if args.use_hedge_mixer: + mixer = HedgeMixer(args.vocab_size, device, eta=args.mixer_eta, min_tokens=args.mixer_min_tokens) + log0(f"hedge_mixer:enabled K={mixer.K} eta={mixer.eta} min_tokens={mixer.min_tokens}") + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + raw_state: dict[str, Tensor] | None = None + if polyak_state is not None: + raw_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + for n, p in base_model.named_parameters(): + if n in polyak_state: + p.data.copy_(polyak_state[n]) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + if mixer is not None and mixer.tokens_seen >= mixer.min_tokens: + expert_nll = mixer.get_expert_nll(logits, x_batch, y_batch, wlens, args.ttt_temperature) + nll = mixer.mix_and_score(expert_nll) + scored_mask = torch.zeros(bsz, seq_len, device=device, dtype=torch.bool) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_mask[i, s:wlen] = True + if not hasattr(mixer, '_chunk_expert_nll'): + mixer._chunk_expert_nll = [] + mixer._chunk_masks = [] + mixer._chunk_expert_nll.append(expert_nll.detach()) + mixer._chunk_masks.append(scored_mask) + else: + if args.ttt_temperature != 1.0: + logits = logits / args.ttt_temperature + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if mixer is not None: + if hasattr(mixer, '_chunk_expert_nll') and mixer._chunk_expert_nll: + all_enll = torch.cat(mixer._chunk_expert_nll, dim=0) + all_masks = torch.cat(mixer._chunk_masks, dim=0) + mixer.update_weights(all_enll, all_masks) + mixer._chunk_expert_nll = [] + mixer._chunk_masks = [] + chunk_tokens = val_tokens[chunk_start:chunk_end].to(device) + mixer.update(chunk_tokens) + if raw_state is not None: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if args.adaptive_lr: + progress = min(ci / (num_chunks * 0.3), 1.0) + lr_mult = 1.0 + (args.adaptive_lr_max - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + distributed = dist.is_available() and dist.is_initialized() + my_seq_s = (chunk_seqs * rank) // world_size if distributed else 0 + my_seq_e = (chunk_seqs * (rank + 1)) // world_size if distributed else chunk_seqs + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_t = base_model.forward_logits(x) + if args.byte_weighted_ttt: + per_tok_nll = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), reduction="none", + ) + byte_weights = base_bytes_lut[y.reshape(-1)].float() + byte_weights = byte_weights / byte_weights.mean().clamp(min=1e-6) + loss = (per_tok_nll * byte_weights).mean() + else: + loss = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), + ) + loss.backward() + if distributed and world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if polyak_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in polyak_state: + polyak_state[n].lerp_(p.data, 1.0 - args.polyak_decay) + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +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 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.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 _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +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 + + _gpu_name = torch.cuda.get_device_name(0) + _is_high_end = "H100" in _gpu_name or "A100" in _gpu_name + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + if _is_high_end: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(True) + enable_flash_sdp(True) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + log0(f"flash_attn_version:{_FA_VERSION} gpu:{_gpu_name} high_end:{_is_high_end}") + 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 + CastedLinear._soft_round_qat = args.soft_round_qat + CastedLinear._soft_round_temp = args.soft_round_temp_start + qat_start_step = 0 if args.qat_enabled else -1 + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, + leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if distributed: + torch._dynamo.config.optimize_ddp = False + 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) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + 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) + if base_model.vrl_enabled: + for s in base_model.vrl_scales: + scalar_params.append(s) + 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()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"sdp_backends:high_end={_is_high_end}") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0(f"vrl_enabled:{args.vrl_enabled} leaky_relu:{args.leaky_relu} ttt_enabled:{args.ttt_enabled}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0 and not args.eval_only: + 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) + if args.eval_only: + log0(f"eval_only:loading checkpoint from {args.checkpoint_path}") + ckpt_state = torch.load(args.checkpoint_path, map_location="cpu") + base_model.load_state_dict(ckpt_state, strict=True) + log0(f"eval_only:checkpoint loaded, {sum(p.numel() for p in base_model.parameters())} params") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + log0(f"eval_only:quantized model size: {len(quant_blob)} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_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, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, gated_attention=args.gated_attention, + ).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 = effective_eval_seq_len + if not args.skip_sliding_window and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0(f"eval_only:sliding_window val_bpb:{sw_val_bpb:.4f} stride:{args.eval_stride} " + f"eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms") + elif args.skip_sliding_window: + log0("eval_only:skipping sliding window eval") + if args.ttt_enabled: + log0(f"eval_only:ttt lr={args.ttt_lr} epochs={args.ttt_epochs} " + f"chunk={args.ttt_chunk_tokens} freeze_blocks={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_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, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"eval_only:ttt val_bpb:{ttt_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + if distributed: + dist.destroy_process_group() + return + + 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 = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + qat_start_step = step + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + if CastedLinear._qat_enabled and CastedLinear._soft_round_qat and qat_start_step >= 0: + qat_total = max(args.iterations - qat_start_step, 1) + qat_progress = min((step - qat_start_step) / qat_total, 1.0) + log_start = math.log(args.soft_round_temp_start) + log_end = math.log(args.soft_round_temp_end) + CastedLinear._soft_round_temp = math.exp(log_start + qat_progress * (log_end - log_start)) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + 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 = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_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, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, gated_attention=args.gated_attention, + ).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) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + + if args.ttt_enabled: + log0("ttt:starting test-time training evaluation") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + 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, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0( + f"ttt_eval val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"ttt_eval_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.11.10 (main, Sep 7 2024, 18:35:41) [GCC 11.4.0] +Running PyTorch 2.11.0+cu128 +Wed Mar 25 14:31:24 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 565.57.01 Driver Version: 565.57.01 CUDA Version: 12.7 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 37C P0 150W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:2A:00.0 Off | 0 | +| N/A 41C P0 153W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3A:00.0 Off | 0 | +| N/A 43C P0 159W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 38C P0 154W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9A:00.0 Off | 0 | +| N/A 37C P0 149W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:AB:00.0 Off | 0 | +| N/A 42C P0 149W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BA:00.0 Off | 0 | +| N/A 40C P0 149W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 38C P0 153W / 700W | 1502MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +flash_attn_version:0 gpu:NVIDIA H100 80GB HBM3 high_end:True +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26993766 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:high_end=True +attention_mode:gqa num_heads:8 num_kv_heads:4 +vrl_enabled:True leaky_relu:True ttt_enabled:True +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9318 train_time:148ms step_avg:147.68ms +step:2/20000 train_loss:8.6411 train_time:241ms step_avg:120.29ms +step:3/20000 train_loss:7.8696 train_time:334ms step_avg:111.38ms +step:4/20000 train_loss:7.2557 train_time:428ms step_avg:107.05ms +step:5/20000 train_loss:7.0230 train_time:522ms step_avg:104.39ms +step:6/20000 train_loss:6.9000 train_time:620ms step_avg:103.29ms +step:7/20000 train_loss:6.7652 train_time:716ms step_avg:102.34ms +step:8/20000 train_loss:6.6954 train_time:810ms step_avg:101.31ms +step:9/20000 train_loss:6.4121 train_time:904ms step_avg:100.48ms +step:10/20000 train_loss:6.0757 train_time:998ms step_avg:99.82ms +step:500/20000 train_loss:2.3833 train_time:46087ms step_avg:92.17ms +step:1000/20000 train_loss:2.2583 train_time:92208ms step_avg:92.21ms +step:1500/20000 train_loss:2.2062 train_time:138345ms step_avg:92.23ms +step:2000/20000 train_loss:2.0493 train_time:184513ms step_avg:92.26ms +step:2500/20000 train_loss:2.1584 train_time:230682ms step_avg:92.27ms +step:3000/20000 train_loss:2.1477 train_time:280652ms step_avg:93.55ms +step:3500/20000 train_loss:2.1564 train_time:331609ms step_avg:94.75ms +step:4000/20000 train_loss:1.9445 train_time:383278ms step_avg:95.82ms +step:4000/20000 val_loss:2.0338 val_bpb:1.2045 train_time:383287ms step_avg:95.82ms +step:4500/20000 train_loss:2.0893 train_time:434336ms step_avg:96.52ms +step:5000/20000 train_loss:2.0658 train_time:486036ms step_avg:97.21ms +swa:start step:5500 +step:5500/20000 train_loss:1.9752 train_time:537167ms step_avg:97.67ms +late_qat:enabled step:5606 scale:0.1499 +step:6000/20000 train_loss:1.8975 train_time:589305ms step_avg:98.22ms +step:6104/20000 val_loss:1.9304 val_bpb:1.1433 train_time:600411ms step_avg:98.36ms +stopping_early: wallclock_cap train_time:600411ms step:6104/20000 +peak memory allocated: 22016 MiB reserved: 22362 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9290 val_bpb:1.1425 eval_time:2064ms +Serialized model: 106181533 bytes +Code size: 94378 bytes +Serialized model int6+lzma: 15905672 bytes +Total submission size int6+lzma: 16000050 bytes +final_int6_roundtrip val_loss:1.9416 val_bpb:1.1499 eval_time:18497ms +final_int6_roundtrip_exact val_loss:1.94159290 val_bpb:1.14992068 +final_int6_sliding_window val_loss:1.9017 val_bpb:1.1263 stride:64 eval_time:100854ms +final_int6_sliding_window_exact val_loss:1.90171167 val_bpb:1.12630375 +ttt:starting test-time training evaluation +ttt606:start chunks=1893 chunk_tokens=32768 total_windows=969088 stride=64 ttt_lr=0.0005 ttt_epochs=3 freeze_blocks=2 optimizer=adamw polyak=True(0.998) byte_weighted=True adaptive_lr=True(max=3.0) temp=0.98 +ttt606:params unfrozen=5256222 frozen=21737544 +hedge_mixer:enabled K=5 eta=0.1 min_tokens=10000 + ttt_chunk [1/1893] bpb=1.168129 time=2.7s + ttt_chunk [11/1893] bpb=1.103920 time=5.3s + ttt_chunk [21/1893] bpb=1.063779 time=7.8s + ttt_chunk [31/1893] bpb=1.051392 time=10.4s + ttt_chunk [41/1893] bpb=1.034915 time=12.9s + ttt_chunk [51/1893] bpb=1.026711 time=15.4s + ttt_chunk [61/1893] bpb=1.030645 time=17.9s + ttt_chunk [71/1893] bpb=1.027674 time=20.5s + ttt_chunk [81/1893] bpb=1.025878 time=23.0s + ttt_chunk [91/1893] bpb=1.025990 time=25.6s + ttt_chunk [101/1893] bpb=1.028326 time=28.1s + ttt_chunk [111/1893] bpb=1.029977 time=30.6s + ttt_chunk [121/1893] bpb=1.023106 time=33.1s + ttt_chunk [131/1893] bpb=1.022599 time=35.7s + ttt_chunk [141/1893] bpb=1.027312 time=38.2s + ttt_chunk [151/1893] bpb=1.028392 time=40.8s + ttt_chunk [161/1893] bpb=1.027186 time=43.5s + ttt_chunk [171/1893] bpb=1.030331 time=46.1s + ttt_chunk [181/1893] bpb=1.031935 time=48.6s + ttt_chunk [191/1893] bpb=1.037781 time=51.1s + ttt_chunk [201/1893] bpb=1.036042 time=53.6s + ttt_chunk [211/1893] bpb=1.033527 time=56.2s + ttt_chunk [221/1893] bpb=1.034428 time=58.7s + ttt_chunk [231/1893] bpb=1.032793 time=61.2s + ttt_chunk [241/1893] bpb=1.032768 time=63.8s + ttt_chunk [251/1893] bpb=1.031678 time=66.3s + ttt_chunk [261/1893] bpb=1.028408 time=68.8s + ttt_chunk [271/1893] bpb=1.027114 time=71.4s + ttt_chunk [281/1893] bpb=1.028005 time=73.9s + ttt_chunk [291/1893] bpb=1.029290 time=76.5s + ttt_chunk [301/1893] bpb=1.029702 time=79.0s + ttt_chunk [311/1893] bpb=1.031474 time=81.5s + ttt_chunk [321/1893] bpb=1.033428 time=84.1s + ttt_chunk [331/1893] bpb=1.033718 time=86.6s + ttt_chunk [341/1893] bpb=1.032800 time=89.2s + ttt_chunk [351/1893] bpb=1.035235 time=91.8s + ttt_chunk [361/1893] bpb=1.036036 time=94.3s + ttt_chunk [371/1893] bpb=1.035861 time=96.9s + ttt_chunk [381/1893] bpb=1.036876 time=99.4s + ttt_chunk [391/1893] bpb=1.037283 time=102.0s + ttt_chunk [401/1893] bpb=1.036092 time=104.6s + ttt_chunk [411/1893] bpb=1.035772 time=107.1s + ttt_chunk [421/1893] bpb=1.035866 time=109.6s + ttt_chunk [431/1893] bpb=1.036612 time=112.2s + ttt_chunk [441/1893] bpb=1.037689 time=114.8s + ttt_chunk [451/1893] bpb=1.039155 time=117.3s + ttt_chunk [461/1893] bpb=1.039054 time=119.9s + ttt_chunk [471/1893] bpb=1.040528 time=122.4s + ttt_chunk [481/1893] bpb=1.041153 time=125.0s + ttt_chunk [491/1893] bpb=1.041071 time=127.5s + ttt_chunk [501/1893] bpb=1.041340 time=130.1s + ttt_chunk [511/1893] bpb=1.041531 time=132.7s + ttt_chunk [521/1893] bpb=1.040209 time=135.2s + ttt_chunk [531/1893] bpb=1.042190 time=137.8s + ttt_chunk [541/1893] bpb=1.043232 time=140.3s + ttt_chunk [551/1893] bpb=1.043034 time=142.9s + ttt_chunk [561/1893] bpb=1.044418 time=145.4s + ttt_chunk [571/1893] bpb=1.044232 time=148.0s + ttt_chunk [581/1893] bpb=1.044174 time=150.5s + ttt_chunk [591/1893] bpb=1.044440 time=153.1s + ttt_chunk [601/1893] bpb=1.045435 time=155.6s + ttt_chunk [611/1893] bpb=1.046031 time=158.2s + ttt_chunk [621/1893] bpb=1.046526 time=160.7s + ttt_chunk [631/1893] bpb=1.047820 time=163.3s + ttt_chunk [641/1893] bpb=1.048080 time=165.8s + ttt_chunk [651/1893] bpb=1.048668 time=168.4s + ttt_chunk [661/1893] bpb=1.048615 time=170.9s + ttt_chunk [671/1893] bpb=1.049353 time=173.5s + ttt_chunk [681/1893] bpb=1.050351 time=176.0s + ttt_chunk [691/1893] bpb=1.051714 time=178.6s + ttt_chunk [701/1893] bpb=1.051695 time=181.1s + ttt_chunk [711/1893] bpb=1.052345 time=183.7s + ttt_chunk [721/1893] bpb=1.052239 time=186.2s + ttt_chunk [731/1893] bpb=1.052731 time=188.8s + ttt_chunk [741/1893] bpb=1.053145 time=191.3s + ttt_chunk [751/1893] bpb=1.053226 time=193.8s + ttt_chunk [761/1893] bpb=1.053376 time=196.4s + ttt_chunk [771/1893] bpb=1.053501 time=199.0s + ttt_chunk [781/1893] bpb=1.054492 time=201.5s + ttt_chunk [791/1893] bpb=1.054253 time=204.0s + ttt_chunk [801/1893] bpb=1.054692 time=206.6s + ttt_chunk [811/1893] bpb=1.054761 time=209.1s + ttt_chunk [821/1893] bpb=1.054779 time=211.7s + ttt_chunk [831/1893] bpb=1.054834 time=214.2s + ttt_chunk [841/1893] bpb=1.054409 time=216.8s + ttt_chunk [851/1893] bpb=1.054398 time=219.3s + ttt_chunk [861/1893] bpb=1.054236 time=221.8s + ttt_chunk [871/1893] bpb=1.054678 time=224.4s + ttt_chunk [881/1893] bpb=1.055060 time=227.0s + ttt_chunk [891/1893] bpb=1.054857 time=229.5s + ttt_chunk [901/1893] bpb=1.054794 time=232.0s + ttt_chunk [911/1893] bpb=1.055076 time=234.6s + ttt_chunk [921/1893] bpb=1.055599 time=237.2s + ttt_chunk [931/1893] bpb=1.055640 time=239.7s + ttt_chunk [941/1893] bpb=1.055444 time=242.2s + ttt_chunk [951/1893] bpb=1.055943 time=244.8s + ttt_chunk [961/1893] bpb=1.056053 time=247.4s + ttt_chunk [971/1893] bpb=1.056784 time=249.9s + ttt_chunk [981/1893] bpb=1.056882 time=252.4s + ttt_chunk [991/1893] bpb=1.056946 time=255.0s + ttt_chunk [1001/1893] bpb=1.056861 time=257.6s + ttt_chunk [1011/1893] bpb=1.056736 time=260.1s + ttt_chunk [1021/1893] bpb=1.056981 time=262.6s + ttt_chunk [1031/1893] bpb=1.057391 time=265.2s + ttt_chunk [1041/1893] bpb=1.057011 time=267.7s + ttt_chunk [1051/1893] bpb=1.056680 time=270.3s + ttt_chunk [1061/1893] bpb=1.056699 time=272.8s + ttt_chunk [1071/1893] bpb=1.057232 time=275.4s + ttt_chunk [1081/1893] bpb=1.057416 time=277.9s + ttt_chunk [1091/1893] bpb=1.058024 time=280.4s + ttt_chunk [1101/1893] bpb=1.057965 time=283.0s + ttt_chunk [1111/1893] bpb=1.057696 time=285.5s + ttt_chunk [1121/1893] bpb=1.057414 time=288.1s + ttt_chunk [1131/1893] bpb=1.057279 time=290.6s + ttt_chunk [1141/1893] bpb=1.056975 time=293.2s + ttt_chunk [1151/1893] bpb=1.056869 time=295.7s + ttt_chunk [1161/1893] bpb=1.056483 time=298.3s + ttt_chunk [1171/1893] bpb=1.056595 time=300.8s + ttt_chunk [1181/1893] bpb=1.055808 time=303.4s + ttt_chunk [1191/1893] bpb=1.055560 time=305.9s + ttt_chunk [1201/1893] bpb=1.055782 time=308.5s + ttt_chunk [1211/1893] bpb=1.055281 time=311.0s + ttt_chunk [1221/1893] bpb=1.054890 time=313.6s + ttt_chunk [1231/1893] bpb=1.054556 time=316.1s + ttt_chunk [1241/1893] bpb=1.054072 time=318.7s + ttt_chunk [1251/1893] bpb=1.053422 time=321.2s + ttt_chunk [1261/1893] bpb=1.053236 time=323.8s + ttt_chunk [1271/1893] bpb=1.052824 time=326.3s + ttt_chunk [1281/1893] bpb=1.052583 time=328.9s + ttt_chunk [1291/1893] bpb=1.052334 time=331.4s + ttt_chunk [1301/1893] bpb=1.051651 time=334.0s + ttt_chunk [1311/1893] bpb=1.051239 time=336.5s + ttt_chunk [1321/1893] bpb=1.050854 time=339.0s + ttt_chunk [1331/1893] bpb=1.050717 time=341.6s + ttt_chunk [1341/1893] bpb=1.050512 time=344.2s + ttt_chunk [1351/1893] bpb=1.050335 time=346.7s + ttt_chunk [1361/1893] bpb=1.050317 time=349.3s + ttt_chunk [1371/1893] bpb=1.050082 time=351.9s + ttt_chunk [1381/1893] bpb=1.049970 time=354.4s + ttt_chunk [1391/1893] bpb=1.049539 time=357.0s + ttt_chunk [1401/1893] bpb=1.049440 time=359.5s + ttt_chunk [1411/1893] bpb=1.049465 time=362.0s + ttt_chunk [1421/1893] bpb=1.049625 time=364.6s + ttt_chunk [1431/1893] bpb=1.049236 time=367.1s + ttt_chunk [1441/1893] bpb=1.049540 time=369.6s + ttt_chunk [1451/1893] bpb=1.049832 time=372.4s + ttt_chunk [1461/1893] bpb=1.049419 time=375.4s + ttt_chunk [1471/1893] bpb=1.050251 time=378.3s + ttt_chunk [1481/1893] bpb=1.049759 time=380.9s + ttt_chunk [1491/1893] bpb=1.049542 time=383.4s + ttt_chunk [1501/1893] bpb=1.049404 time=386.0s + ttt_chunk [1511/1893] bpb=1.049251 time=388.5s + ttt_chunk [1521/1893] bpb=1.049203 time=391.1s + ttt_chunk [1531/1893] bpb=1.048659 time=393.6s + ttt_chunk [1541/1893] bpb=1.048384 time=396.1s + ttt_chunk [1551/1893] bpb=1.048614 time=398.7s + ttt_chunk [1561/1893] bpb=1.048533 time=401.2s + ttt_chunk [1571/1893] bpb=1.048320 time=403.7s + ttt_chunk [1581/1893] bpb=1.048366 time=406.2s + ttt_chunk [1591/1893] bpb=1.048198 time=408.7s + ttt_chunk [1601/1893] bpb=1.048306 time=411.3s + ttt_chunk [1611/1893] bpb=1.048165 time=413.8s + ttt_chunk [1621/1893] bpb=1.047693 time=416.4s + ttt_chunk [1631/1893] bpb=1.047858 time=418.9s + ttt_chunk [1641/1893] bpb=1.047762 time=421.5s + ttt_chunk [1651/1893] bpb=1.047579 time=424.0s + ttt_chunk [1661/1893] bpb=1.047347 time=426.6s + ttt_chunk [1671/1893] bpb=1.047652 time=429.1s + ttt_chunk [1681/1893] bpb=1.047706 time=431.7s + ttt_chunk [1691/1893] bpb=1.047457 time=434.2s + ttt_chunk [1701/1893] bpb=1.047542 time=436.8s + ttt_chunk [1711/1893] bpb=1.047475 time=439.3s + ttt_chunk [1721/1893] bpb=1.047314 time=441.9s + ttt_chunk [1731/1893] bpb=1.047122 time=444.4s + ttt_chunk [1741/1893] bpb=1.046913 time=446.9s + ttt_chunk [1751/1893] bpb=1.046634 time=449.5s + ttt_chunk [1761/1893] bpb=1.046654 time=452.0s + ttt_chunk [1771/1893] bpb=1.046463 time=454.6s + ttt_chunk [1781/1893] bpb=1.046411 time=457.1s + ttt_chunk [1791/1893] bpb=1.045892 time=459.7s + ttt_chunk [1801/1893] bpb=1.045693 time=462.2s + ttt_chunk [1811/1893] bpb=1.045508 time=464.8s + ttt_chunk [1821/1893] bpb=1.045470 time=467.4s + ttt_chunk [1831/1893] bpb=1.044806 time=469.9s + ttt_chunk [1841/1893] bpb=1.044695 time=472.5s + ttt_chunk [1851/1893] bpb=1.044402 time=475.0s + ttt_chunk [1861/1893] bpb=1.043958 time=477.7s + ttt_chunk [1871/1893] bpb=1.043804 time=480.2s + ttt_chunk [1881/1893] bpb=1.043271 time=482.8s + ttt_chunk [1891/1893] bpb=1.042948 time=485.3s + ttt_chunk [1893/1893] bpb=1.042976 time=486.0s +ttt_sliding:done val_loss=1.757687 val_bpb=1.041004 elapsed=486.0s +ttt_eval val_loss:1.7577 val_bpb:1.0410 eval_time:486383ms +ttt_eval_exact val_loss:1.75768689 val_bpb:1.04100394