diff --git a/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/README.md b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/README.md new file mode 100644 index 000000000..03eb0e7af --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/README.md @@ -0,0 +1,64 @@ +# Record: Curriculum Learning + LeakyReLU(0.9)^2 + 7-gram Backoff (val_bpb=0.9633) + +**val_bpb = 0.9633** (seed 42, additional seeds pending compute grant) | **15.56 MB** | 8xH100 SXM, 600s + +## Approach + +Built on PR #753 (Podracing II) with two additions: + +### 1. Curriculum Learning (Shard Reordering) + +Training shards reordered by model perplexity — hardest shards first. Based on PR #650 by @abaybektursun which demonstrated -0.003 BPB from shard ordering alone. Zero code change, just environment variable. + +### 2. LeakyReLU(0.9)^2 Slope Optimization + +Following @MatoTeziTanka's controlled slope sweep on issue #140, replaced standard slope=0.5 with slope=0.9. The sweep showed monotonic improvement from 0.1 to 0.9, with 0.9 giving -0.013 BPB vs 0.5 on the same stack. + +## Results + +| Metric | Value | +|--------|-------| +| Sliding window (stride=64) | 1.1216 | +| **Sliding + 7-gram backoff** | **0.9633** | +| Legal TTT (score-first, 3ep) | 1.1216 | +| Artifact | 15,560,351 bytes | +| Steps | 6,647 at 90.3ms/step | +| Training time | 600s | + +## Architecture (from PR #753) + +- 11L, 512d, GQA 8/4, MLP 3x +- LeakyReLU(0.9)^2 activation +- XSA on all 11 layers +- BigramHash, SmearGate, SWA, EMA +- Int6 QAT + GPTQ (within training budget, issue #677 compliant) +- 7-gram backoff eval cache (backward-looking, no weight updates) + +## Eval-time Techniques + +**7-gram backoff cache** (from PR #753): Multi-order n-gram model built from already-scored tokens. Linear interpolation with entropy-adaptive alpha. Fully backward-looking — each token scored before its statistics enter the cache. + +**Legal score-first TTT** (from PR #753): SGD with 3 epochs, freeze last 2 blocks. Every token scored under inference_mode before any weight update. + +## Reproduction + +```bash +SEED=42 bash run.sh +``` + +Environment variables set in run.sh: +- `SHARD_ORDER=44,63,65,42,...` (curriculum learning) +- `MLP_LEAKY_SLOPE=0.9` +- `NGRAM_EVAL_ORDER=7` + +## Acknowledgments + +- @newjordan (PR #753, Podracing II base) +- @abaybektursun (PR #650, curriculum learning / shard reordering) +- @MatoTeziTanka (LeakyReLU slope sweep, issue #140) +- @Asukabot0 (PR #715/#727, n-gram backoff technique) + +## Status + +1 seed submitted. 2 additional seeds pending OpenAI compute grant ($1000 applied). +Previously PR #486 (formerly #2 on leaderboard, TrigramHash originator). $339 personal compute spent. diff --git a/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/run.sh b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/run.sh new file mode 100644 index 000000000..fd0d13e2e --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/run.sh @@ -0,0 +1,22 @@ +#!/bin/bash +set -euo pipefail +export PYTHONUNBUFFERED=1 + +SEED="${SEED:-42}" + +# Production-ready: PR #753 base + curriculum learning +export SEED +export SHARD_ORDER="${SHARD_ORDER:-44,63,65,42,18,67,30,69,61,3,13,19,50,49,56,45,73,79,57,32,28,68,66,34,46,38,17,77,0,14,26,74,59,62,41,9,58,22,78,4,48,8,12,27,75,36,16,43,52,15,33,47,25,55,54,23,37,51,31,21,60,1,20,72,24,53,39,35,71,76,40,5,10,2,7,6,70,11,64,29}" +# N-gram backoff defaults from PR #753 +export NGRAM_EVAL_ORDER="${NGRAM_EVAL_ORDER:-7}" +# LeakyReLU slope 0.9 > 0.5 (MatoTeziTanka sweep, -0.013 BPP) +export MLP_LEAKY_SLOPE="${MLP_LEAKY_SLOPE:-0.9}" + +NGPU=$(nvidia-smi -L 2>/dev/null | wc -l) +echo "GPUs: $NGPU | Seed: $SEED | Ngram: $NGRAM_EVAL_ORDER | Shard order: ${SHARD_ORDER:+yes}" + +if [ "$NGPU" -gt 1 ]; then + torchrun --standalone --nproc_per_node="$NGPU" train_gpt.py +else + python train_gpt.py +fi diff --git a/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/submission.json b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/submission.json new file mode 100644 index 000000000..e69a26f72 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/submission.json @@ -0,0 +1,17 @@ +{ + "name": "Curriculum Learning + LeakyReLU(0.9)² + 7-gram Backoff", + "author": "ndokutovich", + "github_id": "ndokutovich", + "val_bpb": 0.9633, + "val_loss": 1.6265, + "bytes_total": 15560351, + "artifact_bytes": 15560351, + "training_time_seconds": 600, + "eval_time_seconds": 131, + "hardware": "8xH100 SXM", + "seed": 42, + "num_seeds": 1, + "date": "2026-03-25", + "blurb": "PR #753 base + curriculum learning (hardest-first shard reorder, PR #650) + LeakyReLU(0.9)² slope optimization (MatoTeziTanka sweep) + 7-gram backoff eval cache. 1 seed, 2 additional pending compute grant.", + "notes": "Previously PR #486 (formerly #2 on leaderboard, TrigramHash originator). $360 personal compute spent." +} diff --git a/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/train_gpt.py b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/train_gpt.py new file mode 100644 index 000000000..1d5014396 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/train_gpt.py @@ -0,0 +1,2145 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func +except ImportError: + def flash_attn_3_func(q, k, v, causal=False): + # q: (B, T, Hq, D), k/v: (B, T, Hkv, D) — expand KV for GQA + q2 = q.transpose(1, 2) # (B, Hq, T, D) + k2 = k.transpose(1, 2) # (B, Hkv, T, D) + v2 = v.transpose(1, 2) + if k2.size(1) != q2.size(1): + rep = q2.size(1) // k2.size(1) + k2 = k2.repeat_interleave(rep, dim=1) + v2 = v2.repeat_interleave(rep, dim=1) + out = torch.nn.functional.scaled_dot_product_attention(q2, k2, v2, is_causal=causal) + return out.transpose(1, 2) +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)) + mlp_act = os.environ.get("MLP_ACT", "relu_sq").lower() + mlp_leaky_slope = float(os.environ.get("MLP_LEAKY_SLOPE", 0.5)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 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)) # tighter: collect more recent checkpoints + 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", 11)) # XSA on ALL 11 layers + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.5)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + # F1 capacity add-on: low-rank correction head (active at inference). + # Approx extra params ~= rank * (model_dim + vocab_size). + f1_corr_rank = int(os.environ.get("F1_CORR_RANK", 0)) + f1_corr_scale_init = float(os.environ.get("F1_CORR_SCALE_INIT", 0.10)) + # Post-train self-distillation: EMA teacher -> student. + distill_enabled = bool(int(os.environ.get("DISTILL_ENABLED", "0"))) + distill_steps = int(os.environ.get("DISTILL_STEPS", 24)) + distill_lr_factor = float(os.environ.get("DISTILL_LR_FACTOR", 0.02)) + distill_temperature = float(os.environ.get("DISTILL_TEMPERATURE", 1.5)) + distill_alpha = float(os.environ.get("DISTILL_ALPHA", 0.60)) + distill_kl_clip = float(os.environ.get("DISTILL_KL_CLIP", 10.0)) + # Legal score-first TTT eval (PR #461 recipe) + ttt_eval_enabled = bool(int(os.environ.get("TTT_EVAL_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 2)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_max_train_chunks = int(os.environ.get("TTT_MAX_TRAIN_CHUNKS", 200)) # stop training after N chunks, keep scoring + ttt_ema_decay = float(os.environ.get("TTT_EMA_DECAY", 0.995)) # EMA decay for TTT weight smoothing (0 = disabled) + ttt_freeze_embed = bool(int(os.environ.get("TTT_FREEZE_EMBED", "1"))) # freeze tok_emb/bigram/ve during TTT + # Optional legal score-first hashed n-gram interpolation at eval time. + # Multi-order backoff (2..max_order) with entropy-adaptive alpha. + # Alpha depends only on model entropy (no target/label access). + ngram_eval_order = int(os.environ.get("NGRAM_EVAL_ORDER", 0)) # 0=off, max order for backoff + ngram_eval_min_order = int(os.environ.get("NGRAM_EVAL_MIN_ORDER", 2)) # min order for backoff + ngram_eval_alpha = float(os.environ.get("NGRAM_EVAL_ALPHA", 0.30)) # base alpha (or fixed if adaptive off) + ngram_eval_adaptive = bool(int(os.environ.get("NGRAM_EVAL_ADAPTIVE", "1"))) # entropy-adaptive alpha + ngram_eval_alpha_min = float(os.environ.get("NGRAM_EVAL_ALPHA_MIN", 0.05)) # alpha floor (confident model) + ngram_eval_alpha_max = float(os.environ.get("NGRAM_EVAL_ALPHA_MAX", 0.60)) # alpha ceiling (uncertain model) + ngram_eval_entropy_center = float(os.environ.get("NGRAM_EVAL_ENTROPY_CENTER", 4.0)) # sigmoid center + ngram_eval_entropy_scale = float(os.environ.get("NGRAM_EVAL_ENTROPY_SCALE", 2.0)) # sigmoid steepness + ngram_eval_min_count = int(os.environ.get("NGRAM_EVAL_MIN_COUNT", 2)) + ngram_eval_buckets = int(os.environ.get("NGRAM_EVAL_BUCKETS", 4_194_304)) + ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) + compile_enabled = bool(int(os.environ.get("COMPILE_ENABLED", "1"))) + compile_fullgraph = bool(int(os.environ.get("COMPILE_FULLGRAPH", "1"))) +def maybe_torch_compile(obj, args: Hyperparameters): + if not args.compile_enabled: + return obj + return torch.compile(obj, dynamic=False, fullgraph=args.compile_fullgraph) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale", + ).split(",") + if pattern +) +INT8_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 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + # Use 99.95th percentile clipping to match GPTQ export quantizer + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / 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, + ): + 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 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] — broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, 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] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int, mlp_act: str = "relu_sq", mlp_leaky_slope: float = 0.5): + 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.mlp_act = mlp_act + self.mlp_leaky_slope = mlp_leaky_slope + if self.mlp_act not in {"relu_sq", "leaky_relu_sq"}: + raise ValueError(f"Unsupported MLP_ACT '{self.mlp_act}'. Use 'relu_sq' or 'leaky_relu_sq'.") + def forward(self, x: Tensor) -> Tensor: + x = self.fc(x) + if self.mlp_act == "leaky_relu_sq": + x = F.leaky_relu(x, negative_slope=self.mlp_leaky_slope) + else: + x = F.relu(x) + 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, + mlp_act: str = "relu_sq", + mlp_leaky_slope: float = 0.5, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult, mlp_act=mlp_act, mlp_leaky_slope=mlp_leaky_slope) + 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", + mlp_act: str = "relu_sq", + mlp_leaky_slope: float = 0.5, + f1_corr_rank: int = 0, + f1_corr_scale_init: float = 0.10, + ): + 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.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, + mlp_act=mlp_act, + mlp_leaky_slope=mlp_leaky_slope, + ) + 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() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + # Low-rank correction path for extra capacity under size budget. + self.f1_corr_rank = f1_corr_rank + if f1_corr_rank > 0: + self.f1_corr_in = CastedLinear(model_dim, f1_corr_rank, bias=False) + self.f1_corr_out = CastedLinear(f1_corr_rank, vocab_size, bias=False) + self.f1_corr_out._zero_init = True + self.f1_corr_scale = nn.Parameter(torch.tensor(f1_corr_scale_init, dtype=torch.float32)) + else: + self.f1_corr_in = None + self.f1_corr_out = None + self.f1_corr_scale = None + 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: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + if self.f1_corr_in is not None and self.f1_corr_out is not None and self.f1_corr_scale is not None: + corr_hidden = F.silu(self.f1_corr_in(x_flat)) + corr_proj = self.f1_corr_out(corr_hidden) + logits_proj = logits_proj + self.f1_corr_scale.to(dtype=logits_proj.dtype) * corr_proj + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + if self.f1_corr_in is not None and self.f1_corr_out is not None and self.f1_corr_scale is not None: + corr_hidden = F.silu(self.f1_corr_in(x)) + corr_proj = self.f1_corr_out(corr_hidden) + logits_proj = logits_proj + self.f1_corr_scale.to(dtype=logits_proj.dtype) * corr_proj + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = maybe_torch_compile(base_model.forward_logits, args) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def eval_val_sliding_hashed_ngram( + 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, + order: int, + alpha: float, + min_count: int, + buckets: int, + max_seconds: float = 0.0, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float, float]: + """Score-first sliding eval with multi-order backoff n-gram + entropy-adaptive alpha. + + Legal behavior: + - per-token score is computed before that token updates the cache + - alpha depends only on model entropy (no target/label access) + - backoff tries longest context first, falls back to shorter + """ + min_order = max(args.ngram_eval_min_order, 2) + max_order = max(order, min_order) + adaptive = args.ngram_eval_adaptive + alpha_min = args.ngram_eval_alpha_min + alpha_max = args.ngram_eval_alpha_max + ent_center = args.ngram_eval_entropy_center + ent_scale = args.ngram_eval_entropy_scale + + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + all_window_starts = [ws for ws in range(0, total_tokens, stride) if min(ws + seq_len, total_tokens) - ws >= 1] + total_scored_tokens = 0.0 + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + total_scored_tokens += float(max(wlen - s, 0)) + # Distribute windows across ranks + my_s = (len(all_window_starts) * rank) // world_size + my_e = (len(all_window_starts) * (rank + 1)) // world_size + window_starts = all_window_starts[my_s:my_e] + + val_np = val_tokens.numpy() + # Per-order hash tables for backoff + ctx_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + full_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + mask = np.uint64(buckets - 1) + primes = np.array( + [np.uint64(36313), np.uint64(27191), np.uint64(51647), np.uint64(81929), + np.uint64(131071), np.uint64(174763), np.uint64(233017)], + dtype=np.uint64, + ) + + loss_sum = 0.0 + token_count = 0.0 + byte_count = 0.0 + + base_model.eval() + compiled_logits = maybe_torch_compile(base_model.forward_logits, args) + t0 = time.perf_counter() + deadline = (t0 + max_seconds) if max_seconds > 0.0 else None + cutoff_hit = False + with torch.inference_mode(): + for bi in range(0, len(window_starts), batch_seqs): + if deadline is not None and time.perf_counter() >= deadline: + cutoff_hit = True + break + batch_ws = window_starts[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) + logits_f = logits.float() + nll = F.cross_entropy( + logits_f.reshape(-1, logits_f.size(-1)), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + seg_len = wlen - s + if seg_len <= 0: + continue + + seg_nll = nll[i, s:wlen].to(torch.float64).cpu().numpy() + seg_model_p = np.exp(-seg_nll) + + # Entropy-adaptive alpha (uses model output only, not target) + if adaptive: + log_probs = F.log_softmax(logits_f[i, s:wlen], dim=-1) + probs = log_probs.exp() + entropy = -(probs * log_probs).sum(dim=-1).cpu().numpy() # per-token entropy + sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy - ent_center))) + per_token_alpha = alpha_min + (alpha_max - alpha_min) * sig + else: + per_token_alpha = np.full(seg_len, alpha) + + global_j = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + + # Multi-order backoff: try highest order first, fall back + p_ng = np.zeros(seg_len, dtype=np.float64) + ng_matched = np.zeros(seg_len, dtype=np.bool_) + tgt_np = val_np[global_j].astype(np.uint64) + + for n in range(max_order, min_order - 1, -1): + ctx_width = n - 1 + valid = (global_j >= ctx_width) & (~ng_matched) + if not valid.any(): + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_width): + tok = val_np[jv - (ctx_width - k)].astype(np.uint64) + ctx_hash ^= tok * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + full_key = ((ctx_hash ^ (tgt_np[v_idx] * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + + ctx_counts = ctx_tables[n][ctx_key].astype(np.float64) + full_counts = full_tables[n][full_key].astype(np.float64) + has_data = ctx_counts >= float(min_count) + if has_data.any(): + p = np.minimum(full_counts, ctx_counts) / np.maximum(ctx_counts, 1.0) + p = np.clip(p, 0.0, 1.0) + hit_idx = v_idx[has_data] + p_ng[hit_idx] = p[has_data] + ng_matched[hit_idx] = True + + # Mix where n-gram matched + if ng_matched.any(): + m_idx = np.nonzero(ng_matched)[0] + a = per_token_alpha[m_idx] + seg_model_p[m_idx] = (1.0 - a) * seg_model_p[m_idx] + a * p_ng[m_idx] + + seg_nll = -np.log(np.clip(seg_model_p, 1e-12, 1.0)) + + # Score-first legality: update ALL order caches after segment scoring + for n in range(min_order, max_order + 1): + ctx_width = n - 1 + valid = global_j >= ctx_width + if not valid.any(): + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_width): + tok = val_np[jv - (ctx_width - k)].astype(np.uint64) + ctx_hash ^= tok * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + full_key = ((ctx_hash ^ (tgt_np[v_idx] * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + np.add.at(ctx_tables[n], ctx_key, 1) + np.add.at(full_tables[n], full_key, 1) + + loss_sum += float(seg_nll.sum()) + token_count += float(seg_len) + 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 += float(tb.sum().item()) + + if (bi // batch_seqs) % 2000 == 0 and bi > 0: + elapsed = time.perf_counter() - t0 + prog = min((bi + bsz) / max(len(window_starts), 1), 1.0) + cur_bpb = (loss_sum / max(token_count, 1.0)) / math.log(2.0) * (token_count / max(byte_count, 1.0)) + print( + f"ngram_eval:progress windows={bi + bsz}/{len(window_starts)} " + f"({prog*100:.1f}%) bpb={cur_bpb:.6f} t={elapsed:.0f}s", + flush=True, + ) + # All-reduce across ranks + _loss = torch.tensor(loss_sum, device=device, dtype=torch.float64) + _toks = torch.tensor(token_count, device=device, dtype=torch.float64) + _bytes = torch.tensor(byte_count, device=device, dtype=torch.float64) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(_loss, op=dist.ReduceOp.SUM) + dist.all_reduce(_toks, op=dist.ReduceOp.SUM) + dist.all_reduce(_bytes, op=dist.ReduceOp.SUM) + loss_sum = _loss.item() + token_count = _toks.item() + byte_count = _bytes.item() + + coverage = token_count / max(total_scored_tokens, 1.0) + if cutoff_hit: + elapsed = time.perf_counter() - t0 + print( + f"ngram_eval:cutoff max_seconds={max_seconds:.1f} " + f"coverage={coverage*100:.2f}% elapsed={elapsed:.0f}s", + flush=True, + ) + + val_loss = loss_sum / max(token_count, 1.0) + val_bpb = val_loss / math.log(2.0) * (token_count / max(byte_count, 1.0)) + base_model.train() + return val_loss, val_bpb, coverage +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, +) -> tuple[float, float]: + seq_len, total_tokens, ttt_chunk = args.train_seq_len, val_tokens.numel() - 1, args.ttt_chunk_tokens + master = (rank == 0) + log0 = (lambda msg: print(msg, flush=True)) if master else (lambda msg: None) + window_starts = [ws for ws in range(0, total_tokens, stride) if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end, wlen = min(ws + seq_len, total_tokens), min(ws + seq_len, total_tokens) - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + chunk_windows[min((ws + s) // ttt_chunk, num_chunks - 1)].append(ws) + log0(f"ttt_sliding:start chunks={num_chunks} windows={len(window_starts)} lr={args.ttt_lr} epochs={args.ttt_epochs} freeze={args.ttt_freeze_blocks}") + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + embed_names = {"tok_emb", "bigram", "ve_shared"} if args.ttt_freeze_embed else set() + ttt_params = [] + for name, p in base_model.named_parameters(): + if any(f"blocks.{bi}." in name for bi in frozen_ids): + p.requires_grad_(False) + elif any(en in name for en in embed_names): + p.requires_grad_(False) + else: + p.requires_grad_(True); ttt_params.append(p) + log0(f"ttt_sliding:unfrozen={sum(p.numel() for p in ttt_params)} freeze_embed={args.ttt_freeze_embed}") + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # TTT-EMA: maintain smoothed weights for scoring + ema_decay = args.ttt_ema_decay + ema_state = None + raw_state = None + if ema_decay > 0: + ema_state = {n: p.data.clone() for n, p in base_model.named_parameters() if p.requires_grad} + raw_state = {n: torch.empty_like(p.data) for n, p in base_model.named_parameters() if n in ema_state} + log0(f"ttt_sliding:ema_decay={ema_decay} ema_params={len(ema_state)}") + t0 = time.perf_counter() + cur_lr = args.ttt_lr + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start, chunk_end = ci * ttt_chunk, min((ci + 1) * ttt_chunk, total_tokens) + my_windows = windows[(len(windows) * rank) // world_size:(len(windows) * (rank + 1)) // world_size] + # Swap to EMA weights for scoring (if enabled and past first chunk) + if ema_state is not None and ci > 0: + for n, p in base_model.named_parameters(): + if n in ema_state: + raw_state[n].copy_(p.data) + p.data.copy_(ema_state[n]) + 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 = [] + for i, ws in enumerate(batch_ws): + wlen = min(ws + seq_len, total_tokens) - ws; wlens.append(wlen) + ct = val_tokens[ws:ws + wlen + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = ct[:-1]; y_batch[i, :wlen] = ct[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen, s = wlens[i], 0 if ws == 0 else max(wlens[i] - stride, 0) + loss_sum += nll[i, s:wlen].to(torch.float64).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) + (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + # Restore raw weights after scoring (for training phase) + if ema_state is not None and ci > 0: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + # Phase 2: TRAIN on this chunk (already scored = legal) + if ci < num_chunks - 1 and ci < args.ttt_max_train_chunks and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cur_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(args.ttt_max_train_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + ms, me = (chunk_seqs * rank) // world_size, (chunk_seqs * (rank + 1)) // world_size + for _ep in range(args.ttt_epochs): + for bs in range(0, me - ms, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, me - ms) + start_tok = chunk_start + (ms + bs) * seq_len + end_tok = chunk_start + (ms + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + # Update EMA after this chunk's training + if ema_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in ema_state: + ema_state[n].mul_(ema_decay).add_(p.data, alpha=1.0 - ema_decay) + # Once training stops, load EMA weights permanently for remaining score-only chunks + if ema_state is not None and ci == args.ttt_max_train_chunks: + log0(f" ttt:loading EMA weights permanently at chunk {ci}") + for n, p in base_model.named_parameters(): + if n in ema_state: + p.data.copy_(ema_state[n]) + ema_state = None + raw_state = None + if master and (ci % 5 == 0 or ci == num_chunks - 1): + rl = loss_sum.item() / max(token_count.item(), 1) + cur_bpb = rl / math.log(2) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0 + lr_str = f" lr={cur_lr:.6f}" if ci < args.ttt_max_train_chunks else " lr=done" + log0(f" ttt[{ci+1}/{num_chunks}] bpb={cur_bpb:.6f}{lr_str} t={time.perf_counter()-t0:.0f}s") + if dist.is_available() and dist.is_initialized(): + for t in [loss_sum, token_count, byte_count]: + dist.all_reduce(t, 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 loss={val_loss:.6f} bpb={val_bpb:.6f} time={time.perf_counter()-t0:.0f}s") + return val_loss, val_bpb +# --------------------------------------------------------------------------- +# GPTQ: Hessian-aware quantization with column-wise error compensation +# --------------------------------------------------------------------------- +def _find_best_row_scales(W: Tensor, clip_range: int = 31) -> Tensor: + """Find optimal per-row scales by searching percentile clipping thresholds.""" + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), 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) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 31, + block_size: int = 64, percdamp: float = 0.002) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation. + Uses pre-computed per-row scales and column reordering by Hessian diagonal. + Returns (quantized_int8, scale_fp16) in int6 range [-clip_range, clip_range].""" + W = W.float().clone() + rows, cols = W.shape + # Pre-compute optimal per-row scales from the original weight matrix + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + # Column reordering: process least-important columns first (ascending H_diag) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + # Quantize using pre-computed per-row scales + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + # Undo column reordering + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + return hessians +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor]) -> tuple[dict, dict]: + """Like mixed_quantize_int6 but uses GPTQ for int6 categories when Hessian available.""" + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count = 0, 0 + 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 == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu()) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + elif 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"} + naive_count += 1 + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers", flush=True) + return result, meta +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() + if args.compile_enabled: + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + mlp_act=args.mlp_act, + mlp_leaky_slope=args.mlp_leaky_slope, + f1_corr_rank=args.f1_corr_rank, + f1_corr_scale_init=args.f1_corr_scale_init, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = maybe_torch_compile(base_model, args) + 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]) + if base_model.f1_corr_in is not None and base_model.f1_corr_out is not None: + matrix_params.append(base_model.f1_corr_in.weight) + matrix_params.append(base_model.f1_corr_out.weight) + 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.f1_corr_scale is not None: + scalar_params.append(base_model.f1_corr_scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + f1_corr_params = 0 + if base_model.f1_corr_in is not None and base_model.f1_corr_out is not None: + f1_corr_params = int(base_model.f1_corr_in.weight.numel() + base_model.f1_corr_out.weight.numel()) + est_corr_int6_bytes = 0 + if args.f1_corr_rank > 0: + # int8 payload stores int6 values + per-row fp16 scales. + est_corr_int6_bytes = ( + args.f1_corr_rank * (args.model_dim + args.vocab_size) + + 2 * (args.f1_corr_rank + args.vocab_size) + ) + log0(f"model_params:{n_params}") + log0( + f"f1_corr:rank={args.f1_corr_rank} params={f1_corr_params} " + f"est_int6_bytes~{est_corr_int6_bytes}" + ) + log0(f"mlp_act:{args.mlp_act} mlp_leaky_slope:{args.mlp_leaky_slope}") + log0(f"XSA:last_{args.xsa_last_n} world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads} embed_lr:{token_lr} matrix_lr:{args.matrix_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"compile:enabled={int(args.compile_enabled)} fullgraph={int(args.compile_fullgraph)}") + log0(f"seed:{args.seed}") + if args.ngram_eval_order >= 2: + log0( + f"ngram_eval:order={args.ngram_eval_order} alpha={args.ngram_eval_alpha} " + f"min_count={args.ngram_eval_min_count} buckets={args.ngram_eval_buckets}" + ) + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + 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 + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + loss.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() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + 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" + ) + # GPTQ calibration: collect Hessians from training data DURING training phase + # (must happen before training ends to comply with eval-time data access rules) + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=256, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + if args.distill_enabled and args.distill_steps > 0: + log0( + f"distill:start steps:{args.distill_steps} lr_factor:{args.distill_lr_factor} " + f"temp:{args.distill_temperature} alpha:{args.distill_alpha} kl_clip:{args.distill_kl_clip}" + ) + current_state = base_model.state_dict() + teacher_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + teacher_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, + mlp_act=args.mlp_act, mlp_leaky_slope=args.mlp_leaky_slope, + f1_corr_rank=args.f1_corr_rank, f1_corr_scale_init=args.f1_corr_scale_init, + ).to(device).bfloat16() + for m in teacher_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(teacher_model) + teacher_model.load_state_dict(teacher_state, strict=True) + teacher_model.eval() + for p in teacher_model.parameters(): + p.requires_grad_(False) + compiled_teacher_logits = maybe_torch_compile(teacher_model.forward_logits, args) + model.train() + T = args.distill_temperature + alpha = args.distill_alpha + for d_step in range(args.distill_steps): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * args.distill_lr_factor + 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): + student_logits = base_model.forward_logits(x) + with torch.no_grad(): + teacher_logits = compiled_teacher_logits(x) + student_log_probs = F.log_softmax(student_logits.float() / T, dim=-1) + teacher_probs = F.softmax(teacher_logits.float() / T, dim=-1) + token_kl = F.kl_div(student_log_probs, teacher_probs, reduction="none").sum(dim=-1) + kl_loss = token_kl.mean() * (T * T) + if args.distill_kl_clip > 0: + kl_loss = torch.clamp(kl_loss, max=args.distill_kl_clip) + ce_loss = F.cross_entropy( + student_logits.reshape(-1, student_logits.size(-1)).float(), + y.reshape(-1), + reduction="mean", + ) + loss = alpha * kl_loss + (1.0 - alpha) * ce_loss + (loss * grad_scale).backward() + if world_size > 1: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + 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) + if (d_step + 1) % 8 == 0 or d_step == 0: + log0( + f"distill:step:{d_step + 1}/{args.distill_steps} " + f"kl:{kl_loss.item():.4f} ce:{ce_loss.item():.4f} total:{loss.item():.4f}" + ) + del teacher_model, compiled_teacher_logits + torch.cuda.empty_cache() + log0("distill:done") + # Apply EMA weights (better than SWA alone per PR#401) + 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()} + # GPTQ quantization using Hessians collected during training phase (no training data access here) + quant_result, quant_meta = mixed_quantize_int6_gptq(sd_cpu, {"mlp", "attn", "aux"}, gptq_hessians) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + 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, # must match training model + 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, + mlp_act=args.mlp_act, mlp_leaky_slope=args.mlp_leaky_slope, + f1_corr_rank=args.f1_corr_rank, f1_corr_scale_init=args.f1_corr_scale_init, + ).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 = maybe_torch_compile(eval_model, args) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.ngram_eval_order >= 2: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ng = time.perf_counter() + ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( + args, + eval_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + stride=args.eval_stride, + order=args.ngram_eval_order, + alpha=args.ngram_eval_alpha, + min_count=args.ngram_eval_min_count, + buckets=args.ngram_eval_buckets, + max_seconds=args.ngram_eval_max_seconds, + eval_seq_len=sw_seq_len, + ) + if rank == 0: + torch.cuda.synchronize() + ng_eval_ms = 1000.0 * (time.perf_counter() - t_ng) + if ng_coverage >= 0.999999: + log0( + f"final_int6_sliding_window_ngram{args.ngram_eval_order} val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_int6_sliding_window_ngram{args.ngram_eval_order}_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f}" + ) + else: + log0( + f"final_int6_sliding_window_ngram{args.ngram_eval_order}_partial val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} coverage:{ng_coverage:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_int6_sliding_window_ngram{args.ngram_eval_order}_partial_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f} coverage:{ng_coverage:.8f}" + ) + if distributed: + dist.barrier() + # Legal score-first TTT eval + if args.ttt_eval_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/train_seed42.log b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/train_seed42.log new file mode 100644 index 000000000..02f64d04d --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CurriculumLearning_LeakyReLU09_Ngram7/train_seed42.log @@ -0,0 +1,488 @@ +nohup: ignoring input +GPUs: 8 | Seed: 42 | Ngram: 7 | Shard order: yes +W0325 19:46:07.108000 1881 torch/distributed/run.py:803] +W0325 19:46:07.108000 1881 torch/distributed/run.py:803] ***************************************** +W0325 19:46:07.108000 1881 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 19:46:07.108000 1881 torch/distributed/run.py:803] ***************************************** +logs/64aceaa1-c0ff-4758-ab09-95d21f004c1b.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:26993756 +f1_corr:rank=0 params=0 est_int6_bytes~0 +mlp_act:relu_sq mlp_leaky_slope:0.9 +XSA:last_11 world_size:8 grad_accum_steps:1 +num_heads:8 num_kv_heads:4 embed_lr:0.035 matrix_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +compile:enabled=1 fullgraph=1 +seed:42 +ngram_eval:order=7 alpha=0.3 min_count=2 buckets=4194304 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9279 train_time:140ms step_avg:140.47ms +step:2/20000 train_loss:8.7759 train_time:223ms step_avg:111.47ms +step:3/20000 train_loss:7.7554 train_time:310ms step_avg:103.21ms +step:4/20000 train_loss:7.1970 train_time:396ms step_avg:99.12ms +step:5/20000 train_loss:6.9573 train_time:483ms step_avg:96.50ms +step:6/20000 train_loss:6.8961 train_time:569ms step_avg:94.85ms +step:7/20000 train_loss:6.7367 train_time:656ms step_avg:93.67ms +step:8/20000 train_loss:6.5068 train_time:742ms step_avg:92.74ms +step:9/20000 train_loss:6.2641 train_time:828ms step_avg:92.05ms +step:10/20000 train_loss:5.9256 train_time:916ms step_avg:91.56ms +step:500/20000 train_loss:2.4963 train_time:44466ms step_avg:88.93ms +step:1000/20000 train_loss:2.1549 train_time:89180ms step_avg:89.18ms +step:1500/20000 train_loss:2.0946 train_time:134190ms step_avg:89.46ms +step:2000/20000 train_loss:2.1647 train_time:179213ms step_avg:89.61ms +step:2500/20000 train_loss:2.1655 train_time:224386ms step_avg:89.75ms +step:3000/20000 train_loss:2.1022 train_time:269588ms step_avg:89.86ms +step:3500/20000 train_loss:2.0958 train_time:314837ms step_avg:89.95ms +step:4000/20000 train_loss:2.0166 train_time:360045ms step_avg:90.01ms +step:4000/20000 val_loss:2.0446 val_bpb:1.2109 train_time:360050ms step_avg:90.01ms +step:4500/20000 train_loss:2.0485 train_time:405263ms step_avg:90.06ms +late_qat:enabled step:4910 scale:0.4998 +step:5000/20000 train_loss:1.8073 train_time:450564ms step_avg:90.11ms +step:5500/20000 train_loss:2.0082 train_time:495900ms step_avg:90.16ms +swa:start step:6000 +step:6000/20000 train_loss:2.0299 train_time:541196ms step_avg:90.20ms +step:6500/20000 train_loss:1.8957 train_time:586708ms step_avg:90.26ms +step:6647/20000 val_loss:1.9264 val_bpb:1.1409 train_time:600079ms step_avg:90.28ms +stopping_early: wallclock_cap train_time:600079ms step:6647/20000 +peak memory allocated: 22050 MiB reserved: 22100 MiB +gptq:calibrating with training data... +gptq:calibrated 68 layers in 2.9s +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9248 val_bpb:1.1400 eval_time:2275ms +Serialized model: 106178569 bytes +Code size: 106444 bytes +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +Serialized model int6+zstd: 15453907 bytes +Total submission size int6+zstd: 15560351 bytes +Total submission size int8+zlib: 15560351 bytes +final_int6_roundtrip val_loss:1.9339 val_bpb:1.1454 eval_time:34088ms +final_int6_roundtrip_exact val_loss:1.93391857 val_bpb:1.14537551 +final_int6_sliding_window val_loss:1.8937 val_bpb:1.1216 stride:64 eval_time:95302ms +final_int6_sliding_window_exact val_loss:1.89370169 val_bpb:1.12155977 +final_int8_zlib_roundtrip_exact val_loss:1.89370169 val_bpb:1.12155977 +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.059083 t=68s +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.062669 t=68s +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.039360 t=68s +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.074461 t=68s +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.055979 t=68s +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.071941 t=68s +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.050885 t=69s +ngram_eval:progress windows=64032/121136 (52.9%) bpb=1.052691 t=69s +final_int6_sliding_window_ngram7 val_loss:1.6265 val_bpb:0.9633 eval_time:130562ms +final_int6_sliding_window_ngram7_exact val_loss:1.62654122 val_bpb:0.96333188 +ttt_sliding:start chunks=1893 windows=969088 lr=0.002 epochs=3 freeze=2 +ttt_sliding:unfrozen=21255242 freeze_embed=True +ttt_sliding:ema_decay=0.995 ema_params=94 + ttt[1/1893] bpb=1.152377 lr=0.002000 t=0s + ttt[6/1893] bpb=1.134984 lr=0.001997 t=2s + ttt[11/1893] bpb=1.147888 lr=0.001988 t=3s + ttt[16/1893] bpb=1.133512 lr=0.001972 t=4s + ttt[21/1893] bpb=1.133312 lr=0.001951 t=5s + ttt[26/1893] bpb=1.115730 lr=0.001923 t=6s + ttt[31/1893] bpb=1.130789 lr=0.001890 t=8s + ttt[36/1893] bpb=1.119034 lr=0.001851 t=9s + ttt[41/1893] bpb=1.116946 lr=0.001807 t=10s + ttt[46/1893] bpb=1.114737 lr=0.001758 t=11s + ttt[51/1893] bpb=1.111171 lr=0.001704 t=12s + ttt[56/1893] bpb=1.125816 lr=0.001646 t=14s + ttt[61/1893] bpb=1.118160 lr=0.001584 t=15s + ttt[66/1893] bpb=1.112134 lr=0.001518 t=16s + ttt[71/1893] bpb=1.116360 lr=0.001449 t=17s + ttt[76/1893] bpb=1.119053 lr=0.001377 t=18s + ttt[81/1893] bpb=1.115657 lr=0.001303 t=20s + ttt[86/1893] bpb=1.115521 lr=0.001227 t=21s + ttt[91/1893] bpb=1.116749 lr=0.001149 t=22s + ttt[96/1893] bpb=1.116827 lr=0.001071 t=23s + ttt[101/1893] bpb=1.120127 lr=0.000992 t=24s + ttt[106/1893] bpb=1.122598 lr=0.000913 t=26s + ttt[111/1893] bpb=1.122722 lr=0.000835 t=27s + ttt[116/1893] bpb=1.116436 lr=0.000758 t=28s + ttt[121/1893] bpb=1.116040 lr=0.000682 t=29s + ttt[126/1893] bpb=1.117012 lr=0.000608 t=31s + ttt[131/1893] bpb=1.116291 lr=0.000537 t=32s + ttt[136/1893] bpb=1.117789 lr=0.000468 t=33s + ttt[141/1893] bpb=1.122113 lr=0.000403 t=34s + ttt[146/1893] bpb=1.122343 lr=0.000342 t=35s + ttt[151/1893] bpb=1.123925 lr=0.000285 t=36s + ttt[156/1893] bpb=1.123925 lr=0.000232 t=38s + ttt[161/1893] bpb=1.123516 lr=0.000184 t=39s + ttt[166/1893] bpb=1.128759 lr=0.000141 t=40s + ttt[171/1893] bpb=1.127912 lr=0.000103 t=41s + ttt[176/1893] bpb=1.129128 lr=0.000071 t=42s + ttt[181/1893] bpb=1.130088 lr=0.000045 t=44s + ttt[186/1893] bpb=1.132364 lr=0.000024 t=45s + ttt[191/1893] bpb=1.137522 lr=0.000010 t=46s + ttt[196/1893] bpb=1.135571 lr=0.000002 t=47s + ttt:loading EMA weights permanently at chunk 200 + ttt[201/1893] bpb=1.136281 lr=done t=48s + ttt[206/1893] bpb=1.134038 lr=done t=49s + ttt[211/1893] bpb=1.134172 lr=done t=49s + ttt[216/1893] bpb=1.136447 lr=done t=50s + ttt[221/1893] bpb=1.135683 lr=done t=50s + ttt[226/1893] bpb=1.136074 lr=done t=51s + ttt[231/1893] bpb=1.134364 lr=done t=51s + ttt[236/1893] bpb=1.135414 lr=done t=52s + ttt[241/1893] bpb=1.134697 lr=done t=52s + ttt[246/1893] bpb=1.133832 lr=done t=53s + ttt[251/1893] bpb=1.134132 lr=done t=54s + ttt[256/1893] bpb=1.133859 lr=done t=54s + ttt[261/1893] bpb=1.131289 lr=done t=55s + ttt[266/1893] bpb=1.129756 lr=done t=55s + ttt[271/1893] bpb=1.130141 lr=done t=56s + ttt[276/1893] bpb=1.130316 lr=done t=56s + ttt[281/1893] bpb=1.131629 lr=done t=57s + ttt[286/1893] bpb=1.132806 lr=done t=57s + ttt[291/1893] bpb=1.133409 lr=done t=58s + ttt[296/1893] bpb=1.134319 lr=done t=58s + ttt[301/1893] bpb=1.134154 lr=done t=59s + ttt[306/1893] bpb=1.134985 lr=done t=59s + ttt[311/1893] bpb=1.136261 lr=done t=60s + ttt[316/1893] bpb=1.137590 lr=done t=60s + ttt[321/1893] bpb=1.138234 lr=done t=61s + ttt[326/1893] bpb=1.138882 lr=done t=61s + ttt[331/1893] bpb=1.138107 lr=done t=62s + ttt[336/1893] bpb=1.137965 lr=done t=63s + ttt[341/1893] bpb=1.137147 lr=done t=63s + ttt[346/1893] bpb=1.137605 lr=done t=64s + ttt[351/1893] bpb=1.139426 lr=done t=64s + ttt[356/1893] bpb=1.139166 lr=done t=65s + ttt[361/1893] bpb=1.139617 lr=done t=65s + ttt[366/1893] bpb=1.138252 lr=done t=66s + ttt[371/1893] bpb=1.138992 lr=done t=66s + ttt[376/1893] bpb=1.138791 lr=done t=67s + ttt[381/1893] bpb=1.139170 lr=done t=67s + ttt[386/1893] bpb=1.138536 lr=done t=68s + ttt[391/1893] bpb=1.138968 lr=done t=68s + ttt[396/1893] bpb=1.138130 lr=done t=69s + ttt[401/1893] bpb=1.136891 lr=done t=69s + ttt[406/1893] bpb=1.136497 lr=done t=70s + ttt[411/1893] bpb=1.135724 lr=done t=70s + ttt[416/1893] bpb=1.134798 lr=done t=71s + ttt[421/1893] bpb=1.134874 lr=done t=71s + ttt[426/1893] bpb=1.134750 lr=done t=72s + ttt[431/1893] bpb=1.134820 lr=done t=73s + ttt[436/1893] bpb=1.134530 lr=done t=73s + ttt[441/1893] bpb=1.135140 lr=done t=74s + ttt[446/1893] bpb=1.135608 lr=done t=74s + ttt[451/1893] bpb=1.135431 lr=done t=75s + ttt[456/1893] bpb=1.134624 lr=done t=75s + ttt[461/1893] bpb=1.134329 lr=done t=76s + ttt[466/1893] bpb=1.134882 lr=done t=76s + ttt[471/1893] bpb=1.134898 lr=done t=77s + ttt[476/1893] bpb=1.134739 lr=done t=77s + ttt[481/1893] bpb=1.134486 lr=done t=78s + ttt[486/1893] bpb=1.133786 lr=done t=78s + ttt[491/1893] bpb=1.133416 lr=done t=79s + ttt[496/1893] bpb=1.133229 lr=done t=79s + ttt[501/1893] bpb=1.132898 lr=done t=80s + ttt[506/1893] bpb=1.132339 lr=done t=80s + ttt[511/1893] bpb=1.132228 lr=done t=81s + ttt[516/1893] bpb=1.130508 lr=done t=82s + ttt[521/1893] bpb=1.130018 lr=done t=82s + ttt[526/1893] bpb=1.130712 lr=done t=83s + ttt[531/1893] bpb=1.131161 lr=done t=83s + ttt[536/1893] bpb=1.130590 lr=done t=84s + ttt[541/1893] bpb=1.131507 lr=done t=84s + ttt[546/1893] bpb=1.130825 lr=done t=85s + ttt[551/1893] bpb=1.130474 lr=done t=85s + ttt[556/1893] bpb=1.131491 lr=done t=86s + ttt[561/1893] bpb=1.131002 lr=done t=86s + ttt[566/1893] bpb=1.130469 lr=done t=87s + ttt[571/1893] bpb=1.129964 lr=done t=87s + ttt[576/1893] bpb=1.129295 lr=done t=88s + ttt[581/1893] bpb=1.129178 lr=done t=88s + ttt[586/1893] bpb=1.129440 lr=done t=89s + ttt[591/1893] bpb=1.128564 lr=done t=89s + ttt[596/1893] bpb=1.129048 lr=done t=90s + ttt[601/1893] bpb=1.129056 lr=done t=90s + ttt[606/1893] bpb=1.129047 lr=done t=91s + ttt[611/1893] bpb=1.128988 lr=done t=92s + ttt[616/1893] bpb=1.129258 lr=done t=92s + ttt[621/1893] bpb=1.128864 lr=done t=93s + ttt[626/1893] bpb=1.129331 lr=done t=93s + ttt[631/1893] bpb=1.129607 lr=done t=94s + ttt[636/1893] bpb=1.128891 lr=done t=94s + ttt[641/1893] bpb=1.129370 lr=done t=95s + ttt[646/1893] bpb=1.130105 lr=done t=95s + ttt[651/1893] bpb=1.129462 lr=done t=96s + ttt[656/1893] bpb=1.129546 lr=done t=96s + ttt[661/1893] bpb=1.128934 lr=done t=97s + ttt[666/1893] bpb=1.129435 lr=done t=97s + ttt[671/1893] bpb=1.129316 lr=done t=98s + ttt[676/1893] bpb=1.129616 lr=done t=98s + ttt[681/1893] bpb=1.130015 lr=done t=99s + ttt[686/1893] bpb=1.129933 lr=done t=99s + ttt[691/1893] bpb=1.131000 lr=done t=100s + ttt[696/1893] bpb=1.130812 lr=done t=101s + ttt[701/1893] bpb=1.130446 lr=done t=101s + ttt[706/1893] bpb=1.130028 lr=done t=102s + ttt[711/1893] bpb=1.130414 lr=done t=102s + ttt[716/1893] bpb=1.129785 lr=done t=103s + ttt[721/1893] bpb=1.130038 lr=done t=103s + ttt[726/1893] bpb=1.129770 lr=done t=104s + ttt[731/1893] bpb=1.130101 lr=done t=104s + ttt[736/1893] bpb=1.130431 lr=done t=105s + ttt[741/1893] bpb=1.130223 lr=done t=105s + ttt[746/1893] bpb=1.130366 lr=done t=106s + ttt[751/1893] bpb=1.130081 lr=done t=106s + ttt[756/1893] bpb=1.130176 lr=done t=107s + ttt[761/1893] bpb=1.130012 lr=done t=107s + ttt[766/1893] bpb=1.129962 lr=done t=108s + ttt[771/1893] bpb=1.129697 lr=done t=108s + ttt[776/1893] bpb=1.130025 lr=done t=109s + ttt[781/1893] bpb=1.130423 lr=done t=109s + ttt[786/1893] bpb=1.129940 lr=done t=110s + ttt[791/1893] bpb=1.130055 lr=done t=111s + ttt[796/1893] bpb=1.130263 lr=done t=111s + ttt[801/1893] bpb=1.130379 lr=done t=112s + ttt[806/1893] bpb=1.130255 lr=done t=112s + ttt[811/1893] bpb=1.130162 lr=done t=113s + ttt[816/1893] bpb=1.129918 lr=done t=113s + ttt[821/1893] bpb=1.129933 lr=done t=114s + ttt[826/1893] bpb=1.130124 lr=done t=114s + ttt[831/1893] bpb=1.129757 lr=done t=115s + ttt[836/1893] bpb=1.129125 lr=done t=115s + ttt[841/1893] bpb=1.129113 lr=done t=116s + ttt[846/1893] bpb=1.129122 lr=done t=116s + ttt[851/1893] bpb=1.128881 lr=done t=117s + ttt[856/1893] bpb=1.128693 lr=done t=117s + ttt[861/1893] bpb=1.128649 lr=done t=118s + ttt[866/1893] bpb=1.128844 lr=done t=118s + ttt[871/1893] bpb=1.128922 lr=done t=119s + ttt[876/1893] bpb=1.128933 lr=done t=120s + ttt[881/1893] bpb=1.129114 lr=done t=120s + ttt[886/1893] bpb=1.128934 lr=done t=121s + ttt[891/1893] bpb=1.128705 lr=done t=121s + ttt[896/1893] bpb=1.128663 lr=done t=122s + ttt[901/1893] bpb=1.128436 lr=done t=122s + ttt[906/1893] bpb=1.128253 lr=done t=123s + ttt[911/1893] bpb=1.128597 lr=done t=123s + ttt[916/1893] bpb=1.128973 lr=done t=124s + ttt[921/1893] bpb=1.129069 lr=done t=124s + ttt[926/1893] bpb=1.129278 lr=done t=125s + ttt[931/1893] bpb=1.129035 lr=done t=125s + ttt[936/1893] bpb=1.128673 lr=done t=126s + ttt[941/1893] bpb=1.128720 lr=done t=126s + ttt[946/1893] bpb=1.129231 lr=done t=127s + ttt[951/1893] bpb=1.129101 lr=done t=127s + ttt[956/1893] bpb=1.129525 lr=done t=128s + ttt[961/1893] bpb=1.129199 lr=done t=128s + ttt[966/1893] bpb=1.129661 lr=done t=129s + ttt[971/1893] bpb=1.130096 lr=done t=130s + ttt[976/1893] bpb=1.130315 lr=done t=130s + ttt[981/1893] bpb=1.130189 lr=done t=131s + ttt[986/1893] bpb=1.130276 lr=done t=131s + ttt[991/1893] bpb=1.130198 lr=done t=132s + ttt[996/1893] bpb=1.130134 lr=done t=132s + ttt[1001/1893] bpb=1.130185 lr=done t=133s + ttt[1006/1893] bpb=1.129952 lr=done t=133s + ttt[1011/1893] bpb=1.129988 lr=done t=134s + ttt[1016/1893] bpb=1.130392 lr=done t=134s + ttt[1021/1893] bpb=1.130316 lr=done t=135s + ttt[1026/1893] bpb=1.130485 lr=done t=135s + ttt[1031/1893] bpb=1.130762 lr=done t=136s + ttt[1036/1893] bpb=1.130626 lr=done t=136s + ttt[1041/1893] bpb=1.130422 lr=done t=137s + ttt[1046/1893] bpb=1.130273 lr=done t=137s + ttt[1051/1893] bpb=1.130174 lr=done t=138s + ttt[1056/1893] bpb=1.130348 lr=done t=139s + ttt[1061/1893] bpb=1.130233 lr=done t=139s + ttt[1066/1893] bpb=1.130497 lr=done t=140s + ttt[1071/1893] bpb=1.130854 lr=done t=140s + ttt[1076/1893] bpb=1.131163 lr=done t=141s + ttt[1081/1893] bpb=1.131145 lr=done t=141s + ttt[1086/1893] bpb=1.132021 lr=done t=142s + ttt[1091/1893] bpb=1.131858 lr=done t=142s + ttt[1096/1893] bpb=1.131890 lr=done t=143s + ttt[1101/1893] bpb=1.131881 lr=done t=143s + ttt[1106/1893] bpb=1.131835 lr=done t=144s + ttt[1111/1893] bpb=1.131712 lr=done t=144s + ttt[1116/1893] bpb=1.131547 lr=done t=145s + ttt[1121/1893] bpb=1.131520 lr=done t=145s + ttt[1126/1893] bpb=1.131145 lr=done t=146s + ttt[1131/1893] bpb=1.131374 lr=done t=146s + ttt[1136/1893] bpb=1.131149 lr=done t=147s + ttt[1141/1893] bpb=1.131063 lr=done t=147s + ttt[1146/1893] bpb=1.131336 lr=done t=148s + ttt[1151/1893] bpb=1.131077 lr=done t=149s + ttt[1156/1893] bpb=1.130840 lr=done t=149s + ttt[1161/1893] bpb=1.130711 lr=done t=150s + ttt[1166/1893] bpb=1.131073 lr=done t=150s + ttt[1171/1893] bpb=1.131025 lr=done t=151s + ttt[1176/1893] bpb=1.130552 lr=done t=151s + ttt[1181/1893] bpb=1.130284 lr=done t=152s + ttt[1186/1893] bpb=1.130370 lr=done t=152s + ttt[1191/1893] bpb=1.130162 lr=done t=153s + ttt[1196/1893] bpb=1.130334 lr=done t=153s + ttt[1201/1893] bpb=1.130576 lr=done t=154s + ttt[1206/1893] bpb=1.130497 lr=done t=154s + ttt[1211/1893] bpb=1.130097 lr=done t=155s + ttt[1216/1893] bpb=1.130312 lr=done t=155s + ttt[1221/1893] bpb=1.129813 lr=done t=156s + ttt[1226/1893] bpb=1.129817 lr=done t=156s + ttt[1231/1893] bpb=1.129540 lr=done t=157s + ttt[1236/1893] bpb=1.129497 lr=done t=157s + ttt[1241/1893] bpb=1.129202 lr=done t=158s + ttt[1246/1893] bpb=1.129017 lr=done t=159s + ttt[1251/1893] bpb=1.128608 lr=done t=159s + ttt[1256/1893] bpb=1.128597 lr=done t=160s + ttt[1261/1893] bpb=1.128584 lr=done t=160s + ttt[1266/1893] bpb=1.128450 lr=done t=161s + ttt[1271/1893] bpb=1.128211 lr=done t=161s + ttt[1276/1893] bpb=1.127933 lr=done t=162s + ttt[1281/1893] bpb=1.128022 lr=done t=162s + ttt[1286/1893] bpb=1.128036 lr=done t=163s + ttt[1291/1893] bpb=1.127807 lr=done t=163s + ttt[1296/1893] bpb=1.127624 lr=done t=164s + ttt[1301/1893] bpb=1.127202 lr=done t=164s + ttt[1306/1893] bpb=1.127020 lr=done t=165s + ttt[1311/1893] bpb=1.126821 lr=done t=165s + ttt[1316/1893] bpb=1.126821 lr=done t=166s + ttt[1321/1893] bpb=1.126496 lr=done t=166s + ttt[1326/1893] bpb=1.126531 lr=done t=167s + ttt[1331/1893] bpb=1.126456 lr=done t=168s + ttt[1336/1893] bpb=1.126434 lr=done t=168s + ttt[1341/1893] bpb=1.126328 lr=done t=169s + ttt[1346/1893] bpb=1.126030 lr=done t=169s + ttt[1351/1893] bpb=1.126263 lr=done t=170s + ttt[1356/1893] bpb=1.126357 lr=done t=170s + ttt[1361/1893] bpb=1.126322 lr=done t=171s + ttt[1366/1893] bpb=1.126217 lr=done t=171s + ttt[1371/1893] bpb=1.126206 lr=done t=172s + ttt[1376/1893] bpb=1.126310 lr=done t=172s + ttt[1381/1893] bpb=1.126185 lr=done t=173s + ttt[1386/1893] bpb=1.126188 lr=done t=173s + ttt[1391/1893] bpb=1.125795 lr=done t=174s + ttt[1396/1893] bpb=1.125819 lr=done t=174s + ttt[1401/1893] bpb=1.125775 lr=done t=175s + ttt[1406/1893] bpb=1.125714 lr=done t=175s + ttt[1411/1893] bpb=1.125890 lr=done t=176s + ttt[1416/1893] bpb=1.126027 lr=done t=176s + ttt[1421/1893] bpb=1.126144 lr=done t=177s + ttt[1426/1893] bpb=1.126074 lr=done t=178s + ttt[1431/1893] bpb=1.125859 lr=done t=178s + ttt[1436/1893] bpb=1.126065 lr=done t=179s + ttt[1441/1893] bpb=1.126372 lr=done t=179s + ttt[1446/1893] bpb=1.126592 lr=done t=180s + ttt[1451/1893] bpb=1.126719 lr=done t=180s + ttt[1456/1893] bpb=1.126675 lr=done t=181s + ttt[1461/1893] bpb=1.126259 lr=done t=181s + ttt[1466/1893] bpb=1.126613 lr=done t=182s + ttt[1471/1893] bpb=1.127294 lr=done t=182s + ttt[1476/1893] bpb=1.127231 lr=done t=183s + ttt[1481/1893] bpb=1.126828 lr=done t=183s + ttt[1486/1893] bpb=1.126870 lr=done t=184s + ttt[1491/1893] bpb=1.126655 lr=done t=184s + ttt[1496/1893] bpb=1.126657 lr=done t=185s + ttt[1501/1893] bpb=1.126593 lr=done t=185s + ttt[1506/1893] bpb=1.126668 lr=done t=186s + ttt[1511/1893] bpb=1.126638 lr=done t=187s + ttt[1516/1893] bpb=1.126700 lr=done t=187s + ttt[1521/1893] bpb=1.126666 lr=done t=188s + ttt[1526/1893] bpb=1.126477 lr=done t=188s + ttt[1531/1893] bpb=1.126169 lr=done t=189s + ttt[1536/1893] bpb=1.126113 lr=done t=189s + ttt[1541/1893] bpb=1.126031 lr=done t=190s + ttt[1546/1893] bpb=1.126184 lr=done t=190s + ttt[1551/1893] bpb=1.126351 lr=done t=191s + ttt[1556/1893] bpb=1.126250 lr=done t=191s + ttt[1561/1893] bpb=1.126369 lr=done t=192s + ttt[1566/1893] bpb=1.126259 lr=done t=192s + ttt[1571/1893] bpb=1.126226 lr=done t=193s + ttt[1576/1893] bpb=1.126222 lr=done t=193s + ttt[1581/1893] bpb=1.126351 lr=done t=194s + ttt[1586/1893] bpb=1.126400 lr=done t=194s + ttt[1591/1893] bpb=1.126216 lr=done t=195s + ttt[1596/1893] bpb=1.126129 lr=done t=195s + ttt[1601/1893] bpb=1.126365 lr=done t=196s + ttt[1606/1893] bpb=1.126412 lr=done t=197s + ttt[1611/1893] bpb=1.126310 lr=done t=197s + ttt[1616/1893] bpb=1.126217 lr=done t=198s + ttt[1621/1893] bpb=1.125919 lr=done t=198s + ttt[1626/1893] bpb=1.125971 lr=done t=199s + ttt[1631/1893] bpb=1.126234 lr=done t=199s + ttt[1636/1893] bpb=1.126270 lr=done t=200s + ttt[1641/1893] bpb=1.126230 lr=done t=200s + ttt[1646/1893] bpb=1.126354 lr=done t=201s + ttt[1651/1893] bpb=1.126178 lr=done t=201s + ttt[1656/1893] bpb=1.126142 lr=done t=202s + ttt[1661/1893] bpb=1.126052 lr=done t=202s + ttt[1666/1893] bpb=1.126303 lr=done t=203s + ttt[1671/1893] bpb=1.126527 lr=done t=203s + ttt[1676/1893] bpb=1.126624 lr=done t=204s + ttt[1681/1893] bpb=1.126693 lr=done t=204s + ttt[1686/1893] bpb=1.126194 lr=done t=205s + ttt[1691/1893] bpb=1.126538 lr=done t=206s + ttt[1696/1893] bpb=1.126704 lr=done t=206s + ttt[1701/1893] bpb=1.126693 lr=done t=207s + ttt[1706/1893] bpb=1.126697 lr=done t=207s + ttt[1711/1893] bpb=1.126707 lr=done t=208s + ttt[1716/1893] bpb=1.126550 lr=done t=208s + ttt[1721/1893] bpb=1.126711 lr=done t=209s + ttt[1726/1893] bpb=1.126777 lr=done t=209s + ttt[1731/1893] bpb=1.126591 lr=done t=210s + ttt[1736/1893] bpb=1.126426 lr=done t=210s + ttt[1741/1893] bpb=1.126393 lr=done t=211s + ttt[1746/1893] bpb=1.126185 lr=done t=211s + ttt[1751/1893] bpb=1.126216 lr=done t=212s + ttt[1756/1893] bpb=1.126218 lr=done t=212s + ttt[1761/1893] bpb=1.126359 lr=done t=213s + ttt[1766/1893] bpb=1.126183 lr=done t=213s + ttt[1771/1893] bpb=1.126254 lr=done t=214s + ttt[1776/1893] bpb=1.126209 lr=done t=214s + ttt[1781/1893] bpb=1.126286 lr=done t=215s + ttt[1786/1893] bpb=1.126058 lr=done t=216s + ttt[1791/1893] bpb=1.125885 lr=done t=216s + ttt[1796/1893] bpb=1.125668 lr=done t=217s + ttt[1801/1893] bpb=1.125769 lr=done t=217s + ttt[1806/1893] bpb=1.125785 lr=done t=218s + ttt[1811/1893] bpb=1.125659 lr=done t=218s + ttt[1816/1893] bpb=1.125809 lr=done t=219s + ttt[1821/1893] bpb=1.125711 lr=done t=219s + ttt[1826/1893] bpb=1.125563 lr=done t=220s + ttt[1831/1893] bpb=1.125102 lr=done t=220s + ttt[1836/1893] bpb=1.125085 lr=done t=221s + ttt[1841/1893] bpb=1.125135 lr=done t=221s + ttt[1846/1893] bpb=1.124966 lr=done t=222s + ttt[1851/1893] bpb=1.124914 lr=done t=222s + ttt[1856/1893] bpb=1.124736 lr=done t=223s + ttt[1861/1893] bpb=1.124559 lr=done t=223s + ttt[1866/1893] bpb=1.124636 lr=done t=224s + ttt[1871/1893] bpb=1.124551 lr=done t=225s + ttt[1876/1893] bpb=1.124419 lr=done t=225s + ttt[1881/1893] bpb=1.124092 lr=done t=226s + ttt[1886/1893] bpb=1.123974 lr=done t=226s + ttt[1891/1893] bpb=1.123847 lr=done t=227s + ttt[1893/1893] bpb=1.123893 lr=done t=227s +ttt_sliding:done loss=1.893689 bpb=1.121552 time=228s +legal_ttt val_loss:1.8937 val_bpb:1.1216 eval_time:227978ms +legal_ttt_exact val_loss:1.89368867 val_bpb:1.12155207