diff --git a/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/README.md b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/README.md new file mode 100644 index 000000000..625a363a2 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/README.md @@ -0,0 +1,72 @@ +# Submission: Depth Recurrence + LoRA TTT + BigramHash(3072) + +Built on PR #549 (LeakyReLU² + Legal TTT + Parallel Muon, 1.1194 BPB) with: + +## Key Improvements + +1. **Depth Recurrence** (layers 4,5): Repeat layers 4 and 5 → 13 virtual layers from 11 physical. Zero parameter cost, just extra compute per step. + +2. **LoRA TTT** (rank 8): Replace full-param SGD TTT with LoRA adapters on Q/V projections. Uses Adam optimizer instead of SGD. ~24x more effective in score-first framework per community benchmarks. + +3. **BigramHash(2048)**: Keep proven bucket count to stay within 16MB artifact budget. + +4. **SDPA Fallback**: Automatically falls back to PyTorch SDPA when FlashAttention 3 is not available. + +## Architecture + +- 11 physical layers (13 virtual via depth recurrence on layers 4,5) +- 512d, 8H/4KV GQA, 3x MLP with LeakyReLU(0.5)² +- XSA on last 4 layers, Partial RoPE (16/64 dims), LN Scale 1/√(layer+1) +- BigramHash(2048, dim=128) + SmearGate + ValueEmbedding(layers 9,10) +- Parameter Banking + Parallel Muon optimizer +- EMA(0.997) + Tight SWA weight averaging +- Int6 GPTQ-lite quantization + lzma compression +- Legal score-first LoRA TTT (rank 8, Adam, 3 epochs per 32K chunk) + +## Quick Test (1xH100) + +```bash +cd /workspace/parameter-golf + +# Copy submission script +cp submission/train_gpt.py train_gpt_sub.py + +# Download data if not already done +python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 10 + +# Quick smoke test (1000 steps, ~10 min on 1xH100) +RUN_ID=test_v1 \ +NUM_LAYERS=11 BIGRAM_VOCAB_SIZE=2048 XSA_LAST_N=4 \ +DEPTH_RECURRENCE=4,5 \ +EMA_ENABLED=1 SWA_ENABLED=1 SWA_EVERY=50 \ +ROPE_DIMS=16 LN_SCALE=1 \ +VE_ENABLED=1 VE_DIM=128 VE_LAYERS=9,10 \ +TTT_ENABLED=0 \ +MUON_WD=0.04 ADAM_WD=0.04 \ +MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 \ +MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 \ +MUON_MOMENTUM_WARMUP_STEPS=1500 WARMDOWN_ITERS=3500 \ +ITERATIONS=20000 MAX_WALLCLOCK_SECONDS=600 EVAL_STRIDE=64 \ +SEED=1337 \ +torchrun --standalone --nproc_per_node=1 train_gpt_sub.py +``` + +## Full Run (8xH100) + +```bash +RUN_ID=submission_v1 \ +NUM_LAYERS=11 BIGRAM_VOCAB_SIZE=2048 XSA_LAST_N=4 \ +DEPTH_RECURRENCE=4,5 \ +EMA_ENABLED=1 SWA_ENABLED=1 SWA_EVERY=50 \ +ROPE_DIMS=16 LN_SCALE=1 LATE_QAT=1 LATE_QAT_THRESHOLD=0.15 \ +VE_ENABLED=1 VE_DIM=128 VE_LAYERS=9,10 \ +TTT_ENABLED=1 TTT_USE_LORA=1 TTT_LORA_RANK=8 TTT_LORA_LR=0.01 \ +TTT_EPOCHS=3 TTT_CHUNK_TOKENS=32768 TTT_FREEZE_BLOCKS=0 TTT_GRAD_CLIP=1.0 \ +MUON_WD=0.04 ADAM_WD=0.04 \ +MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 \ +MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 \ +MUON_MOMENTUM_WARMUP_STEPS=1500 WARMDOWN_ITERS=3500 \ +ITERATIONS=9000 MAX_WALLCLOCK_SECONDS=600 EVAL_STRIDE=64 \ +SEED=1337 \ +torchrun --standalone --nproc_per_node=8 train_gpt_sub.py +``` diff --git a/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/submission.json b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/submission.json new file mode 100644 index 000000000..75a17c301 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/submission.json @@ -0,0 +1,12 @@ +{ + "author": "naazimsnh02", + "github_id": "naazimsnh02", + "val_bpb": null, + "date": "2026-03-25", + "summary": "Depth Recurrence (layers 4,5) + LoRA TTT (rank 8) + Parallel Muon on PR #549 stack", + "architecture": "11L 512d 8H/4KV GQA, 3x MLP LeakyReLU(0.5)², depth recurrence layers 4+5, XSA4, PartialRoPE 16/64, LN Scale, BigramHash(2048), SmearGate, VE128, EMA+SWA, Int6 GPTQ-lite + lzma", + "training_time_seconds": null, + "artifact_bytes": null, + "num_seeds": 1, + "gpu": "8xH100 SXM" +} diff --git a/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/train_gpt.py b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/train_gpt.py new file mode 100644 index 000000000..67963a279 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/train_gpt.py @@ -0,0 +1,2003 @@ +from __future__ import annotations +import copy +import glob +import io +import lzma +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _FA3_AVAILABLE = True +except ImportError: + _FA3_AVAILABLE = False +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)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + 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") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 2)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + # LoRA TTT (replaces full-param SGD; 24x more effective in score-first framework) + ttt_lora_rank = int(os.environ.get("TTT_LORA_RANK", 8)) + ttt_use_lora = bool(int(os.environ.get("TTT_USE_LORA", "1"))) + ttt_lora_lr = float(os.environ.get("TTT_LORA_LR", 0.01)) + # Depth recurrence: comma-separated layer indices to repeat (e.g. "4,5") + depth_recurrence = os.environ.get("DEPTH_RECURRENCE", "4,5") + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +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) + +# --- Quantization helpers --- + +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,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + +# --- Data loading --- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (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, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _FA3_AVAILABLE: + y = flash_attn_3_func(q, k, v, causal=True) + else: + # SDPA fallback: [B, T, H, D] -> [B, H, T, D] + q_t = q.transpose(1, 2) + # GQA: repeat KV heads + if self.num_kv_heads != self.num_heads: + group = self.num_heads // self.num_kv_heads + k = k[:, :, :, None, :].expand(-1, -1, -1, group, -1).reshape(bsz, seqlen, self.num_heads, self.head_dim) + v = v[:, :, :, None, :].expand(-1, -1, -1, group, -1).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k_t = k.transpose(1, 2) + v_t = v.transpose(1, 2) + y = F.scaled_dot_product_attention(q_t, k_t, v_t, is_causal=True) + y = y.transpose(1, 2) # back to [B, T, H, D] + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), 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__() + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + 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, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + 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", + gated_attention: bool = False, + value_residual: bool = False, + depth_recurrence: str = "", + ): + super().__init__() + # Depth recurrence: which physical layers to repeat + self.recurrence_layers = [int(x) for x in depth_recurrence.split(",") if x.strip()] if depth_recurrence else [] + 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.value_residual = value_residual + 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)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.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) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + def _run_layer(i, x, x0, v0, ve_cache): + ve = self._get_ve(i, input_ids, ve_cache) + x_out, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + return x_out, raw_v + + for i in range(self.num_encoder_layers): + x, raw_v = _run_layer(i, x, x0, v0, ve_cache) + if v0 is None and raw_v is not None: + v0 = raw_v + # Depth recurrence: repeat this layer + if i in self.recurrence_layers: + x, _ = _run_layer(i, x, x0, v0, ve_cache) + 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() + x, _ = _run_layer(bi, x, x0, v0, ve_cache) + # Depth recurrence: repeat this layer + if bi in self.recurrence_layers: + x, _ = _run_layer(bi, x, x0, v0, ve_cache) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + def _run_layer(i, x, x0, v0, ve_cache): + ve = self._get_ve(i, input_ids, ve_cache) + x_out, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + return x_out, raw_v + for i in range(self.num_encoder_layers): + x, raw_v = _run_layer(i, x, x0, v0, ve_cache) + if v0 is None and raw_v is not None: + v0 = raw_v + if i in self.recurrence_layers: + x, _ = _run_layer(i, x, x0, v0, ve_cache) + 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() + x, _ = _run_layer(bi, x, x0, v0, ve_cache) + if bi in self.recurrence_layers: + x, _ = _run_layer(bi, x, x0, v0, ve_cache) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +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 + + +class LoRAAdapter: + """Lightweight LoRA adapters for TTT on Q/V bank projections.""" + def __init__(self, model: nn.Module, rank: int, device: torch.device): + self.adapters: list[tuple[nn.Parameter, nn.Parameter, int, str]] = [] # (A, B, layer_idx, bank_type) + n = model.num_layers + for i in range(n): + # LoRA on Q projection (qo_bank[i]) + out_dim, in_dim = model.qo_bank[i].shape + A = nn.Parameter(torch.randn(out_dim, rank, device=device, dtype=torch.float32) * 0.01) + B = nn.Parameter(torch.zeros(rank, in_dim, device=device, dtype=torch.float32)) + self.adapters.append((A, B, i, 'q')) + # LoRA on V projection (kv_bank[n + i]) + out_dim_v, in_dim_v = model.kv_bank[n + i].shape + A_v = nn.Parameter(torch.randn(out_dim_v, rank, device=device, dtype=torch.float32) * 0.01) + B_v = nn.Parameter(torch.zeros(rank, in_dim_v, device=device, dtype=torch.float32)) + self.adapters.append((A_v, B_v, i, 'v')) + + def parameters(self) -> list[nn.Parameter]: + params = [] + for A, B, _, _ in self.adapters: + params.extend([A, B]) + return params + + def apply_to_model(self, model: nn.Module): + """Add LoRA deltas to the model banks (modifies in-place).""" + n = model.num_layers + for A, B, i, bank_type in self.adapters: + delta = (A @ B).to(dtype=model.qo_bank.dtype) + if bank_type == 'q': + model.qo_bank.data[i] += delta + elif bank_type == 'v': + model.kv_bank.data[n + i] += delta + + def remove_from_model(self, model: nn.Module): + """Remove LoRA deltas from the model banks.""" + n = model.num_layers + for A, B, i, bank_type in self.adapters: + delta = (A @ B).to(dtype=model.qo_bank.dtype) + if bank_type == 'q': + model.qo_bank.data[i] -= delta + elif bank_type == 'v': + model.kv_bank.data[n + i] -= delta + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT with LoRA or full-param SGD.""" + 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) + + use_lora = args.ttt_use_lora + ttt_lr = args.ttt_lora_lr if use_lora else args.ttt_lr + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={ttt_lr} ttt_epochs={args.ttt_epochs} " + f"use_lora={use_lora} lora_rank={args.ttt_lora_rank} " + 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) + + # Setup LoRA or full-param TTT + if use_lora: + lora = LoRAAdapter(base_model, args.ttt_lora_rank, device) + ttt_params = lora.parameters() + # Freeze all base model params + for p in base_model.parameters(): + p.requires_grad_(False) + lora.apply_to_model(base_model) + optimizer = torch.optim.Adam(ttt_params, lr=ttt_lr) + log0(f"ttt_lora:params={sum(p.numel() for p in ttt_params)}") + else: + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + optimizer = torch.optim.SGD(ttt_params, lr=ttt_lr, momentum=args.ttt_momentum) + log0(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)}") + + 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 (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 scored chunk (legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + if use_lora: + # Remove LoRA before training step (apply fresh each time) + lora.remove_from_model(base_model) + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + if use_lora: + lora.apply_to_model(base_model) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if use_lora: + lora.remove_from_model(base_model) + 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 use_lora: + lora.apply_to_model(base_model) + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + if use_lora: + lora.remove_from_model(base_model) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +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 _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "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) + try: + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + except FileNotFoundError: + log0("nvidia-smi not found", console=False) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + depth_recurrence=args.depth_recurrence, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("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.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + 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 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + 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" + ) + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, 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, + gated_attention=args.gated_attention, value_residual=args.value_residual, + depth_recurrence=args.depth_recurrence, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, 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}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/upload_and_run.sh b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/upload_and_run.sh new file mode 100644 index 000000000..537361d38 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_DepthRecurrence_LoRATTT_ParallelMuon/upload_and_run.sh @@ -0,0 +1,44 @@ +#!/bin/bash +# Upload this script and train_gpt.py to your RunPod, then run from /workspace/parameter-golf/ + +set -e + +# ============================================================ +# STEP 1: Copy the submission script into position +# ============================================================ +cp /workspace/parameter-golf/submission/train_gpt.py /workspace/parameter-golf/train_gpt_sub.py + +cd /workspace/parameter-golf + +# ============================================================ +# STEP 2: Make sure data is downloaded +# ============================================================ +if [ ! -d "data/datasets/fineweb10B_sp1024" ]; then + echo "Downloading dataset..." + python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 10 +fi + +# ============================================================ +# STEP 3: Quick 1xH100 smoke test (disable TTT, ~10 min) +# ============================================================ +echo "" +echo "==========================================" +echo "RUNNING: 1xH100 smoke test (no TTT)" +echo "==========================================" + +RUN_ID=smoke_depth_recurrence_v1 \ +NUM_LAYERS=11 BIGRAM_VOCAB_SIZE=2048 XSA_LAST_N=4 \ +DEPTH_RECURRENCE=4,5 \ +EMA_ENABLED=1 SWA_ENABLED=1 SWA_EVERY=50 \ +ROPE_DIMS=16 LN_SCALE=1 \ +VE_ENABLED=1 VE_DIM=128 VE_LAYERS=9,10 \ +TTT_ENABLED=0 \ +MUON_WD=0.04 ADAM_WD=0.04 \ +MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 \ +MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 \ +MUON_MOMENTUM_WARMUP_STEPS=1500 WARMDOWN_ITERS=3500 \ +ITERATIONS=20000 MAX_WALLCLOCK_SECONDS=600 EVAL_STRIDE=64 \ +GRAD_CLIP_NORM=0.3 \ +TRAIN_LOG_EVERY=100 VAL_LOSS_EVERY=0 \ +SEED=1337 \ +torchrun --standalone --nproc_per_node=1 train_gpt_sub.py diff --git a/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/README.md b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/README.md new file mode 100644 index 000000000..d69bd902d --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/README.md @@ -0,0 +1,78 @@ +# Record: Complementary Training + Backoff N-gram Mixer + Legal TTT + +## Summary + +- **BPB: 0.4311** (3-seed average: seeds 42, 1337, 2024) +- 11L transformer (26.99M params) with VRL, LeakyReLU(0.5)², XSA-4 +- **Complementary training**: model trained with bigram-weighted loss (COMPLEMENT_ALPHA=0.5) to specialize on tokens n-gram caches can't predict +- **BackoffNgramMixer**: orders 2-10, 4M flat hash buckets, greedy cascade (highest order wins) +- **Entropy-adaptive alpha** (0.20 + 0.55*sigmoid(2*(H-3.0))): n-gram gets 20-75% weight based on model uncertainty +- AdamW TTT (lr=5e-4, **3 epochs**, Polyak EMA 0.998, freeze first 9/11 blocks) +- Int6 mixed quantization + lzma compression +- Artifact: max 15,962,841B across seeds (under 16,000,000 byte limit) + +## Architecture + +- 11 layers, 512 model dim, 8 attention heads, 4 KV heads (GQA) +- 3x MLP expansion with LeakyReLU(0.5)² activation +- BigramHash(2048, dim=128), ValueEmbedding(128, layers 9,10) +- Value Residual Learning (VRL) across all layers +- XSA (Exclusive Self-Attention) on last 4 layers +- U-Net skip connections (encoder-decoder with skip weights) +- SmearGate (learned 1-token look-back) +- Partial RoPE (16 dims), LN Scale + +## Key Innovation: Complementary Training + +Standard approach: train model on uniform cross-entropy, bolt on n-gram cache at eval. + +Our approach: during training, downweight tokens that a bigram predictor would get right (COMPLEMENT_ALPHA=0.5). The model learns to focus its 27M parameters on tokens that statistical caches can't predict — novel word choices, long-range dependencies, semantic surprises. + +| Config | BPB | +|--------|-----| +| Base model only | ~1.139 | +| + Standard backoff (alpha=0.05) | ~0.700 | +| + Complementary training + alpha=0.20 | **0.4311** | + +## Validated Results (3-Seed) + +| Seed | BPB | Artifact (bytes) | TTT Eval Time | +|------|-----|-----------------|---------------| +| 1337 | 0.431107 | 15,916,181 | 477s | +| 42 | 0.431062 | 15,962,841 | 477s | +| 2024 | 0.431112 | 15,958,961 | 475s | +| **Mean** | **0.431094** | **max 15,962,841** | **~476s** | + +All runs: training stopped at 600s, full eval (diag+q_rt+q_sw+TTT+ngram) completed in ~562s ≈ 9.37 min. + +## Eval Stack + +- **BackoffNgramMixer**: orders 2-10, 4M flat hash buckets, greedy cascade +- **Entropy-adaptive alpha**: `0.20 + 0.55 * sigmoid(2*(H - 3.0))` +- **AdamW TTT**: lr=5e-4, 3 epochs/chunk, Polyak EMA 0.998, freeze first 9/11 blocks +- **Sliding window**: stride=64 + +## Legality + +1. **Complementary training**: reweights training loss using training-data bigram statistics only. No validation data accessed during training. +2. **N-gram cache**: built from already-scored tokens only (score-first, backward-looking). +3. **Alpha formula**: fixed function of model entropy, computed before seeing target token. +4. **TTT**: score-first legal TTT on already-evaluated chunks. +5. **Committed distribution**: (1-α)·P_neural + α·P_ngram — proper mixture, all tokens have nonzero probability. + +## Reproduction + +```bash +# Single seed +VRL_ENABLED=1 LEAKY_RELU=1 TTT_ENABLED=1 TTT_OPTIMIZER=adamw TTT_LR=0.0005 TTT_EPOCHS=3 TTT_FREEZE_BLOCKS=2 TTT_TEMPERATURE=0.98 USE_HEDGE_MIXER=1 NGRAM_ORDER=10 NGRAM_BUCKETS=4194304 ALPHA_BASE=0.20 ALPHA_RANGE=0.55 ALPHA_CENTER=3.0 COMPLEMENT_ALPHA=0.5 SEED=42 torchrun --standalone --nproc_per_node=8 train_gpt.py + +# Multi-seed (3-seed validation) +for SEED in 42 1337 2024; do + VRL_ENABLED=1 LEAKY_RELU=1 TTT_ENABLED=1 TTT_OPTIMIZER=adamw TTT_LR=0.0005 TTT_EPOCHS=3 TTT_FREEZE_BLOCKS=2 TTT_TEMPERATURE=0.98 USE_HEDGE_MIXER=1 NGRAM_ORDER=10 NGRAM_BUCKETS=4194304 ALPHA_BASE=0.20 ALPHA_RANGE=0.55 ALPHA_CENTER=3.0 COMPLEMENT_ALPHA=0.5 SEED=$SEED torchrun --standalone --nproc_per_node=8 train_gpt.py +done +``` + +## Credits + +Based on PR #803 (pentxayc) — Complementary Training + BackoffNgramMixer. +Builds on techniques from: PR #779 (BackoffNgramMixer), PR #549 (LeakyReLU² + TTT), PR #287 (XSA + EMA), PR #413 (VRL), PR #414 (GPTQ-lite base). diff --git a/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed1337.txt b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed1337.txt new file mode 100644 index 000000000..1f66a60fc --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed1337.txt @@ -0,0 +1,2234 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +_FA_VERSION = 0 +_fa_func = None +try: + from flash_attn_interface import flash_attn_func as _fa_func + _FA_VERSION = 3 +except ImportError: + try: + from flash_attn import flash_attn_func as _fa_func + _FA_VERSION = 2 + except ImportError: + _FA_VERSION = 0 + _fa_func = None +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + soft_round_qat = bool(int(os.environ.get("SOFT_ROUND_QAT", "1"))) + soft_round_temp_start = float(os.environ.get("SOFT_ROUND_TEMP_START", 1.0)) + soft_round_temp_end = float(os.environ.get("SOFT_ROUND_TEMP_END", 0.05)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + leaky_relu = bool(int(os.environ.get("LEAKY_RELU", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adamw") + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + polyak_decay = float(os.environ.get("POLYAK_DECAY", 0.998)) + use_polyak = bool(int(os.environ.get("USE_POLYAK", "1"))) + byte_weighted_ttt = bool(int(os.environ.get("BYTE_WEIGHTED_TTT", "1"))) + adaptive_lr = bool(int(os.environ.get("ADAPTIVE_LR", "1"))) + adaptive_lr_max = float(os.environ.get("ADAPTIVE_LR_MAX", 3.0)) + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + checkpoint_path = os.environ.get("CHECKPOINT_PATH", "final_model.pt") + ttt_max_chunks = int(os.environ.get("TTT_MAX_CHUNKS", 0)) + skip_sliding_window = bool(int(os.environ.get("SKIP_SLIDING_WINDOW", "0"))) + use_hedge_mixer = bool(int(os.environ.get("USE_HEDGE_MIXER", "1"))) + mixer_eta = float(os.environ.get("MIXER_ETA", 0.1)) + mixer_min_tokens = int(os.environ.get("MIXER_MIN_TOKENS", 10000)) +class BackoffNgramMixer: + PRIMES = [36313, 27191, 51647, 81929, 131071, 174763, 233017] + def __init__(self, vocab_size: int, device: torch.device, num_buckets: int = 4_000_000, + max_order: int = 7, min_count: int = 2, min_tokens: int = 5000, + alpha_base: float = 0.05, alpha_range: float = 0.55, alpha_center: float = 4.0): + self.V = vocab_size + self.B = num_buckets + self.MASK = num_buckets - 1 if (num_buckets & (num_buckets - 1)) == 0 else None + self.max_order = max_order + self.min_count = min_count + self.min_tokens = min_tokens + self.device = device + self.tokens_seen = 0 + self.alpha_base = alpha_base + self.alpha_range = alpha_range + self.alpha_center = alpha_center + self.uni_counts = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.uni_total = 0.0 + self.ctx_counts = [] + self.full_counts = [] + for _ in range(max_order - 1): + self.ctx_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + self.full_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + def _bucket(self, h: Tensor) -> Tensor: + if self.MASK is not None: + return h & self.MASK + return h.abs() % self.B + def update(self, tokens: Tensor): + t = tokens.to(self.device).long() + n = t.numel() + self.tokens_seen += n + ones = torch.ones(n, device=self.device, dtype=torch.float32) + self.uni_counts.scatter_add_(0, t, ones) + self.uni_total += n + for order in range(2, self.max_order + 1): + if n < order: + continue + oi = order - 2 + nxt = t[order - 1:] + ctx_h = t[0:n - order + 1] * self.PRIMES[0] + for k in range(1, order - 1): + ctx_h = ctx_h ^ (t[k:n - order + 1 + k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (nxt * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + self.ctx_counts[oi].scatter_add_(0, ctx_key, ones[:n - order + 1]) + self.full_counts[oi].scatter_add_(0, full_key, ones[:n - order + 1]) + def score(self, logits: Tensor, x_batch: Tensor, y_batch: Tensor, + temperature: float = 1.0) -> Tensor: + bsz, slen, V = logits.shape + if temperature != 1.0: + logits = logits / temperature + log_probs_neural = F.log_softmax(logits.float(), dim=-1) + neural_p = log_probs_neural.gather(-1, y_batch.unsqueeze(-1)).squeeze(-1).exp() + neural_nll = -neural_p.clamp(min=1e-12).log() + if self.tokens_seen < self.min_tokens: + return neural_nll + ctx_stack = [x_batch] + for k in range(1, self.max_order - 1): + shifted = torch.zeros_like(x_batch) + if k < slen: + shifted[:, k:] = x_batch[:, :-k] + ctx_stack.append(shifted) + if self.uni_total > 0: + uni_p = (self.uni_counts[y_batch] + 0.5) / (self.uni_total + 0.5 * V) + ngram_p = uni_p + else: + ngram_p = torch.full((bsz, slen), 1.0 / V, device=self.device) + ngram_hit = torch.zeros(bsz, slen, device=self.device, dtype=torch.bool) + for order in range(self.max_order, 1, -1): + oi = order - 2 + cw = order - 1 + ctx_h = ctx_stack[cw - 1] * self.PRIMES[0] + for k in range(1, cw): + ctx_h = ctx_h ^ (ctx_stack[cw - 1 - k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (y_batch * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + ctx_c = self.ctx_counts[oi][ctx_key] + full_c = self.full_counts[oi][full_key] + valid = (ctx_c >= self.min_count) & (~ngram_hit) + min_pos = order - 2 + if min_pos > 0: + valid[:, :min_pos] = False + p = torch.where(valid, full_c.clamp(max=ctx_c) / ctx_c.clamp(min=1), torch.zeros_like(ctx_c)) + p = p.clamp(0, 1) + ngram_p = torch.where(valid, p, ngram_p) + ngram_hit = ngram_hit | valid + ngram_nll = -ngram_p.clamp(min=1e-12).log() + probs_neural = log_probs_neural.exp() + entropy = -(probs_neural * log_probs_neural).sum(dim=-1) + alpha = self.alpha_base + self.alpha_range * torch.sigmoid( + 2.0 * (entropy - self.alpha_center)) + mixed_p = (1.0 - alpha) * neural_p + alpha * ngram_p + return -mixed_p.clamp(min=1e-12).log() +class TrainNgramTracker: + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) +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:{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"val too short for {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 too small; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_scales", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + _soft_round_qat: bool = True + _soft_round_temp: float = 1.0 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._soft_round_qat: + w32 = self.weight.float() + row_max = w32.detach().abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_s = w32 / scale[:, None] + residual = w_s - w_s.detach().round() + temp = CastedLinear._soft_round_temp + w_soft = w_s.detach().round() + 0.5 * torch.tanh(residual / temp) + w = (w_soft.clamp(-32, 31) * scale[:, None]).to(x.dtype) + else: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim%num_heads!=0") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads%num_kv_heads!=0") + 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("odd head_dim") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _FA_VERSION == 3: + y = _fa_func(q, k, v, causal=True) + elif _FA_VERSION == 2: + y = _fa_func(q.bfloat16(), k.bfloat16(), v.bfloat16(), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, enable_gqa=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int, leaky: bool = False): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + self._neg_slope = 0.5 if leaky else 0.0 + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), self._neg_slope) + return self.proj(x.square()) +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + **kwargs, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=kwargs.get("gated_attention", False)) + self.mlp = MLP(dim, mlp_mult, leaky=kwargs.get("leaky", False)) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + vrl_enabled: bool = False, + leaky_relu: bool = False, + gated_attention: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + if logit_softcap <= 0.0: + raise ValueError(f"softcap<=0:{logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + leaky=leaky_relu, + gated_attention=gated_attention, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + if self.vrl_enabled: + self.vrl_scales = nn.ParameterList( + [nn.Parameter(torch.zeros(1, dtype=torch.float32)) for _ in range(num_layers - 1)] + ) + else: + self.vrl_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("need lm_head") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) +def eval_val_sliding_ttt( + args, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + if args.ttt_max_chunks > 0: + num_chunks = min(num_chunks, args.ttt_max_chunks) + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + if ci < num_chunks: + chunk_windows[ci].append(ws) + log0(f"ttt:c={num_chunks} ct={ttt_chunk} w={len(window_starts)} s={stride} lr={args.ttt_lr} ep={args.ttt_epochs} fb={args.ttt_freeze_blocks} o={args.ttt_optimizer} pk={args.use_polyak}({args.polyak_decay}) bw={args.byte_weighted_ttt} alr={args.adaptive_lr}({args.adaptive_lr_max}) t={args.ttt_temperature}") + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + num_blocks = len(base_model.blocks) + unfrozen_block_start = max(0, num_blocks - args.ttt_freeze_blocks) if args.ttt_freeze_blocks > 0 else 0 + ttt_params = [] + for name, p in base_model.named_parameters(): + unfreeze = False + if args.ttt_freeze_blocks <= 0: + unfreeze = True + elif "norm" in name or "scale" in name or "lm_head" in name or "tok_emb" in name: + unfreeze = True + else: + for bi in range(unfrozen_block_start, num_blocks): + if f"blocks.{bi}." in name: + unfreeze = True + break + if unfreeze: + p.requires_grad_(True) + ttt_params.append(p) + else: + p.requires_grad_(False) + log0(f"ttt:uf={sum(p.numel() for p in ttt_params)} f={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + polyak_state: dict[str, Tensor] | None = None + if args.use_polyak: + polyak_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + mixer: BackoffNgramMixer | None = None + if args.use_hedge_mixer: + ngram_order = int(os.environ.get("NGRAM_ORDER", "7")) + ngram_buckets = int(os.environ.get("NGRAM_BUCKETS", "4000000")) + alpha_base = float(os.environ.get("ALPHA_BASE", "0.05")) + alpha_range = float(os.environ.get("ALPHA_RANGE", "0.55")) + alpha_center = float(os.environ.get("ALPHA_CENTER", "4.0")) + min_count = int(os.environ.get("MIN_COUNT", "2")) + mixer = BackoffNgramMixer(args.vocab_size, device, num_buckets=ngram_buckets, + max_order=ngram_order, min_count=min_count, + min_tokens=args.mixer_min_tokens, + alpha_base=alpha_base, alpha_range=alpha_range, + alpha_center=alpha_center) + mem_mb = ngram_buckets * 4 * 2 * (ngram_order - 1) / 1e6 + log0(f"bo:o={ngram_order} b={ngram_buckets} m={mem_mb:.0f}M a={alpha_base}+{alpha_range}*s(H-{alpha_center}) mc={min_count}") + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + raw_state: dict[str, Tensor] | None = None + if polyak_state is not None: + raw_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + for n, p in base_model.named_parameters(): + if n in polyak_state: + p.data.copy_(polyak_state[n]) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + if mixer is not None and mixer.tokens_seen >= mixer.min_tokens: + nll = mixer.score(logits, x_batch, y_batch, args.ttt_temperature) + else: + if args.ttt_temperature != 1.0: + logits = logits / args.ttt_temperature + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if mixer is not None: + chunk_tokens = val_tokens[chunk_start:chunk_end].to(device) + mixer.update(chunk_tokens) + if raw_state is not None: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if args.adaptive_lr: + progress = min(ci / (num_chunks * 0.3), 1.0) + lr_mult = 1.0 + (args.adaptive_lr_max - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + distributed = dist.is_available() and dist.is_initialized() + my_seq_s = (chunk_seqs * rank) // world_size if distributed else 0 + my_seq_e = (chunk_seqs * (rank + 1)) // world_size if distributed else chunk_seqs + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_t = base_model.forward_logits(x) + if args.byte_weighted_ttt: + per_tok_nll = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), reduction="none", + ) + byte_weights = base_bytes_lut[y.reshape(-1)].float() + byte_weights = byte_weights / byte_weights.mean().clamp(min=1e-6) + loss = (per_tok_nll * byte_weights).mean() + else: + loss = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), + ) + loss.backward() + if distributed and world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if polyak_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in polyak_state: + polyak_state[n].lerp_(p.data, 1.0 - args.polyak_decay) + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" tc[{ci+1}/{num_chunks}]bpb={rbpb:.6f} t={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + log0(f"ttt:vl={val_loss:.6f} bpb={val_bpb:.6f} t={time.perf_counter()-t0:.1f}s") + return val_loss, val_bpb +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"bad WORLD_SIZE:{world_size}") + if 8 % world_size != 0: + raise ValueError(f"8%WORLD_SIZE={world_size}!=0") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("no CUDA") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + _gpu_name = torch.cuda.get_device_name(0) + _is_high_end = "H100" in _gpu_name or "A100" in _gpu_name + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + if _is_high_end: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(True) + enable_flash_sdp(True) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("="*60,console=False) + log0(f"py:{sys.version}",console=False) + log0(f"pt:{torch.__version__}",console=False) + log0(subprocess.run(["nvidia-smi"],stdout=subprocess.PIPE,stderr=subprocess.PIPE,text=True,check=False).stdout,console=False) + log0("="*60,console=False) + log0(f"fa:{_FA_VERSION} gpu:{_gpu_name} he:{_is_high_end}") + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"need .model:{args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"vocab mismatch:{args.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"bpb:sp={args.tokenizer_path}") + log0(f"train:{dataset_dir.name} shards:{actual_train_files}") + log0(f"val:{args.val_files} n:{val_tokens.numel()-1}") + CastedLinear._qat_enabled = args.qat_enabled + CastedLinear._soft_round_qat = args.soft_round_qat + CastedLinear._soft_round_temp = args.soft_round_temp_start + qat_start_step = 0 if args.qat_enabled else -1 + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, + leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + if complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=complement_alpha) + base_model._ngram_tracker = tracker + log0(f"compl:{complement_alpha}") + else: + base_model._ngram_tracker = None + if distributed: + torch._dynamo.config.optimize_ddp = False + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + if base_model.vrl_enabled: + for s in base_model.vrl_scales: + scalar_params.append(s) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"p:{n_params}") + log0(f"mtp:{args.mtp_num_heads} w:{args.mtp_loss_weight} p:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"xsa:{args.xsa_last_n} l:{xsa_layers}") + log0(f"ws:{world_size} ga:{grad_accum_steps}") + log0(f"sdp:{_is_high_end}") + log0(f"attn:h={args.num_heads} kv={args.num_kv_heads}") + log0(f"vrl:{args.vrl_enabled} lrelu:{args.leaky_relu} ttt:{args.ttt_enabled}") + log0(f"tie:{args.tie_embeddings} elr:{token_lr} hlr:{args.head_lr if base_model.lm_head is not None else 0.0} mlr:{args.matrix_lr} slr:{args.scalar_lr}") + log0(f"tbt:{args.train_batch_tokens} tsl:{args.train_seq_len} it:{args.iterations} wu:{args.warmup_steps} mws:{args.max_wallclock_seconds:.3f}") + log0(f"s:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0 and not args.eval_only: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"wu:{warmup_step+1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.eval_only: + log0(f"eval:load {args.checkpoint_path}") + ckpt_state = torch.load(args.checkpoint_path, map_location="cpu") + base_model.load_state_dict(ckpt_state, strict=True) + log0(f"eval:loaded {sum(p.numel() for p in base_model.parameters())}p") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + log0(f"eval:qsize:{len(quant_blob)}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = effective_eval_seq_len + if not args.skip_sliding_window and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0(f"eval:sw bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + elif args.skip_sliding_window: + log0("eval:skip_sw") + if args.ttt_enabled: + log0(f"eval:ttt lr={args.ttt_lr} ep={args.ttt_epochs} c={args.ttt_chunk_tokens} fb={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"eval:ttt bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + if distributed: + dist.destroy_process_group() + return + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0(f"s:{step}/{args.iterations} vl:{val_loss:.4f} bpb:{val_bpb:.4f} tt:{training_time_ms:.0f}ms sa:{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"stop tt:{training_time_ms:.0f}ms s:{step}/{args.iterations}") + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + qat_start_step = step + log0(f"qat:{step} s:{scale:.4f}") + if CastedLinear._qat_enabled and CastedLinear._soft_round_qat and qat_start_step >= 0: + qat_total = max(args.iterations - qat_start_step, 1) + qat_progress = min((step - qat_start_step) / qat_total, 1.0) + log_start = math.log(args.soft_round_temp_start) + log_end = math.log(args.soft_round_temp_end) + CastedLinear._soft_round_temp = math.exp(log_start + qat_progress * (log_end - log_start)) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + 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:{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"s:{step}/{args.iterations} tl:{train_loss.item():.4f} tt:{approx_training_time_ms:.0f}ms sa:{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"mem:{torch.cuda.max_memory_allocated()//1024//1024}M R:{torch.cuda.max_memory_reserved()//1024//1024}M") + log0("ema:apply") + 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"diag vl:{diag_val_loss:.4f} bpb:{diag_val_bpb:.4f} t:{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"excl_mtp:{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"model:{model_bytes}B") + log0(f"code:{code_bytes}B") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"q:{quant_file_bytes}B") + log0(f"total:{quant_file_bytes+code_bytes}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0(f"q_rt vl:{q_val_loss:.4f} bpb:{q_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_qeval):.0f}ms") + log0(f"q_rt_x vl:{q_val_loss:.8f} 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"q_sw vl:{sw_val_loss:.4f} bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + log0(f"q_sw_x vl:{sw_val_loss:.8f} bpb:{sw_val_bpb:.8f}") + log0(f"q8_x vl:{sw_val_loss:.8f} 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"q_s64 vl:{sw64_val_loss:.4f} bpb:{sw64_val_bpb:.4f} s:64 t:{1000.0*(time.perf_counter()-t_slide64):.0f}ms") + log0(f"q_s64_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + log0(f"q8_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + if args.ttt_enabled: + log0("ttt:start") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"ttt vl:{ttt_val_loss:.4f} bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + log0(f"ttt_x vl:{ttt_val_loss:.8f} bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() +============================================================ +py:3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +pt:2.9.1+cu128 +Sat Mar 28 17:21:10 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 575.57.08 Driver Version: 575.57.08 CUDA Version: 12.9 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 23C P0 112W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:2A:00.0 Off | 0 | +| N/A 25C P0 115W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3A:00.0 Off | 0 | +| N/A 25C P0 116W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 22C P0 110W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9A:00.0 Off | 0 | +| N/A 23C P0 115W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:AB:00.0 Off | 0 | +| N/A 26C P0 114W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BA:00.0 Off | 0 | +| N/A 25C P0 116W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 23C P0 112W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 42587 C /usr/local/bin/python 1510MiB | +| 1 N/A N/A 42588 C /usr/local/bin/python 1510MiB | +| 2 N/A N/A 42589 C /usr/local/bin/python 1510MiB | +| 3 N/A N/A 42590 C /usr/local/bin/python 1510MiB | +| 4 N/A N/A 42591 C /usr/local/bin/python 1510MiB | +| 5 N/A N/A 42592 C /usr/local/bin/python 1510MiB | +| 6 N/A N/A 42593 C /usr/local/bin/python 1510MiB | +| 7 N/A N/A 42594 C /usr/local/bin/python 1510MiB | ++-----------------------------------------------------------------------------------------+ + +============================================================ +fa:3 gpu:NVIDIA H100 80GB HBM3 he:True +bpb:sp=./data/tokenizers/fineweb_1024_bpe.model +train:fineweb10B_sp1024 shards:80 +val:./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin n:62021632 +compl:0.5 +p:26993766 +mtp:0 w:0.2 p:0 +xsa:4 l:[7, 8, 9, 10] +ws:8 ga:1 +sdp:True +attn:h=8 kv=4 +vrl:True lrelu:True ttt:True +tie:True elr:0.035 hlr:0.0 mlr:0.025 slr:0.025 +tbt:786432 tsl:2048 it:20000 wu:20 mws:600.000 +s:1337 +wu:1/20 +wu:2/20 +wu:3/20 +wu:4/20 +wu:5/20 +wu:6/20 +wu:7/20 +wu:8/20 +wu:9/20 +wu:10/20 +wu:11/20 +wu:12/20 +wu:13/20 +wu:14/20 +wu:15/20 +wu:16/20 +wu:17/20 +wu:18/20 +wu:19/20 +wu:20/20 +s:0/20000 vl:6.9279 bpb:4.1031 tt:0ms sa:0.01ms +s:1/20000 tl:6.9299 tt:141ms sa:140.78ms +s:2/20000 tl:8.4023 tt:226ms sa:112.78ms +s:3/20000 tl:7.6863 tt:315ms sa:105.10ms +s:4/20000 tl:7.1121 tt:404ms sa:101.12ms +s:5/20000 tl:6.9430 tt:494ms sa:98.72ms +s:6/20000 tl:6.7009 tt:582ms sa:97.00ms +s:7/20000 tl:6.5880 tt:671ms sa:95.87ms +s:8/20000 tl:6.6290 tt:760ms sa:94.99ms +s:9/20000 tl:6.3012 tt:850ms sa:94.41ms +s:10/20000 tl:5.9668 tt:939ms sa:93.90ms +s:500/20000 tl:2.3517 tt:42777ms sa:85.55ms +s:1000/20000 tl:2.2343 tt:85641ms sa:85.64ms +s:1500/20000 tl:2.1866 tt:128571ms sa:85.71ms +s:2000/20000 tl:2.0326 tt:171599ms sa:85.80ms +s:2500/20000 tl:2.1350 tt:214622ms sa:85.85ms +s:3000/20000 tl:2.1273 tt:257656ms sa:85.89ms +s:3500/20000 tl:2.1500 tt:300706ms sa:85.92ms +s:4000/20000 tl:1.9410 tt:343735ms sa:85.93ms +s:4000/20000 vl:2.0531 bpb:1.2159 tt:343748ms sa:85.94ms +s:4500/20000 tl:2.0901 tt:386759ms sa:85.95ms +s:5000/20000 tl:2.0724 tt:429734ms sa:85.95ms +s:5500/20000 tl:1.9874 tt:472713ms sa:85.95ms +s:6000/20000 tl:1.9131 tt:515702ms sa:85.95ms +swa:6300 +qat:6454 s:0.1500 +s:6500/20000 tl:2.0519 tt:558892ms sa:85.98ms +s:6976/20000 vl:1.9228 bpb:1.1388 tt:600048ms sa:86.02ms +stop tt:600048ms s:6976/20000 +mem:21441M R:22162M +ema:apply +diag vl:1.9211 bpb:1.1378 t:2002ms +model:106181533B +code:94053B +q:15822128B +total:15916181B +q_rt vl:1.9344 bpb:1.1457 t:7509ms +q_rt_x vl:1.93440132 bpb:1.14566142 +q_sw vl:1.8945 bpb:1.1220 s:64 t:76214ms +q_sw_x vl:1.89452350 bpb:1.12204650 +q8_x vl:1.89452350 bpb:1.12204650 +ttt:start +ttt:c=1893 ct=32768 w=969088 s=64 lr=0.0005 ep=3 fb=2 o=adamw pk=True(0.998) bw=True alr=True(3.0) t=0.98 +ttt:uf=5256222 f=21737544 +bo:o=10 b=4194304 m=302M a=0.2+0.55*s(H-3.0) mc=2 + tc[1/1893]bpb=1.165537 t=0.6s + tc[11/1893]bpb=1.298618 t=3.4s + tc[21/1893]bpb=1.280793 t=6.1s + tc[31/1893]bpb=1.266802 t=8.6s + tc[41/1893]bpb=1.240077 t=11.1s + tc[51/1893]bpb=1.221587 t=13.6s + tc[61/1893]bpb=1.212953 t=16.1s + tc[71/1893]bpb=1.195028 t=18.6s + tc[81/1893]bpb=1.177854 t=21.1s + tc[91/1893]bpb=1.162006 t=23.6s + tc[101/1893]bpb=1.147398 t=26.2s + tc[111/1893]bpb=1.131680 t=28.7s + tc[121/1893]bpb=1.108420 t=31.2s + tc[131/1893]bpb=1.090080 t=33.7s + tc[141/1893]bpb=1.076736 t=36.2s + tc[151/1893]bpb=1.059955 t=38.7s + tc[161/1893]bpb=1.043343 t=41.2s + tc[171/1893]bpb=1.028306 t=44.1s + tc[181/1893]bpb=1.013761 t=47.1s + tc[191/1893]bpb=1.001233 t=49.8s + tc[201/1893]bpb=0.985468 t=52.4s + tc[211/1893]bpb=0.968488 t=54.9s + tc[221/1893]bpb=0.953677 t=57.4s + tc[231/1893]bpb=0.938641 t=59.9s + tc[241/1893]bpb=0.925069 t=62.4s + tc[251/1893]bpb=0.911910 t=64.9s + tc[261/1893]bpb=0.896771 t=67.4s + tc[271/1893]bpb=0.883897 t=69.9s + tc[281/1893]bpb=0.871509 t=72.4s + tc[291/1893]bpb=0.860447 t=74.9s + tc[301/1893]bpb=0.848994 t=77.4s + tc[311/1893]bpb=0.838391 t=79.9s + tc[321/1893]bpb=0.827846 t=82.4s + tc[331/1893]bpb=0.817512 t=85.0s + tc[341/1893]bpb=0.806599 t=87.5s + tc[351/1893]bpb=0.797524 t=90.0s + tc[361/1893]bpb=0.788820 t=92.5s + tc[371/1893]bpb=0.779430 t=95.0s + tc[381/1893]bpb=0.770923 t=97.5s + tc[391/1893]bpb=0.762503 t=100.0s + tc[401/1893]bpb=0.753675 t=102.5s + tc[411/1893]bpb=0.745689 t=105.0s + tc[421/1893]bpb=0.737622 t=107.6s + tc[431/1893]bpb=0.729990 t=110.1s + tc[441/1893]bpb=0.722860 t=112.6s + tc[451/1893]bpb=0.715686 t=115.1s + tc[461/1893]bpb=0.708439 t=117.6s + tc[471/1893]bpb=0.701641 t=120.1s + tc[481/1893]bpb=0.695378 t=122.6s + tc[491/1893]bpb=0.688691 t=125.1s + tc[501/1893]bpb=0.682750 t=127.6s + tc[511/1893]bpb=0.677030 t=130.1s + tc[521/1893]bpb=0.670948 t=132.6s + tc[531/1893]bpb=0.665552 t=135.2s + tc[541/1893]bpb=0.660474 t=137.7s + tc[551/1893]bpb=0.655024 t=140.2s + tc[561/1893]bpb=0.650029 t=142.7s + tc[571/1893]bpb=0.644885 t=145.2s + tc[581/1893]bpb=0.639938 t=147.7s + tc[591/1893]bpb=0.635252 t=150.2s + tc[601/1893]bpb=0.630816 t=152.7s + tc[611/1893]bpb=0.626558 t=155.2s + tc[621/1893]bpb=0.622307 t=157.8s + tc[631/1893]bpb=0.618280 t=160.3s + tc[641/1893]bpb=0.614362 t=162.8s + tc[651/1893]bpb=0.610324 t=165.3s + tc[661/1893]bpb=0.606549 t=167.8s + tc[671/1893]bpb=0.602916 t=170.3s + tc[681/1893]bpb=0.599193 t=172.8s + tc[691/1893]bpb=0.595991 t=175.3s + tc[701/1893]bpb=0.592479 t=177.8s + tc[711/1893]bpb=0.589372 t=180.5s + tc[721/1893]bpb=0.586168 t=183.1s + tc[731/1893]bpb=0.583137 t=185.6s + tc[741/1893]bpb=0.580061 t=188.1s + tc[751/1893]bpb=0.576991 t=190.6s + tc[761/1893]bpb=0.574085 t=193.1s + tc[771/1893]bpb=0.571305 t=195.6s + tc[781/1893]bpb=0.568900 t=198.2s + tc[791/1893]bpb=0.566151 t=200.7s + tc[801/1893]bpb=0.563423 t=203.2s + tc[811/1893]bpb=0.560860 t=205.7s + tc[821/1893]bpb=0.558299 t=208.2s + tc[831/1893]bpb=0.555988 t=210.7s + tc[841/1893]bpb=0.553476 t=213.2s + tc[851/1893]bpb=0.551128 t=215.7s + tc[861/1893]bpb=0.548802 t=218.2s + tc[871/1893]bpb=0.546545 t=220.7s + tc[881/1893]bpb=0.544431 t=223.2s + tc[891/1893]bpb=0.542383 t=225.7s + tc[901/1893]bpb=0.540490 t=228.2s + tc[911/1893]bpb=0.538543 t=230.7s + tc[921/1893]bpb=0.536600 t=233.2s + tc[931/1893]bpb=0.534664 t=235.7s + tc[941/1893]bpb=0.532666 t=238.3s + tc[951/1893]bpb=0.530828 t=240.8s + tc[961/1893]bpb=0.528862 t=243.3s + tc[971/1893]bpb=0.527167 t=245.8s + tc[981/1893]bpb=0.525333 t=248.3s + tc[991/1893]bpb=0.523604 t=250.8s + tc[1001/1893]bpb=0.521761 t=253.3s + tc[1011/1893]bpb=0.519988 t=255.8s + tc[1021/1893]bpb=0.518370 t=258.3s + tc[1031/1893]bpb=0.516662 t=260.9s + tc[1041/1893]bpb=0.514839 t=263.4s + tc[1051/1893]bpb=0.513129 t=265.9s + tc[1061/1893]bpb=0.511475 t=268.4s + tc[1071/1893]bpb=0.510116 t=270.9s + tc[1081/1893]bpb=0.508592 t=273.4s + tc[1091/1893]bpb=0.507041 t=275.9s + tc[1101/1893]bpb=0.505481 t=278.4s + tc[1111/1893]bpb=0.503918 t=280.9s + tc[1121/1893]bpb=0.502419 t=283.4s + tc[1131/1893]bpb=0.500966 t=285.9s + tc[1141/1893]bpb=0.499526 t=288.4s + tc[1151/1893]bpb=0.498082 t=291.0s + tc[1161/1893]bpb=0.496614 t=293.5s + tc[1171/1893]bpb=0.495235 t=296.0s + tc[1181/1893]bpb=0.493697 t=298.5s + tc[1191/1893]bpb=0.492381 t=301.1s + tc[1201/1893]bpb=0.491074 t=303.6s + tc[1211/1893]bpb=0.489676 t=306.1s + tc[1221/1893]bpb=0.488375 t=308.6s + tc[1231/1893]bpb=0.486969 t=311.1s + tc[1241/1893]bpb=0.485607 t=313.6s + tc[1251/1893]bpb=0.484290 t=316.1s + tc[1261/1893]bpb=0.483112 t=318.6s + tc[1271/1893]bpb=0.481888 t=321.1s + tc[1281/1893]bpb=0.480636 t=323.7s + tc[1291/1893]bpb=0.479487 t=326.2s + tc[1301/1893]bpb=0.478232 t=328.7s + tc[1311/1893]bpb=0.477010 t=331.2s + tc[1321/1893]bpb=0.475828 t=333.7s + tc[1331/1893]bpb=0.474706 t=336.2s + tc[1341/1893]bpb=0.473606 t=338.7s + tc[1351/1893]bpb=0.472596 t=341.2s + tc[1361/1893]bpb=0.471622 t=343.7s + tc[1371/1893]bpb=0.470611 t=346.2s + tc[1381/1893]bpb=0.469707 t=348.7s + tc[1391/1893]bpb=0.468649 t=351.3s + tc[1401/1893]bpb=0.467744 t=353.8s + tc[1411/1893]bpb=0.466891 t=356.3s + tc[1421/1893]bpb=0.465990 t=358.8s + tc[1431/1893]bpb=0.465075 t=361.3s + tc[1441/1893]bpb=0.464264 t=363.9s + tc[1451/1893]bpb=0.463494 t=366.4s + tc[1461/1893]bpb=0.462583 t=368.9s + tc[1471/1893]bpb=0.461847 t=371.4s + tc[1481/1893]bpb=0.460914 t=373.9s + tc[1491/1893]bpb=0.460073 t=376.4s + tc[1501/1893]bpb=0.459284 t=378.9s + tc[1511/1893]bpb=0.458452 t=381.4s + tc[1521/1893]bpb=0.457616 t=383.9s + tc[1531/1893]bpb=0.456813 t=386.4s + tc[1541/1893]bpb=0.455937 t=389.0s + tc[1551/1893]bpb=0.455205 t=391.5s + tc[1561/1893]bpb=0.454465 t=394.0s + tc[1571/1893]bpb=0.453653 t=396.5s + tc[1581/1893]bpb=0.452934 t=399.0s + tc[1591/1893]bpb=0.452143 t=401.5s + tc[1601/1893]bpb=0.451427 t=404.0s + tc[1611/1893]bpb=0.450661 t=406.5s + tc[1621/1893]bpb=0.449864 t=409.0s + tc[1631/1893]bpb=0.449133 t=411.6s + tc[1641/1893]bpb=0.448402 t=414.1s + tc[1651/1893]bpb=0.447653 t=416.6s + tc[1661/1893]bpb=0.446913 t=419.1s + tc[1671/1893]bpb=0.446280 t=421.6s + tc[1681/1893]bpb=0.445580 t=424.1s + tc[1691/1893]bpb=0.444815 t=426.6s + tc[1701/1893]bpb=0.444107 t=429.1s + tc[1711/1893]bpb=0.443379 t=431.6s + tc[1721/1893]bpb=0.442681 t=434.1s + tc[1731/1893]bpb=0.442013 t=436.6s + tc[1741/1893]bpb=0.441352 t=439.2s + tc[1751/1893]bpb=0.440618 t=441.7s + tc[1761/1893]bpb=0.440007 t=444.2s + tc[1771/1893]bpb=0.439338 t=446.7s + tc[1781/1893]bpb=0.438753 t=449.2s + tc[1791/1893]bpb=0.438028 t=451.7s + tc[1801/1893]bpb=0.437393 t=454.2s + tc[1811/1893]bpb=0.436756 t=456.7s + tc[1821/1893]bpb=0.436117 t=459.2s + tc[1831/1893]bpb=0.435400 t=461.8s + tc[1841/1893]bpb=0.434747 t=464.3s + tc[1851/1893]bpb=0.434130 t=466.8s + tc[1861/1893]bpb=0.433449 t=469.3s + tc[1871/1893]bpb=0.432846 t=471.8s + tc[1881/1893]bpb=0.432214 t=474.3s + tc[1891/1893]bpb=0.431594 t=476.8s + tc[1893/1893]bpb=0.431519 t=477.2s +ttt:vl=0.727904 bpb=0.431107 t=477.2s +ttt vl:0.7279 bpb:0.4311 t:477652ms +ttt_x vl:0.72790353 bpb:0.43110661 diff --git a/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed2024.txt b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed2024.txt new file mode 100644 index 000000000..7a51f3d67 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed2024.txt @@ -0,0 +1,2234 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +_FA_VERSION = 0 +_fa_func = None +try: + from flash_attn_interface import flash_attn_func as _fa_func + _FA_VERSION = 3 +except ImportError: + try: + from flash_attn import flash_attn_func as _fa_func + _FA_VERSION = 2 + except ImportError: + _FA_VERSION = 0 + _fa_func = None +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + soft_round_qat = bool(int(os.environ.get("SOFT_ROUND_QAT", "1"))) + soft_round_temp_start = float(os.environ.get("SOFT_ROUND_TEMP_START", 1.0)) + soft_round_temp_end = float(os.environ.get("SOFT_ROUND_TEMP_END", 0.05)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + leaky_relu = bool(int(os.environ.get("LEAKY_RELU", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adamw") + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + polyak_decay = float(os.environ.get("POLYAK_DECAY", 0.998)) + use_polyak = bool(int(os.environ.get("USE_POLYAK", "1"))) + byte_weighted_ttt = bool(int(os.environ.get("BYTE_WEIGHTED_TTT", "1"))) + adaptive_lr = bool(int(os.environ.get("ADAPTIVE_LR", "1"))) + adaptive_lr_max = float(os.environ.get("ADAPTIVE_LR_MAX", 3.0)) + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + checkpoint_path = os.environ.get("CHECKPOINT_PATH", "final_model.pt") + ttt_max_chunks = int(os.environ.get("TTT_MAX_CHUNKS", 0)) + skip_sliding_window = bool(int(os.environ.get("SKIP_SLIDING_WINDOW", "0"))) + use_hedge_mixer = bool(int(os.environ.get("USE_HEDGE_MIXER", "1"))) + mixer_eta = float(os.environ.get("MIXER_ETA", 0.1)) + mixer_min_tokens = int(os.environ.get("MIXER_MIN_TOKENS", 10000)) +class BackoffNgramMixer: + PRIMES = [36313, 27191, 51647, 81929, 131071, 174763, 233017] + def __init__(self, vocab_size: int, device: torch.device, num_buckets: int = 4_000_000, + max_order: int = 7, min_count: int = 2, min_tokens: int = 5000, + alpha_base: float = 0.05, alpha_range: float = 0.55, alpha_center: float = 4.0): + self.V = vocab_size + self.B = num_buckets + self.MASK = num_buckets - 1 if (num_buckets & (num_buckets - 1)) == 0 else None + self.max_order = max_order + self.min_count = min_count + self.min_tokens = min_tokens + self.device = device + self.tokens_seen = 0 + self.alpha_base = alpha_base + self.alpha_range = alpha_range + self.alpha_center = alpha_center + self.uni_counts = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.uni_total = 0.0 + self.ctx_counts = [] + self.full_counts = [] + for _ in range(max_order - 1): + self.ctx_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + self.full_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + def _bucket(self, h: Tensor) -> Tensor: + if self.MASK is not None: + return h & self.MASK + return h.abs() % self.B + def update(self, tokens: Tensor): + t = tokens.to(self.device).long() + n = t.numel() + self.tokens_seen += n + ones = torch.ones(n, device=self.device, dtype=torch.float32) + self.uni_counts.scatter_add_(0, t, ones) + self.uni_total += n + for order in range(2, self.max_order + 1): + if n < order: + continue + oi = order - 2 + nxt = t[order - 1:] + ctx_h = t[0:n - order + 1] * self.PRIMES[0] + for k in range(1, order - 1): + ctx_h = ctx_h ^ (t[k:n - order + 1 + k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (nxt * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + self.ctx_counts[oi].scatter_add_(0, ctx_key, ones[:n - order + 1]) + self.full_counts[oi].scatter_add_(0, full_key, ones[:n - order + 1]) + def score(self, logits: Tensor, x_batch: Tensor, y_batch: Tensor, + temperature: float = 1.0) -> Tensor: + bsz, slen, V = logits.shape + if temperature != 1.0: + logits = logits / temperature + log_probs_neural = F.log_softmax(logits.float(), dim=-1) + neural_p = log_probs_neural.gather(-1, y_batch.unsqueeze(-1)).squeeze(-1).exp() + neural_nll = -neural_p.clamp(min=1e-12).log() + if self.tokens_seen < self.min_tokens: + return neural_nll + ctx_stack = [x_batch] + for k in range(1, self.max_order - 1): + shifted = torch.zeros_like(x_batch) + if k < slen: + shifted[:, k:] = x_batch[:, :-k] + ctx_stack.append(shifted) + if self.uni_total > 0: + uni_p = (self.uni_counts[y_batch] + 0.5) / (self.uni_total + 0.5 * V) + ngram_p = uni_p + else: + ngram_p = torch.full((bsz, slen), 1.0 / V, device=self.device) + ngram_hit = torch.zeros(bsz, slen, device=self.device, dtype=torch.bool) + for order in range(self.max_order, 1, -1): + oi = order - 2 + cw = order - 1 + ctx_h = ctx_stack[cw - 1] * self.PRIMES[0] + for k in range(1, cw): + ctx_h = ctx_h ^ (ctx_stack[cw - 1 - k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (y_batch * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + ctx_c = self.ctx_counts[oi][ctx_key] + full_c = self.full_counts[oi][full_key] + valid = (ctx_c >= self.min_count) & (~ngram_hit) + min_pos = order - 2 + if min_pos > 0: + valid[:, :min_pos] = False + p = torch.where(valid, full_c.clamp(max=ctx_c) / ctx_c.clamp(min=1), torch.zeros_like(ctx_c)) + p = p.clamp(0, 1) + ngram_p = torch.where(valid, p, ngram_p) + ngram_hit = ngram_hit | valid + ngram_nll = -ngram_p.clamp(min=1e-12).log() + probs_neural = log_probs_neural.exp() + entropy = -(probs_neural * log_probs_neural).sum(dim=-1) + alpha = self.alpha_base + self.alpha_range * torch.sigmoid( + 2.0 * (entropy - self.alpha_center)) + mixed_p = (1.0 - alpha) * neural_p + alpha * ngram_p + return -mixed_p.clamp(min=1e-12).log() +class TrainNgramTracker: + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) +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:{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"val too short for {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 too small; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_scales", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + _soft_round_qat: bool = True + _soft_round_temp: float = 1.0 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._soft_round_qat: + w32 = self.weight.float() + row_max = w32.detach().abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_s = w32 / scale[:, None] + residual = w_s - w_s.detach().round() + temp = CastedLinear._soft_round_temp + w_soft = w_s.detach().round() + 0.5 * torch.tanh(residual / temp) + w = (w_soft.clamp(-32, 31) * scale[:, None]).to(x.dtype) + else: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim%num_heads!=0") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads%num_kv_heads!=0") + 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("odd head_dim") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _FA_VERSION == 3: + y = _fa_func(q, k, v, causal=True) + elif _FA_VERSION == 2: + y = _fa_func(q.bfloat16(), k.bfloat16(), v.bfloat16(), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, enable_gqa=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int, leaky: bool = False): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + self._neg_slope = 0.5 if leaky else 0.0 + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), self._neg_slope) + return self.proj(x.square()) +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + **kwargs, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=kwargs.get("gated_attention", False)) + self.mlp = MLP(dim, mlp_mult, leaky=kwargs.get("leaky", False)) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + vrl_enabled: bool = False, + leaky_relu: bool = False, + gated_attention: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + if logit_softcap <= 0.0: + raise ValueError(f"softcap<=0:{logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + leaky=leaky_relu, + gated_attention=gated_attention, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + if self.vrl_enabled: + self.vrl_scales = nn.ParameterList( + [nn.Parameter(torch.zeros(1, dtype=torch.float32)) for _ in range(num_layers - 1)] + ) + else: + self.vrl_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("need lm_head") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) +def eval_val_sliding_ttt( + args, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + if args.ttt_max_chunks > 0: + num_chunks = min(num_chunks, args.ttt_max_chunks) + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + if ci < num_chunks: + chunk_windows[ci].append(ws) + log0(f"ttt:c={num_chunks} ct={ttt_chunk} w={len(window_starts)} s={stride} lr={args.ttt_lr} ep={args.ttt_epochs} fb={args.ttt_freeze_blocks} o={args.ttt_optimizer} pk={args.use_polyak}({args.polyak_decay}) bw={args.byte_weighted_ttt} alr={args.adaptive_lr}({args.adaptive_lr_max}) t={args.ttt_temperature}") + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + num_blocks = len(base_model.blocks) + unfrozen_block_start = max(0, num_blocks - args.ttt_freeze_blocks) if args.ttt_freeze_blocks > 0 else 0 + ttt_params = [] + for name, p in base_model.named_parameters(): + unfreeze = False + if args.ttt_freeze_blocks <= 0: + unfreeze = True + elif "norm" in name or "scale" in name or "lm_head" in name or "tok_emb" in name: + unfreeze = True + else: + for bi in range(unfrozen_block_start, num_blocks): + if f"blocks.{bi}." in name: + unfreeze = True + break + if unfreeze: + p.requires_grad_(True) + ttt_params.append(p) + else: + p.requires_grad_(False) + log0(f"ttt:uf={sum(p.numel() for p in ttt_params)} f={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + polyak_state: dict[str, Tensor] | None = None + if args.use_polyak: + polyak_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + mixer: BackoffNgramMixer | None = None + if args.use_hedge_mixer: + ngram_order = int(os.environ.get("NGRAM_ORDER", "7")) + ngram_buckets = int(os.environ.get("NGRAM_BUCKETS", "4000000")) + alpha_base = float(os.environ.get("ALPHA_BASE", "0.05")) + alpha_range = float(os.environ.get("ALPHA_RANGE", "0.55")) + alpha_center = float(os.environ.get("ALPHA_CENTER", "4.0")) + min_count = int(os.environ.get("MIN_COUNT", "2")) + mixer = BackoffNgramMixer(args.vocab_size, device, num_buckets=ngram_buckets, + max_order=ngram_order, min_count=min_count, + min_tokens=args.mixer_min_tokens, + alpha_base=alpha_base, alpha_range=alpha_range, + alpha_center=alpha_center) + mem_mb = ngram_buckets * 4 * 2 * (ngram_order - 1) / 1e6 + log0(f"bo:o={ngram_order} b={ngram_buckets} m={mem_mb:.0f}M a={alpha_base}+{alpha_range}*s(H-{alpha_center}) mc={min_count}") + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + raw_state: dict[str, Tensor] | None = None + if polyak_state is not None: + raw_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + for n, p in base_model.named_parameters(): + if n in polyak_state: + p.data.copy_(polyak_state[n]) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + if mixer is not None and mixer.tokens_seen >= mixer.min_tokens: + nll = mixer.score(logits, x_batch, y_batch, args.ttt_temperature) + else: + if args.ttt_temperature != 1.0: + logits = logits / args.ttt_temperature + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if mixer is not None: + chunk_tokens = val_tokens[chunk_start:chunk_end].to(device) + mixer.update(chunk_tokens) + if raw_state is not None: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if args.adaptive_lr: + progress = min(ci / (num_chunks * 0.3), 1.0) + lr_mult = 1.0 + (args.adaptive_lr_max - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + distributed = dist.is_available() and dist.is_initialized() + my_seq_s = (chunk_seqs * rank) // world_size if distributed else 0 + my_seq_e = (chunk_seqs * (rank + 1)) // world_size if distributed else chunk_seqs + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_t = base_model.forward_logits(x) + if args.byte_weighted_ttt: + per_tok_nll = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), reduction="none", + ) + byte_weights = base_bytes_lut[y.reshape(-1)].float() + byte_weights = byte_weights / byte_weights.mean().clamp(min=1e-6) + loss = (per_tok_nll * byte_weights).mean() + else: + loss = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), + ) + loss.backward() + if distributed and world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if polyak_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in polyak_state: + polyak_state[n].lerp_(p.data, 1.0 - args.polyak_decay) + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" tc[{ci+1}/{num_chunks}]bpb={rbpb:.6f} t={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + log0(f"ttt:vl={val_loss:.6f} bpb={val_bpb:.6f} t={time.perf_counter()-t0:.1f}s") + return val_loss, val_bpb +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"bad WORLD_SIZE:{world_size}") + if 8 % world_size != 0: + raise ValueError(f"8%WORLD_SIZE={world_size}!=0") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("no CUDA") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + _gpu_name = torch.cuda.get_device_name(0) + _is_high_end = "H100" in _gpu_name or "A100" in _gpu_name + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + if _is_high_end: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(True) + enable_flash_sdp(True) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("="*60,console=False) + log0(f"py:{sys.version}",console=False) + log0(f"pt:{torch.__version__}",console=False) + log0(subprocess.run(["nvidia-smi"],stdout=subprocess.PIPE,stderr=subprocess.PIPE,text=True,check=False).stdout,console=False) + log0("="*60,console=False) + log0(f"fa:{_FA_VERSION} gpu:{_gpu_name} he:{_is_high_end}") + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"need .model:{args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"vocab mismatch:{args.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"bpb:sp={args.tokenizer_path}") + log0(f"train:{dataset_dir.name} shards:{actual_train_files}") + log0(f"val:{args.val_files} n:{val_tokens.numel()-1}") + CastedLinear._qat_enabled = args.qat_enabled + CastedLinear._soft_round_qat = args.soft_round_qat + CastedLinear._soft_round_temp = args.soft_round_temp_start + qat_start_step = 0 if args.qat_enabled else -1 + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, + leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + if complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=complement_alpha) + base_model._ngram_tracker = tracker + log0(f"compl:{complement_alpha}") + else: + base_model._ngram_tracker = None + if distributed: + torch._dynamo.config.optimize_ddp = False + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + if base_model.vrl_enabled: + for s in base_model.vrl_scales: + scalar_params.append(s) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"p:{n_params}") + log0(f"mtp:{args.mtp_num_heads} w:{args.mtp_loss_weight} p:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"xsa:{args.xsa_last_n} l:{xsa_layers}") + log0(f"ws:{world_size} ga:{grad_accum_steps}") + log0(f"sdp:{_is_high_end}") + log0(f"attn:h={args.num_heads} kv={args.num_kv_heads}") + log0(f"vrl:{args.vrl_enabled} lrelu:{args.leaky_relu} ttt:{args.ttt_enabled}") + log0(f"tie:{args.tie_embeddings} elr:{token_lr} hlr:{args.head_lr if base_model.lm_head is not None else 0.0} mlr:{args.matrix_lr} slr:{args.scalar_lr}") + log0(f"tbt:{args.train_batch_tokens} tsl:{args.train_seq_len} it:{args.iterations} wu:{args.warmup_steps} mws:{args.max_wallclock_seconds:.3f}") + log0(f"s:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0 and not args.eval_only: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"wu:{warmup_step+1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.eval_only: + log0(f"eval:load {args.checkpoint_path}") + ckpt_state = torch.load(args.checkpoint_path, map_location="cpu") + base_model.load_state_dict(ckpt_state, strict=True) + log0(f"eval:loaded {sum(p.numel() for p in base_model.parameters())}p") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + log0(f"eval:qsize:{len(quant_blob)}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = effective_eval_seq_len + if not args.skip_sliding_window and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0(f"eval:sw bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + elif args.skip_sliding_window: + log0("eval:skip_sw") + if args.ttt_enabled: + log0(f"eval:ttt lr={args.ttt_lr} ep={args.ttt_epochs} c={args.ttt_chunk_tokens} fb={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"eval:ttt bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + if distributed: + dist.destroy_process_group() + return + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0(f"s:{step}/{args.iterations} vl:{val_loss:.4f} bpb:{val_bpb:.4f} tt:{training_time_ms:.0f}ms sa:{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"stop tt:{training_time_ms:.0f}ms s:{step}/{args.iterations}") + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + qat_start_step = step + log0(f"qat:{step} s:{scale:.4f}") + if CastedLinear._qat_enabled and CastedLinear._soft_round_qat and qat_start_step >= 0: + qat_total = max(args.iterations - qat_start_step, 1) + qat_progress = min((step - qat_start_step) / qat_total, 1.0) + log_start = math.log(args.soft_round_temp_start) + log_end = math.log(args.soft_round_temp_end) + CastedLinear._soft_round_temp = math.exp(log_start + qat_progress * (log_end - log_start)) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + 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:{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"s:{step}/{args.iterations} tl:{train_loss.item():.4f} tt:{approx_training_time_ms:.0f}ms sa:{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"mem:{torch.cuda.max_memory_allocated()//1024//1024}M R:{torch.cuda.max_memory_reserved()//1024//1024}M") + log0("ema:apply") + 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"diag vl:{diag_val_loss:.4f} bpb:{diag_val_bpb:.4f} t:{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"excl_mtp:{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"model:{model_bytes}B") + log0(f"code:{code_bytes}B") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"q:{quant_file_bytes}B") + log0(f"total:{quant_file_bytes+code_bytes}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0(f"q_rt vl:{q_val_loss:.4f} bpb:{q_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_qeval):.0f}ms") + log0(f"q_rt_x vl:{q_val_loss:.8f} 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"q_sw vl:{sw_val_loss:.4f} bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + log0(f"q_sw_x vl:{sw_val_loss:.8f} bpb:{sw_val_bpb:.8f}") + log0(f"q8_x vl:{sw_val_loss:.8f} 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"q_s64 vl:{sw64_val_loss:.4f} bpb:{sw64_val_bpb:.4f} s:64 t:{1000.0*(time.perf_counter()-t_slide64):.0f}ms") + log0(f"q_s64_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + log0(f"q8_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + if args.ttt_enabled: + log0("ttt:start") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"ttt vl:{ttt_val_loss:.4f} bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + log0(f"ttt_x vl:{ttt_val_loss:.8f} bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() +============================================================ +py:3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +pt:2.9.1+cu128 +Sat Mar 28 18:06:49 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 575.57.08 Driver Version: 575.57.08 CUDA Version: 12.9 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 24C P0 113W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:2A:00.0 Off | 0 | +| N/A 27C P0 115W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3A:00.0 Off | 0 | +| N/A 28C P0 118W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 23C P0 111W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9A:00.0 Off | 0 | +| N/A 24C P0 117W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:AB:00.0 Off | 0 | +| N/A 28C P0 114W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BA:00.0 Off | 0 | +| N/A 26C P0 116W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 23C P0 112W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 44818 C /usr/local/bin/python 1510MiB | +| 1 N/A N/A 44819 C /usr/local/bin/python 1510MiB | +| 2 N/A N/A 44820 C /usr/local/bin/python 1510MiB | +| 3 N/A N/A 44821 C /usr/local/bin/python 1510MiB | +| 4 N/A N/A 44822 C /usr/local/bin/python 1510MiB | +| 5 N/A N/A 44823 C /usr/local/bin/python 1510MiB | +| 6 N/A N/A 44824 C /usr/local/bin/python 1510MiB | +| 7 N/A N/A 44825 C /usr/local/bin/python 1510MiB | ++-----------------------------------------------------------------------------------------+ + +============================================================ +fa:3 gpu:NVIDIA H100 80GB HBM3 he:True +bpb:sp=./data/tokenizers/fineweb_1024_bpe.model +train:fineweb10B_sp1024 shards:80 +val:./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin n:62021632 +compl:0.5 +p:26993766 +mtp:0 w:0.2 p:0 +xsa:4 l:[7, 8, 9, 10] +ws:8 ga:1 +sdp:True +attn:h=8 kv=4 +vrl:True lrelu:True ttt:True +tie:True elr:0.035 hlr:0.0 mlr:0.025 slr:0.025 +tbt:786432 tsl:2048 it:20000 wu:20 mws:600.000 +s:2024 +wu:1/20 +wu:2/20 +wu:3/20 +wu:4/20 +wu:5/20 +wu:6/20 +wu:7/20 +wu:8/20 +wu:9/20 +wu:10/20 +wu:11/20 +wu:12/20 +wu:13/20 +wu:14/20 +wu:15/20 +wu:16/20 +wu:17/20 +wu:18/20 +wu:19/20 +wu:20/20 +s:0/20000 vl:6.9317 bpb:4.1053 tt:0ms sa:0.02ms +s:1/20000 tl:6.9330 tt:139ms sa:139.36ms +s:2/20000 tl:8.5604 tt:225ms sa:112.42ms +s:3/20000 tl:7.7503 tt:314ms sa:104.58ms +s:4/20000 tl:7.0656 tt:403ms sa:100.64ms +s:5/20000 tl:6.8499 tt:492ms sa:98.39ms +s:6/20000 tl:6.7502 tt:581ms sa:96.82ms +s:7/20000 tl:6.5784 tt:670ms sa:95.78ms +s:8/20000 tl:6.5176 tt:760ms sa:94.95ms +s:9/20000 tl:6.2999 tt:849ms sa:94.33ms +s:10/20000 tl:5.9759 tt:937ms sa:93.74ms +s:500/20000 tl:2.3603 tt:42783ms sa:85.57ms +s:1000/20000 tl:2.2375 tt:85671ms sa:85.67ms +s:1500/20000 tl:2.1885 tt:128632ms sa:85.75ms +s:2000/20000 tl:2.0337 tt:171645ms sa:85.82ms +s:2500/20000 tl:2.1397 tt:214598ms sa:85.84ms +s:3000/20000 tl:2.1265 tt:257603ms sa:85.87ms +s:3500/20000 tl:2.1455 tt:300627ms sa:85.89ms +s:4000/20000 tl:1.9409 tt:343610ms sa:85.90ms +s:4000/20000 vl:2.0534 bpb:1.2161 tt:343622ms sa:85.91ms +s:4500/20000 tl:2.0926 tt:386604ms sa:85.91ms +s:5000/20000 tl:2.0750 tt:429580ms sa:85.92ms +s:5500/20000 tl:1.9878 tt:472545ms sa:85.92ms +s:6000/20000 tl:1.9126 tt:515533ms sa:85.92ms +swa:6300 +qat:6457 s:0.1498 +s:6500/20000 tl:2.0521 tt:558690ms sa:85.95ms +s:6978/20000 vl:1.9236 bpb:1.1393 tt:600013ms sa:85.99ms +stop tt:600013ms s:6978/20000 +mem:21441M R:22162M +ema:apply +diag vl:1.9220 bpb:1.1383 t:2006ms +model:106181533B +code:94053B +q:15864908B +total:15958961B +q_rt vl:1.9356 bpb:1.1463 t:7602ms +q_rt_x vl:1.93556157 bpb:1.14634859 +q_sw vl:1.8957 bpb:1.1227 s:64 t:75975ms +q_sw_x vl:1.89566226 bpb:1.12272094 +q8_x vl:1.89566226 bpb:1.12272094 +ttt:start +ttt:c=1893 ct=32768 w=969088 s=64 lr=0.0005 ep=3 fb=2 o=adamw pk=True(0.998) bw=True alr=True(3.0) t=0.98 +ttt:uf=5256222 f=21737544 +bo:o=10 b=4194304 m=302M a=0.2+0.55*s(H-3.0) mc=2 + tc[1/1893]bpb=1.160859 t=0.6s + tc[11/1893]bpb=1.300438 t=3.1s + tc[21/1893]bpb=1.281103 t=5.6s + tc[31/1893]bpb=1.267638 t=8.1s + tc[41/1893]bpb=1.240888 t=10.6s + tc[51/1893]bpb=1.221912 t=13.1s + tc[61/1893]bpb=1.213181 t=15.6s + tc[71/1893]bpb=1.195257 t=18.1s + tc[81/1893]bpb=1.178250 t=20.6s + tc[91/1893]bpb=1.162466 t=23.1s + tc[101/1893]bpb=1.147838 t=25.6s + tc[111/1893]bpb=1.132120 t=28.1s + tc[121/1893]bpb=1.108718 t=30.6s + tc[131/1893]bpb=1.090351 t=33.1s + tc[141/1893]bpb=1.077093 t=35.7s + tc[151/1893]bpb=1.060346 t=38.2s + tc[161/1893]bpb=1.043791 t=40.7s + tc[171/1893]bpb=1.028706 t=43.2s + tc[181/1893]bpb=1.014181 t=45.7s + tc[191/1893]bpb=1.001713 t=48.2s + tc[201/1893]bpb=0.986008 t=50.7s + tc[211/1893]bpb=0.969036 t=53.2s + tc[221/1893]bpb=0.954120 t=55.7s + tc[231/1893]bpb=0.939048 t=58.2s + tc[241/1893]bpb=0.925450 t=60.7s + tc[251/1893]bpb=0.912312 t=63.2s + tc[261/1893]bpb=0.897124 t=65.7s + tc[271/1893]bpb=0.884211 t=68.2s + tc[281/1893]bpb=0.871813 t=70.7s + tc[291/1893]bpb=0.860730 t=73.2s + tc[301/1893]bpb=0.849240 t=75.7s + tc[311/1893]bpb=0.838658 t=78.2s + tc[321/1893]bpb=0.828118 t=80.7s + tc[331/1893]bpb=0.817785 t=83.2s + tc[341/1893]bpb=0.806843 t=85.7s + tc[351/1893]bpb=0.797775 t=88.2s + tc[361/1893]bpb=0.789084 t=90.7s + tc[371/1893]bpb=0.779681 t=93.2s + tc[381/1893]bpb=0.771174 t=95.7s + tc[391/1893]bpb=0.762735 t=98.2s + tc[401/1893]bpb=0.753925 t=100.7s + tc[411/1893]bpb=0.745926 t=103.2s + tc[421/1893]bpb=0.737867 t=105.7s + tc[431/1893]bpb=0.730220 t=108.2s + tc[441/1893]bpb=0.723078 t=110.7s + tc[451/1893]bpb=0.715914 t=113.2s + tc[461/1893]bpb=0.708622 t=115.7s + tc[471/1893]bpb=0.701848 t=118.2s + tc[481/1893]bpb=0.695575 t=120.7s + tc[491/1893]bpb=0.688887 t=123.2s + tc[501/1893]bpb=0.682940 t=125.7s + tc[511/1893]bpb=0.677219 t=128.2s + tc[521/1893]bpb=0.671149 t=130.7s + tc[531/1893]bpb=0.665760 t=133.2s + tc[541/1893]bpb=0.660689 t=135.7s + tc[551/1893]bpb=0.655241 t=138.2s + tc[561/1893]bpb=0.650245 t=140.7s + tc[571/1893]bpb=0.645104 t=143.2s + tc[581/1893]bpb=0.640156 t=145.7s + tc[591/1893]bpb=0.635460 t=148.2s + tc[601/1893]bpb=0.631024 t=150.7s + tc[611/1893]bpb=0.626768 t=153.2s + tc[621/1893]bpb=0.622509 t=155.7s + tc[631/1893]bpb=0.618496 t=158.2s + tc[641/1893]bpb=0.614566 t=160.7s + tc[651/1893]bpb=0.610515 t=163.4s + tc[661/1893]bpb=0.606732 t=165.9s + tc[671/1893]bpb=0.603112 t=168.4s + tc[681/1893]bpb=0.599375 t=170.9s + tc[691/1893]bpb=0.596183 t=173.4s + tc[701/1893]bpb=0.592676 t=175.9s + tc[711/1893]bpb=0.589580 t=178.4s + tc[721/1893]bpb=0.586379 t=180.9s + tc[731/1893]bpb=0.583357 t=183.4s + tc[741/1893]bpb=0.580287 t=185.9s + tc[751/1893]bpb=0.577222 t=188.4s + tc[761/1893]bpb=0.574311 t=191.0s + tc[771/1893]bpb=0.571540 t=193.5s + tc[781/1893]bpb=0.569131 t=196.0s + tc[791/1893]bpb=0.566397 t=198.6s + tc[801/1893]bpb=0.563672 t=201.1s + tc[811/1893]bpb=0.561108 t=203.6s + tc[821/1893]bpb=0.558542 t=206.1s + tc[831/1893]bpb=0.556221 t=208.6s + tc[841/1893]bpb=0.553715 t=211.1s + tc[851/1893]bpb=0.551368 t=213.6s + tc[861/1893]bpb=0.549048 t=216.1s + tc[871/1893]bpb=0.546789 t=218.6s + tc[881/1893]bpb=0.544681 t=221.1s + tc[891/1893]bpb=0.542627 t=223.7s + tc[901/1893]bpb=0.540739 t=226.2s + tc[911/1893]bpb=0.538803 t=228.7s + tc[921/1893]bpb=0.536850 t=231.2s + tc[931/1893]bpb=0.534913 t=233.7s + tc[941/1893]bpb=0.532916 t=236.2s + tc[951/1893]bpb=0.531074 t=238.7s + tc[961/1893]bpb=0.529103 t=241.2s + tc[971/1893]bpb=0.527414 t=243.7s + tc[981/1893]bpb=0.525592 t=246.3s + tc[991/1893]bpb=0.523870 t=248.8s + tc[1001/1893]bpb=0.522030 t=251.3s + tc[1011/1893]bpb=0.520255 t=253.8s + tc[1021/1893]bpb=0.518635 t=256.3s + tc[1031/1893]bpb=0.516932 t=258.8s + tc[1041/1893]bpb=0.515114 t=261.3s + tc[1051/1893]bpb=0.513400 t=263.8s + tc[1061/1893]bpb=0.511744 t=266.3s + tc[1071/1893]bpb=0.510389 t=268.9s + tc[1081/1893]bpb=0.508856 t=271.4s + tc[1091/1893]bpb=0.507295 t=273.9s + tc[1101/1893]bpb=0.505725 t=276.4s + tc[1111/1893]bpb=0.504157 t=278.9s + tc[1121/1893]bpb=0.502643 t=281.4s + tc[1131/1893]bpb=0.501190 t=284.0s + tc[1141/1893]bpb=0.499749 t=286.5s + tc[1151/1893]bpb=0.498309 t=289.0s + tc[1161/1893]bpb=0.496844 t=291.5s + tc[1171/1893]bpb=0.495458 t=294.0s + tc[1181/1893]bpb=0.493917 t=296.5s + tc[1191/1893]bpb=0.492597 t=299.0s + tc[1201/1893]bpb=0.491280 t=301.5s + tc[1211/1893]bpb=0.489876 t=304.0s + tc[1221/1893]bpb=0.488574 t=306.5s + tc[1231/1893]bpb=0.487178 t=309.1s + tc[1241/1893]bpb=0.485817 t=311.6s + tc[1251/1893]bpb=0.484501 t=314.1s + tc[1261/1893]bpb=0.483317 t=316.6s + tc[1271/1893]bpb=0.482093 t=319.1s + tc[1281/1893]bpb=0.480831 t=321.6s + tc[1291/1893]bpb=0.479680 t=324.2s + tc[1301/1893]bpb=0.478422 t=326.7s + tc[1311/1893]bpb=0.477201 t=329.2s + tc[1321/1893]bpb=0.476019 t=331.7s + tc[1331/1893]bpb=0.474889 t=334.2s + tc[1341/1893]bpb=0.473787 t=336.7s + tc[1351/1893]bpb=0.472773 t=339.2s + tc[1361/1893]bpb=0.471793 t=341.7s + tc[1371/1893]bpb=0.470780 t=344.2s + tc[1381/1893]bpb=0.469876 t=346.7s + tc[1391/1893]bpb=0.468811 t=349.3s + tc[1401/1893]bpb=0.467906 t=351.8s + tc[1411/1893]bpb=0.467052 t=354.3s + tc[1421/1893]bpb=0.466155 t=356.8s + tc[1431/1893]bpb=0.465238 t=359.3s + tc[1441/1893]bpb=0.464413 t=361.8s + tc[1451/1893]bpb=0.463656 t=364.3s + tc[1461/1893]bpb=0.462743 t=366.8s + tc[1471/1893]bpb=0.462010 t=369.3s + tc[1481/1893]bpb=0.461077 t=371.8s + tc[1491/1893]bpb=0.460242 t=374.3s + tc[1501/1893]bpb=0.459461 t=376.8s + tc[1511/1893]bpb=0.458629 t=379.3s + tc[1521/1893]bpb=0.457795 t=381.8s + tc[1531/1893]bpb=0.456991 t=384.3s + tc[1541/1893]bpb=0.456112 t=386.8s + tc[1551/1893]bpb=0.455376 t=389.3s + tc[1561/1893]bpb=0.454636 t=391.8s + tc[1571/1893]bpb=0.453818 t=394.3s + tc[1581/1893]bpb=0.453099 t=396.8s + tc[1591/1893]bpb=0.452301 t=399.3s + tc[1601/1893]bpb=0.451585 t=401.8s + tc[1611/1893]bpb=0.450820 t=404.4s + tc[1621/1893]bpb=0.450023 t=406.9s + tc[1631/1893]bpb=0.449292 t=409.4s + tc[1641/1893]bpb=0.448558 t=411.9s + tc[1651/1893]bpb=0.447809 t=414.4s + tc[1661/1893]bpb=0.447068 t=416.9s + tc[1671/1893]bpb=0.446430 t=419.4s + tc[1681/1893]bpb=0.445727 t=421.9s + tc[1691/1893]bpb=0.444961 t=424.4s + tc[1701/1893]bpb=0.444246 t=426.9s + tc[1711/1893]bpb=0.443508 t=429.5s + tc[1721/1893]bpb=0.442798 t=432.0s + tc[1731/1893]bpb=0.442125 t=434.6s + tc[1741/1893]bpb=0.441459 t=437.1s + tc[1751/1893]bpb=0.440719 t=439.6s + tc[1761/1893]bpb=0.440104 t=442.1s + tc[1771/1893]bpb=0.439434 t=444.6s + tc[1781/1893]bpb=0.438847 t=447.1s + tc[1791/1893]bpb=0.438118 t=449.7s + tc[1801/1893]bpb=0.437481 t=452.2s + tc[1811/1893]bpb=0.436840 t=454.7s + tc[1821/1893]bpb=0.436193 t=457.2s + tc[1831/1893]bpb=0.435474 t=459.7s + tc[1841/1893]bpb=0.434813 t=462.2s + tc[1851/1893]bpb=0.434187 t=464.8s + tc[1861/1893]bpb=0.433507 t=467.3s + tc[1871/1893]bpb=0.432899 t=469.8s + tc[1881/1893]bpb=0.432265 t=472.3s + tc[1891/1893]bpb=0.431640 t=474.8s + tc[1893/1893]bpb=0.431566 t=475.1s +ttt:vl=0.727912 bpb=0.431112 t=475.1s +ttt vl:0.7279 bpb:0.4311 t:475615ms +ttt_x vl:0.72791201 bpb:0.43111164 diff --git a/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed42.txt b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed42.txt new file mode 100644 index 000000000..1ee357607 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/seed42.txt @@ -0,0 +1,2234 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +_FA_VERSION = 0 +_fa_func = None +try: + from flash_attn_interface import flash_attn_func as _fa_func + _FA_VERSION = 3 +except ImportError: + try: + from flash_attn import flash_attn_func as _fa_func + _FA_VERSION = 2 + except ImportError: + _FA_VERSION = 0 + _fa_func = None +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + soft_round_qat = bool(int(os.environ.get("SOFT_ROUND_QAT", "1"))) + soft_round_temp_start = float(os.environ.get("SOFT_ROUND_TEMP_START", 1.0)) + soft_round_temp_end = float(os.environ.get("SOFT_ROUND_TEMP_END", 0.05)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + leaky_relu = bool(int(os.environ.get("LEAKY_RELU", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adamw") + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + polyak_decay = float(os.environ.get("POLYAK_DECAY", 0.998)) + use_polyak = bool(int(os.environ.get("USE_POLYAK", "1"))) + byte_weighted_ttt = bool(int(os.environ.get("BYTE_WEIGHTED_TTT", "1"))) + adaptive_lr = bool(int(os.environ.get("ADAPTIVE_LR", "1"))) + adaptive_lr_max = float(os.environ.get("ADAPTIVE_LR_MAX", 3.0)) + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + checkpoint_path = os.environ.get("CHECKPOINT_PATH", "final_model.pt") + ttt_max_chunks = int(os.environ.get("TTT_MAX_CHUNKS", 0)) + skip_sliding_window = bool(int(os.environ.get("SKIP_SLIDING_WINDOW", "0"))) + use_hedge_mixer = bool(int(os.environ.get("USE_HEDGE_MIXER", "1"))) + mixer_eta = float(os.environ.get("MIXER_ETA", 0.1)) + mixer_min_tokens = int(os.environ.get("MIXER_MIN_TOKENS", 10000)) +class BackoffNgramMixer: + PRIMES = [36313, 27191, 51647, 81929, 131071, 174763, 233017] + def __init__(self, vocab_size: int, device: torch.device, num_buckets: int = 4_000_000, + max_order: int = 7, min_count: int = 2, min_tokens: int = 5000, + alpha_base: float = 0.05, alpha_range: float = 0.55, alpha_center: float = 4.0): + self.V = vocab_size + self.B = num_buckets + self.MASK = num_buckets - 1 if (num_buckets & (num_buckets - 1)) == 0 else None + self.max_order = max_order + self.min_count = min_count + self.min_tokens = min_tokens + self.device = device + self.tokens_seen = 0 + self.alpha_base = alpha_base + self.alpha_range = alpha_range + self.alpha_center = alpha_center + self.uni_counts = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.uni_total = 0.0 + self.ctx_counts = [] + self.full_counts = [] + for _ in range(max_order - 1): + self.ctx_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + self.full_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + def _bucket(self, h: Tensor) -> Tensor: + if self.MASK is not None: + return h & self.MASK + return h.abs() % self.B + def update(self, tokens: Tensor): + t = tokens.to(self.device).long() + n = t.numel() + self.tokens_seen += n + ones = torch.ones(n, device=self.device, dtype=torch.float32) + self.uni_counts.scatter_add_(0, t, ones) + self.uni_total += n + for order in range(2, self.max_order + 1): + if n < order: + continue + oi = order - 2 + nxt = t[order - 1:] + ctx_h = t[0:n - order + 1] * self.PRIMES[0] + for k in range(1, order - 1): + ctx_h = ctx_h ^ (t[k:n - order + 1 + k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (nxt * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + self.ctx_counts[oi].scatter_add_(0, ctx_key, ones[:n - order + 1]) + self.full_counts[oi].scatter_add_(0, full_key, ones[:n - order + 1]) + def score(self, logits: Tensor, x_batch: Tensor, y_batch: Tensor, + temperature: float = 1.0) -> Tensor: + bsz, slen, V = logits.shape + if temperature != 1.0: + logits = logits / temperature + log_probs_neural = F.log_softmax(logits.float(), dim=-1) + neural_p = log_probs_neural.gather(-1, y_batch.unsqueeze(-1)).squeeze(-1).exp() + neural_nll = -neural_p.clamp(min=1e-12).log() + if self.tokens_seen < self.min_tokens: + return neural_nll + ctx_stack = [x_batch] + for k in range(1, self.max_order - 1): + shifted = torch.zeros_like(x_batch) + if k < slen: + shifted[:, k:] = x_batch[:, :-k] + ctx_stack.append(shifted) + if self.uni_total > 0: + uni_p = (self.uni_counts[y_batch] + 0.5) / (self.uni_total + 0.5 * V) + ngram_p = uni_p + else: + ngram_p = torch.full((bsz, slen), 1.0 / V, device=self.device) + ngram_hit = torch.zeros(bsz, slen, device=self.device, dtype=torch.bool) + for order in range(self.max_order, 1, -1): + oi = order - 2 + cw = order - 1 + ctx_h = ctx_stack[cw - 1] * self.PRIMES[0] + for k in range(1, cw): + ctx_h = ctx_h ^ (ctx_stack[cw - 1 - k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (y_batch * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + ctx_c = self.ctx_counts[oi][ctx_key] + full_c = self.full_counts[oi][full_key] + valid = (ctx_c >= self.min_count) & (~ngram_hit) + min_pos = order - 2 + if min_pos > 0: + valid[:, :min_pos] = False + p = torch.where(valid, full_c.clamp(max=ctx_c) / ctx_c.clamp(min=1), torch.zeros_like(ctx_c)) + p = p.clamp(0, 1) + ngram_p = torch.where(valid, p, ngram_p) + ngram_hit = ngram_hit | valid + ngram_nll = -ngram_p.clamp(min=1e-12).log() + probs_neural = log_probs_neural.exp() + entropy = -(probs_neural * log_probs_neural).sum(dim=-1) + alpha = self.alpha_base + self.alpha_range * torch.sigmoid( + 2.0 * (entropy - self.alpha_center)) + mixed_p = (1.0 - alpha) * neural_p + alpha * ngram_p + return -mixed_p.clamp(min=1e-12).log() +class TrainNgramTracker: + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) +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:{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"val too short for {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 too small; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_scales", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + _soft_round_qat: bool = True + _soft_round_temp: float = 1.0 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._soft_round_qat: + w32 = self.weight.float() + row_max = w32.detach().abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_s = w32 / scale[:, None] + residual = w_s - w_s.detach().round() + temp = CastedLinear._soft_round_temp + w_soft = w_s.detach().round() + 0.5 * torch.tanh(residual / temp) + w = (w_soft.clamp(-32, 31) * scale[:, None]).to(x.dtype) + else: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim%num_heads!=0") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads%num_kv_heads!=0") + 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("odd head_dim") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _FA_VERSION == 3: + y = _fa_func(q, k, v, causal=True) + elif _FA_VERSION == 2: + y = _fa_func(q.bfloat16(), k.bfloat16(), v.bfloat16(), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, enable_gqa=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int, leaky: bool = False): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + self._neg_slope = 0.5 if leaky else 0.0 + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), self._neg_slope) + return self.proj(x.square()) +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + **kwargs, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=kwargs.get("gated_attention", False)) + self.mlp = MLP(dim, mlp_mult, leaky=kwargs.get("leaky", False)) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + vrl_enabled: bool = False, + leaky_relu: bool = False, + gated_attention: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + if logit_softcap <= 0.0: + raise ValueError(f"softcap<=0:{logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + leaky=leaky_relu, + gated_attention=gated_attention, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + if self.vrl_enabled: + self.vrl_scales = nn.ParameterList( + [nn.Parameter(torch.zeros(1, dtype=torch.float32)) for _ in range(num_layers - 1)] + ) + else: + self.vrl_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("need lm_head") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) +def eval_val_sliding_ttt( + args, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + if args.ttt_max_chunks > 0: + num_chunks = min(num_chunks, args.ttt_max_chunks) + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + if ci < num_chunks: + chunk_windows[ci].append(ws) + log0(f"ttt:c={num_chunks} ct={ttt_chunk} w={len(window_starts)} s={stride} lr={args.ttt_lr} ep={args.ttt_epochs} fb={args.ttt_freeze_blocks} o={args.ttt_optimizer} pk={args.use_polyak}({args.polyak_decay}) bw={args.byte_weighted_ttt} alr={args.adaptive_lr}({args.adaptive_lr_max}) t={args.ttt_temperature}") + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + num_blocks = len(base_model.blocks) + unfrozen_block_start = max(0, num_blocks - args.ttt_freeze_blocks) if args.ttt_freeze_blocks > 0 else 0 + ttt_params = [] + for name, p in base_model.named_parameters(): + unfreeze = False + if args.ttt_freeze_blocks <= 0: + unfreeze = True + elif "norm" in name or "scale" in name or "lm_head" in name or "tok_emb" in name: + unfreeze = True + else: + for bi in range(unfrozen_block_start, num_blocks): + if f"blocks.{bi}." in name: + unfreeze = True + break + if unfreeze: + p.requires_grad_(True) + ttt_params.append(p) + else: + p.requires_grad_(False) + log0(f"ttt:uf={sum(p.numel() for p in ttt_params)} f={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + polyak_state: dict[str, Tensor] | None = None + if args.use_polyak: + polyak_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + mixer: BackoffNgramMixer | None = None + if args.use_hedge_mixer: + ngram_order = int(os.environ.get("NGRAM_ORDER", "7")) + ngram_buckets = int(os.environ.get("NGRAM_BUCKETS", "4000000")) + alpha_base = float(os.environ.get("ALPHA_BASE", "0.05")) + alpha_range = float(os.environ.get("ALPHA_RANGE", "0.55")) + alpha_center = float(os.environ.get("ALPHA_CENTER", "4.0")) + min_count = int(os.environ.get("MIN_COUNT", "2")) + mixer = BackoffNgramMixer(args.vocab_size, device, num_buckets=ngram_buckets, + max_order=ngram_order, min_count=min_count, + min_tokens=args.mixer_min_tokens, + alpha_base=alpha_base, alpha_range=alpha_range, + alpha_center=alpha_center) + mem_mb = ngram_buckets * 4 * 2 * (ngram_order - 1) / 1e6 + log0(f"bo:o={ngram_order} b={ngram_buckets} m={mem_mb:.0f}M a={alpha_base}+{alpha_range}*s(H-{alpha_center}) mc={min_count}") + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + raw_state: dict[str, Tensor] | None = None + if polyak_state is not None: + raw_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + for n, p in base_model.named_parameters(): + if n in polyak_state: + p.data.copy_(polyak_state[n]) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + if mixer is not None and mixer.tokens_seen >= mixer.min_tokens: + nll = mixer.score(logits, x_batch, y_batch, args.ttt_temperature) + else: + if args.ttt_temperature != 1.0: + logits = logits / args.ttt_temperature + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if mixer is not None: + chunk_tokens = val_tokens[chunk_start:chunk_end].to(device) + mixer.update(chunk_tokens) + if raw_state is not None: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if args.adaptive_lr: + progress = min(ci / (num_chunks * 0.3), 1.0) + lr_mult = 1.0 + (args.adaptive_lr_max - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + distributed = dist.is_available() and dist.is_initialized() + my_seq_s = (chunk_seqs * rank) // world_size if distributed else 0 + my_seq_e = (chunk_seqs * (rank + 1)) // world_size if distributed else chunk_seqs + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_t = base_model.forward_logits(x) + if args.byte_weighted_ttt: + per_tok_nll = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), reduction="none", + ) + byte_weights = base_bytes_lut[y.reshape(-1)].float() + byte_weights = byte_weights / byte_weights.mean().clamp(min=1e-6) + loss = (per_tok_nll * byte_weights).mean() + else: + loss = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), + ) + loss.backward() + if distributed and world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if polyak_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in polyak_state: + polyak_state[n].lerp_(p.data, 1.0 - args.polyak_decay) + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" tc[{ci+1}/{num_chunks}]bpb={rbpb:.6f} t={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + log0(f"ttt:vl={val_loss:.6f} bpb={val_bpb:.6f} t={time.perf_counter()-t0:.1f}s") + return val_loss, val_bpb +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"bad WORLD_SIZE:{world_size}") + if 8 % world_size != 0: + raise ValueError(f"8%WORLD_SIZE={world_size}!=0") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("no CUDA") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + _gpu_name = torch.cuda.get_device_name(0) + _is_high_end = "H100" in _gpu_name or "A100" in _gpu_name + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + if _is_high_end: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(True) + enable_flash_sdp(True) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("="*60,console=False) + log0(f"py:{sys.version}",console=False) + log0(f"pt:{torch.__version__}",console=False) + log0(subprocess.run(["nvidia-smi"],stdout=subprocess.PIPE,stderr=subprocess.PIPE,text=True,check=False).stdout,console=False) + log0("="*60,console=False) + log0(f"fa:{_FA_VERSION} gpu:{_gpu_name} he:{_is_high_end}") + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"need .model:{args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"vocab mismatch:{args.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"bpb:sp={args.tokenizer_path}") + log0(f"train:{dataset_dir.name} shards:{actual_train_files}") + log0(f"val:{args.val_files} n:{val_tokens.numel()-1}") + CastedLinear._qat_enabled = args.qat_enabled + CastedLinear._soft_round_qat = args.soft_round_qat + CastedLinear._soft_round_temp = args.soft_round_temp_start + qat_start_step = 0 if args.qat_enabled else -1 + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, + leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + if complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=complement_alpha) + base_model._ngram_tracker = tracker + log0(f"compl:{complement_alpha}") + else: + base_model._ngram_tracker = None + if distributed: + torch._dynamo.config.optimize_ddp = False + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + if base_model.vrl_enabled: + for s in base_model.vrl_scales: + scalar_params.append(s) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"p:{n_params}") + log0(f"mtp:{args.mtp_num_heads} w:{args.mtp_loss_weight} p:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"xsa:{args.xsa_last_n} l:{xsa_layers}") + log0(f"ws:{world_size} ga:{grad_accum_steps}") + log0(f"sdp:{_is_high_end}") + log0(f"attn:h={args.num_heads} kv={args.num_kv_heads}") + log0(f"vrl:{args.vrl_enabled} lrelu:{args.leaky_relu} ttt:{args.ttt_enabled}") + log0(f"tie:{args.tie_embeddings} elr:{token_lr} hlr:{args.head_lr if base_model.lm_head is not None else 0.0} mlr:{args.matrix_lr} slr:{args.scalar_lr}") + log0(f"tbt:{args.train_batch_tokens} tsl:{args.train_seq_len} it:{args.iterations} wu:{args.warmup_steps} mws:{args.max_wallclock_seconds:.3f}") + log0(f"s:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0 and not args.eval_only: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"wu:{warmup_step+1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.eval_only: + log0(f"eval:load {args.checkpoint_path}") + ckpt_state = torch.load(args.checkpoint_path, map_location="cpu") + base_model.load_state_dict(ckpt_state, strict=True) + log0(f"eval:loaded {sum(p.numel() for p in base_model.parameters())}p") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + log0(f"eval:qsize:{len(quant_blob)}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = effective_eval_seq_len + if not args.skip_sliding_window and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0(f"eval:sw bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + elif args.skip_sliding_window: + log0("eval:skip_sw") + if args.ttt_enabled: + log0(f"eval:ttt lr={args.ttt_lr} ep={args.ttt_epochs} c={args.ttt_chunk_tokens} fb={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"eval:ttt bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + if distributed: + dist.destroy_process_group() + return + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0(f"s:{step}/{args.iterations} vl:{val_loss:.4f} bpb:{val_bpb:.4f} tt:{training_time_ms:.0f}ms sa:{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"stop tt:{training_time_ms:.0f}ms s:{step}/{args.iterations}") + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + qat_start_step = step + log0(f"qat:{step} s:{scale:.4f}") + if CastedLinear._qat_enabled and CastedLinear._soft_round_qat and qat_start_step >= 0: + qat_total = max(args.iterations - qat_start_step, 1) + qat_progress = min((step - qat_start_step) / qat_total, 1.0) + log_start = math.log(args.soft_round_temp_start) + log_end = math.log(args.soft_round_temp_end) + CastedLinear._soft_round_temp = math.exp(log_start + qat_progress * (log_end - log_start)) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + 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:{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"s:{step}/{args.iterations} tl:{train_loss.item():.4f} tt:{approx_training_time_ms:.0f}ms sa:{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"mem:{torch.cuda.max_memory_allocated()//1024//1024}M R:{torch.cuda.max_memory_reserved()//1024//1024}M") + log0("ema:apply") + 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"diag vl:{diag_val_loss:.4f} bpb:{diag_val_bpb:.4f} t:{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"excl_mtp:{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"model:{model_bytes}B") + log0(f"code:{code_bytes}B") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"q:{quant_file_bytes}B") + log0(f"total:{quant_file_bytes+code_bytes}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0(f"q_rt vl:{q_val_loss:.4f} bpb:{q_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_qeval):.0f}ms") + log0(f"q_rt_x vl:{q_val_loss:.8f} 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"q_sw vl:{sw_val_loss:.4f} bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + log0(f"q_sw_x vl:{sw_val_loss:.8f} bpb:{sw_val_bpb:.8f}") + log0(f"q8_x vl:{sw_val_loss:.8f} 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"q_s64 vl:{sw64_val_loss:.4f} bpb:{sw64_val_bpb:.4f} s:64 t:{1000.0*(time.perf_counter()-t_slide64):.0f}ms") + log0(f"q_s64_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + log0(f"q8_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + if args.ttt_enabled: + log0("ttt:start") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"ttt vl:{ttt_val_loss:.4f} bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + log0(f"ttt_x vl:{ttt_val_loss:.8f} bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() +============================================================ +py:3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +pt:2.9.1+cu128 +Sat Mar 28 17:44:51 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 575.57.08 Driver Version: 575.57.08 CUDA Version: 12.9 | +|-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 23C P0 112W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:2A:00.0 Off | 0 | +| N/A 26C P0 115W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3A:00.0 Off | 0 | +| N/A 26C P0 117W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 23C P0 111W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9A:00.0 Off | 0 | +| N/A 23C P0 116W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:AB:00.0 Off | 0 | +| N/A 27C P0 114W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BA:00.0 Off | 0 | +| N/A 25C P0 116W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 23C P0 112W / 700W | 1519MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 43751 C /usr/local/bin/python 1510MiB | +| 1 N/A N/A 43752 C /usr/local/bin/python 1510MiB | +| 2 N/A N/A 43753 C /usr/local/bin/python 1510MiB | +| 3 N/A N/A 43754 C /usr/local/bin/python 1510MiB | +| 4 N/A N/A 43755 C /usr/local/bin/python 1510MiB | +| 5 N/A N/A 43756 C /usr/local/bin/python 1510MiB | +| 6 N/A N/A 43757 C /usr/local/bin/python 1510MiB | +| 7 N/A N/A 43758 C /usr/local/bin/python 1510MiB | ++-----------------------------------------------------------------------------------------+ + +============================================================ +fa:3 gpu:NVIDIA H100 80GB HBM3 he:True +bpb:sp=./data/tokenizers/fineweb_1024_bpe.model +train:fineweb10B_sp1024 shards:80 +val:./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin n:62021632 +compl:0.5 +p:26993766 +mtp:0 w:0.2 p:0 +xsa:4 l:[7, 8, 9, 10] +ws:8 ga:1 +sdp:True +attn:h=8 kv=4 +vrl:True lrelu:True ttt:True +tie:True elr:0.035 hlr:0.0 mlr:0.025 slr:0.025 +tbt:786432 tsl:2048 it:20000 wu:20 mws:600.000 +s:42 +wu:1/20 +wu:2/20 +wu:3/20 +wu:4/20 +wu:5/20 +wu:6/20 +wu:7/20 +wu:8/20 +wu:9/20 +wu:10/20 +wu:11/20 +wu:12/20 +wu:13/20 +wu:14/20 +wu:15/20 +wu:16/20 +wu:17/20 +wu:18/20 +wu:19/20 +wu:20/20 +s:0/20000 vl:6.9301 bpb:4.1044 tt:0ms sa:0.01ms +s:1/20000 tl:6.9318 tt:140ms sa:139.72ms +s:2/20000 tl:8.4789 tt:224ms sa:112.16ms +s:3/20000 tl:7.7246 tt:313ms sa:104.21ms +s:4/20000 tl:7.1363 tt:402ms sa:100.48ms +s:5/20000 tl:6.8906 tt:491ms sa:98.10ms +s:6/20000 tl:6.7574 tt:579ms sa:96.55ms +s:7/20000 tl:6.6339 tt:668ms sa:95.42ms +s:8/20000 tl:6.5757 tt:757ms sa:94.57ms +s:9/20000 tl:6.2880 tt:846ms sa:93.99ms +s:10/20000 tl:5.9478 tt:934ms sa:93.43ms +s:500/20000 tl:2.3523 tt:42782ms sa:85.56ms +s:1000/20000 tl:2.2425 tt:85722ms sa:85.72ms +s:1500/20000 tl:2.1848 tt:128626ms sa:85.75ms +s:2000/20000 tl:2.0330 tt:171630ms sa:85.82ms +s:2500/20000 tl:2.1310 tt:214656ms sa:85.86ms +s:3000/20000 tl:2.1301 tt:257634ms sa:85.88ms +s:3500/20000 tl:2.1460 tt:300614ms sa:85.89ms +s:4000/20000 tl:1.9418 tt:343619ms sa:85.90ms +s:4000/20000 vl:2.0525 bpb:1.2156 tt:343632ms sa:85.91ms +s:4500/20000 tl:2.0894 tt:386622ms sa:85.92ms +s:5000/20000 tl:2.0727 tt:429637ms sa:85.93ms +s:5500/20000 tl:1.9855 tt:472634ms sa:85.93ms +s:6000/20000 tl:1.9127 tt:515558ms sa:85.93ms +swa:6300 +qat:6456 s:0.1500 +s:6500/20000 tl:2.0495 tt:558730ms sa:85.96ms +s:6979/20000 vl:1.9223 bpb:1.1385 tt:600057ms sa:85.98ms +stop tt:600057ms s:6979/20000 +mem:21441M R:22162M +ema:apply +diag vl:1.9205 bpb:1.1375 t:2003ms +model:106181533B +code:94053B +q:15868788B +total:15962841B +q_rt vl:1.9338 bpb:1.1453 t:7809ms +q_rt_x vl:1.93378768 bpb:1.14529799 +q_sw vl:1.8940 bpb:1.1217 s:64 t:75497ms +q_sw_x vl:1.89396969 bpb:1.12171850 +q8_x vl:1.89396969 bpb:1.12171850 +ttt:start +ttt:c=1893 ct=32768 w=969088 s=64 lr=0.0005 ep=3 fb=2 o=adamw pk=True(0.998) bw=True alr=True(3.0) t=0.98 +ttt:uf=5256222 f=21737544 +bo:o=10 b=4194304 m=302M a=0.2+0.55*s(H-3.0) mc=2 + tc[1/1893]bpb=1.158968 t=0.6s + tc[11/1893]bpb=1.299932 t=3.1s + tc[21/1893]bpb=1.282005 t=5.6s + tc[31/1893]bpb=1.268115 t=8.0s + tc[41/1893]bpb=1.242339 t=10.5s + tc[51/1893]bpb=1.223521 t=13.0s + tc[61/1893]bpb=1.214404 t=15.6s + tc[71/1893]bpb=1.196230 t=18.1s + tc[81/1893]bpb=1.179177 t=20.6s + tc[91/1893]bpb=1.163334 t=23.1s + tc[101/1893]bpb=1.148475 t=25.6s + tc[111/1893]bpb=1.132706 t=28.1s + tc[121/1893]bpb=1.109240 t=30.6s + tc[131/1893]bpb=1.090856 t=33.1s + tc[141/1893]bpb=1.077436 t=35.6s + tc[151/1893]bpb=1.060598 t=38.1s + tc[161/1893]bpb=1.043927 t=40.6s + tc[171/1893]bpb=1.028810 t=43.1s + tc[181/1893]bpb=1.014332 t=45.6s + tc[191/1893]bpb=1.001857 t=48.1s + tc[201/1893]bpb=0.986090 t=50.6s + tc[211/1893]bpb=0.969135 t=53.0s + tc[221/1893]bpb=0.954269 t=55.5s + tc[231/1893]bpb=0.939247 t=58.0s + tc[241/1893]bpb=0.925669 t=60.5s + tc[251/1893]bpb=0.912470 t=63.0s + tc[261/1893]bpb=0.897320 t=65.5s + tc[271/1893]bpb=0.884391 t=68.0s + tc[281/1893]bpb=0.871949 t=70.5s + tc[291/1893]bpb=0.860849 t=73.0s + tc[301/1893]bpb=0.849371 t=75.5s + tc[311/1893]bpb=0.838742 t=78.0s + tc[321/1893]bpb=0.828162 t=80.5s + tc[331/1893]bpb=0.817836 t=83.0s + tc[341/1893]bpb=0.806923 t=85.5s + tc[351/1893]bpb=0.797844 t=88.0s + tc[361/1893]bpb=0.789127 t=90.5s + tc[371/1893]bpb=0.779705 t=93.0s + tc[381/1893]bpb=0.771159 t=95.5s + tc[391/1893]bpb=0.762694 t=98.0s + tc[401/1893]bpb=0.753873 t=100.5s + tc[411/1893]bpb=0.745877 t=103.0s + tc[421/1893]bpb=0.737801 t=105.5s + tc[431/1893]bpb=0.730163 t=108.0s + tc[441/1893]bpb=0.723031 t=110.5s + tc[451/1893]bpb=0.715854 t=112.9s + tc[461/1893]bpb=0.708577 t=115.4s + tc[471/1893]bpb=0.701808 t=117.9s + tc[481/1893]bpb=0.695557 t=120.4s + tc[491/1893]bpb=0.688889 t=122.9s + tc[501/1893]bpb=0.682947 t=125.4s + tc[511/1893]bpb=0.677217 t=127.9s + tc[521/1893]bpb=0.671118 t=130.4s + tc[531/1893]bpb=0.665734 t=132.9s + tc[541/1893]bpb=0.660638 t=135.4s + tc[551/1893]bpb=0.655185 t=137.9s + tc[561/1893]bpb=0.650173 t=140.4s + tc[571/1893]bpb=0.645017 t=142.9s + tc[581/1893]bpb=0.640043 t=145.4s + tc[591/1893]bpb=0.635344 t=147.9s + tc[601/1893]bpb=0.630906 t=150.6s + tc[611/1893]bpb=0.626633 t=153.1s + tc[621/1893]bpb=0.622355 t=155.5s + tc[631/1893]bpb=0.618330 t=158.0s + tc[641/1893]bpb=0.614421 t=160.5s + tc[651/1893]bpb=0.610373 t=163.1s + tc[661/1893]bpb=0.606595 t=165.8s + tc[671/1893]bpb=0.602969 t=168.8s + tc[681/1893]bpb=0.599233 t=171.8s + tc[691/1893]bpb=0.596029 t=174.7s + tc[701/1893]bpb=0.592503 t=177.6s + tc[711/1893]bpb=0.589402 t=180.1s + tc[721/1893]bpb=0.586197 t=182.6s + tc[731/1893]bpb=0.583167 t=185.1s + tc[741/1893]bpb=0.580083 t=187.7s + tc[751/1893]bpb=0.576994 t=190.6s + tc[761/1893]bpb=0.574088 t=193.6s + tc[771/1893]bpb=0.571296 t=196.3s + tc[781/1893]bpb=0.568886 t=198.7s + tc[791/1893]bpb=0.566151 t=201.2s + tc[801/1893]bpb=0.563434 t=203.7s + tc[811/1893]bpb=0.560865 t=206.2s + tc[821/1893]bpb=0.558304 t=208.7s + tc[831/1893]bpb=0.555967 t=211.2s + tc[841/1893]bpb=0.553462 t=213.7s + tc[851/1893]bpb=0.551105 t=216.2s + tc[861/1893]bpb=0.548774 t=218.7s + tc[871/1893]bpb=0.546515 t=221.2s + tc[881/1893]bpb=0.544391 t=223.7s + tc[891/1893]bpb=0.542345 t=226.1s + tc[901/1893]bpb=0.540455 t=228.6s + tc[911/1893]bpb=0.538515 t=231.1s + tc[921/1893]bpb=0.536571 t=233.6s + tc[931/1893]bpb=0.534643 t=236.1s + tc[941/1893]bpb=0.532661 t=238.6s + tc[951/1893]bpb=0.530811 t=241.1s + tc[961/1893]bpb=0.528852 t=243.5s + tc[971/1893]bpb=0.527154 t=246.0s + tc[981/1893]bpb=0.525324 t=248.5s + tc[991/1893]bpb=0.523592 t=251.0s + tc[1001/1893]bpb=0.521753 t=253.5s + tc[1011/1893]bpb=0.519977 t=256.0s + tc[1021/1893]bpb=0.518366 t=258.5s + tc[1031/1893]bpb=0.516660 t=261.0s + tc[1041/1893]bpb=0.514843 t=263.5s + tc[1051/1893]bpb=0.513133 t=266.0s + tc[1061/1893]bpb=0.511485 t=268.4s + tc[1071/1893]bpb=0.510129 t=270.9s + tc[1081/1893]bpb=0.508615 t=273.4s + tc[1091/1893]bpb=0.507064 t=275.9s + tc[1101/1893]bpb=0.505497 t=278.4s + tc[1111/1893]bpb=0.503946 t=280.9s + tc[1121/1893]bpb=0.502451 t=283.4s + tc[1131/1893]bpb=0.500998 t=285.9s + tc[1141/1893]bpb=0.499562 t=288.4s + tc[1151/1893]bpb=0.498121 t=290.9s + tc[1161/1893]bpb=0.496659 t=293.4s + tc[1171/1893]bpb=0.495273 t=295.9s + tc[1181/1893]bpb=0.493737 t=298.3s + tc[1191/1893]bpb=0.492424 t=300.8s + tc[1201/1893]bpb=0.491112 t=303.3s + tc[1211/1893]bpb=0.489710 t=305.8s + tc[1221/1893]bpb=0.488411 t=308.3s + tc[1231/1893]bpb=0.487016 t=310.8s + tc[1241/1893]bpb=0.485656 t=313.3s + tc[1251/1893]bpb=0.484337 t=315.8s + tc[1261/1893]bpb=0.483150 t=318.3s + tc[1271/1893]bpb=0.481929 t=320.8s + tc[1281/1893]bpb=0.480667 t=323.3s + tc[1291/1893]bpb=0.479515 t=325.8s + tc[1301/1893]bpb=0.478262 t=328.3s + tc[1311/1893]bpb=0.477050 t=330.8s + tc[1321/1893]bpb=0.475875 t=333.3s + tc[1331/1893]bpb=0.474746 t=335.8s + tc[1341/1893]bpb=0.473649 t=338.3s + tc[1351/1893]bpb=0.472633 t=340.8s + tc[1361/1893]bpb=0.471657 t=343.3s + tc[1371/1893]bpb=0.470643 t=345.8s + tc[1381/1893]bpb=0.469739 t=348.3s + tc[1391/1893]bpb=0.468672 t=350.8s + tc[1401/1893]bpb=0.467768 t=353.3s + tc[1411/1893]bpb=0.466916 t=355.8s + tc[1421/1893]bpb=0.466017 t=358.3s + tc[1431/1893]bpb=0.465103 t=360.7s + tc[1441/1893]bpb=0.464285 t=363.2s + tc[1451/1893]bpb=0.463525 t=365.7s + tc[1461/1893]bpb=0.462612 t=368.2s + tc[1471/1893]bpb=0.461878 t=370.7s + tc[1481/1893]bpb=0.460946 t=373.2s + tc[1491/1893]bpb=0.460104 t=375.7s + tc[1501/1893]bpb=0.459320 t=378.2s + tc[1511/1893]bpb=0.458486 t=380.7s + tc[1521/1893]bpb=0.457643 t=383.2s + tc[1531/1893]bpb=0.456836 t=385.7s + tc[1541/1893]bpb=0.455967 t=388.2s + tc[1551/1893]bpb=0.455227 t=390.7s + tc[1561/1893]bpb=0.454478 t=393.2s + tc[1571/1893]bpb=0.453658 t=395.7s + tc[1581/1893]bpb=0.452937 t=398.2s + tc[1591/1893]bpb=0.452145 t=400.7s + tc[1601/1893]bpb=0.451431 t=403.2s + tc[1611/1893]bpb=0.450667 t=405.7s + tc[1621/1893]bpb=0.449867 t=408.2s + tc[1631/1893]bpb=0.449138 t=410.7s + tc[1641/1893]bpb=0.448406 t=413.2s + tc[1651/1893]bpb=0.447656 t=415.7s + tc[1661/1893]bpb=0.446913 t=418.6s + tc[1671/1893]bpb=0.446282 t=421.5s + tc[1681/1893]bpb=0.445589 t=424.3s + tc[1691/1893]bpb=0.444823 t=426.8s + tc[1701/1893]bpb=0.444113 t=429.3s + tc[1711/1893]bpb=0.443380 t=431.8s + tc[1721/1893]bpb=0.442675 t=434.3s + tc[1731/1893]bpb=0.442007 t=436.8s + tc[1741/1893]bpb=0.441351 t=439.3s + tc[1751/1893]bpb=0.440613 t=441.8s + tc[1761/1893]bpb=0.439999 t=444.3s + tc[1771/1893]bpb=0.439328 t=446.8s + tc[1781/1893]bpb=0.438745 t=449.3s + tc[1791/1893]bpb=0.438018 t=451.8s + tc[1801/1893]bpb=0.437383 t=454.3s + tc[1811/1893]bpb=0.436741 t=456.8s + tc[1821/1893]bpb=0.436099 t=459.3s + tc[1831/1893]bpb=0.435378 t=461.8s + tc[1841/1893]bpb=0.434725 t=464.3s + tc[1851/1893]bpb=0.434106 t=466.8s + tc[1861/1893]bpb=0.433425 t=469.3s + tc[1871/1893]bpb=0.432817 t=471.8s + tc[1881/1893]bpb=0.432184 t=474.3s + tc[1891/1893]bpb=0.431561 t=476.8s + tc[1893/1893]bpb=0.431484 t=477.1s +ttt:vl=0.727829 bpb=0.431062 t=477.1s +ttt vl:0.7278 bpb:0.4311 t:477619ms +ttt_x vl:0.72782858 bpb:0.43106222 diff --git a/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/submission.json b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/submission.json new file mode 100644 index 000000000..b77b3089e --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/submission.json @@ -0,0 +1,26 @@ +{ + "author": "naazimsnh02", + "github_id": "naazimsnh02", + "val_bpb": 0.431094, + "blurb": "Complementary Training + Backoff N-gram Mixer + Legal Score-First TTT", + "model_params": 26993766, + "artifact_bytes": 15962841, + "gpu": "8xH100 SXM 80GB", + "seeds": [ + {"seed": 1337, "val_bpb": 0.431107, "artifact_bytes": 15916181}, + {"seed": 42, "val_bpb": 0.431062, "artifact_bytes": 15962841}, + {"seed": 2024, "val_bpb": 0.431112, "artifact_bytes": 15958961} + ], + "techniques": [ + "Complementary training (COMPLEMENT_ALPHA=0.5)", + "BackoffNgramMixer (orders 2-10, 4M flat hash buckets)", + "Entropy-adaptive alpha (0.20 + 0.55*sigmoid(2*(H-3.0)))", + "AdamW TTT (lr=5e-4, 3 epochs, freeze 2 blocks, Polyak 0.998)", + "VRL (Value Residual Learning)", + "LeakyReLU(0.5)^2", + "XSA-4 (last 4 layers)", + "Int6 mixed quantization + lzma", + "Sliding window eval stride=64" + ], + "run_command": "VRL_ENABLED=1 LEAKY_RELU=1 TTT_ENABLED=1 TTT_OPTIMIZER=adamw TTT_LR=0.0005 TTT_EPOCHS=3 TTT_FREEZE_BLOCKS=2 TTT_TEMPERATURE=0.98 USE_HEDGE_MIXER=1 NGRAM_ORDER=10 NGRAM_BUCKETS=4194304 ALPHA_BASE=0.20 ALPHA_RANGE=0.55 ALPHA_CENTER=3.0 COMPLEMENT_ALPHA=0.5 SEED=42 torchrun --standalone --nproc_per_node=8 train_gpt.py" +} diff --git a/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/train_gpt.py b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/train_gpt.py new file mode 100644 index 000000000..0817b95dd --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_ComplementaryBackoff_NgramMixer/train_gpt.py @@ -0,0 +1,1900 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +_FA_VERSION = 0 +_fa_func = None +try: + from flash_attn_interface import flash_attn_func as _fa_func + _FA_VERSION = 3 +except ImportError: + try: + from flash_attn import flash_attn_func as _fa_func + _FA_VERSION = 2 + except ImportError: + _FA_VERSION = 0 + _fa_func = None +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + soft_round_qat = bool(int(os.environ.get("SOFT_ROUND_QAT", "1"))) + soft_round_temp_start = float(os.environ.get("SOFT_ROUND_TEMP_START", 1.0)) + soft_round_temp_end = float(os.environ.get("SOFT_ROUND_TEMP_END", 0.05)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + leaky_relu = bool(int(os.environ.get("LEAKY_RELU", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adamw") + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + polyak_decay = float(os.environ.get("POLYAK_DECAY", 0.998)) + use_polyak = bool(int(os.environ.get("USE_POLYAK", "1"))) + byte_weighted_ttt = bool(int(os.environ.get("BYTE_WEIGHTED_TTT", "1"))) + adaptive_lr = bool(int(os.environ.get("ADAPTIVE_LR", "1"))) + adaptive_lr_max = float(os.environ.get("ADAPTIVE_LR_MAX", 3.0)) + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) + checkpoint_path = os.environ.get("CHECKPOINT_PATH", "final_model.pt") + ttt_max_chunks = int(os.environ.get("TTT_MAX_CHUNKS", 0)) + skip_sliding_window = bool(int(os.environ.get("SKIP_SLIDING_WINDOW", "0"))) + use_hedge_mixer = bool(int(os.environ.get("USE_HEDGE_MIXER", "1"))) + mixer_eta = float(os.environ.get("MIXER_ETA", 0.1)) + mixer_min_tokens = int(os.environ.get("MIXER_MIN_TOKENS", 10000)) +class BackoffNgramMixer: + PRIMES = [36313, 27191, 51647, 81929, 131071, 174763, 233017] + def __init__(self, vocab_size: int, device: torch.device, num_buckets: int = 4_000_000, + max_order: int = 7, min_count: int = 2, min_tokens: int = 5000, + alpha_base: float = 0.05, alpha_range: float = 0.55, alpha_center: float = 4.0): + self.V = vocab_size + self.B = num_buckets + self.MASK = num_buckets - 1 if (num_buckets & (num_buckets - 1)) == 0 else None + self.max_order = max_order + self.min_count = min_count + self.min_tokens = min_tokens + self.device = device + self.tokens_seen = 0 + self.alpha_base = alpha_base + self.alpha_range = alpha_range + self.alpha_center = alpha_center + self.uni_counts = torch.zeros(vocab_size, device=device, dtype=torch.float32) + self.uni_total = 0.0 + self.ctx_counts = [] + self.full_counts = [] + for _ in range(max_order - 1): + self.ctx_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + self.full_counts.append(torch.zeros(num_buckets, device=device, dtype=torch.float32)) + def _bucket(self, h: Tensor) -> Tensor: + if self.MASK is not None: + return h & self.MASK + return h.abs() % self.B + def update(self, tokens: Tensor): + t = tokens.to(self.device).long() + n = t.numel() + self.tokens_seen += n + ones = torch.ones(n, device=self.device, dtype=torch.float32) + self.uni_counts.scatter_add_(0, t, ones) + self.uni_total += n + for order in range(2, self.max_order + 1): + if n < order: + continue + oi = order - 2 + nxt = t[order - 1:] + ctx_h = t[0:n - order + 1] * self.PRIMES[0] + for k in range(1, order - 1): + ctx_h = ctx_h ^ (t[k:n - order + 1 + k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (nxt * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + self.ctx_counts[oi].scatter_add_(0, ctx_key, ones[:n - order + 1]) + self.full_counts[oi].scatter_add_(0, full_key, ones[:n - order + 1]) + def score(self, logits: Tensor, x_batch: Tensor, y_batch: Tensor, + temperature: float = 1.0) -> Tensor: + bsz, slen, V = logits.shape + if temperature != 1.0: + logits = logits / temperature + log_probs_neural = F.log_softmax(logits.float(), dim=-1) + neural_p = log_probs_neural.gather(-1, y_batch.unsqueeze(-1)).squeeze(-1).exp() + neural_nll = -neural_p.clamp(min=1e-12).log() + if self.tokens_seen < self.min_tokens: + return neural_nll + ctx_stack = [x_batch] + for k in range(1, self.max_order - 1): + shifted = torch.zeros_like(x_batch) + if k < slen: + shifted[:, k:] = x_batch[:, :-k] + ctx_stack.append(shifted) + if self.uni_total > 0: + uni_p = (self.uni_counts[y_batch] + 0.5) / (self.uni_total + 0.5 * V) + ngram_p = uni_p + else: + ngram_p = torch.full((bsz, slen), 1.0 / V, device=self.device) + ngram_hit = torch.zeros(bsz, slen, device=self.device, dtype=torch.bool) + for order in range(self.max_order, 1, -1): + oi = order - 2 + cw = order - 1 + ctx_h = ctx_stack[cw - 1] * self.PRIMES[0] + for k in range(1, cw): + ctx_h = ctx_h ^ (ctx_stack[cw - 1 - k] * self.PRIMES[k % len(self.PRIMES)]) + ctx_key = self._bucket(ctx_h) + full_h = ctx_h ^ (y_batch * self.PRIMES[(order - 1) % len(self.PRIMES)]) + full_key = self._bucket(full_h) + ctx_c = self.ctx_counts[oi][ctx_key] + full_c = self.full_counts[oi][full_key] + valid = (ctx_c >= self.min_count) & (~ngram_hit) + min_pos = order - 2 + if min_pos > 0: + valid[:, :min_pos] = False + p = torch.where(valid, full_c.clamp(max=ctx_c) / ctx_c.clamp(min=1), torch.zeros_like(ctx_c)) + p = p.clamp(0, 1) + ngram_p = torch.where(valid, p, ngram_p) + ngram_hit = ngram_hit | valid + ngram_nll = -ngram_p.clamp(min=1e-12).log() + probs_neural = log_probs_neural.exp() + entropy = -(probs_neural * log_probs_neural).sum(dim=-1) + alpha = self.alpha_base + self.alpha_range * torch.sigmoid( + 2.0 * (entropy - self.alpha_center)) + mixed_p = (1.0 - alpha) * neural_p + alpha * ngram_p + return -mixed_p.clamp(min=1e-12).log() +class TrainNgramTracker: + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) +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:{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"val too short for {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 too small; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,vrl_scales", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + _soft_round_qat: bool = True + _soft_round_temp: float = 1.0 + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._soft_round_qat: + w32 = self.weight.float() + row_max = w32.detach().abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_s = w32 / scale[:, None] + residual = w_s - w_s.detach().round() + temp = CastedLinear._soft_round_temp + w_soft = w_s.detach().round() + 0.5 * torch.tanh(residual / temp) + w = (w_soft.clamp(-32, 31) * scale[:, None]).to(x.dtype) + else: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim%num_heads!=0") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads%num_kv_heads!=0") + 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("odd head_dim") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _FA_VERSION == 3: + y = _fa_func(q, k, v, causal=True) + elif _FA_VERSION == 2: + y = _fa_func(q.bfloat16(), k.bfloat16(), v.bfloat16(), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + is_causal=True, enable_gqa=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int, leaky: bool = False): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + self._neg_slope = 0.5 if leaky else 0.0 + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), self._neg_slope) + return self.proj(x.square()) +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + **kwargs, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=kwargs.get("gated_attention", False)) + self.mlp = MLP(dim, mlp_mult, leaky=kwargs.get("leaky", False)) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + vrl_enabled: bool = False, + leaky_relu: bool = False, + gated_attention: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + if logit_softcap <= 0.0: + raise ValueError(f"softcap<=0:{logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + leaky=leaky_relu, + gated_attention=gated_attention, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + if self.vrl_enabled: + self.vrl_scales = nn.ParameterList( + [nn.Parameter(torch.zeros(1, dtype=torch.float32)) for _ in range(num_layers - 1)] + ) + else: + self.vrl_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("need lm_head") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + if self.vrl_enabled: + mix0 = self.blocks[0].resid_mix.to(dtype=x0.dtype) + x_in_0 = mix0[0][None, None, :] * x0 + mix0[1][None, None, :] * x0 + n0 = F.rms_norm(x_in_0, (x_in_0.size(-1),)) * self.blocks[0].ln_scale_factor + v0_raw = self.blocks[0].attn.c_v(n0) + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.vrl_enabled and i > 0: + vr = v0_raw * self.vrl_scales[i - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[i](x, x0, v_embed=v_extra) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + if self.vrl_enabled: + vr = v0_raw * self.vrl_scales[bi - 1].to(dtype=v0_raw.dtype) + v_extra = (ve + vr) if ve is not None else vr + else: + v_extra = ve + x = self.blocks[bi](x, x0, v_embed=v_extra) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) +def eval_val_sliding_ttt( + args, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + if args.ttt_max_chunks > 0: + num_chunks = min(num_chunks, args.ttt_max_chunks) + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + if ci < num_chunks: + chunk_windows[ci].append(ws) + log0(f"ttt:c={num_chunks} ct={ttt_chunk} w={len(window_starts)} s={stride} lr={args.ttt_lr} ep={args.ttt_epochs} fb={args.ttt_freeze_blocks} o={args.ttt_optimizer} pk={args.use_polyak}({args.polyak_decay}) bw={args.byte_weighted_ttt} alr={args.adaptive_lr}({args.adaptive_lr_max}) t={args.ttt_temperature}") + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + num_blocks = len(base_model.blocks) + unfrozen_block_start = max(0, num_blocks - args.ttt_freeze_blocks) if args.ttt_freeze_blocks > 0 else 0 + ttt_params = [] + for name, p in base_model.named_parameters(): + unfreeze = False + if args.ttt_freeze_blocks <= 0: + unfreeze = True + elif "norm" in name or "scale" in name or "lm_head" in name or "tok_emb" in name: + unfreeze = True + else: + for bi in range(unfrozen_block_start, num_blocks): + if f"blocks.{bi}." in name: + unfreeze = True + break + if unfreeze: + p.requires_grad_(True) + ttt_params.append(p) + else: + p.requires_grad_(False) + log0(f"ttt:uf={sum(p.numel() for p in ttt_params)} f={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + if args.ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + polyak_state: dict[str, Tensor] | None = None + if args.use_polyak: + polyak_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + mixer: BackoffNgramMixer | None = None + if args.use_hedge_mixer: + ngram_order = int(os.environ.get("NGRAM_ORDER", "7")) + ngram_buckets = int(os.environ.get("NGRAM_BUCKETS", "4000000")) + alpha_base = float(os.environ.get("ALPHA_BASE", "0.05")) + alpha_range = float(os.environ.get("ALPHA_RANGE", "0.55")) + alpha_center = float(os.environ.get("ALPHA_CENTER", "4.0")) + min_count = int(os.environ.get("MIN_COUNT", "2")) + mixer = BackoffNgramMixer(args.vocab_size, device, num_buckets=ngram_buckets, + max_order=ngram_order, min_count=min_count, + min_tokens=args.mixer_min_tokens, + alpha_base=alpha_base, alpha_range=alpha_range, + alpha_center=alpha_center) + mem_mb = ngram_buckets * 4 * 2 * (ngram_order - 1) / 1e6 + log0(f"bo:o={ngram_order} b={ngram_buckets} m={mem_mb:.0f}M a={alpha_base}+{alpha_range}*s(H-{alpha_center}) mc={min_count}") + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + raw_state: dict[str, Tensor] | None = None + if polyak_state is not None: + raw_state = {n: p.data.detach().clone() for n, p in base_model.named_parameters() if p.requires_grad} + for n, p in base_model.named_parameters(): + if n in polyak_state: + p.data.copy_(polyak_state[n]) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + if mixer is not None and mixer.tokens_seen >= mixer.min_tokens: + nll = mixer.score(logits, x_batch, y_batch, args.ttt_temperature) + else: + if args.ttt_temperature != 1.0: + logits = logits / args.ttt_temperature + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if mixer is not None: + chunk_tokens = val_tokens[chunk_start:chunk_end].to(device) + mixer.update(chunk_tokens) + if raw_state is not None: + for n, p in base_model.named_parameters(): + if n in raw_state: + p.data.copy_(raw_state[n]) + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if args.adaptive_lr: + progress = min(ci / (num_chunks * 0.3), 1.0) + lr_mult = 1.0 + (args.adaptive_lr_max - 1.0) * progress + cos_lr = cos_lr * lr_mult + for pg in optimizer.param_groups: + pg['lr'] = cos_lr + distributed = dist.is_available() and dist.is_initialized() + my_seq_s = (chunk_seqs * rank) // world_size if distributed else 0 + my_seq_e = (chunk_seqs * (rank + 1)) // world_size if distributed else chunk_seqs + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_t = base_model.forward_logits(x) + if args.byte_weighted_ttt: + per_tok_nll = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), reduction="none", + ) + byte_weights = base_bytes_lut[y.reshape(-1)].float() + byte_weights = byte_weights / byte_weights.mean().clamp(min=1e-6) + loss = (per_tok_nll * byte_weights).mean() + else: + loss = F.cross_entropy( + logits_t.reshape(-1, logits_t.size(-1)).float(), + y.reshape(-1), + ) + loss.backward() + if distributed and world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + if polyak_state is not None: + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if n in polyak_state: + polyak_state[n].lerp_(p.data, 1.0 - args.polyak_decay) + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" tc[{ci+1}/{num_chunks}]bpb={rbpb:.6f} t={elapsed:.1f}s") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + log0(f"ttt:vl={val_loss:.6f} bpb={val_bpb:.6f} t={time.perf_counter()-t0:.1f}s") + return val_loss, val_bpb +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out +def main() -> None: + global zeropower_via_newtonschulz5 + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"bad WORLD_SIZE:{world_size}") + if 8 % world_size != 0: + raise ValueError(f"8%WORLD_SIZE={world_size}!=0") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("no CUDA") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + _gpu_name = torch.cuda.get_device_name(0) + _is_high_end = "H100" in _gpu_name or "A100" in _gpu_name + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + if _is_high_end: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(True) + enable_flash_sdp(True) + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("="*60,console=False) + log0(f"py:{sys.version}",console=False) + log0(f"pt:{torch.__version__}",console=False) + log0(subprocess.run(["nvidia-smi"],stdout=subprocess.PIPE,stderr=subprocess.PIPE,text=True,check=False).stdout,console=False) + log0("="*60,console=False) + log0(f"fa:{_FA_VERSION} gpu:{_gpu_name} he:{_is_high_end}") + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"need .model:{args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"vocab mismatch:{args.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"bpb:sp={args.tokenizer_path}") + log0(f"train:{dataset_dir.name} shards:{actual_train_files}") + log0(f"val:{args.val_files} n:{val_tokens.numel()-1}") + CastedLinear._qat_enabled = args.qat_enabled + CastedLinear._soft_round_qat = args.soft_round_qat + CastedLinear._soft_round_temp = args.soft_round_temp_start + qat_start_step = 0 if args.qat_enabled else -1 + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, + leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + if complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=complement_alpha) + base_model._ngram_tracker = tracker + log0(f"compl:{complement_alpha}") + else: + base_model._ngram_tracker = None + if distributed: + torch._dynamo.config.optimize_ddp = False + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + if base_model.vrl_enabled: + for s in base_model.vrl_scales: + scalar_params.append(s) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"p:{n_params}") + log0(f"mtp:{args.mtp_num_heads} w:{args.mtp_loss_weight} p:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"xsa:{args.xsa_last_n} l:{xsa_layers}") + log0(f"ws:{world_size} ga:{grad_accum_steps}") + log0(f"sdp:{_is_high_end}") + log0(f"attn:h={args.num_heads} kv={args.num_kv_heads}") + log0(f"vrl:{args.vrl_enabled} lrelu:{args.leaky_relu} ttt:{args.ttt_enabled}") + log0(f"tie:{args.tie_embeddings} elr:{token_lr} hlr:{args.head_lr if base_model.lm_head is not None else 0.0} mlr:{args.matrix_lr} slr:{args.scalar_lr}") + log0(f"tbt:{args.train_batch_tokens} tsl:{args.train_seq_len} it:{args.iterations} wu:{args.warmup_steps} mws:{args.max_wallclock_seconds:.3f}") + log0(f"s:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0 and not args.eval_only: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"wu:{warmup_step+1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.eval_only: + log0(f"eval:load {args.checkpoint_path}") + ckpt_state = torch.load(args.checkpoint_path, map_location="cpu") + base_model.load_state_dict(ckpt_state, strict=True) + log0(f"eval:loaded {sum(p.numel() for p in base_model.parameters())}p") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + log0(f"eval:qsize:{len(quant_blob)}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = effective_eval_seq_len + if not args.skip_sliding_window and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0(f"eval:sw bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + elif args.skip_sliding_window: + log0("eval:skip_sw") + if args.ttt_enabled: + log0(f"eval:ttt lr={args.ttt_lr} ep={args.ttt_epochs} c={args.ttt_chunk_tokens} fb={args.ttt_freeze_blocks}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"eval:ttt bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + if distributed: + dist.destroy_process_group() + return + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0(f"s:{step}/{args.iterations} vl:{val_loss:.4f} bpb:{val_bpb:.4f} tt:{training_time_ms:.0f}ms sa:{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"stop tt:{training_time_ms:.0f}ms s:{step}/{args.iterations}") + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + qat_start_step = step + log0(f"qat:{step} s:{scale:.4f}") + if CastedLinear._qat_enabled and CastedLinear._soft_round_qat and qat_start_step >= 0: + qat_total = max(args.iterations - qat_start_step, 1) + qat_progress = min((step - qat_start_step) / qat_total, 1.0) + log_start = math.log(args.soft_round_temp_start) + log_end = math.log(args.soft_round_temp_end) + CastedLinear._soft_round_temp = math.exp(log_start + qat_progress * (log_end - log_start)) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + 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:{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"s:{step}/{args.iterations} tl:{train_loss.item():.4f} tt:{approx_training_time_ms:.0f}ms sa:{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"mem:{torch.cuda.max_memory_allocated()//1024//1024}M R:{torch.cuda.max_memory_reserved()//1024//1024}M") + log0("ema:apply") + 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"diag vl:{diag_val_loss:.4f} bpb:{diag_val_bpb:.4f} t:{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"excl_mtp:{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"model:{model_bytes}B") + log0(f"code:{code_bytes}B") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"q:{quant_file_bytes}B") + log0(f"total:{quant_file_bytes+code_bytes}B") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_raw_disk = lzma.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + vrl_enabled=args.vrl_enabled, leaky_relu=args.leaky_relu, + gated_attention=args.gated_attention, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0(f"q_rt vl:{q_val_loss:.4f} bpb:{q_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_qeval):.0f}ms") + log0(f"q_rt_x vl:{q_val_loss:.8f} 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"q_sw vl:{sw_val_loss:.4f} bpb:{sw_val_bpb:.4f} s:{args.eval_stride} t:{1000.0*(time.perf_counter()-t_slide):.0f}ms") + log0(f"q_sw_x vl:{sw_val_loss:.8f} bpb:{sw_val_bpb:.8f}") + log0(f"q8_x vl:{sw_val_loss:.8f} 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"q_s64 vl:{sw64_val_loss:.4f} bpb:{sw64_val_bpb:.4f} s:64 t:{1000.0*(time.perf_counter()-t_slide64):.0f}ms") + log0(f"q_s64_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + log0(f"q8_x vl:{sw64_val_loss:.8f} bpb:{sw64_val_bpb:.8f}") + if args.ttt_enabled: + log0("ttt:start") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.ttt_batch_seqs, log0=log0, + ) + torch.cuda.synchronize() + log0(f"ttt vl:{ttt_val_loss:.4f} bpb:{ttt_val_bpb:.4f} t:{1000.0*(time.perf_counter()-t_ttt):.0f}ms") + log0(f"ttt_x vl:{ttt_val_loss:.8f} bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() \ No newline at end of file