diff --git a/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/README.md b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/README.md new file mode 100644 index 0000000000..dc74052887 --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/README.md @@ -0,0 +1,40 @@ +11L 512d int6+zstd, 7-gram cache, 100ep cosine TTT, GPTQ. + +## setup + +```bash +pip install -r requirements.txt +pip install flash_attn_3 --find-links https://windreamer.github.io/flash-attention3-wheels/cu128_torch291 +python3 data/cached_challenge_fineweb.py --variant sp1024 +``` + +## run + +```bash +SEED=1337 NUM_LAYERS=11 MODEL_DIM=512 MLP_MULT=3.0 \ +LEAKY_RELU=0.5 XSA_LAST_N=11 VRL_ENABLED=1 GATED_ATTN=1 \ +BIGRAM_VOCAB_SIZE=4096 BIGRAM_DIM=128 \ +EMA_ENABLED=1 EMA_DECAY=0.997 SWA_ENABLED=0 \ +ROPE_DIMS=16 LN_SCALE=1 \ +LATE_QAT=1 QAT_ENABLED=0 \ +TTT_ENABLED=1 TTT_LR=0.001 TTT_EPOCHS=100 TTT_COSINE=1 TTT_ADAMW=1 TTT_PER_LAYER_LR=1 TTT_ETA_MIN_RATIO=0.01 \ +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=3000 \ +ITERATIONS=9000 MAX_WALLCLOCK_SECONDS=600 EVAL_STRIDE=64 \ +PRUNE_PCT=0.07 GPTQ_ENABLED=1 GPTQ_BATCHES=256 \ +NGRAM_ALPHA=0.40 NGRAM_ORDER=7 NGRAM_BUCKETS=4000000 NGRAM_ADAPTIVE=1 NGRAM_CONF_SCALE=1 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## toggles + +- `GPTQ_ENABLED=0` - skip hessian-aware quant, use naive per-row int6 +- `NGRAM_ALPHA=0` - disable n-gram eval cache +- `NGRAM_CONF_SCALE=0` - disable count-confidence weighting +- `LEAKY_RELU=0` - standard ReLU (before squaring) +- `XSA_LAST_N=0` - no exclusive self-attention +- `VRL_ENABLED=0` - no value residual +- `GATED_ATTN=0` - no per-head attention gate +- `TTT_ENABLED=0` - skip test-time training diff --git a/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/requirements.txt b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/requirements.txt new file mode 100644 index 0000000000..864700d2b3 --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/requirements.txt @@ -0,0 +1 @@ +zstandard diff --git a/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/submission.json b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/submission.json new file mode 100644 index 0000000000..02f139d709 --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/submission.json @@ -0,0 +1,14 @@ +{ + "author": "bopmite", + "github_id": "bopmite", + "val_bpb": null, + "val_loss": null, + "bytes_model": null, + "bytes_code": 83461, + "bytes_total": null, + "architecture": "11L 512d 3xMLP LeakyReLU^2 XSA-all VRL GatedAttn GPTQ int6+zstd 7-gram 100ep-cosine-TTT", + "tokenizer": "sp1024", + "training_time_minutes": 10, + "gpu_config": "8xH100 SXM", + "seed": 1337 +} diff --git a/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/train_gpt.py b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/train_gpt.py new file mode 100644 index 0000000000..395bd62fce --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_OnlineLogitBias_11L_Int6/train_gpt.py @@ -0,0 +1,2095 @@ +"""Parameter Golf submission: 11L 512d int6+zstd.""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func +except ImportError: + try: + from flash_attn import flash_attn_func as flash_attn_3_func + except ImportError: + flash_attn_3_func = None + +_LEAKY_RELU_NEG_SLOPE = float(os.environ.get("LEAKY_RELU", "0")) + +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", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + 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", 9)) + 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", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + 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", 200)) + 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"))) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 0)) + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + rope_dims = int(os.environ.get("ROPE_DIMS", 0)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "0"))) + late_qat = bool(int(os.environ.get("LATE_QAT", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 4096)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + vrl_enabled = bool(int(os.environ.get("VRL_ENABLED", "0"))) + gated_attn = bool(int(os.environ.get("GATED_ATTN", "0"))) + + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.001)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 100)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,vr_lambda,attn_gate", + ).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 + _qat_clip_pct: float = 0.9999 + _qat_alpha: float = 3.0 + + 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_det = self.weight.float() + row_clip = torch.quantile(w32_det.abs(), CastedLinear._qat_clip_pct, dim=1) + scale = (row_clip / 31.0).clamp_min(1.0 / 31.0) + w32 = self.weight.float() + y = w32 / scale[:, None] + alpha = CastedLinear._qat_alpha + y_floor = torch.floor(y).detach() + frac = y - y_floor + tanh_half = math.tanh(alpha * 0.5) + soft_frac = 0.5 * torch.tanh(alpha * (frac - 0.5)) / tanh_half + 0.5 + y_soft = y_floor + soft_frac + w = (torch.clamp(y_soft, -31, 31) * scale[:, None]).to(x.dtype) + 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.rope_dims = rope_dims if rope_dims > 0 else dim + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + rd = self.rope_dims + inv_freq = 1.0 / (base ** (torch.arange(0, rd, 2, dtype=torch.float32) / rd)) + 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) -> Tensor: + rd = cos.size(-1) * 2 + if rd < x.size(-1): + x_rope, x_pass = x[..., :rd], x[..., rd:] + half = rd // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rot = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rot, 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, + rope_dims: int = 0, + layer_idx: int = 0, + vrl_enabled: bool = False, + gated_attn: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = rope_dims + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.use_xsa = False + self.vrl_enabled = vrl_enabled + self.vr_lambda = nn.Parameter(torch.tensor([0.9, 0.1], dtype=torch.float32)) if (vrl_enabled and layer_idx > 0) else None + self.attn_gate = nn.Parameter(torch.full((num_heads,), 4.0, dtype=torch.float32)) if gated_attn else None + + 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, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + 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).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.vrl_enabled else None + if self.vr_lambda is not None and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v + lam[1] * v0 + 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) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + fa_dtype = torch.bfloat16 + if flash_attn_3_func is not None: + y = flash_attn_3_func(q.to(fa_dtype), k.to(fa_dtype), v.to(fa_dtype), causal=True) + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2).to(fa_dtype), k.transpose(1, 2).to(fa_dtype), v.transpose(1, 2).to(fa_dtype), + is_causal=True, enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.attn_gate is not None: + gate = torch.sigmoid(self.attn_gate.to(dtype=y.dtype))[None, None, :, None] + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v + + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = self.fc(x) + if _LEAKY_RELU_NEG_SLOPE > 0: + x = F.leaky_relu(x, negative_slope=_LEAKY_RELU_NEG_SLOPE) + else: + x = torch.relu(x) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + rope_dims: int = 0, + layer_idx: int = 0, + ln_scale: bool = False, + vrl_enabled: bool = False, + gated_attn: 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, + rope_dims=rope_dims, layer_idx=layer_idx, + vrl_enabled=vrl_enabled, gated_attn=gated_attn) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x: Tensor, x0: Tensor, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + s = self.ln_scale_factor + attn_out, raw_v = self.attn(self.attn_norm(x) * s, v0=v0) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x) * s) + return x, 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, + vrl_enabled: bool = False, + gated_attn: bool = False, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.vrl_enabled = vrl_enabled + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + rope_dims=rope_dims, + layer_idx=i, + ln_scale=ln_scale, + vrl_enabled=vrl_enabled, + gated_attn=gated_attn, + ) + for i in range(num_layers) + ] + ) + 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 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] = [] + v0: Tensor | None = None + + for i in range(self.num_encoder_layers): + x, raw_v = self.blocks[i](x, x0, v0=v0) + if i == 0 and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, _ = self.blocks[self.num_encoder_layers + i](x, x0, v0=v0) + + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + + return main_loss + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + v0: Tensor | None = None + for i in range(self.num_encoder_layers): + x, raw_v = self.blocks[i](x, x0, v0=v0) + if i == 0 and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, _ = self.blocks[self.num_encoder_layers + i](x, x0, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, + ngram_alpha: float = 0.0, + ngram_order: int = 7, + ngram_buckets: int = 4_000_000, +) -> 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) + _PRIMES = [36313, 27191, 48611, 59369, 73721, 87671, 91813] + _knn_enabled = bool(int(os.environ.get("KNN_LM", "0"))) + _knn_k = int(os.environ.get("KNN_K", "32")) + _knn_temp = float(os.environ.get("KNN_TEMP", "10.0")) + _knn_lambda = float(os.environ.get("KNN_LAMBDA", "0.1")) + _knn_buf_size = int(os.environ.get("KNN_BUF", "100000")) + if _knn_enabled: + _hdim = base_model.blocks[0].attn.c_q.weight.shape[1] + _knn_keys = torch.zeros(_knn_buf_size, _hdim, device=device, dtype=torch.float16) + _knn_vals = torch.zeros(_knn_buf_size, device=device, dtype=torch.long) + _knn_ptr = 0 + _knn_fill = 0 + _hidden_capture = [None] + def _capture_hook(module, inp, out): + _hidden_capture[0] = out.detach() + _knn_hook = base_model.final_norm.register_forward_hook(_capture_hook) + use_ngram = ngram_alpha > 0 + if use_ngram: + ng_ctx = [torch.zeros(ngram_buckets, dtype=torch.int32, device=device) for _ in range(ngram_order - 1)] + ng_jnt = [torch.zeros(ngram_buckets, dtype=torch.int32, device=device) for _ in range(ngram_order - 1)] + val_long = val_tokens.to(dtype=torch.int64, device=device) + _ng_bloom = bool(int(os.environ.get("NGRAM_BLOOM", "0"))) + _BLOOM_PRIMES = [104729, 131071, 174763] # independent of _PRIMES + _bloom_bits = ngram_buckets * 8 # 8 bits per count bucket + if _ng_bloom: + bloom = [torch.zeros(_bloom_bits, dtype=torch.uint8, device=device) for _ in range(ngram_order - 1)] + _apm_enabled = bool(int(os.environ.get("APM_ENABLED", "0"))) + _apm_bins = int(os.environ.get("APM_BINS", "64")) + _apm_decay = float(os.environ.get("APM_DECAY", "0.995")) + if _apm_enabled: + _apm_num = torch.zeros(1024, _apm_bins, device=device) # hits + _apm_den = torch.ones(1024, _apm_bins, device=device) # total, init 1 for Laplace + _ng_learned = bool(int(os.environ.get("NGRAM_LEARNED_MIX", "0"))) + if _ng_learned: + _mix_w = torch.tensor([-2.0, 0.5, 0.3, 0.5], device=device) + _mix_lr = float(os.environ.get("NGRAM_MIX_LR", "0.01")) + _mix_mom = torch.zeros(4, device=device) + _ng_logistic = bool(int(os.environ.get("NGRAM_LOGISTIC", "0"))) + if _ng_logistic: + _log_w = torch.tensor([0.8, 0.2], device=device) # [model_weight, ngram_weight] + _log_lr = float(os.environ.get("NGRAM_LOG_LR", "0.005")) + + 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) + + if not use_ngram: + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + else: + nll = torch.zeros(bsz, seq_len, dtype=torch.float64, device=device) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + + if use_ngram: + scored_logits = logits[i, s:wlen].float() + scored_tgt = y_batch[i, s:wlen] + p_model = F.softmax(scored_logits, dim=-1) + p_model_t = p_model.gather(1, scored_tgt.unsqueeze(1)).squeeze(1) + ns = scored_tgt.shape[0] + p_ng = torch.zeros(ns, device=device) + ng_cc = torch.zeros(ns, device=device) # context count for matched order + for oi in range(ngram_order - 2, -1, -1): + order = oi + 2 + global_pos = torch.arange(ws + s + 1, ws + s + 1 + ns, device=device) + h = torch.zeros(ns, dtype=torch.long, device=device) + valid = torch.ones(ns, dtype=torch.bool, device=device) + for k in range(order - 1): + idx = global_pos - (order - 1) + k + valid &= (idx >= 0) & (idx < val_long.numel()) + tok = torch.where(valid, val_long[idx.clamp(0, val_long.numel()-1)], torch.zeros_like(h)) + h = (h * _PRIMES[k % len(_PRIMES)] + tok) % ngram_buckets + jh = (h * 27191 + scored_tgt.long()) % ngram_buckets + cc = ng_ctx[oi][h]; jc = ng_jnt[oi][jh] + has = valid & (cc >= 2) & (p_ng == 0) + if _ng_bloom and has.any(): + bloom_ok = torch.ones(ns, dtype=torch.bool, device=device) + for bp in _BLOOM_PRIMES: + bh = torch.zeros(ns, dtype=torch.long, device=device) + for k in range(order - 1): + idx = global_pos - (order - 1) + k + tok = torch.where(valid, val_long[idx.clamp(0, val_long.numel()-1)], torch.zeros_like(bh)) + bh = (bh * bp + tok) % _bloom_bits + bloom_ok &= (bloom[oi][bh] > 0) + has = has & bloom_ok + if has.any(): + p_ng[has] = jc[has].float() / cc[has].float() + ng_cc[has] = cc[has].float() + entropy = -(p_model * torch.log(p_model + 1e-10)).sum(dim=-1) + _ng_adaptive = bool(int(os.environ.get("NGRAM_ADAPTIVE", "1"))) + if _ng_learned: + feats = torch.stack([torch.ones(ns, device=device), + entropy, + torch.log(ng_cc + 1.0), + p_ng], dim=1) # (ns, 4) + alpha_raw = feats @ _mix_w + alpha = torch.sigmoid(alpha_raw) + alpha = torch.where(p_ng > 0, alpha, torch.zeros_like(alpha)) + elif _ng_adaptive: + _ng_alpha_lo = float(os.environ.get("NGRAM_ALPHA_LO", "0.05")) + _ng_alpha_hi = float(os.environ.get("NGRAM_ALPHA_HI", "0.55")) + alpha = _ng_alpha_lo + (_ng_alpha_hi - _ng_alpha_lo) * torch.sigmoid(2.0 * (entropy - 4.0)) + _ng_conf_scale = bool(int(os.environ.get("NGRAM_CONF_SCALE", "1"))) + if _ng_conf_scale: + count_conf = torch.clamp(ng_cc / 8.0, 0.2, 1.0) + alpha = alpha * count_conf + alpha = torch.where(p_ng > 0, alpha, torch.zeros_like(alpha)) + else: + alpha = torch.where(p_ng > 0, torch.full_like(p_ng, ngram_alpha), torch.zeros_like(p_ng)) + if _ng_logistic and (p_ng > 0).any(): + _eps = 1e-7 + pm_c = p_model_t.clamp(_eps, 1 - _eps) + pn_c = p_ng.clamp(_eps, 1 - _eps) + s_m = torch.log(pm_c / (1 - pm_c)) + s_n = torch.log(pn_c / (1 - pn_c)) + has_ng = p_ng > 0 + logit_mix = _log_w[0] * s_m + torch.where(has_ng, _log_w[1] * s_n, torch.zeros_like(s_n)) + p_mixed = torch.sigmoid(logit_mix) + nll[i, s:wlen] = -torch.log(p_mixed.clamp(min=1e-10)).to(torch.float64) + # PAQ-style weight update: w += lr * stretch(p) * (y - p_mixed) + err = (1.0 - p_mixed) # target=1 for correct token + if has_ng.any(): + _log_w[0] += _log_lr * (s_m[has_ng] * err[has_ng]).mean() + _log_w[1] += _log_lr * (s_n[has_ng] * err[has_ng]).mean() + else: + p_mixed = (1 - alpha) * p_model_t + alpha * p_ng + nll[i, s:wlen] = -torch.log(p_mixed.clamp(min=1e-10)).to(torch.float64) + if _ng_learned and (p_ng > 0).any(): + mask = p_ng > 0 + dnll_dalpha = -(p_ng[mask] - p_model_t[mask]) / p_mixed[mask] + dalpha_draw = alpha[mask] * (1 - alpha[mask]) + grad_w = (dnll_dalpha * dalpha_draw).unsqueeze(1) * feats[mask] + g = grad_w.mean(dim=0) + _mix_mom.mul_(0.9).add_(g, alpha=0.1) + _mix_w -= _mix_lr * _mix_mom + if _apm_enabled: + prev_tok = x_batch[i, s:wlen].long() + cur_p = torch.exp(-nll[i, s:wlen].float()).clamp(1e-7, 1 - 1e-7) + logit_p = torch.log(cur_p / (1 - cur_p)) + bin_idx = ((logit_p + 8.0) / 16.0 * _apm_bins).long().clamp(0, _apm_bins - 1) + apm_p = _apm_num[prev_tok, bin_idx] / _apm_den[prev_tok, bin_idx] + corrected = 0.7 * cur_p + 0.3 * apm_p.clamp(0.01, 0.99) + nll[i, s:wlen] = -torch.log(corrected.clamp(min=1e-10)).to(torch.float64) + _apm_num.mul_(_apm_decay) + _apm_den.mul_(_apm_decay) + _apm_num[prev_tok, bin_idx] += 1.0 # target=correct token, so hit=1 + _apm_den[prev_tok, bin_idx] += 1.0 + for oi, order in enumerate(range(2, ngram_order + 1)): + global_pos = torch.arange(ws + s + 1, ws + s + 1 + ns, device=device) + h = torch.zeros(ns, dtype=torch.long, device=device) + vld = torch.ones(ns, dtype=torch.bool, device=device) + for k in range(order - 1): + idx = global_pos - (order - 1) + k + vld &= (idx >= 0) & (idx < val_long.numel()) + tok = torch.where(vld, val_long[idx.clamp(0, val_long.numel()-1)], torch.zeros_like(h)) + h = (h * _PRIMES[k % len(_PRIMES)] + tok) % ngram_buckets + jh = (h * 27191 + scored_tgt.long()) % ngram_buckets + m = vld + if m.any(): + ng_ctx[oi].scatter_add_(0, h[m], torch.ones(m.sum(), dtype=torch.int32, device=device)) + ng_jnt[oi].scatter_add_(0, jh[m], torch.ones(m.sum(), dtype=torch.int32, device=device)) + if _ng_bloom: + for bp in _BLOOM_PRIMES: + bh = torch.zeros(ns, dtype=torch.long, device=device) + for k in range(order - 1): + idx = global_pos - (order - 1) + k + tok = torch.where(vld, val_long[idx.clamp(0, val_long.numel()-1)], torch.zeros_like(bh)) + bh = (bh * bp + tok) % _bloom_bits + bloom[oi][bh[m]] = 1 + + if _knn_enabled and _knn_fill >= _knn_k and _hidden_capture[0] is not None: + h_scored = _hidden_capture[0][i, s:wlen].float() # (ns, hdim) + ns_knn = h_scored.shape[0] + buf_len = min(_knn_fill, _knn_buf_size) + keys_buf = _knn_keys[:buf_len].float() + dists = torch.cdist(h_scored, keys_buf) # (ns_knn, buf_len) + topk_d, topk_i = dists.topk(_knn_k, dim=1, largest=False) + nn_tokens = _knn_vals[topk_i] # (ns_knn, k) + per_token_temp = topk_d[:, 0:1].clamp(min=1.0) * (_knn_temp / 10.0) + nn_weights = F.softmax(-topk_d / per_token_temp, dim=1) # (ns_knn, k) + tgt_knn = y_batch[i, s:wlen] + p_knn = torch.zeros(ns_knn, device=device) + for ki in range(_knn_k): + p_knn += nn_weights[:, ki] * (nn_tokens[:, ki] == tgt_knn).float() + has_knn = p_knn > 0 + if has_knn.any(): + cur_p = torch.exp(-nll[i, s:wlen].float()) + mixed_p = (1 - _knn_lambda) * cur_p + _knn_lambda * p_knn + nll[i, s:wlen] = torch.where(has_knn, + -torch.log(mixed_p.clamp(min=1e-10)).to(torch.float64), + nll[i, s:wlen]) + end_ptr = _knn_ptr + ns_knn + if end_ptr <= _knn_buf_size: + _knn_keys[_knn_ptr:end_ptr] = h_scored.half() + _knn_vals[_knn_ptr:end_ptr] = tgt_knn + else: + first = _knn_buf_size - _knn_ptr + _knn_keys[_knn_ptr:] = h_scored[:first].half() + _knn_vals[_knn_ptr:] = tgt_knn[:first] + _knn_keys[:ns_knn - first] = h_scored[first:].half() + _knn_vals[:ns_knn - first] = tgt_knn[first:] + _knn_ptr = end_ptr % _knn_buf_size + _knn_fill += ns_knn + + 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 _knn_enabled: + _knn_hook.remove() + + 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 LoRALinear(nn.Module): + def __init__(self, base: nn.Linear, rank: int = 4): + super().__init__() + self.base = base + self.lora_a = nn.Parameter(torch.randn(base.in_features, rank, device=base.weight.device, dtype=torch.float32) * 0.01) + self.lora_b = nn.Parameter(torch.zeros(rank, base.out_features, device=base.weight.device, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + base_out = self.base(x) + lora_out = (x.float() @ self.lora_a) @ self.lora_b + return base_out + lora_out.to(base_out.dtype) + + +def inject_lora(model: nn.Module, rank: int = 4, target_names: tuple = (".c_q", ".c_k", ".c_v", ".proj")): + lora_params = [] + for name, module in list(model.named_modules()): + if isinstance(module, CastedLinear) and any(t in name for t in target_names): + parent_name, attr_name = name.rsplit(".", 1) + parent = dict(model.named_modules())[parent_name] + module.weight.requires_grad_(False) + if module.bias is not None: + module.bias.requires_grad_(False) + lora = LoRALinear(module, rank=rank) + setattr(parent, attr_name, lora) + lora_params.extend([lora.lora_a, lora.lora_b]) + return lora_params + + +def remove_lora(model: nn.Module): + for name, module in list(model.named_modules()): + if isinstance(module, LoRALinear): + parent_name, attr_name = name.rsplit(".", 1) + parent = dict(model.named_modules())[parent_name] + setattr(parent, attr_name, module.base) + module.base.weight.requires_grad_(True) + + +def ttt_adapt(args: Hyperparameters, base_model: nn.Module, device: torch.device, + val_tokens: Tensor, rank: int = 0, world_size: int = 1, + log_fn=None) -> None: + seq_len = args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + batch_seqs = args.ttt_batch_seqs + + use_lora = bool(int(os.environ.get("TTT_LORA", "0"))) + lora_rank = int(os.environ.get("TTT_LORA_RANK", "4")) + lora_params = [] + + if use_lora: + for p in base_model.parameters(): + p.requires_grad_(False) + lora_params = inject_lora(base_model, rank=lora_rank) + if log_fn: + n_lora = sum(p.numel() for p in lora_params) + log_fn(f"ttt_lora:injected rank={lora_rank} params={n_lora}") + ttt_params = lora_params + param_groups = [{'params': ttt_params, 'lr': args.ttt_lr}] + else: + frozen_params: set[int] = set() + if args.ttt_freeze_blocks > 0: + for i, block in enumerate(base_model.blocks): + if i < args.ttt_freeze_blocks: + for p in block.parameters(): + p.requires_grad_(False) + frozen_params.add(id(p)) + + per_layer_lr = bool(int(os.environ.get("TTT_PER_LAYER_LR", "1"))) + if per_layer_lr: + param_groups = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: continue + lr_mul = 1.0 + if 'mlp.proj' in name: lr_mul = 3.0 + elif 'mlp.fc' in name: lr_mul = 0.5 + param_groups.append({'params': [p], 'lr': args.ttt_lr * lr_mul}) + ttt_params = [p for pg in param_groups for p in pg['params']] + else: + ttt_params = [p for p in base_model.parameters() if p.requires_grad] + param_groups = [{'params': ttt_params, 'lr': args.ttt_lr}] + + use_adamw = bool(int(os.environ.get("TTT_ADAMW", "1"))) + cosine_ttt = bool(int(os.environ.get("TTT_COSINE", "1"))) + progressive = bool(int(os.environ.get("TTT_PROGRESSIVE", "0"))) + if use_adamw: + optimizer = torch.optim.AdamW(param_groups, weight_decay=0.0) + else: + optimizer = torch.optim.SGD(param_groups, momentum=args.ttt_momentum) + + my_start = (total_seqs * rank) // world_size + my_end = (total_seqs * (rank + 1)) // world_size + total_steps = args.ttt_epochs * ((my_end - my_start + batch_seqs - 1) // batch_seqs) + step_count = 0 + + base_model.train() + t0 = time.perf_counter() + + for epoch in range(args.ttt_epochs): + # Progressive unfreezing: gradually unfreeze more blocks + if progressive and len(base_model.blocks) > 0: + n_blocks = len(base_model.blocks) + frac = (epoch + 1) / args.ttt_epochs + unfreeze_from = max(0, int(n_blocks * (1 - frac))) + for i, blk in enumerate(base_model.blocks): + for p in blk.parameters(): + p.requires_grad_(i >= unfreeze_from) + ttt_params = [p for p in base_model.parameters() if p.requires_grad] + + epoch_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + epoch_tokens = torch.zeros((), device=device, dtype=torch.float64) + + for batch_start in range(my_start, my_end, batch_seqs): + batch_end = min(batch_start + batch_seqs, my_end) + raw_start = batch_start * seq_len + raw_end = batch_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) + + if cosine_ttt and total_steps > 0: + progress = step_count / total_steps + _ttt_eta_min_ratio = float(os.environ.get("TTT_ETA_MIN_RATIO", "0.01")) + lr_mul = _ttt_eta_min_ratio + (1 - _ttt_eta_min_ratio) * 0.5 * (1 + math.cos(math.pi * progress)) + for pg in optimizer.param_groups: + pg['lr'] = args.ttt_lr * lr_mul + step_count += 1 + + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + + epoch_loss_sum += loss.detach().to(torch.float64) * y.numel() + epoch_tokens += float(y.numel()) + + if world_size > 1: + dist.all_reduce(epoch_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(epoch_tokens, op=dist.ReduceOp.SUM) + + elapsed = time.perf_counter() - t0 + if log_fn: + log_fn(f"ttt_epoch:{epoch+1}/{args.ttt_epochs} " + f"loss:{epoch_loss_sum.item()/max(epoch_tokens.item(),1):.4f} time:{elapsed:.1f}s") + + if use_lora: + remove_lora(base_model) + for p in base_model.parameters(): + p.requires_grad_(True) + + if log_fn: + log_fn(f"ttt:done elapsed={time.perf_counter()-t0:.1f}s") + + +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) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(amax / 31.0 if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8) + return q, scale + +def quantize_int6_gptq(weight: Tensor, hessian: Tensor | None, clip_range: int = 31, block_size: int = 128): + t32 = weight.float() + if t32.ndim != 2 or hessian is None: + return quantize_int6_per_row(t32) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0; H[dead, dead] = 1 + damp = 0.01 * torch.mean(torch.diag(H)) + H.diagonal().add_(damp) + perm = torch.argsort(torch.diag(H), descending=True) + W = t32[:, perm].clone(); W[:, dead[perm]] = 0; H = H[perm][:, perm] + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except torch.linalg.LinAlgError: + return quantize_int6_per_row(t32) + row_clip = torch.quantile(t32.abs(), 0.9999, dim=1) + scale = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + sf = scale.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols); count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i]; d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q[:, i1+i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + if i2 < cols: W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + inv_perm = torch.argsort(perm) + return Q[:, inv_perm], scale + + +def collect_gptq_hessians(model: nn.Module, tokens: Tensor, device: str, + num_batches: int = 256, seq_len: int = 512) -> dict[str, Tensor]: + hessians: dict[str, Tensor] = {} + handles = [] + nsamples: dict[str, int] = {} + + def make_hook(name): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device) + nsamples[name] = 0 + hessians[name].addmm_(x.t(), x) + nsamples[name] += x.shape[0] + return hook_fn + + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + param_name = name + ".weight" + handles.append(module.register_forward_hook(make_hook(param_name))) + + model.eval() + total_tokens = tokens.numel() + with torch.inference_mode(), torch.autocast("cuda", torch.bfloat16): + for bi in range(num_batches): + start = (bi * seq_len) % (total_tokens - seq_len - 1) + x = tokens[start:start + seq_len].unsqueeze(0).to(device=device, dtype=torch.long) + y = tokens[start + 1:start + seq_len + 1].unsqueeze(0).to(device=device, dtype=torch.long) + model(x, y) + + for h in handles: + h.remove() + + for name in hessians: + if nsamples[name] > 0: + hessians[name] /= nsamples[name] + + return hessians + + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + h = hessians.get(name) if hessians else None + if h is not None: + q, s = quantize_int6_gptq(t, h.cpu()) + meta[name] = {"type": "int6", "gptq": True} + else: + q, s = quantize_int6_per_row(t) + meta[name] = {"type": "int6"} + result[name + ".q"] = q + result[name + ".scale"] = s + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + + CastedLinear._qat_enabled = args.qat_enabled + + 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, + vrl_enabled=args.vrl_enabled, + gated_attn=args.gated_attn, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + 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): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + + ema_state: dict[str, Tensor] | None = None + if args.ema_enabled: + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + + 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) + qat_threshold = float(os.environ.get("QAT_THRESHOLD", "0.1")) + if args.late_qat and scale < qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + if ema_state is not None: + d = args.ema_decay + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(d).add_(t.detach().float(), alpha=1.0 - d) + + if args.swa_enabled and not args.ema_enabled and scale < 0.5 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().float().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].add_(t.detach().float()) + swa_count += 1 + + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + 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" + ) + + if ema_state is not None: + log0("ema:applying EMA weights") + avg_state = {name: t.to(dtype=base_model.state_dict()[name].dtype) + for name, t in ema_state.items()} + del ema_state + base_model.load_state_dict(avg_state, strict=True) + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + avg_state = {name: (t / swa_count).to(dtype=base_model.state_dict()[name].dtype) + for name, t in swa_state.items()} + del swa_state + base_model.load_state_dict(avg_state, strict=True) + + + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + _prune_pct = float(os.environ.get("PRUNE_PCT", "0.07")) + if _prune_pct > 0: + for name, t in sd_cpu.items(): + if t.ndim == 2 and t.numel() > 65536: + threshold = torch.quantile(t.abs().flatten().float(), _prune_pct) + mask = t.abs() >= threshold + sd_cpu[name] = t * mask + log0(f"pruned {_prune_pct*100:.0f}% of weights by magnitude") + + _gptq_enabled = int(os.environ.get("GPTQ_ENABLED", "0")) + gptq_hessians = None + if _gptq_enabled: + log0("gptq:collecting hessians...") + t_gptq = time.perf_counter() + gptq_hessians = collect_gptq_hessians( + base_model, val_tokens, device, + num_batches=int(os.environ.get("GPTQ_BATCHES", "256")), + seq_len=args.train_seq_len, + ) + log0(f"gptq:collected {len(gptq_hessians)} hessians in {time.perf_counter() - t_gptq:.1f}s") + + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}, hessians=gptq_hessians) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + vrl_enabled=args.vrl_enabled, + gated_attn=args.gated_attn, + ).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) + + if args.ttt_enabled: + if distributed: + dist.barrier() + for block in eval_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + log0(f"ttt:start lr={args.ttt_lr} momentum={args.ttt_momentum} " + f"epochs={args.ttt_epochs} freeze_blocks={args.ttt_freeze_blocks}") + t_ttt = time.perf_counter() + ttt_adapt(args, eval_model, device, val_tokens, + rank=rank, world_size=world_size, log_fn=log0) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + if distributed: + dist.barrier() + + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + + _ng_alpha = float(os.environ.get("NGRAM_ALPHA", "0")) + _ng_order = int(os.environ.get("NGRAM_ORDER", "7")) + _ng_buckets = int(os.environ.get("NGRAM_BUCKETS", "4000000")) + if _ng_alpha > 0 and args.eval_stride > 0: + log0(f"ngram: alpha={_ng_alpha} order={_ng_order} buckets={_ng_buckets}") + torch.cuda.synchronize(); t_ng = time.perf_counter() + ng_loss, ng_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=effective_eval_seq_len, + ngram_alpha=_ng_alpha, ngram_order=_ng_order, ngram_buckets=_ng_buckets, + ) + torch.cuda.synchronize() + log0(f"ngram val_loss:{ng_loss:.4f} val_bpb:{ng_bpb:.8f} time:{1000*(time.perf_counter()-t_ng):.0f}ms") + + _olb_lr = float(os.environ.get("OLB_LR", "0")) + if _olb_lr > 0 and args.eval_stride > 0: + V = args.vocab_size + olb_bias = torch.zeros(V, device=device) + olb_mom = torch.zeros(V, device=device) + olb_momentum = float(os.environ.get("OLB_MOMENTUM", "0.9")) + log0(f"olb: learning bias (lr={_olb_lr}, momentum={olb_momentum})") + esl = effective_eval_seq_len + eval_model.eval() + compiled_lg = torch.compile(eval_model.forward_logits, dynamic=False, fullgraph=True) + total_tok = val_tokens.numel() - 1 + wins = [ws for ws in range(0, total_tok, 64) if min(ws + esl, total_tok) - ws >= 1] + nw = len(wins) + my_s, my_e = (nw * rank) // world_size, (nw * (rank + 1)) // world_size + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + tok_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + torch.cuda.synchronize(); t_olb = time.perf_counter() + with torch.inference_mode(): + for bi in range(my_s, my_e, 32): + bws = wins[bi:min(bi+32, my_e)] + bsz = len(bws) + xb = torch.zeros(bsz, esl, dtype=torch.int64, device=device) + yb = torch.zeros(bsz, esl, dtype=torch.int64, device=device) + wls = [] + for i, ws in enumerate(bws): + end = min(ws + esl, total_tok); wl = end - ws; wls.append(wl) + ch = val_tokens[ws:end+1].to(torch.int64, device=device) + xb[i,:wl] = ch[:-1]; yb[i,:wl] = ch[1:] + with torch.autocast("cuda", torch.bfloat16): + lg = compiled_lg(xb) + lg = lg.float() + olb_bias[None, None, :] + nll = F.cross_entropy(lg.reshape(-1, V), yb.reshape(-1), reduction="none").reshape(bsz, esl) + for i, ws in enumerate(bws): + wl = wls[i]; s = 0 if ws == 0 else max(wl - 64, 0) + loss_sum += nll[i, s:wl].to(torch.float64).sum() + tok_count += float(wl - s) + tgt = yb[i, s:wl]; prev = xb[i, s:wl] + 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() + p = torch.softmax(lg[i, s:wl], dim=-1) + oh = torch.zeros_like(p); oh.scatter_(1, tgt.unsqueeze(1), 1.0) + g = (p - oh).mean(dim=0) + olb_mom.mul_(olb_momentum).add_(g, alpha=1-olb_momentum) + olb_bias.sub_(olb_mom, alpha=_olb_lr) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum); dist.all_reduce(tok_count); dist.all_reduce(byte_count) + vl = (loss_sum / tok_count).item() + bpb1 = vl / math.log(2) * (tok_count.item() / byte_count.item()) + torch.cuda.synchronize() + log0(f"olb_pass1 val_bpb:{bpb1:.8f} time:{1000*(time.perf_counter()-t_olb):.0f}ms") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main()