diff --git a/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/README.md b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/README.md new file mode 100644 index 000000000..51f81f06d --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/README.md @@ -0,0 +1,106 @@ +# Full GPTQ + XSA-4 + SWA/EMA + Score-First TTT + +**val_bpb: 1.1198** (3-seed mean, std 0.0006) | **~15.9 MB** | 8xH100 SXM + +## Results (8xH100 SXM, PyTorch 2.9.1+cu128) + +| Seed | Steps | ms/step | Pre-TTT bpb | **Post-TTT bpb** | TTT gain | TTT time | Artifact | +|------|-------|---------|-------------|-----------------|----------|----------|----------| +| 1337 | 6,461 | 86.67 | 1.1193 | **1.1193** | -0.00004 | ~236s | 15,899,061 | +| 42 | 6,457 | 86.73 | 1.1197 | **1.1196** | -0.00004 | ~236s | 15,954,941 | +| 2025 | 6,457 | 86.74 | 1.1206 | **1.1205** | -0.00006 | ~236s | 15,907,769 | + +**Mean: 1.1198 | Std: 0.0006** + +## Timing Budget + +| Phase | Time | Notes | +|-------|------|-------| +| Training loop | 560s | Main training (Muon + Adam, ~6,460 steps) | +| GPTQ calibration + quantization | ~40s | Hessian calibration on training data, within training budget | +| **Total artifact construction** | **~600s** | **Within 10-min training limit** | +| Standard eval (roundtrip + sliding) | ~82s | No training data access | +| Score-first TTT | ~236s | Legal: score chunk, then adapt, never re-score | +| **Total eval** | **~318s** | **Within 10-min eval limit** | + +Note: GPTQ calibration uses training data and is counted as part of the training/artifact construction phase, not the eval phase. No training data is accessed during evaluation. + +## Key Techniques + +### Full Hessian GPTQ +- 256-batch calibration from training data for per-layer Hessian approximation +- Column-wise int6 quantization with Cholesky error compensation, block size 128 +- Percentile clip search over [0.999, 0.9995, 0.9999, 0.99999, 1.0] per layer +- Act-order column permutation (quantize most-activated columns first) + +### Legal Score-First TTT (PR #461/#549 recipe) +Backward-looking, score-first TTT: +1. Val tokens processed in non-overlapping 128K-token chunks +2. **For each chunk**: + - **SCORE**: Sliding window eval under `torch.inference_mode()` — no gradients, no weight mutation + - **TRAIN**: AdamW(lr=1e-4, wd=0) on the already-scored chunk. 3 epochs, first 9/11 blocks frozen, grad clip 1.0 +3. Last chunk scored but never trained on +4. Chunk N scored by model adapted only on chunks 0..N-1 + +`inference_mode()` provides a hard guarantee that scoring is stateless. + +### XSA on Last 4 Layers +Cross-Sequence Attention on transformer layers 7-10. Extended context beyond training sequence length at eval time. + +### SWA/EMA Weight Blending +EMA (decay=0.997) + Stochastic Weight Averaging (every 50 steps during warmdown), blended 50/50. Smooths weight landscape before quantization. + +## Architecture + +| Component | Setting | +|-----------|---------| +| Layers | 11 (512d, 8H, 4KV GQA) | +| MLP | 3x with LeakyReLU(0.5)^2 | +| BigramHash | 3072 buckets, 128-dim | +| XSA | Last 4 layers | +| RoPE | Partial (16/64 dims) | +| LN Scale | 1/sqrt(layer+1) | +| VE128 | Layers 9-10 | +| Weight avg | EMA(0.997) + Tight SWA(every 50) | +| Quantization | Full GPTQ int6 + LZMA | +| Optimizer | Muon (matrices) + AdamW (scalars/embeddings) | + +## Training + +- Muon optimizer (matrices): lr=0.025, momentum=0.99, WD=0.04, 5 Newton-Schulz steps +- AdamW (embeddings): lr=0.035, (scalars): lr=0.025, WD=0.04 +- Gradient clip: 0.3 +- Batch: 786,432 tokens/step, seq_len=2048 +- Warmdown: 4,000 iters (wallclock-based) +- Late QAT: disabled (LATE_QAT_THRESHOLD=0) + +## Run Command + +```bash +DATA_PATH=./data/datasets/fineweb10B_sp1024/ \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 MAX_WALLCLOCK_SECONDS=560 \ +XSA_LAST_N=4 WARMDOWN_ITERS=4000 \ +CLIP_RANGE=31 COMPRESSOR=lzma \ +LATE_QAT_THRESHOLD=0 PRUNE_PCT=0.005 \ +GPTQ_ENABLED=1 GPTQ_CALIB_BATCHES=256 GPTQ_BLOCK_SIZE=128 \ +SWA_ENABLED=1 \ +TTT_ENABLED=1 TTT_ADAMW=1 TTT_LR=0.0001 TTT_EPOCHS=3 TTT_WD=0 \ +TTT_FREEZE_BLOCKS=9 TTT_FREEZE_EMBEDDINGS=0 TTT_CHUNK_TOKENS=131072 \ +EVAL_STRIDE=64 NUM_KV_HEADS=4 SEED=1337 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## Requirements + +```bash +pip install flash_attn_3 --find-links https://windreamer.github.io/flash-attention3-wheels/cu128_torch291 +pip install zstandard sentencepiece lzma +``` + +## Credits + +- **Base architecture**: PR #589 by @RoyiRa (11L GEPA, GQA, VE128, Late QAT) +- **GPTQ reference**: PR #609 by @saml212, PR #626 by @kshitizz36 +- **Score-first TTT**: PR #461 by @Christopher-Lee-McClendon +- **XSA, BigramHash, SmearGate**: Various community contributors diff --git a/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/submission.json b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/submission.json new file mode 100644 index 000000000..2909660d9 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/submission.json @@ -0,0 +1,19 @@ +{ + "author": "Robby Sneiderman", + "github_id": "Robby955", + "name": "Full GPTQ + XSA-4 + SWA/EMA + Score-First TTT", + "blurb": "Hessian-aware GPTQ with percentile clip search, XSA on last 4 layers, EMA(0.997) + Tight SWA, legal score-first TTT (3 epochs, AdamW lr=1e-4, freeze first 9/11 blocks). LeakyReLU(0.5)^2, GQA 8H/4KV, BigramHash(3072), VE128, Partial RoPE 16/64, LN Scale, U-Net skips. LZMA compression.", + "date": "2026-03-25", + "val_loss": 1.89074159, + "val_bpb": 1.11980664, + "pre_quant_val_loss": 1.9232, + "pre_quant_val_bpb": 1.1390, + "bytes_total": 15920590, + "seeds": { + "1337": {"val_bpb": 1.11929318, "val_loss": 1.88987464, "bytes_total": 15899061}, + "42": {"val_bpb": 1.11963121, "val_loss": 1.89044540, "bytes_total": 15954941}, + "2025": {"val_bpb": 1.12049552, "val_loss": 1.89190474, "bytes_total": 15907769} + }, + "mean_val_bpb": 1.11980664, + "std_val_bpb": 0.00062007 +} diff --git a/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_gpt.py b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_gpt.py new file mode 100644 index 000000000..0a863df1d --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_gpt.py @@ -0,0 +1,1499 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) + compressor = os.environ.get("COMPRESSOR", "lzma") + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) + ttt_enabled = bool(int(os.environ.get("TTT_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_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + _clip_range: int = 31 + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + self.use_vrl = False + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, v_embed: Tensor | None = None, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + return x_out, raw_v +class GPT(nn.Module): + def __init__(self, vocab_size: int, num_layers: int, model_dim: int, num_heads: int, + num_kv_heads: int, mlp_mult: int, tie_embeddings: bool, tied_embed_init_std: float, + logit_softcap: float, rope_base: float, qk_gain_init: float, + bigram_vocab_size: int = 0, bigram_dim: int = 128, xsa_last_n: int = 0, + rope_dims: int = 0, ln_scale: bool = False, + ve_enabled: bool = False, ve_dim: int = 128, ve_layers: str = "9,10", + use_vrl: bool = False): + super().__init__() + self.use_vrl = use_vrl + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init, + layer_idx=i, ln_scale=ln_scale) + for i in range(num_layers) + ]) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor, lora=None) -> 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 = {} + v0 = None + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + return F.cross_entropy(logits.float(), targets, reduction="mean") + def forward_logits(self, input_ids: Tensor, return_hidden: bool = False): + 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 = {} + v0 = None + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + 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 _make_gpt(args, device): + a = args + m = GPT(vocab_size=a.vocab_size, num_layers=a.num_layers, model_dim=a.model_dim, num_heads=a.num_heads, num_kv_heads=a.num_kv_heads, mlp_mult=a.mlp_mult, tie_embeddings=a.tie_embeddings, tied_embed_init_std=a.tied_embed_init_std, logit_softcap=a.logit_softcap, rope_base=a.rope_base, qk_gain_init=a.qk_gain_init, bigram_vocab_size=a.bigram_vocab_size, bigram_dim=a.bigram_dim, xsa_last_n=a.xsa_last_n, rope_dims=a.rope_dims, ln_scale=a.ln_scale, ve_enabled=a.ve_enabled, ve_dim=a.ve_dim, ve_layers=a.ve_layers, use_vrl=a.vrl).to(device).bfloat16() + for mod in m.modules(): + if isinstance(mod, CastedLinear): mod.float() + restore_low_dim_params_to_fp32(m) + return m +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True; torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False); enable_flash_sdp(True); enable_mem_efficient_sdp(False); enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 = False + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + base_model = _make_gpt(args, device) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [p for n, p in block_named_params if p.ndim == 2 and not any(pat in n for pat in CONTROL_TENSOR_NAME_PATTERNS)] + scalar_params = [p for n, p in block_named_params if p.ndim < 2 or any(pat in n for pat in CONTROL_TENSOR_NAME_PATTERNS)] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + ab = (args.beta1, args.beta2) + optimizer_tok = torch.optim.AdamW(tok_params, betas=ab, 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=ab, 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=ab, eps=args.adam_eps, fused=True) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0(f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}") + log0(f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} iterations:{args.iterations} warmup_steps:{args.warmup_steps} max_wallclock_seconds:{args.max_wallclock_seconds:.3f}") + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + # Warmup phase + 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} 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 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}") + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} 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 reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB") + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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} eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms") + export_sd = base_model.state_dict() + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians(base_model, calib_loader, args, device, grad_accum_steps, num_batches=args.gptq_calib_batches) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range) + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) + log0(f"gptq:over by {len(test_blob) - target_model_bytes} bytes, max_prune={max_prune}") + pc = [] # (cost, qkey, idx_tuple) + for name, info in quant_meta.items(): + if not (isinstance(info, dict) and info.get("type") == "int6"): + continue + qk, q, s = name + ".q", quant_result[name + ".q"], quant_result[name + ".scale"] + H = hessians.get(name) if hessians else None + hd = torch.diag(H).float() if H is not None else None + for idx in (q.abs() == 1).nonzero(as_tuple=False): + r, c = idx[0].item(), (idx[1].item() if len(idx) > 1 else 0) + sc = s[r].float().item() if s.ndim > 0 else s.float().item() + pc.append((sc * sc * (hd[c].item() if hd is not None and c < len(hd) else 1.0), qk, tuple(idx.tolist()))) + pc.sort(key=lambda x: x[0]) + pc = pc[:max_prune] + log0(f"gptq:pruning candidates={len(pc)}") + def _try_prune(n): + qr = {k: v.clone() for k, v in quant_result.items()} + for i in range(n): + qr[pc[i][1]][pc[i][2]] = 0 + return qr + lo, hi, best_n = 0, len(pc), 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: lo = 1; continue + blob = _serialize_and_compress(_try_prune(mid), quant_meta, fast=True) + if len(blob) <= int(target_model_bytes * 0.997): best_n = mid; hi = mid - 1 + else: lo = mid + 1 + if best_n > 0: + final_blob = _serialize_and_compress(_try_prune(best_n), quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(pc): + best_n = min(best_n + max(1, best_n // 10), len(pc)) + final_blob = _serialize_and_compress(_try_prune(best_n), quant_meta) + for i in range(best_n): + quant_result[pc[i][1]][pc[i][2]] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}%)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = _make_gpt(args, device) + eval_model.load_state_dict(deq_state, strict=True) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val(args, compiled_eval, rank, world_size, device, grad_accum_steps, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, eval_seq_len=effective_eval_seq_len) + torch.cuda.synchronize() + log0(f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} 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} 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}") + # Score-first TTT (PR#461/549 recipe) + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt(args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=64, batch_seqs=32, log_fn=log0) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed1337.log b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed1337.log new file mode 100644 index 000000000..ac4d8934f --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed1337.log @@ -0,0 +1,2396 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # 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", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) # 31=int6 (63 levels), 15=int5 (31 levels) + compressor = os.environ.get("COMPRESSOR", "lzma") # lzma or zstd + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs, then eval + ttt_full_epochs = int(os.environ.get("TTT_FULL_EPOCHS", 0)) + ttt_proj_lr_mult = float(os.environ.get("TTT_PROJ_LR_MULT", 3.0)) + ttt_fc_lr_mult = float(os.environ.get("TTT_FC_LR_MULT", 0.5)) + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + # Requant-only mode: skip training, load fp32 model, re-quantize + eval + requant_only = bool(int(os.environ.get("REQUANT_ONLY", "0"))) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False # bool: only 2 values, max 2 torch.compile graphs + _clip_range: int = 31 # 31=int6, 15=int5 — set once before compile, not changed during training + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # Registered buffer: torch.compile treats as dynamic input (not frozen at trace time) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range # int, set once before compile + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + # Soft-Round: fully differentiable rounding (no STE, matching #606) + # alpha buffer: 1.0 (soft, linear) -> 16.0 (hard, nearly round()) + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + # Fully differentiable: gradients flow through tanh soft-round + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + 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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + 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, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + 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, + ) + 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) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + 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 + 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, lora=None) -> 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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def ttt_full_pretrain( + args, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + log_fn=None, +) -> None: + """Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs with + cosine LR decay and differentiated per-layer LRs. All params unfrozen. + Data sharded across GPUs with gradient all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_epochs = args.ttt_full_epochs + batch_seqs = args.ttt_train_batch_seqs # 32 + + # Shard val data across GPUs + rank_tokens = total_tokens // world_size + rank_start = rank * rank_tokens + rank_end = rank_start + rank_tokens + + # Unfreeze params with differentiated LR groups + proj_params, fc_params, other_params = [], [], [] + for name, p in base_model.named_parameters(): + # Respect freeze settings + freeze = False + if args.ttt_freeze_blocks > 0: + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + freeze = any(f"blocks.{bi}." in name for bi in frozen_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + continue + p.requires_grad_(True) + if "mlp.proj" in name: + proj_params.append(p) + elif "mlp.fc" in name: + fc_params.append(p) + else: + other_params.append(p) + + base_lr = args.ttt_lr + optimizer = torch.optim.AdamW([ + {"params": proj_params, "lr": base_lr * args.ttt_proj_lr_mult}, + {"params": fc_params, "lr": base_lr * args.ttt_fc_lr_mult}, + {"params": other_params, "lr": base_lr}, + ], weight_decay=args.ttt_wd) + + total_trainable = sum(p.numel() for pg in [proj_params, fc_params, other_params] for p in pg) + n_rank_seqs = max(1, (rank_end - rank_start - seq_len) // (batch_seqs * seq_len)) + steps_per_epoch = n_rank_seqs + total_steps = ttt_epochs * steps_per_epoch + + if log_fn: + log_fn(f"ttt_full:start epochs={ttt_epochs} trainable={total_trainable} " + f"rank_tokens={rank_tokens} steps_per_ep={steps_per_epoch} " + f"total_steps={total_steps} base_lr={base_lr}") + + base_model.train() + t0 = time.perf_counter() + global_step = 0 + + for epoch in range(ttt_epochs): + for bs_idx in range(0, rank_end - rank_start - seq_len, batch_seqs * seq_len): + # Cosine LR schedule across all TTT steps + progress = global_step / max(total_steps, 1) + cos_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('initial_lr', pg['lr']) * cos_mul if global_step == 0 else pg['lr'] + # Set initial_lr on first step + if global_step == 0: + for pg in optimizer.param_groups: + pg['initial_lr'] = pg['lr'] + else: + for pg in optimizer.param_groups: + pg['lr'] = pg['initial_lr'] * cos_mul + + start_tok = rank_start + bs_idx + end_tok = min(start_tok + batch_seqs * seq_len + 1, val_tokens.numel()) + if end_tok - start_tok < seq_len + 1: + continue + + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + n_seqs = (local.numel() - 1) // seq_len + if n_seqs == 0: + continue + x = local[:n_seqs * seq_len].reshape(n_seqs, seq_len) + y = local[1:n_seqs * seq_len + 1].reshape(n_seqs, 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() + + # DDP-style gradient sync + 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) + + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.ttt_grad_clip) + optimizer.step() + global_step += 1 + + if log_fn and (epoch % 5 == 0 or epoch == ttt_epochs - 1): + elapsed = time.perf_counter() - t0 + log_fn(f" ttt_full:epoch {epoch+1}/{ttt_epochs} loss={loss.item():.4f} " + f"lr={optimizer.param_groups[2]['lr']:.6f} time={elapsed:.1f}s") + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(True) + if log_fn: + log_fn(f"ttt_full:done total_steps={global_step} time={time.perf_counter() - t0:.1f}s") + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + stride = args.eval_stride if args.eval_stride > 0 else 64 + log0(f"eval_only:TTT disabled, computing sliding window BPB stride={stride}") + 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=stride, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f} stride:{stride}") + if distributed: + dist.destroy_process_group() + return + if args.requant_only: + log0(f"requant_only:loading fp32 model, re-quantizing with damp={args.gptq_damp_factor}") + fp32_state = torch.load("final_model.pt", map_location="cpu") + base_model.load_state_dict(fp32_state, strict=False) + base_model.eval() + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + sd_cpu = {k: v for k, v in sd_cpu.items() if "mtp_heads" not in k} + # Re-collect Hessians and re-quantize + hessians = None + if args.gptq_enabled: + log0(f"requant:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"requant:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Pruning + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + log0(f"requant:artifact={len(quant_blob)} code={code_bytes} total={len(quant_blob)+code_bytes}") + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + if distributed: + dist.barrier() + # Load and eval + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"requant_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 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} soft_round_qat:alpha_1_to_16") + # Anneal Soft-Round alpha: 1 (soft) -> 16 (hard) over QAT window + # Uses registered buffer (tensor) so torch.compile sees dynamic value, not frozen float + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold # 0->1 as QAT runs + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + 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) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + 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() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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()} + # Pre-GPTQ magnitude pruning (like #606): zero out smallest weights → better compression + prune_pct = float(os.environ.get("PRUNE_PCT", 0)) + if prune_pct > 0: + n_pruned = 0 + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), prune_pct) + mask = v.abs() < thresh + n_pruned += mask.sum().item() + v[mask] = 0.0 + log0(f"pre_gptq_prune: zeroed bottom {prune_pct*100:.1f}% → {n_pruned} values") + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + over_bytes = len(test_blob) - target_model_bytes + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) # cap at 0.5% of params + log0(f"gptq:over by {over_bytes} bytes, total_params={total_params}, max_prune={max_prune}") + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + if len(prune_candidates) > max_prune: + log0(f"gptq:WARNING model too large! Need to prune {len(prune_candidates)} but cap={max_prune}. Reduce model size.") + log0(f"gptq:capping pruning at {max_prune} to avoid catastrophic quality loss") + prune_candidates = prune_candidates[:max_prune] + # Use fast preset (6) for binary search with safety margin, then verify with final preset + search_target = int(target_model_bytes * 0.997) # 0.3% safety margin for preset difference + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta, fast=True) + if len(blob) <= search_target: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + # Verify with final preset and increase pruning if needed + if best_n > 0: + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(prune_candidates): + best_n += max(1, best_n // 10) # 10% increments + best_n = min(best_n, len(prune_candidates)) + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}% of params)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load( + io.BytesIO(decompressed), + 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, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Full-epoch TTT (PR #672 recipe): train on ALL val data, then eval + if args.ttt_full_epochs > 0: + if distributed: + dist.barrier() + log0(f"ttt_full:start epochs={args.ttt_full_epochs} lr={args.ttt_lr} " + f"proj_mult={args.ttt_proj_lr_mult} fc_mult={args.ttt_fc_lr_mult}") + t_ttt = time.perf_counter() + ttt_full_pretrain(args, eval_model, rank, world_size, device, val_tokens, log_fn=log0) + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt_full:elapsed={ttt_elapsed:.1f}s") + # Now do sliding window eval on the TTT-adapted model + t_slide = time.perf_counter() + stride = args.eval_stride + sw_seq_len = getattr(args, 'eval_seq_len', args.train_seq_len) + ttt_val_loss, ttt_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=stride, eval_seq_len=sw_seq_len, + ) + log0(f"ttt_full_sliding:elapsed={time.perf_counter() - t_slide:.1f}s") + log0(f"final_ttt_full val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_full_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + # Legacy score-first TTT (PR#461/549 recipe) + elif args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] +Running PyTorch 2.9.1+cu128 +Wed Mar 25 07:31:05 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 39C P0 126W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 34C P0 120W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 32C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 40C P0 123W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 40C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 33C P0 113W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 38C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 33C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: clip_range=31 (int6) compressor=lzma +model_params:27124828 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:False active_layers:[] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:560.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9313 train_time:146ms step_avg:146.13ms +step:2/20000 train_loss:8.6866 train_time:226ms step_avg:112.82ms +step:3/20000 train_loss:7.9001 train_time:310ms step_avg:103.32ms +step:4/20000 train_loss:7.2185 train_time:395ms step_avg:98.70ms +step:5/20000 train_loss:6.9821 train_time:479ms step_avg:95.79ms +step:6/20000 train_loss:6.9115 train_time:564ms step_avg:93.98ms +step:7/20000 train_loss:6.8338 train_time:648ms step_avg:92.57ms +step:8/20000 train_loss:6.7689 train_time:732ms step_avg:91.56ms +step:9/20000 train_loss:6.3977 train_time:817ms step_avg:90.78ms +step:10/20000 train_loss:6.1341 train_time:902ms step_avg:90.18ms +step:500/20000 train_loss:2.3804 train_time:43136ms step_avg:86.27ms +step:1000/20000 train_loss:2.2571 train_time:86506ms step_avg:86.51ms +step:1500/20000 train_loss:2.2083 train_time:129882ms step_avg:86.59ms +step:2000/20000 train_loss:2.0517 train_time:173267ms step_avg:86.63ms +step:2500/20000 train_loss:2.1574 train_time:216607ms step_avg:86.64ms +step:3000/20000 train_loss:2.1348 train_time:259916ms step_avg:86.64ms +step:3500/20000 train_loss:2.1477 train_time:303214ms step_avg:86.63ms +step:4000/20000 train_loss:1.9384 train_time:346504ms step_avg:86.63ms +step:4000/20000 val_loss:2.0300 val_bpb:1.2023 train_time:346509ms step_avg:86.63ms +step:4500/20000 train_loss:2.0902 train_time:389796ms step_avg:86.62ms +step:5000/20000 train_loss:2.0677 train_time:433059ms step_avg:86.61ms +step:5500/20000 train_loss:1.9872 train_time:476312ms step_avg:86.60ms +swa:start step:5700 +step:6000/20000 train_loss:1.9088 train_time:519880ms step_avg:86.65ms +step:6461/20000 val_loss:1.9242 val_bpb:1.1396 train_time:560005ms step_avg:86.67ms +stopping_early: wallclock_cap train_time:560005ms step:6461/20000 +peak memory allocated: 20671 MiB reserved: 20724 MiB +swa:applying 16 snapshots, blending with EMA (0.50/0.50) +DIAGNOSTIC post_ema val_loss:1.9232 val_bpb:1.1390 eval_time:1997ms +Serialized model: 106440713 bytes +Code size: 110105 bytes +pre_gptq_prune: zeroed bottom 0.5% → 135033 values +gptq:collecting hessians batches=256 +gptq:hessians collected layers=68 time=39.1s +gptq:pre_prune artifact=15788956 target=15839895 +Saved quantized model to final_int6_model.pt +Serialized model int63+lzma: 15788956 bytes +Total submission size: 15899061 bytes +final_int6_roundtrip val_loss:1.9297 val_bpb:1.1429 eval_time:6679ms +final_int6_roundtrip_exact val_loss:1.92971162 val_bpb:1.14288392 +final_int6_sliding_window val_loss:1.8899 val_bpb:1.1193 stride:64 eval_time:75504ms +final_int6_sliding_window_exact val_loss:1.88994378 val_bpb:1.11933413 +final_int8_zlib_roundtrip_exact val_loss:1.88994378 val_bpb:1.11933413 +ttt:start lr=0.0001 epochs=3 chunks=131072 +ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.0001 ttt_epochs=3 freeze_blocks=9 +ttt_sliding:params unfrozen=5872660 frozen=21252168 + ttt_chunk [1/474] bpb=1.196617 time=0.7s + ttt_chunk [11/474] bpb=1.116559 time=5.7s + ttt_chunk [21/474] bpb=1.112152 time=10.7s + ttt_chunk [31/474] bpb=1.110862 time=15.7s + ttt_chunk [41/474] bpb=1.119142 time=20.7s + ttt_chunk [51/474] bpb=1.126728 time=25.6s + ttt_chunk [61/474] bpb=1.125006 time=30.6s + ttt_chunk [71/474] bpb=1.126600 time=35.6s + ttt_chunk [81/474] bpb=1.127177 time=40.6s + ttt_chunk [91/474] bpb=1.129259 time=45.6s + ttt_chunk [101/474] bpb=1.125823 time=50.6s + ttt_chunk [111/474] bpb=1.126224 time=55.6s + ttt_chunk [121/474] bpb=1.129620 time=60.6s + ttt_chunk [131/474] bpb=1.130398 time=65.6s + ttt_chunk [141/474] bpb=1.130335 time=70.5s + ttt_chunk [151/474] bpb=1.128617 time=75.5s + ttt_chunk [161/474] bpb=1.129564 time=80.5s + ttt_chunk [171/474] bpb=1.128329 time=85.5s + ttt_chunk [181/474] bpb=1.129114 time=90.5s + ttt_chunk [191/474] bpb=1.128090 time=95.5s + ttt_chunk [201/474] bpb=1.127323 time=100.5s + ttt_chunk [211/474] bpb=1.126048 time=105.5s + ttt_chunk [221/474] bpb=1.126192 time=110.4s + ttt_chunk [231/474] bpb=1.125677 time=115.4s + ttt_chunk [241/474] bpb=1.124596 time=120.4s + ttt_chunk [251/474] bpb=1.125662 time=125.5s + ttt_chunk [261/474] bpb=1.126347 time=130.5s + ttt_chunk [271/474] bpb=1.124921 time=135.4s + ttt_chunk [281/474] bpb=1.124596 time=140.4s + ttt_chunk [291/474] bpb=1.123143 time=145.4s + ttt_chunk [301/474] bpb=1.123564 time=150.4s + ttt_chunk [311/474] bpb=1.123019 time=155.4s + ttt_chunk [321/474] bpb=1.121416 time=160.3s + ttt_chunk [331/474] bpb=1.120385 time=165.4s + ttt_chunk [341/474] bpb=1.119662 time=170.3s + ttt_chunk [351/474] bpb=1.118090 time=175.3s + ttt_chunk [361/474] bpb=1.118631 time=180.3s + ttt_chunk [371/474] bpb=1.118314 time=185.3s + ttt_chunk [381/474] bpb=1.119079 time=190.3s + ttt_chunk [391/474] bpb=1.120151 time=195.3s + ttt_chunk [401/474] bpb=1.120607 time=200.2s + ttt_chunk [411/474] bpb=1.121022 time=205.2s + ttt_chunk [421/474] bpb=1.122460 time=210.2s + ttt_chunk [431/474] bpb=1.121030 time=215.2s + ttt_chunk [441/474] bpb=1.120658 time=220.2s + ttt_chunk [451/474] bpb=1.120013 time=225.1s + ttt_chunk [461/474] bpb=1.120218 time=230.1s + ttt_chunk [471/474] bpb=1.120311 time=235.1s + ttt_chunk [474/474] bpb=1.120180 time=236.2s +ttt_sliding:done val_loss=1.889875 val_bpb=1.119293 elapsed=236.2s +ttt:elapsed=236.6s +final_ttt val_loss:1.8899 val_bpb:1.1193 +final_ttt_exact val_loss:1.88987464 val_bpb:1.11929318 diff --git a/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed2025.log b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed2025.log new file mode 100644 index 000000000..2c56d6549 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed2025.log @@ -0,0 +1,11598 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # 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", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) # 31=int6 (63 levels), 15=int5 (31 levels) + compressor = os.environ.get("COMPRESSOR", "lzma") # lzma or zstd + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs, then eval + ttt_full_epochs = int(os.environ.get("TTT_FULL_EPOCHS", 0)) + ttt_proj_lr_mult = float(os.environ.get("TTT_PROJ_LR_MULT", 3.0)) + ttt_fc_lr_mult = float(os.environ.get("TTT_FC_LR_MULT", 0.5)) + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + # Requant-only mode: skip training, load fp32 model, re-quantize + eval + requant_only = bool(int(os.environ.get("REQUANT_ONLY", "0"))) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False # bool: only 2 values, max 2 torch.compile graphs + _clip_range: int = 31 # 31=int6, 15=int5 — set once before compile, not changed during training + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # Registered buffer: torch.compile treats as dynamic input (not frozen at trace time) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range # int, set once before compile + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + # Soft-Round: fully differentiable rounding (no STE, matching #606) + # alpha buffer: 1.0 (soft, linear) -> 16.0 (hard, nearly round()) + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + # Fully differentiable: gradients flow through tanh soft-round + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + 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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + 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, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + 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, + ) + 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) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + 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 + 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, lora=None) -> 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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def ttt_full_pretrain( + args, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + log_fn=None, +) -> None: + """Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs with + cosine LR decay and differentiated per-layer LRs. All params unfrozen. + Data sharded across GPUs with gradient all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_epochs = args.ttt_full_epochs + batch_seqs = args.ttt_train_batch_seqs # 32 + + # Shard val data across GPUs + rank_tokens = total_tokens // world_size + rank_start = rank * rank_tokens + rank_end = rank_start + rank_tokens + + # Unfreeze params with differentiated LR groups + proj_params, fc_params, other_params = [], [], [] + for name, p in base_model.named_parameters(): + # Respect freeze settings + freeze = False + if args.ttt_freeze_blocks > 0: + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + freeze = any(f"blocks.{bi}." in name for bi in frozen_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + continue + p.requires_grad_(True) + if "mlp.proj" in name: + proj_params.append(p) + elif "mlp.fc" in name: + fc_params.append(p) + else: + other_params.append(p) + + base_lr = args.ttt_lr + optimizer = torch.optim.AdamW([ + {"params": proj_params, "lr": base_lr * args.ttt_proj_lr_mult}, + {"params": fc_params, "lr": base_lr * args.ttt_fc_lr_mult}, + {"params": other_params, "lr": base_lr}, + ], weight_decay=args.ttt_wd) + + total_trainable = sum(p.numel() for pg in [proj_params, fc_params, other_params] for p in pg) + n_rank_seqs = max(1, (rank_end - rank_start - seq_len) // (batch_seqs * seq_len)) + steps_per_epoch = n_rank_seqs + total_steps = ttt_epochs * steps_per_epoch + + if log_fn: + log_fn(f"ttt_full:start epochs={ttt_epochs} trainable={total_trainable} " + f"rank_tokens={rank_tokens} steps_per_ep={steps_per_epoch} " + f"total_steps={total_steps} base_lr={base_lr}") + + base_model.train() + t0 = time.perf_counter() + global_step = 0 + + for epoch in range(ttt_epochs): + for bs_idx in range(0, rank_end - rank_start - seq_len, batch_seqs * seq_len): + # Cosine LR schedule across all TTT steps + progress = global_step / max(total_steps, 1) + cos_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('initial_lr', pg['lr']) * cos_mul if global_step == 0 else pg['lr'] + # Set initial_lr on first step + if global_step == 0: + for pg in optimizer.param_groups: + pg['initial_lr'] = pg['lr'] + else: + for pg in optimizer.param_groups: + pg['lr'] = pg['initial_lr'] * cos_mul + + start_tok = rank_start + bs_idx + end_tok = min(start_tok + batch_seqs * seq_len + 1, val_tokens.numel()) + if end_tok - start_tok < seq_len + 1: + continue + + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + n_seqs = (local.numel() - 1) // seq_len + if n_seqs == 0: + continue + x = local[:n_seqs * seq_len].reshape(n_seqs, seq_len) + y = local[1:n_seqs * seq_len + 1].reshape(n_seqs, 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() + + # DDP-style gradient sync + 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) + + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.ttt_grad_clip) + optimizer.step() + global_step += 1 + + if log_fn and (epoch % 5 == 0 or epoch == ttt_epochs - 1): + elapsed = time.perf_counter() - t0 + log_fn(f" ttt_full:epoch {epoch+1}/{ttt_epochs} loss={loss.item():.4f} " + f"lr={optimizer.param_groups[2]['lr']:.6f} time={elapsed:.1f}s") + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(True) + if log_fn: + log_fn(f"ttt_full:done total_steps={global_step} time={time.perf_counter() - t0:.1f}s") + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + stride = args.eval_stride if args.eval_stride > 0 else 64 + log0(f"eval_only:TTT disabled, computing sliding window BPB stride={stride}") + 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=stride, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f} stride:{stride}") + if distributed: + dist.destroy_process_group() + return + if args.requant_only: + log0(f"requant_only:loading fp32 model, re-quantizing with damp={args.gptq_damp_factor}") + fp32_state = torch.load("final_model.pt", map_location="cpu") + base_model.load_state_dict(fp32_state, strict=False) + base_model.eval() + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + sd_cpu = {k: v for k, v in sd_cpu.items() if "mtp_heads" not in k} + # Re-collect Hessians and re-quantize + hessians = None + if args.gptq_enabled: + log0(f"requant:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"requant:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Pruning + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + log0(f"requant:artifact={len(quant_blob)} code={code_bytes} total={len(quant_blob)+code_bytes}") + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + if distributed: + dist.barrier() + # Load and eval + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"requant_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 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} soft_round_qat:alpha_1_to_16") + # Anneal Soft-Round alpha: 1 (soft) -> 16 (hard) over QAT window + # Uses registered buffer (tensor) so torch.compile sees dynamic value, not frozen float + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold # 0->1 as QAT runs + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + 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) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + 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() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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()} + # Pre-GPTQ magnitude pruning (like #606): zero out smallest weights → better compression + prune_pct = float(os.environ.get("PRUNE_PCT", 0)) + if prune_pct > 0: + n_pruned = 0 + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), prune_pct) + mask = v.abs() < thresh + n_pruned += mask.sum().item() + v[mask] = 0.0 + log0(f"pre_gptq_prune: zeroed bottom {prune_pct*100:.1f}% → {n_pruned} values") + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + over_bytes = len(test_blob) - target_model_bytes + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) # cap at 0.5% of params + log0(f"gptq:over by {over_bytes} bytes, total_params={total_params}, max_prune={max_prune}") + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + if len(prune_candidates) > max_prune: + log0(f"gptq:WARNING model too large! Need to prune {len(prune_candidates)} but cap={max_prune}. Reduce model size.") + log0(f"gptq:capping pruning at {max_prune} to avoid catastrophic quality loss") + prune_candidates = prune_candidates[:max_prune] + # Use fast preset (6) for binary search with safety margin, then verify with final preset + search_target = int(target_model_bytes * 0.997) # 0.3% safety margin for preset difference + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta, fast=True) + if len(blob) <= search_target: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + # Verify with final preset and increase pruning if needed + if best_n > 0: + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(prune_candidates): + best_n += max(1, best_n // 10) # 10% increments + best_n = min(best_n, len(prune_candidates)) + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}% of params)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load( + io.BytesIO(decompressed), + 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, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Full-epoch TTT (PR #672 recipe): train on ALL val data, then eval + if args.ttt_full_epochs > 0: + if distributed: + dist.barrier() + log0(f"ttt_full:start epochs={args.ttt_full_epochs} lr={args.ttt_lr} " + f"proj_mult={args.ttt_proj_lr_mult} fc_mult={args.ttt_fc_lr_mult}") + t_ttt = time.perf_counter() + ttt_full_pretrain(args, eval_model, rank, world_size, device, val_tokens, log_fn=log0) + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt_full:elapsed={ttt_elapsed:.1f}s") + # Now do sliding window eval on the TTT-adapted model + t_slide = time.perf_counter() + stride = args.eval_stride + sw_seq_len = getattr(args, 'eval_seq_len', args.train_seq_len) + ttt_val_loss, ttt_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=stride, eval_seq_len=sw_seq_len, + ) + log0(f"ttt_full_sliding:elapsed={time.perf_counter() - t_slide:.1f}s") + log0(f"final_ttt_full val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_full_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + # Legacy score-first TTT (PR#461/549 recipe) + elif args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] +Running PyTorch 2.9.1+cu128 +Wed Mar 25 06:52:28 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 45C P0 130W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 38C P0 123W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 36C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 46C P0 128W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 47C P0 125W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 37C P0 115W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 44C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 36C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: clip_range=31 (int6) compressor=zstd +model_params:27124828 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:False active_layers:[] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:560.000 +seed:2025 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9306 val_bpb:4.1047 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9326 train_time:145ms step_avg:145.29ms +step:2/20000 train_loss:8.8150 train_time:227ms step_avg:113.29ms +step:3/20000 train_loss:7.9592 train_time:312ms step_avg:103.85ms +step:4/20000 train_loss:7.1654 train_time:397ms step_avg:99.17ms +step:5/20000 train_loss:6.9463 train_time:481ms step_avg:96.26ms +step:6/20000 train_loss:6.8441 train_time:566ms step_avg:94.39ms +step:7/20000 train_loss:6.6819 train_time:652ms step_avg:93.07ms +step:8/20000 train_loss:6.6619 train_time:737ms step_avg:92.13ms +step:9/20000 train_loss:6.3738 train_time:822ms step_avg:91.31ms +step:10/20000 train_loss:6.0574 train_time:906ms step_avg:90.59ms +step:500/20000 train_loss:2.3894 train_time:45738ms step_avg:91.48ms +step:1000/20000 train_loss:2.2499 train_time:125384ms step_avg:125.38ms +step:1500/20000 train_loss:2.1654 train_time:232740ms step_avg:155.16ms +step:2000/20000 train_loss:2.0076 train_time:340714ms step_avg:170.36ms +swa:start step:2350 +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # 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", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) # 31=int6 (63 levels), 15=int5 (31 levels) + compressor = os.environ.get("COMPRESSOR", "lzma") # lzma or zstd + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs, then eval + ttt_full_epochs = int(os.environ.get("TTT_FULL_EPOCHS", 0)) + ttt_proj_lr_mult = float(os.environ.get("TTT_PROJ_LR_MULT", 3.0)) + ttt_fc_lr_mult = float(os.environ.get("TTT_FC_LR_MULT", 0.5)) + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + # Requant-only mode: skip training, load fp32 model, re-quantize + eval + requant_only = bool(int(os.environ.get("REQUANT_ONLY", "0"))) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False # bool: only 2 values, max 2 torch.compile graphs + _clip_range: int = 31 # 31=int6, 15=int5 — set once before compile, not changed during training + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # Registered buffer: torch.compile treats as dynamic input (not frozen at trace time) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range # int, set once before compile + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + # Soft-Round: fully differentiable rounding (no STE, matching #606) + # alpha buffer: 1.0 (soft, linear) -> 16.0 (hard, nearly round()) + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + # Fully differentiable: gradients flow through tanh soft-round + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + 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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + 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, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + 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, + ) + 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) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + 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 + 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, lora=None) -> 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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def ttt_full_pretrain( + args, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + log_fn=None, +) -> None: + """Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs with + cosine LR decay and differentiated per-layer LRs. All params unfrozen. + Data sharded across GPUs with gradient all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_epochs = args.ttt_full_epochs + batch_seqs = args.ttt_train_batch_seqs # 32 + + # Shard val data across GPUs + rank_tokens = total_tokens // world_size + rank_start = rank * rank_tokens + rank_end = rank_start + rank_tokens + + # Unfreeze params with differentiated LR groups + proj_params, fc_params, other_params = [], [], [] + for name, p in base_model.named_parameters(): + # Respect freeze settings + freeze = False + if args.ttt_freeze_blocks > 0: + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + freeze = any(f"blocks.{bi}." in name for bi in frozen_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + continue + p.requires_grad_(True) + if "mlp.proj" in name: + proj_params.append(p) + elif "mlp.fc" in name: + fc_params.append(p) + else: + other_params.append(p) + + base_lr = args.ttt_lr + optimizer = torch.optim.AdamW([ + {"params": proj_params, "lr": base_lr * args.ttt_proj_lr_mult}, + {"params": fc_params, "lr": base_lr * args.ttt_fc_lr_mult}, + {"params": other_params, "lr": base_lr}, + ], weight_decay=args.ttt_wd) + + total_trainable = sum(p.numel() for pg in [proj_params, fc_params, other_params] for p in pg) + n_rank_seqs = max(1, (rank_end - rank_start - seq_len) // (batch_seqs * seq_len)) + steps_per_epoch = n_rank_seqs + total_steps = ttt_epochs * steps_per_epoch + + if log_fn: + log_fn(f"ttt_full:start epochs={ttt_epochs} trainable={total_trainable} " + f"rank_tokens={rank_tokens} steps_per_ep={steps_per_epoch} " + f"total_steps={total_steps} base_lr={base_lr}") + + base_model.train() + t0 = time.perf_counter() + global_step = 0 + + for epoch in range(ttt_epochs): + for bs_idx in range(0, rank_end - rank_start - seq_len, batch_seqs * seq_len): + # Cosine LR schedule across all TTT steps + progress = global_step / max(total_steps, 1) + cos_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('initial_lr', pg['lr']) * cos_mul if global_step == 0 else pg['lr'] + # Set initial_lr on first step + if global_step == 0: + for pg in optimizer.param_groups: + pg['initial_lr'] = pg['lr'] + else: + for pg in optimizer.param_groups: + pg['lr'] = pg['initial_lr'] * cos_mul + + start_tok = rank_start + bs_idx + end_tok = min(start_tok + batch_seqs * seq_len + 1, val_tokens.numel()) + if end_tok - start_tok < seq_len + 1: + continue + + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + n_seqs = (local.numel() - 1) // seq_len + if n_seqs == 0: + continue + x = local[:n_seqs * seq_len].reshape(n_seqs, seq_len) + y = local[1:n_seqs * seq_len + 1].reshape(n_seqs, 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() + + # DDP-style gradient sync + 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) + + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.ttt_grad_clip) + optimizer.step() + global_step += 1 + + if log_fn and (epoch % 5 == 0 or epoch == ttt_epochs - 1): + elapsed = time.perf_counter() - t0 + log_fn(f" ttt_full:epoch {epoch+1}/{ttt_epochs} loss={loss.item():.4f} " + f"lr={optimizer.param_groups[2]['lr']:.6f} time={elapsed:.1f}s") + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(True) + if log_fn: + log_fn(f"ttt_full:done total_steps={global_step} time={time.perf_counter() - t0:.1f}s") + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + stride = args.eval_stride if args.eval_stride > 0 else 64 + log0(f"eval_only:TTT disabled, computing sliding window BPB stride={stride}") + 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=stride, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f} stride:{stride}") + if distributed: + dist.destroy_process_group() + return + if args.requant_only: + log0(f"requant_only:loading fp32 model, re-quantizing with damp={args.gptq_damp_factor}") + fp32_state = torch.load("final_model.pt", map_location="cpu") + base_model.load_state_dict(fp32_state, strict=False) + base_model.eval() + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + sd_cpu = {k: v for k, v in sd_cpu.items() if "mtp_heads" not in k} + # Re-collect Hessians and re-quantize + hessians = None + if args.gptq_enabled: + log0(f"requant:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"requant:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Pruning + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + log0(f"requant:artifact={len(quant_blob)} code={code_bytes} total={len(quant_blob)+code_bytes}") + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + if distributed: + dist.barrier() + # Load and eval + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"requant_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 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} soft_round_qat:alpha_1_to_16") + # Anneal Soft-Round alpha: 1 (soft) -> 16 (hard) over QAT window + # Uses registered buffer (tensor) so torch.compile sees dynamic value, not frozen float + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold # 0->1 as QAT runs + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + 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) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + 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() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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()} + # Pre-GPTQ magnitude pruning (like #606): zero out smallest weights → better compression + prune_pct = float(os.environ.get("PRUNE_PCT", 0)) + if prune_pct > 0: + n_pruned = 0 + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), prune_pct) + mask = v.abs() < thresh + n_pruned += mask.sum().item() + v[mask] = 0.0 + log0(f"pre_gptq_prune: zeroed bottom {prune_pct*100:.1f}% → {n_pruned} values") + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + over_bytes = len(test_blob) - target_model_bytes + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) # cap at 0.5% of params + log0(f"gptq:over by {over_bytes} bytes, total_params={total_params}, max_prune={max_prune}") + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + if len(prune_candidates) > max_prune: + log0(f"gptq:WARNING model too large! Need to prune {len(prune_candidates)} but cap={max_prune}. Reduce model size.") + log0(f"gptq:capping pruning at {max_prune} to avoid catastrophic quality loss") + prune_candidates = prune_candidates[:max_prune] + # Use fast preset (6) for binary search with safety margin, then verify with final preset + search_target = int(target_model_bytes * 0.997) # 0.3% safety margin for preset difference + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta, fast=True) + if len(blob) <= search_target: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + # Verify with final preset and increase pruning if needed + if best_n > 0: + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(prune_candidates): + best_n += max(1, best_n // 10) # 10% increments + best_n = min(best_n, len(prune_candidates)) + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}% of params)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load( + io.BytesIO(decompressed), + 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, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Full-epoch TTT (PR #672 recipe): train on ALL val data, then eval + if args.ttt_full_epochs > 0: + if distributed: + dist.barrier() + log0(f"ttt_full:start epochs={args.ttt_full_epochs} lr={args.ttt_lr} " + f"proj_mult={args.ttt_proj_lr_mult} fc_mult={args.ttt_fc_lr_mult}") + t_ttt = time.perf_counter() + ttt_full_pretrain(args, eval_model, rank, world_size, device, val_tokens, log_fn=log0) + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt_full:elapsed={ttt_elapsed:.1f}s") + # Now do sliding window eval on the TTT-adapted model + t_slide = time.perf_counter() + stride = args.eval_stride + sw_seq_len = getattr(args, 'eval_seq_len', args.train_seq_len) + ttt_val_loss, ttt_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=stride, eval_seq_len=sw_seq_len, + ) + log0(f"ttt_full_sliding:elapsed={time.perf_counter() - t_slide:.1f}s") + log0(f"final_ttt_full val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_full_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + # Legacy score-first TTT (PR#461/549 recipe) + elif args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] +Running PyTorch 2.9.1+cu128 +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # 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", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) # 31=int6 (63 levels), 15=int5 (31 levels) + compressor = os.environ.get("COMPRESSOR", "lzma") # lzma or zstd + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs, then eval + ttt_full_epochs = int(os.environ.get("TTT_FULL_EPOCHS", 0)) + ttt_proj_lr_mult = float(os.environ.get("TTT_PROJ_LR_MULT", 3.0)) + ttt_fc_lr_mult = float(os.environ.get("TTT_FC_LR_MULT", 0.5)) + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + # Requant-only mode: skip training, load fp32 model, re-quantize + eval + requant_only = bool(int(os.environ.get("REQUANT_ONLY", "0"))) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False # bool: only 2 values, max 2 torch.compile graphs + _clip_range: int = 31 # 31=int6, 15=int5 — set once before compile, not changed during training + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # Registered buffer: torch.compile treats as dynamic input (not frozen at trace time) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range # int, set once before compile + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + # Soft-Round: fully differentiable rounding (no STE, matching #606) + # alpha buffer: 1.0 (soft, linear) -> 16.0 (hard, nearly round()) + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + # Fully differentiable: gradients flow through tanh soft-round + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + 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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + 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, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + 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, + ) + 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) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + 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 + 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, lora=None) -> 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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def ttt_full_pretrain( + args, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + log_fn=None, +) -> None: + """Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs with + cosine LR decay and differentiated per-layer LRs. All params unfrozen. + Data sharded across GPUs with gradient all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_epochs = args.ttt_full_epochs + batch_seqs = args.ttt_train_batch_seqs # 32 + + # Shard val data across GPUs + rank_tokens = total_tokens // world_size + rank_start = rank * rank_tokens + rank_end = rank_start + rank_tokens + + # Unfreeze params with differentiated LR groups + proj_params, fc_params, other_params = [], [], [] + for name, p in base_model.named_parameters(): + # Respect freeze settings + freeze = False + if args.ttt_freeze_blocks > 0: + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + freeze = any(f"blocks.{bi}." in name for bi in frozen_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + continue + p.requires_grad_(True) + if "mlp.proj" in name: + proj_params.append(p) + elif "mlp.fc" in name: + fc_params.append(p) + else: + other_params.append(p) + + base_lr = args.ttt_lr + optimizer = torch.optim.AdamW([ + {"params": proj_params, "lr": base_lr * args.ttt_proj_lr_mult}, + {"params": fc_params, "lr": base_lr * args.ttt_fc_lr_mult}, + {"params": other_params, "lr": base_lr}, + ], weight_decay=args.ttt_wd) + + total_trainable = sum(p.numel() for pg in [proj_params, fc_params, other_params] for p in pg) + n_rank_seqs = max(1, (rank_end - rank_start - seq_len) // (batch_seqs * seq_len)) + steps_per_epoch = n_rank_seqs + total_steps = ttt_epochs * steps_per_epoch + + if log_fn: + log_fn(f"ttt_full:start epochs={ttt_epochs} trainable={total_trainable} " + f"rank_tokens={rank_tokens} steps_per_ep={steps_per_epoch} " + f"total_steps={total_steps} base_lr={base_lr}") + + base_model.train() + t0 = time.perf_counter() + global_step = 0 + + for epoch in range(ttt_epochs): + for bs_idx in range(0, rank_end - rank_start - seq_len, batch_seqs * seq_len): + # Cosine LR schedule across all TTT steps + progress = global_step / max(total_steps, 1) + cos_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('initial_lr', pg['lr']) * cos_mul if global_step == 0 else pg['lr'] + # Set initial_lr on first step + if global_step == 0: + for pg in optimizer.param_groups: + pg['initial_lr'] = pg['lr'] + else: + for pg in optimizer.param_groups: + pg['lr'] = pg['initial_lr'] * cos_mul + + start_tok = rank_start + bs_idx + end_tok = min(start_tok + batch_seqs * seq_len + 1, val_tokens.numel()) + if end_tok - start_tok < seq_len + 1: + continue + + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + n_seqs = (local.numel() - 1) // seq_len + if n_seqs == 0: + continue + x = local[:n_seqs * seq_len].reshape(n_seqs, seq_len) + y = local[1:n_seqs * seq_len + 1].reshape(n_seqs, 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() + + # DDP-style gradient sync + 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) + + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.ttt_grad_clip) + optimizer.step() + global_step += 1 + + if log_fn and (epoch % 5 == 0 or epoch == ttt_epochs - 1): + elapsed = time.perf_counter() - t0 + log_fn(f" ttt_full:epoch {epoch+1}/{ttt_epochs} loss={loss.item():.4f} " + f"lr={optimizer.param_groups[2]['lr']:.6f} time={elapsed:.1f}s") + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(True) + if log_fn: + log_fn(f"ttt_full:done total_steps={global_step} time={time.perf_counter() - t0:.1f}s") + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + stride = args.eval_stride if args.eval_stride > 0 else 64 + log0(f"eval_only:TTT disabled, computing sliding window BPB stride={stride}") + 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=stride, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f} stride:{stride}") + if distributed: + dist.destroy_process_group() + return + if args.requant_only: + log0(f"requant_only:loading fp32 model, re-quantizing with damp={args.gptq_damp_factor}") + fp32_state = torch.load("final_model.pt", map_location="cpu") + base_model.load_state_dict(fp32_state, strict=False) + base_model.eval() + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + sd_cpu = {k: v for k, v in sd_cpu.items() if "mtp_heads" not in k} + # Re-collect Hessians and re-quantize + hessians = None + if args.gptq_enabled: + log0(f"requant:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"requant:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Pruning + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + log0(f"requant:artifact={len(quant_blob)} code={code_bytes} total={len(quant_blob)+code_bytes}") + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + if distributed: + dist.barrier() + # Load and eval + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"requant_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 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} soft_round_qat:alpha_1_to_16") + # Anneal Soft-Round alpha: 1 (soft) -> 16 (hard) over QAT window + # Uses registered buffer (tensor) so torch.compile sees dynamic value, not frozen float + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold # 0->1 as QAT runs + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + 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) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + 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() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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()} + # Pre-GPTQ magnitude pruning (like #606): zero out smallest weights → better compression + prune_pct = float(os.environ.get("PRUNE_PCT", 0)) + if prune_pct > 0: + n_pruned = 0 + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), prune_pct) + mask = v.abs() < thresh + n_pruned += mask.sum().item() + v[mask] = 0.0 + log0(f"pre_gptq_prune: zeroed bottom {prune_pct*100:.1f}% → {n_pruned} values") + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + over_bytes = len(test_blob) - target_model_bytes + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) # cap at 0.5% of params + log0(f"gptq:over by {over_bytes} bytes, total_params={total_params}, max_prune={max_prune}") + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + if len(prune_candidates) > max_prune: + log0(f"gptq:WARNING model too large! Need to prune {len(prune_candidates)} but cap={max_prune}. Reduce model size.") + log0(f"gptq:capping pruning at {max_prune} to avoid catastrophic quality loss") + prune_candidates = prune_candidates[:max_prune] + # Use fast preset (6) for binary search with safety margin, then verify with final preset + search_target = int(target_model_bytes * 0.997) # 0.3% safety margin for preset difference + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta, fast=True) + if len(blob) <= search_target: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + # Verify with final preset and increase pruning if needed + if best_n > 0: + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(prune_candidates): + best_n += max(1, best_n // 10) # 10% increments + best_n = min(best_n, len(prune_candidates)) + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}% of params)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load( + io.BytesIO(decompressed), + 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, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Full-epoch TTT (PR #672 recipe): train on ALL val data, then eval + if args.ttt_full_epochs > 0: + if distributed: + dist.barrier() + log0(f"ttt_full:start epochs={args.ttt_full_epochs} lr={args.ttt_lr} " + f"proj_mult={args.ttt_proj_lr_mult} fc_mult={args.ttt_fc_lr_mult}") + t_ttt = time.perf_counter() + ttt_full_pretrain(args, eval_model, rank, world_size, device, val_tokens, log_fn=log0) + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt_full:elapsed={ttt_elapsed:.1f}s") + # Now do sliding window eval on the TTT-adapted model + t_slide = time.perf_counter() + stride = args.eval_stride + sw_seq_len = getattr(args, 'eval_seq_len', args.train_seq_len) + ttt_val_loss, ttt_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=stride, eval_seq_len=sw_seq_len, + ) + log0(f"ttt_full_sliding:elapsed={time.perf_counter() - t_slide:.1f}s") + log0(f"final_ttt_full val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_full_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + # Legacy score-first TTT (PR#461/549 recipe) + elif args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] +Running PyTorch 2.9.1+cu128 +Wed Mar 25 07:04:03 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 47C P0 132W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 38C P0 124W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 36C P0 121W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 48C P0 130W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 48C P0 126W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 36C P0 114W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 46C P0 123W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 36C P0 118W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: clip_range=31 (int6) compressor=lzma +Wed Mar 25 07:04:04 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 47C P0 131W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 38C P0 123W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 36C P0 121W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 47C P0 130W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 48C P0 126W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 36C P0 114W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 45C P0 123W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 36C P0 118W / 700W | 4557MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: clip_range=31 (int6) compressor=lzma +model_params:27124828 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:False active_layers:[] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:560.000 +seed:2025 +model_params:27124828 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:False active_layers:[] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:560.000 +seed:2025 +warmup_step:1/20 +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 +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:6.9306 val_bpb:4.1047 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9326 train_time:175ms step_avg:175.00ms +step:2/20000 train_loss:8.8150 train_time:264ms step_avg:132.25ms +step:3/20000 train_loss:8.0363 train_time:366ms step_avg:121.85ms +step:4/20000 train_loss:7.1227 train_time:458ms step_avg:114.42ms +step:5/20000 train_loss:6.9731 train_time:556ms step_avg:111.14ms +step:6/20000 train_loss:6.9005 train_time:653ms step_avg:108.76ms +step:7/20000 train_loss:6.7004 train_time:737ms step_avg:105.24ms +step:8/20000 train_loss:6.6940 train_time:822ms step_avg:102.80ms +step:9/20000 train_loss:6.3936 train_time:921ms step_avg:102.35ms +step:10/20000 train_loss:6.1039 train_time:1017ms step_avg:101.72ms +step:0/20000 val_loss:6.9306 val_bpb:4.1047 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9326 train_time:741ms step_avg:740.56ms +step:2/20000 train_loss:8.8150 train_time:1108ms step_avg:553.87ms +step:3/20000 train_loss:8.6347 train_time:1316ms step_avg:438.59ms +step:4/20000 train_loss:8.0001 train_time:1505ms step_avg:376.15ms +step:5/20000 train_loss:7.2377 train_time:1697ms step_avg:339.39ms +step:6/20000 train_loss:6.6727 train_time:2007ms step_avg:334.50ms +step:7/20000 train_loss:6.3174 train_time:2385ms step_avg:340.66ms +step:8/20000 train_loss:6.0913 train_time:2749ms step_avg:343.64ms +step:9/20000 train_loss:5.9749 train_time:3103ms step_avg:344.81ms +step:10/20000 train_loss:5.8786 train_time:3477ms step_avg:347.75ms +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # 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", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) # 31=int6 (63 levels), 15=int5 (31 levels) + compressor = os.environ.get("COMPRESSOR", "lzma") # lzma or zstd + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs, then eval + ttt_full_epochs = int(os.environ.get("TTT_FULL_EPOCHS", 0)) + ttt_proj_lr_mult = float(os.environ.get("TTT_PROJ_LR_MULT", 3.0)) + ttt_fc_lr_mult = float(os.environ.get("TTT_FC_LR_MULT", 0.5)) + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + # Requant-only mode: skip training, load fp32 model, re-quantize + eval + requant_only = bool(int(os.environ.get("REQUANT_ONLY", "0"))) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False # bool: only 2 values, max 2 torch.compile graphs + _clip_range: int = 31 # 31=int6, 15=int5 — set once before compile, not changed during training + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # Registered buffer: torch.compile treats as dynamic input (not frozen at trace time) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range # int, set once before compile + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + # Soft-Round: fully differentiable rounding (no STE, matching #606) + # alpha buffer: 1.0 (soft, linear) -> 16.0 (hard, nearly round()) + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + # Fully differentiable: gradients flow through tanh soft-round + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + 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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + 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, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + 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, + ) + 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) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + 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 + 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, lora=None) -> 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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def ttt_full_pretrain( + args, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + log_fn=None, +) -> None: + """Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs with + cosine LR decay and differentiated per-layer LRs. All params unfrozen. + Data sharded across GPUs with gradient all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_epochs = args.ttt_full_epochs + batch_seqs = args.ttt_train_batch_seqs # 32 + + # Shard val data across GPUs + rank_tokens = total_tokens // world_size + rank_start = rank * rank_tokens + rank_end = rank_start + rank_tokens + + # Unfreeze params with differentiated LR groups + proj_params, fc_params, other_params = [], [], [] + for name, p in base_model.named_parameters(): + # Respect freeze settings + freeze = False + if args.ttt_freeze_blocks > 0: + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + freeze = any(f"blocks.{bi}." in name for bi in frozen_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + continue + p.requires_grad_(True) + if "mlp.proj" in name: + proj_params.append(p) + elif "mlp.fc" in name: + fc_params.append(p) + else: + other_params.append(p) + + base_lr = args.ttt_lr + optimizer = torch.optim.AdamW([ + {"params": proj_params, "lr": base_lr * args.ttt_proj_lr_mult}, + {"params": fc_params, "lr": base_lr * args.ttt_fc_lr_mult}, + {"params": other_params, "lr": base_lr}, + ], weight_decay=args.ttt_wd) + + total_trainable = sum(p.numel() for pg in [proj_params, fc_params, other_params] for p in pg) + n_rank_seqs = max(1, (rank_end - rank_start - seq_len) // (batch_seqs * seq_len)) + steps_per_epoch = n_rank_seqs + total_steps = ttt_epochs * steps_per_epoch + + if log_fn: + log_fn(f"ttt_full:start epochs={ttt_epochs} trainable={total_trainable} " + f"rank_tokens={rank_tokens} steps_per_ep={steps_per_epoch} " + f"total_steps={total_steps} base_lr={base_lr}") + + base_model.train() + t0 = time.perf_counter() + global_step = 0 + + for epoch in range(ttt_epochs): + for bs_idx in range(0, rank_end - rank_start - seq_len, batch_seqs * seq_len): + # Cosine LR schedule across all TTT steps + progress = global_step / max(total_steps, 1) + cos_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('initial_lr', pg['lr']) * cos_mul if global_step == 0 else pg['lr'] + # Set initial_lr on first step + if global_step == 0: + for pg in optimizer.param_groups: + pg['initial_lr'] = pg['lr'] + else: + for pg in optimizer.param_groups: + pg['lr'] = pg['initial_lr'] * cos_mul + + start_tok = rank_start + bs_idx + end_tok = min(start_tok + batch_seqs * seq_len + 1, val_tokens.numel()) + if end_tok - start_tok < seq_len + 1: + continue + + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + n_seqs = (local.numel() - 1) // seq_len + if n_seqs == 0: + continue + x = local[:n_seqs * seq_len].reshape(n_seqs, seq_len) + y = local[1:n_seqs * seq_len + 1].reshape(n_seqs, 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() + + # DDP-style gradient sync + 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) + + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.ttt_grad_clip) + optimizer.step() + global_step += 1 + + if log_fn and (epoch % 5 == 0 or epoch == ttt_epochs - 1): + elapsed = time.perf_counter() - t0 + log_fn(f" ttt_full:epoch {epoch+1}/{ttt_epochs} loss={loss.item():.4f} " + f"lr={optimizer.param_groups[2]['lr']:.6f} time={elapsed:.1f}s") + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(True) + if log_fn: + log_fn(f"ttt_full:done total_steps={global_step} time={time.perf_counter() - t0:.1f}s") + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + stride = args.eval_stride if args.eval_stride > 0 else 64 + log0(f"eval_only:TTT disabled, computing sliding window BPB stride={stride}") + 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=stride, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f} stride:{stride}") + if distributed: + dist.destroy_process_group() + return + if args.requant_only: + log0(f"requant_only:loading fp32 model, re-quantizing with damp={args.gptq_damp_factor}") + fp32_state = torch.load("final_model.pt", map_location="cpu") + base_model.load_state_dict(fp32_state, strict=False) + base_model.eval() + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + sd_cpu = {k: v for k, v in sd_cpu.items() if "mtp_heads" not in k} + # Re-collect Hessians and re-quantize + hessians = None + if args.gptq_enabled: + log0(f"requant:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"requant:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Pruning + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + log0(f"requant:artifact={len(quant_blob)} code={code_bytes} total={len(quant_blob)+code_bytes}") + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + if distributed: + dist.barrier() + # Load and eval + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"requant_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 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} soft_round_qat:alpha_1_to_16") + # Anneal Soft-Round alpha: 1 (soft) -> 16 (hard) over QAT window + # Uses registered buffer (tensor) so torch.compile sees dynamic value, not frozen float + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold # 0->1 as QAT runs + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + 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) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + 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() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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()} + # Pre-GPTQ magnitude pruning (like #606): zero out smallest weights → better compression + prune_pct = float(os.environ.get("PRUNE_PCT", 0)) + if prune_pct > 0: + n_pruned = 0 + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), prune_pct) + mask = v.abs() < thresh + n_pruned += mask.sum().item() + v[mask] = 0.0 + log0(f"pre_gptq_prune: zeroed bottom {prune_pct*100:.1f}% → {n_pruned} values") + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + over_bytes = len(test_blob) - target_model_bytes + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) # cap at 0.5% of params + log0(f"gptq:over by {over_bytes} bytes, total_params={total_params}, max_prune={max_prune}") + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + if len(prune_candidates) > max_prune: + log0(f"gptq:WARNING model too large! Need to prune {len(prune_candidates)} but cap={max_prune}. Reduce model size.") + log0(f"gptq:capping pruning at {max_prune} to avoid catastrophic quality loss") + prune_candidates = prune_candidates[:max_prune] + # Use fast preset (6) for binary search with safety margin, then verify with final preset + search_target = int(target_model_bytes * 0.997) # 0.3% safety margin for preset difference + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta, fast=True) + if len(blob) <= search_target: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + # Verify with final preset and increase pruning if needed + if best_n > 0: + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(prune_candidates): + best_n += max(1, best_n // 10) # 10% increments + best_n = min(best_n, len(prune_candidates)) + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}% of params)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load( + io.BytesIO(decompressed), + 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, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Full-epoch TTT (PR #672 recipe): train on ALL val data, then eval + if args.ttt_full_epochs > 0: + if distributed: + dist.barrier() + log0(f"ttt_full:start epochs={args.ttt_full_epochs} lr={args.ttt_lr} " + f"proj_mult={args.ttt_proj_lr_mult} fc_mult={args.ttt_fc_lr_mult}") + t_ttt = time.perf_counter() + ttt_full_pretrain(args, eval_model, rank, world_size, device, val_tokens, log_fn=log0) + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt_full:elapsed={ttt_elapsed:.1f}s") + # Now do sliding window eval on the TTT-adapted model + t_slide = time.perf_counter() + stride = args.eval_stride + sw_seq_len = getattr(args, 'eval_seq_len', args.train_seq_len) + ttt_val_loss, ttt_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=stride, eval_seq_len=sw_seq_len, + ) + log0(f"ttt_full_sliding:elapsed={time.perf_counter() - t_slide:.1f}s") + log0(f"final_ttt_full val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_full_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + # Legacy score-first TTT (PR#461/549 recipe) + elif args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] +Running PyTorch 2.9.1+cu128 +Wed Mar 25 07:05:53 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 70C P0 600W / 700W | 69423MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 56C P0 584W / 700W | 69423MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 57C P0 581W / 700W | 69417MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 72C P0 594W / 700W | 69417MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 72C P0 596W / 700W | 69417MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 57C P0 556W / 700W | 69417MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 69C P0 583W / 700W | 69423MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 57C P0 568W / 700W | 69423MiB / 81559MiB | 100% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: clip_range=31 (int6) compressor=lzma +model_params:27124828 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:False active_layers:[] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:560.000 +seed:2025 +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # 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", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) # 31=int6 (63 levels), 15=int5 (31 levels) + compressor = os.environ.get("COMPRESSOR", "lzma") # lzma or zstd + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs, then eval + ttt_full_epochs = int(os.environ.get("TTT_FULL_EPOCHS", 0)) + ttt_proj_lr_mult = float(os.environ.get("TTT_PROJ_LR_MULT", 3.0)) + ttt_fc_lr_mult = float(os.environ.get("TTT_FC_LR_MULT", 0.5)) + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + # Requant-only mode: skip training, load fp32 model, re-quantize + eval + requant_only = bool(int(os.environ.get("REQUANT_ONLY", "0"))) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False # bool: only 2 values, max 2 torch.compile graphs + _clip_range: int = 31 # 31=int6, 15=int5 — set once before compile, not changed during training + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # Registered buffer: torch.compile treats as dynamic input (not frozen at trace time) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range # int, set once before compile + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + # Soft-Round: fully differentiable rounding (no STE, matching #606) + # alpha buffer: 1.0 (soft, linear) -> 16.0 (hard, nearly round()) + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + # Fully differentiable: gradients flow through tanh soft-round + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + 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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + 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, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + 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, + ) + 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) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + 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 + 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, lora=None) -> 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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def ttt_full_pretrain( + args, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + log_fn=None, +) -> None: + """Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs with + cosine LR decay and differentiated per-layer LRs. All params unfrozen. + Data sharded across GPUs with gradient all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_epochs = args.ttt_full_epochs + batch_seqs = args.ttt_train_batch_seqs # 32 + + # Shard val data across GPUs + rank_tokens = total_tokens // world_size + rank_start = rank * rank_tokens + rank_end = rank_start + rank_tokens + + # Unfreeze params with differentiated LR groups + proj_params, fc_params, other_params = [], [], [] + for name, p in base_model.named_parameters(): + # Respect freeze settings + freeze = False + if args.ttt_freeze_blocks > 0: + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + freeze = any(f"blocks.{bi}." in name for bi in frozen_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + continue + p.requires_grad_(True) + if "mlp.proj" in name: + proj_params.append(p) + elif "mlp.fc" in name: + fc_params.append(p) + else: + other_params.append(p) + + base_lr = args.ttt_lr + optimizer = torch.optim.AdamW([ + {"params": proj_params, "lr": base_lr * args.ttt_proj_lr_mult}, + {"params": fc_params, "lr": base_lr * args.ttt_fc_lr_mult}, + {"params": other_params, "lr": base_lr}, + ], weight_decay=args.ttt_wd) + + total_trainable = sum(p.numel() for pg in [proj_params, fc_params, other_params] for p in pg) + n_rank_seqs = max(1, (rank_end - rank_start - seq_len) // (batch_seqs * seq_len)) + steps_per_epoch = n_rank_seqs + total_steps = ttt_epochs * steps_per_epoch + + if log_fn: + log_fn(f"ttt_full:start epochs={ttt_epochs} trainable={total_trainable} " + f"rank_tokens={rank_tokens} steps_per_ep={steps_per_epoch} " + f"total_steps={total_steps} base_lr={base_lr}") + + base_model.train() + t0 = time.perf_counter() + global_step = 0 + + for epoch in range(ttt_epochs): + for bs_idx in range(0, rank_end - rank_start - seq_len, batch_seqs * seq_len): + # Cosine LR schedule across all TTT steps + progress = global_step / max(total_steps, 1) + cos_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('initial_lr', pg['lr']) * cos_mul if global_step == 0 else pg['lr'] + # Set initial_lr on first step + if global_step == 0: + for pg in optimizer.param_groups: + pg['initial_lr'] = pg['lr'] + else: + for pg in optimizer.param_groups: + pg['lr'] = pg['initial_lr'] * cos_mul + + start_tok = rank_start + bs_idx + end_tok = min(start_tok + batch_seqs * seq_len + 1, val_tokens.numel()) + if end_tok - start_tok < seq_len + 1: + continue + + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + n_seqs = (local.numel() - 1) // seq_len + if n_seqs == 0: + continue + x = local[:n_seqs * seq_len].reshape(n_seqs, seq_len) + y = local[1:n_seqs * seq_len + 1].reshape(n_seqs, 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() + + # DDP-style gradient sync + 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) + + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.ttt_grad_clip) + optimizer.step() + global_step += 1 + + if log_fn and (epoch % 5 == 0 or epoch == ttt_epochs - 1): + elapsed = time.perf_counter() - t0 + log_fn(f" ttt_full:epoch {epoch+1}/{ttt_epochs} loss={loss.item():.4f} " + f"lr={optimizer.param_groups[2]['lr']:.6f} time={elapsed:.1f}s") + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(True) + if log_fn: + log_fn(f"ttt_full:done total_steps={global_step} time={time.perf_counter() - t0:.1f}s") + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + stride = args.eval_stride if args.eval_stride > 0 else 64 + log0(f"eval_only:TTT disabled, computing sliding window BPB stride={stride}") + 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=stride, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f} stride:{stride}") + if distributed: + dist.destroy_process_group() + return + if args.requant_only: + log0(f"requant_only:loading fp32 model, re-quantizing with damp={args.gptq_damp_factor}") + fp32_state = torch.load("final_model.pt", map_location="cpu") + base_model.load_state_dict(fp32_state, strict=False) + base_model.eval() + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + sd_cpu = {k: v for k, v in sd_cpu.items() if "mtp_heads" not in k} + # Re-collect Hessians and re-quantize + hessians = None + if args.gptq_enabled: + log0(f"requant:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"requant:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Pruning + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + log0(f"requant:artifact={len(quant_blob)} code={code_bytes} total={len(quant_blob)+code_bytes}") + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + if distributed: + dist.barrier() + # Load and eval + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"requant_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 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} soft_round_qat:alpha_1_to_16") + # Anneal Soft-Round alpha: 1 (soft) -> 16 (hard) over QAT window + # Uses registered buffer (tensor) so torch.compile sees dynamic value, not frozen float + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold # 0->1 as QAT runs + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + 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) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + 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() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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()} + # Pre-GPTQ magnitude pruning (like #606): zero out smallest weights → better compression + prune_pct = float(os.environ.get("PRUNE_PCT", 0)) + if prune_pct > 0: + n_pruned = 0 + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), prune_pct) + mask = v.abs() < thresh + n_pruned += mask.sum().item() + v[mask] = 0.0 + log0(f"pre_gptq_prune: zeroed bottom {prune_pct*100:.1f}% → {n_pruned} values") + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + over_bytes = len(test_blob) - target_model_bytes + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) # cap at 0.5% of params + log0(f"gptq:over by {over_bytes} bytes, total_params={total_params}, max_prune={max_prune}") + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + if len(prune_candidates) > max_prune: + log0(f"gptq:WARNING model too large! Need to prune {len(prune_candidates)} but cap={max_prune}. Reduce model size.") + log0(f"gptq:capping pruning at {max_prune} to avoid catastrophic quality loss") + prune_candidates = prune_candidates[:max_prune] + # Use fast preset (6) for binary search with safety margin, then verify with final preset + search_target = int(target_model_bytes * 0.997) # 0.3% safety margin for preset difference + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta, fast=True) + if len(blob) <= search_target: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + # Verify with final preset and increase pruning if needed + if best_n > 0: + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(prune_candidates): + best_n += max(1, best_n // 10) # 10% increments + best_n = min(best_n, len(prune_candidates)) + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}% of params)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load( + io.BytesIO(decompressed), + 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, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Full-epoch TTT (PR #672 recipe): train on ALL val data, then eval + if args.ttt_full_epochs > 0: + if distributed: + dist.barrier() + log0(f"ttt_full:start epochs={args.ttt_full_epochs} lr={args.ttt_lr} " + f"proj_mult={args.ttt_proj_lr_mult} fc_mult={args.ttt_fc_lr_mult}") + t_ttt = time.perf_counter() + ttt_full_pretrain(args, eval_model, rank, world_size, device, val_tokens, log_fn=log0) + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt_full:elapsed={ttt_elapsed:.1f}s") + # Now do sliding window eval on the TTT-adapted model + t_slide = time.perf_counter() + stride = args.eval_stride + sw_seq_len = getattr(args, 'eval_seq_len', args.train_seq_len) + ttt_val_loss, ttt_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=stride, eval_seq_len=sw_seq_len, + ) + log0(f"ttt_full_sliding:elapsed={time.perf_counter() - t_slide:.1f}s") + log0(f"final_ttt_full val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_full_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + # Legacy score-first TTT (PR#461/549 recipe) + elif args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] +Running PyTorch 2.9.1+cu128 +Wed Mar 25 07:07:11 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 51C P0 135W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 40C P0 124W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 37C P0 123W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 51C P0 133W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 52C P0 129W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 38C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 49C P0 126W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 38C P0 119W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: clip_range=31 (int6) compressor=lzma +model_params:27124828 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:False active_layers:[] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:560.000 +seed:2025 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9306 val_bpb:4.1047 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9326 train_time:156ms step_avg:156.10ms +step:2/20000 train_loss:8.8150 train_time:238ms step_avg:118.86ms +step:3/20000 train_loss:7.9890 train_time:323ms step_avg:107.79ms +step:4/20000 train_loss:7.1448 train_time:408ms step_avg:102.01ms +step:5/20000 train_loss:6.9585 train_time:492ms step_avg:98.46ms +step:6/20000 train_loss:6.8691 train_time:577ms step_avg:96.12ms +step:7/20000 train_loss:6.6838 train_time:661ms step_avg:94.42ms +step:8/20000 train_loss:6.6529 train_time:745ms step_avg:93.18ms +step:9/20000 train_loss:6.3724 train_time:830ms step_avg:92.25ms +step:10/20000 train_loss:6.0642 train_time:914ms step_avg:91.44ms +step:500/20000 train_loss:2.3761 train_time:43238ms step_avg:86.48ms +step:1000/20000 train_loss:2.2586 train_time:86684ms step_avg:86.68ms +step:1500/20000 train_loss:2.2129 train_time:130135ms step_avg:86.76ms +step:2000/20000 train_loss:2.0523 train_time:173589ms step_avg:86.79ms +step:2500/20000 train_loss:2.1594 train_time:216957ms step_avg:86.78ms +step:3000/20000 train_loss:2.1383 train_time:260298ms step_avg:86.77ms +step:3500/20000 train_loss:2.1496 train_time:303628ms step_avg:86.75ms +step:4000/20000 train_loss:1.9442 train_time:346926ms step_avg:86.73ms +step:4000/20000 val_loss:2.0321 val_bpb:1.2035 train_time:346931ms step_avg:86.73ms +step:4500/20000 train_loss:2.0904 train_time:390228ms step_avg:86.72ms +step:5000/20000 train_loss:2.0720 train_time:433510ms step_avg:86.70ms +step:5500/20000 train_loss:1.9815 train_time:476785ms step_avg:86.69ms +swa:start step:5700 +step:6000/20000 train_loss:1.9102 train_time:520302ms step_avg:86.72ms +step:6457/20000 val_loss:1.9262 val_bpb:1.1408 train_time:560049ms step_avg:86.74ms +stopping_early: wallclock_cap train_time:560049ms step:6457/20000 +peak memory allocated: 20671 MiB reserved: 20724 MiB +swa:applying 16 snapshots, blending with EMA (0.50/0.50) +DIAGNOSTIC post_ema val_loss:1.9252 val_bpb:1.1402 eval_time:1995ms +Serialized model: 106440713 bytes +Code size: 110105 bytes +pre_gptq_prune: zeroed bottom 0.5% → 135031 values +gptq:collecting hessians batches=256 +gptq:hessians collected layers=68 time=38.8s +gptq:pre_prune artifact=15797664 target=15839895 +Saved quantized model to final_int6_model.pt +Serialized model int63+lzma: 15797664 bytes +Total submission size: 15907769 bytes +final_int6_roundtrip val_loss:1.9317 val_bpb:1.1441 eval_time:6677ms +final_int6_roundtrip_exact val_loss:1.93172183 val_bpb:1.14407447 +final_int6_sliding_window val_loss:1.8920 val_bpb:1.1206 stride:64 eval_time:75146ms +final_int6_sliding_window_exact val_loss:1.89200709 val_bpb:1.12055614 +final_int8_zlib_roundtrip_exact val_loss:1.89200709 val_bpb:1.12055614 +ttt:start lr=0.0001 epochs=3 chunks=131072 +ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.0001 ttt_epochs=3 freeze_blocks=9 +ttt_sliding:params unfrozen=5872660 frozen=21252168 + ttt_chunk [1/474] bpb=1.196662 time=0.7s + ttt_chunk [11/474] bpb=1.117841 time=5.7s + ttt_chunk [21/474] bpb=1.113582 time=10.6s + ttt_chunk [31/474] bpb=1.112672 time=15.6s + ttt_chunk [41/474] bpb=1.120738 time=20.6s + ttt_chunk [51/474] bpb=1.128199 time=25.6s + ttt_chunk [61/474] bpb=1.126182 time=30.6s + ttt_chunk [71/474] bpb=1.127851 time=35.5s + ttt_chunk [81/474] bpb=1.128426 time=40.5s + ttt_chunk [91/474] bpb=1.130429 time=45.6s + ttt_chunk [101/474] bpb=1.127067 time=50.6s + ttt_chunk [111/474] bpb=1.127515 time=55.6s + ttt_chunk [121/474] bpb=1.130813 time=60.5s + ttt_chunk [131/474] bpb=1.131611 time=65.5s + ttt_chunk [141/474] bpb=1.131494 time=70.5s + ttt_chunk [151/474] bpb=1.129833 time=75.5s + ttt_chunk [161/474] bpb=1.130794 time=80.4s + ttt_chunk [171/474] bpb=1.129527 time=85.4s + ttt_chunk [181/474] bpb=1.130340 time=90.4s + ttt_chunk [191/474] bpb=1.129347 time=95.4s + ttt_chunk [201/474] bpb=1.128565 time=100.4s + ttt_chunk [211/474] bpb=1.127311 time=105.3s + ttt_chunk [221/474] bpb=1.127458 time=110.3s + ttt_chunk [231/474] bpb=1.126915 time=115.3s + ttt_chunk [241/474] bpb=1.125863 time=120.3s + ttt_chunk [251/474] bpb=1.126959 time=125.3s + ttt_chunk [261/474] bpb=1.127640 time=130.3s + ttt_chunk [271/474] bpb=1.126202 time=135.3s + ttt_chunk [281/474] bpb=1.125863 time=140.3s + ttt_chunk [291/474] bpb=1.124424 time=145.3s + ttt_chunk [301/474] bpb=1.124851 time=150.3s + ttt_chunk [311/474] bpb=1.124278 time=155.3s + ttt_chunk [321/474] bpb=1.122619 time=160.2s + ttt_chunk [331/474] bpb=1.121591 time=165.2s + ttt_chunk [341/474] bpb=1.120850 time=170.2s + ttt_chunk [351/474] bpb=1.119245 time=175.2s + ttt_chunk [361/474] bpb=1.119786 time=180.1s + ttt_chunk [371/474] bpb=1.119489 time=185.1s + ttt_chunk [381/474] bpb=1.120253 time=190.1s + ttt_chunk [391/474] bpb=1.121309 time=195.0s + ttt_chunk [401/474] bpb=1.121765 time=200.0s + ttt_chunk [411/474] bpb=1.122188 time=205.0s + ttt_chunk [421/474] bpb=1.123644 time=209.9s + ttt_chunk [431/474] bpb=1.122234 time=214.9s + ttt_chunk [441/474] bpb=1.121870 time=219.9s + ttt_chunk [451/474] bpb=1.121236 time=224.9s + ttt_chunk [461/474] bpb=1.121403 time=229.9s + ttt_chunk [471/474] bpb=1.121484 time=234.8s + ttt_chunk [474/474] bpb=1.121351 time=235.9s +ttt_sliding:done val_loss=1.891905 val_bpb=1.120496 elapsed=235.9s +ttt:elapsed=236.3s +final_ttt val_loss:1.8919 val_bpb:1.1205 +final_ttt_exact val_loss:1.89190474 val_bpb:1.12049552 diff --git a/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed42.log b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed42.log new file mode 100644 index 000000000..f8ca98ab3 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_FullGPTQ_XSA4_ScoreFirstTTT/train_seed42.log @@ -0,0 +1,2400 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +from flash_attn_interface import flash_attn_func as flash_attn_3_func +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # 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", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + clip_range = int(os.environ.get("CLIP_RANGE", 31)) # 31=int6 (63 levels), 15=int5 (31 levels) + compressor = os.environ.get("COMPRESSOR", "lzma") # lzma or zstd + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs, then eval + ttt_full_epochs = int(os.environ.get("TTT_FULL_EPOCHS", 0)) + ttt_proj_lr_mult = float(os.environ.get("TTT_PROJ_LR_MULT", 3.0)) + ttt_fc_lr_mult = float(os.environ.get("TTT_FC_LR_MULT", 0.5)) + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + # Requant-only mode: skip training, load fp32 model, re-quantize + eval + requant_only = bool(int(os.environ.get("REQUANT_ONLY", "0"))) + gptq_damp_factor = float(os.environ.get("GPTQ_DAMP_FACTOR", "0.01")) + swa_ema_blend = float(os.environ.get("SWA_EMA_BLEND", "0.5")) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False # bool: only 2 values, max 2 torch.compile graphs + _clip_range: int = 31 # 31=int6, 15=int5 — set once before compile, not changed during training + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # Registered buffer: torch.compile treats as dynamic input (not frozen at trace time) + self.register_buffer('_soft_round_alpha', torch.tensor(1.0), persistent=False) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + cr = CastedLinear._clip_range # int, set once before compile + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / float(cr)).clamp_min(1.0 / float(cr)) + x_norm = w32 / scale[:, None] + # Soft-Round: fully differentiable rounding (no STE, matching #606) + # alpha buffer: 1.0 (soft, linear) -> 16.0 (hard, nearly round()) + alpha = self._soft_round_alpha + fl = x_norm.floor() + r = x_norm - fl - 0.5 + tanh_half = torch.tanh(alpha * 0.5) + q_soft = fl + 0.5 * torch.tanh(alpha * r) / (tanh_half + 1e-10) + 0.5 + q_soft = torch.clamp(q_soft, -cr, cr) + # Fully differentiable: gradients flow through tanh soft-round + w_q = (q_soft * scale[:, None]).to(x.dtype) + w = w_q + 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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + 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, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.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 + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + 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, + ) + 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) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + 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 + 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, lora=None) -> 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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """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 = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def ttt_full_pretrain( + args, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + log_fn=None, +) -> None: + """Full-epoch TTT (PR #672 recipe): train on ALL val data for N epochs with + cosine LR decay and differentiated per-layer LRs. All params unfrozen. + Data sharded across GPUs with gradient all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_epochs = args.ttt_full_epochs + batch_seqs = args.ttt_train_batch_seqs # 32 + + # Shard val data across GPUs + rank_tokens = total_tokens // world_size + rank_start = rank * rank_tokens + rank_end = rank_start + rank_tokens + + # Unfreeze params with differentiated LR groups + proj_params, fc_params, other_params = [], [], [] + for name, p in base_model.named_parameters(): + # Respect freeze settings + freeze = False + if args.ttt_freeze_blocks > 0: + frozen_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + freeze = any(f"blocks.{bi}." in name for bi in frozen_ids) + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + continue + p.requires_grad_(True) + if "mlp.proj" in name: + proj_params.append(p) + elif "mlp.fc" in name: + fc_params.append(p) + else: + other_params.append(p) + + base_lr = args.ttt_lr + optimizer = torch.optim.AdamW([ + {"params": proj_params, "lr": base_lr * args.ttt_proj_lr_mult}, + {"params": fc_params, "lr": base_lr * args.ttt_fc_lr_mult}, + {"params": other_params, "lr": base_lr}, + ], weight_decay=args.ttt_wd) + + total_trainable = sum(p.numel() for pg in [proj_params, fc_params, other_params] for p in pg) + n_rank_seqs = max(1, (rank_end - rank_start - seq_len) // (batch_seqs * seq_len)) + steps_per_epoch = n_rank_seqs + total_steps = ttt_epochs * steps_per_epoch + + if log_fn: + log_fn(f"ttt_full:start epochs={ttt_epochs} trainable={total_trainable} " + f"rank_tokens={rank_tokens} steps_per_ep={steps_per_epoch} " + f"total_steps={total_steps} base_lr={base_lr}") + + base_model.train() + t0 = time.perf_counter() + global_step = 0 + + for epoch in range(ttt_epochs): + for bs_idx in range(0, rank_end - rank_start - seq_len, batch_seqs * seq_len): + # Cosine LR schedule across all TTT steps + progress = global_step / max(total_steps, 1) + cos_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('initial_lr', pg['lr']) * cos_mul if global_step == 0 else pg['lr'] + # Set initial_lr on first step + if global_step == 0: + for pg in optimizer.param_groups: + pg['initial_lr'] = pg['lr'] + else: + for pg in optimizer.param_groups: + pg['lr'] = pg['initial_lr'] * cos_mul + + start_tok = rank_start + bs_idx + end_tok = min(start_tok + batch_seqs * seq_len + 1, val_tokens.numel()) + if end_tok - start_tok < seq_len + 1: + continue + + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + n_seqs = (local.numel() - 1) // seq_len + if n_seqs == 0: + continue + x = local[:n_seqs * seq_len].reshape(n_seqs, seq_len) + y = local[1:n_seqs * seq_len + 1].reshape(n_seqs, 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() + + # DDP-style gradient sync + 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) + + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.ttt_grad_clip) + optimizer.step() + global_step += 1 + + if log_fn and (epoch % 5 == 0 or epoch == ttt_epochs - 1): + elapsed = time.perf_counter() - t0 + log_fn(f" ttt_full:epoch {epoch+1}/{ttt_epochs} loss={loss.item():.4f} " + f"lr={optimizer.param_groups[2]['lr']:.6f} time={elapsed:.1f}s") + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(True) + if log_fn: + log_fn(f"ttt_full:done total_steps={global_step} time={time.perf_counter() - t0:.1f}s") + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = args.gptq_damp_factor * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, + damp_factor: float = 0.01, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = damp_factor * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128, gptq_damp_factor: float = 0.01, + clip_range: int = 31): + 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: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, clip_range=clip_range, block_size=gptq_block_size, damp_factor=gptq_damp_factor) + else: + q, s = quantize_int6_per_row(t, clip_range=clip_range) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 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 + CastedLinear._clip_range = args.clip_range + log0(f"mixed_precision: clip_range={args.clip_range} ({'int5' if args.clip_range == 15 else 'int6'}) compressor={args.compressor}") + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + stride = args.eval_stride if args.eval_stride > 0 else 64 + log0(f"eval_only:TTT disabled, computing sliding window BPB stride={stride}") + 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=stride, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f} stride:{stride}") + if distributed: + dist.destroy_process_group() + return + if args.requant_only: + log0(f"requant_only:loading fp32 model, re-quantizing with damp={args.gptq_damp_factor}") + fp32_state = torch.load("final_model.pt", map_location="cpu") + base_model.load_state_dict(fp32_state, strict=False) + base_model.eval() + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + sd_cpu = {k: v for k, v in sd_cpu.items() if "mtp_heads" not in k} + # Re-collect Hessians and re-quantize + hessians = None + if args.gptq_enabled: + log0(f"requant:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"requant:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Pruning + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + log0(f"requant:artifact={len(quant_blob)} code={code_bytes} total={len(quant_blob)+code_bytes}") + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + if distributed: + dist.barrier() + # Load and eval + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"requant_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 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} soft_round_qat:alpha_1_to_16") + # Anneal Soft-Round alpha: 1 (soft) -> 16 (hard) over QAT window + # Uses registered buffer (tensor) so torch.compile sees dynamic value, not frozen float + if CastedLinear._qat_enabled and args.late_qat_threshold > 0: + qat_progress = 1.0 - scale / args.late_qat_threshold # 0->1 as QAT runs + qat_progress = max(0.0, min(1.0, qat_progress)) + new_alpha = 1.0 + 15.0 * qat_progress + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._soft_round_alpha.fill_(new_alpha) + 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) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + 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() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + blend = args.swa_ema_blend + log0(f"swa:applying {swa_count} snapshots, blending with EMA ({blend:.2f}/{1-blend:.2f})") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = blend * ema_w + (1 - blend) * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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()} + # Pre-GPTQ magnitude pruning (like #606): zero out smallest weights → better compression + prune_pct = float(os.environ.get("PRUNE_PCT", 0)) + if prune_pct > 0: + n_pruned = 0 + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), prune_pct) + mask = v.abs() < thresh + n_pruned += mask.sum().item() + v[mask] = 0.0 + log0(f"pre_gptq_prune: zeroed bottom {prune_pct*100:.1f}% → {n_pruned} values") + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + gptq_damp_factor=args.gptq_damp_factor, clip_range=args.clip_range, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm, fast=False): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + raw = buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + level = 10 if fast else 22 + return zstd.ZstdCompressor(level=level).compress(raw) + preset = 6 if fast else (9 | lzma.PRESET_EXTREME) + return lzma.compress(raw, preset=preset) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + over_bytes = len(test_blob) - target_model_bytes + total_params = sum(v.numel() for v in quant_result.values() if v.dtype == torch.int8) + max_prune = max(1000, total_params // 200) # cap at 0.5% of params + log0(f"gptq:over by {over_bytes} bytes, total_params={total_params}, max_prune={max_prune}") + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + if len(prune_candidates) > max_prune: + log0(f"gptq:WARNING model too large! Need to prune {len(prune_candidates)} but cap={max_prune}. Reduce model size.") + log0(f"gptq:capping pruning at {max_prune} to avoid catastrophic quality loss") + prune_candidates = prune_candidates[:max_prune] + # Use fast preset (6) for binary search with safety margin, then verify with final preset + search_target = int(target_model_bytes * 0.997) # 0.3% safety margin for preset difference + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta, fast=True) + if len(blob) <= search_target: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + # Verify with final preset and increase pruning if needed + if best_n > 0: + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + while len(final_blob) > target_model_bytes and best_n < len(prune_candidates): + best_n += max(1, best_n // 10) # 10% increments + best_n = min(best_n, len(prune_candidates)) + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(best_n): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + final_blob = _serialize_and_compress(qr_test, quant_meta) + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values ({100*best_n/total_params:.2f}% of params)") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + if args.compressor == "zstd": + import zstandard as zstd + quant_blob = zstd.ZstdCompressor(level=22).compress(quant_raw) + comp_label = "zstd" + else: + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) + comp_label = "lzma" + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int{args.clip_range*2+1}+{comp_label}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + if args.compressor == "zstd": + import zstandard as zstd + decompressed = zstd.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = lzma.decompress(quant_blob_disk) + quant_state = torch.load( + io.BytesIO(decompressed), + 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, + use_vrl=args.vrl, + ).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) + CastedLinear._qat_enabled = False + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Full-epoch TTT (PR #672 recipe): train on ALL val data, then eval + if args.ttt_full_epochs > 0: + if distributed: + dist.barrier() + log0(f"ttt_full:start epochs={args.ttt_full_epochs} lr={args.ttt_lr} " + f"proj_mult={args.ttt_proj_lr_mult} fc_mult={args.ttt_fc_lr_mult}") + t_ttt = time.perf_counter() + ttt_full_pretrain(args, eval_model, rank, world_size, device, val_tokens, log_fn=log0) + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt_full:elapsed={ttt_elapsed:.1f}s") + # Now do sliding window eval on the TTT-adapted model + t_slide = time.perf_counter() + stride = args.eval_stride + sw_seq_len = getattr(args, 'eval_seq_len', args.train_seq_len) + ttt_val_loss, ttt_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=stride, eval_seq_len=sw_seq_len, + ) + log0(f"ttt_full_sliding:elapsed={time.perf_counter() - t_slide:.1f}s") + log0(f"final_ttt_full val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_full_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + # Legacy score-first TTT (PR#461/549 recipe) + elif args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] +Running PyTorch 2.9.1+cu128 +Wed Mar 25 07:49:48 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 39C P0 125W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 34C P0 120W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 32C P0 120W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 40C P0 123W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 40C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 33C P0 113W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 38C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 33C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: clip_range=31 (int6) compressor=lzma +model_params:27124828 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:False active_layers:[] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:560.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9323 train_time:142ms step_avg:142.38ms +step:2/20000 train_loss:8.6763 train_time:223ms step_avg:111.61ms +step:3/20000 train_loss:7.9016 train_time:308ms step_avg:102.62ms +step:4/20000 train_loss:7.1817 train_time:392ms step_avg:98.08ms +step:5/20000 train_loss:6.9300 train_time:476ms step_avg:95.22ms +step:6/20000 train_loss:6.8825 train_time:560ms step_avg:93.38ms +step:7/20000 train_loss:6.8050 train_time:645ms step_avg:92.12ms +step:8/20000 train_loss:6.7247 train_time:729ms step_avg:91.14ms +step:9/20000 train_loss:6.4340 train_time:813ms step_avg:90.34ms +step:10/20000 train_loss:6.1369 train_time:898ms step_avg:89.78ms +step:500/20000 train_loss:2.3732 train_time:43155ms step_avg:86.31ms +step:1000/20000 train_loss:2.2573 train_time:86586ms step_avg:86.59ms +step:1500/20000 train_loss:2.2063 train_time:130047ms step_avg:86.70ms +step:2000/20000 train_loss:2.0460 train_time:173444ms step_avg:86.72ms +step:2500/20000 train_loss:2.1557 train_time:216820ms step_avg:86.73ms +step:3000/20000 train_loss:2.1346 train_time:260168ms step_avg:86.72ms +step:3500/20000 train_loss:2.1458 train_time:303509ms step_avg:86.72ms +step:4000/20000 train_loss:1.9396 train_time:346826ms step_avg:86.71ms +step:4000/20000 val_loss:2.0303 val_bpb:1.2025 train_time:346830ms step_avg:86.71ms +step:4500/20000 train_loss:2.0885 train_time:390134ms step_avg:86.70ms +step:5000/20000 train_loss:2.0688 train_time:433446ms step_avg:86.69ms +step:5500/20000 train_loss:1.9831 train_time:476733ms step_avg:86.68ms +swa:start step:5700 +step:6000/20000 train_loss:1.9068 train_time:520234ms step_avg:86.71ms +step:6457/20000 val_loss:1.9249 val_bpb:1.1400 train_time:560036ms step_avg:86.73ms +stopping_early: wallclock_cap train_time:560036ms step:6457/20000 +peak memory allocated: 20671 MiB reserved: 20724 MiB +swa:applying 16 snapshots, blending with EMA (0.50/0.50) +DIAGNOSTIC post_ema val_loss:1.9239 val_bpb:1.1394 eval_time:2032ms +Serialized model: 106440713 bytes +Code size: 110105 bytes +pre_gptq_prune: zeroed bottom 0.5% → 135027 values +gptq:collecting hessians batches=256 +gptq:hessians collected layers=68 time=39.0s +gptq:pre_prune artifact=15844836 target=15839895 +gptq:over by 4941 bytes, total_params=27000832, max_prune=135004 +gptq:pruning candidates=4248001 +gptq:WARNING model too large! Need to prune 4248001 but cap=135004. Reduce model size. +gptq:capping pruning at 135004 to avoid catastrophic quality loss +Saved quantized model to final_int6_model.pt +Serialized model int63+lzma: 15844836 bytes +Total submission size: 15954941 bytes +final_int6_roundtrip val_loss:1.9303 val_bpb:1.1432 eval_time:7164ms +final_int6_roundtrip_exact val_loss:1.93027411 val_bpb:1.14321706 +final_int6_sliding_window val_loss:1.8905 val_bpb:1.1197 stride:64 eval_time:75199ms +final_int6_sliding_window_exact val_loss:1.89050610 val_bpb:1.11966716 +final_int8_zlib_roundtrip_exact val_loss:1.89050610 val_bpb:1.11966716 +ttt:start lr=0.0001 epochs=3 chunks=131072 +ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.0001 ttt_epochs=3 freeze_blocks=9 +ttt_sliding:params unfrozen=5872660 frozen=21252168 + ttt_chunk [1/474] bpb=1.196246 time=0.7s + ttt_chunk [11/474] bpb=1.116298 time=5.7s + ttt_chunk [21/474] bpb=1.112208 time=10.7s + ttt_chunk [31/474] bpb=1.110903 time=15.6s + ttt_chunk [41/474] bpb=1.119091 time=20.6s + ttt_chunk [51/474] bpb=1.126708 time=25.6s + ttt_chunk [61/474] bpb=1.124796 time=30.6s + ttt_chunk [71/474] bpb=1.126421 time=35.6s + ttt_chunk [81/474] bpb=1.127054 time=40.5s + ttt_chunk [91/474] bpb=1.129102 time=45.5s + ttt_chunk [101/474] bpb=1.125812 time=50.5s + ttt_chunk [111/474] bpb=1.126195 time=55.4s + ttt_chunk [121/474] bpb=1.129650 time=60.4s + ttt_chunk [131/474] bpb=1.130550 time=65.4s + ttt_chunk [141/474] bpb=1.130510 time=70.4s + ttt_chunk [151/474] bpb=1.128860 time=75.3s + ttt_chunk [161/474] bpb=1.129833 time=80.3s + ttt_chunk [171/474] bpb=1.128625 time=85.3s + ttt_chunk [181/474] bpb=1.129412 time=90.2s + ttt_chunk [191/474] bpb=1.128368 time=95.2s + ttt_chunk [201/474] bpb=1.127543 time=100.2s + ttt_chunk [211/474] bpb=1.126266 time=105.2s + ttt_chunk [221/474] bpb=1.126399 time=110.1s + ttt_chunk [231/474] bpb=1.125855 time=115.1s + ttt_chunk [241/474] bpb=1.124810 time=120.1s + ttt_chunk [251/474] bpb=1.125920 time=125.0s + ttt_chunk [261/474] bpb=1.126605 time=130.0s + ttt_chunk [271/474] bpb=1.125170 time=135.0s + ttt_chunk [281/474] bpb=1.124867 time=140.0s + ttt_chunk [291/474] bpb=1.123477 time=145.0s + ttt_chunk [301/474] bpb=1.123928 time=149.9s + ttt_chunk [311/474] bpb=1.123333 time=154.9s + ttt_chunk [321/474] bpb=1.121727 time=159.9s + ttt_chunk [331/474] bpb=1.120689 time=164.9s + ttt_chunk [341/474] bpb=1.119965 time=169.8s + ttt_chunk [351/474] bpb=1.118366 time=174.8s + ttt_chunk [361/474] bpb=1.118845 time=179.8s + ttt_chunk [371/474] bpb=1.118522 time=184.7s + ttt_chunk [381/474] bpb=1.119275 time=189.7s + ttt_chunk [391/474] bpb=1.120397 time=194.7s + ttt_chunk [401/474] bpb=1.120843 time=199.7s + ttt_chunk [411/474] bpb=1.121274 time=204.6s + ttt_chunk [421/474] bpb=1.122724 time=209.6s + ttt_chunk [431/474] bpb=1.121300 time=214.6s + ttt_chunk [441/474] bpb=1.120932 time=219.6s + ttt_chunk [451/474] bpb=1.120300 time=224.6s + ttt_chunk [461/474] bpb=1.120515 time=229.5s + ttt_chunk [471/474] bpb=1.120621 time=234.5s + ttt_chunk [474/474] bpb=1.120494 time=235.6s +ttt_sliding:done val_loss=1.890445 val_bpb=1.119631 elapsed=235.6s +ttt:elapsed=236.0s +final_ttt val_loss:1.8904 val_bpb:1.1196 +final_ttt_exact val_loss:1.89044540 val_bpb:1.11963121