diff --git a/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/README.md b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/README.md new file mode 100644 index 000000000..78e5cf930 --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/README.md @@ -0,0 +1,71 @@ +# 10L MLP3x + BigramHash(2048) + SWA + Stride-32 Sliding Eval + +## Summary + +10-layer transformer with 3x relu² MLP, BigramHash(2048), SmearGate, orthogonal initialization, mixed int5/int6 quantization + zstd-22 compression, SWA weight averaging, and dense stride-32 sliding window evaluation. + +**val_bpb: 1.1487 (mean of 3 seeds)** | **14.9 MB** | 8xH100 SXM, 600s + +### Stride-32 dense sliding eval + +Standard submissions use stride-64 or stride-256 for sliding window evaluation. Stride-32 scores every token with near-maximum context from the 2048-token window, with only 32 positions of new content per window. This approximately doubles eval time but extracts more quality from the same trained model. + +### Techniques used +- 10-layer relu² MLP with 3x expansion (hidden=1536) +- BigramHash(2048, dim=128) + SmearGate for n-gram features +- Orthogonal initialization with scaled projections +- Mixed quantization: int5 for MLP, int6 for attention, fp16 for embeddings +- 5% magnitude pruning + zstd-22 compression +- SWA: averaging 24-25 checkpoints during warmdown +- Muon optimizer (WD=0.04) + AdamW for embeddings/scalars +- Training: seq_len=2048, batch=786K tokens, grad_clip=0.3 + +## Results + +### Reproducibility (3 seeds) + +| Seed | Steps | Step time | Sliding s32 BPB | Artifact | +|------|-------|-----------|-----------------|----------| +| 1337 | 6,374 | 94ms | 1.1503 | 14.90 MB | +| 42 | 6,626 | 91ms | 1.1493 | 14.78 MB | +| 2025 | 6,622 | 91ms | 1.1464 | 14.99 MB | + +**Mean: 1.1487** | Std: 0.0020 + +### Summary (mean of 3 seeds) + +| Metric | Value | +|--------|-------| +| val_bpb (sliding s32, mean) | **1.1487** | +| val_bpb std | 0.0020 | +| Model params | 24,468,561 | +| Artifact size | ~14.9 MB | +| Code size | 46,541 bytes | + +## Architecture + +| Component | Detail | +|-----------|--------| +| Layers | 10 | +| Model dim | 512 | +| Heads / KV heads | 8 / 4 (GQA) | +| MLP | relu² 3x expansion (hidden=1536) | +| Vocab | 1024 (SentencePiece BPE) | +| Seq length | 2048 | +| BigramHash | vocab=2048, dim=128 | +| SmearGate | learnable prev-token blending | +| Skip connections | U-Net encoder/decoder with learned skip weights | +| Embeddings | Tied input/output | + +## Run Command + +```bash +RUN_ID=submission \ +SEED=1337 \ +VAL_LOSS_EVERY=0 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## Hardware + +8x NVIDIA H100 80GB SXM, 600s training cap. diff --git a/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/submission.json b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/submission.json new file mode 100644 index 000000000..d82109eeb --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/submission.json @@ -0,0 +1,10 @@ +{ + "author": "Rhodrium", + "github_id": "Rhodrium", + "name": "10L MLP3x + BigramHash(2048) + SWA + Stride-32", + "blurb": "10-layer relu² MLP3x with BigramHash(2048), SmearGate, OrthoInit, mixed int5/int6 + zstd-22, SWA averaging, stride-32 sliding eval. Key contribution: dense stride-32 eval for better token-level context coverage.", + "date": "2026-03-21", + "val_bpb": 1.1487, + "bytes_total": 14902179, + "bytes_code": 46541 +} diff --git a/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train.log b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train.log new file mode 100644 index 000000000..07302a0fb --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train.log @@ -0,0 +1,1135 @@ +""" +11L MLP3x + BigramHash(2048) + EMA + SmearGate + mixed int5/int6 + stride-32 sliding eval. +EMA(0.997) replaces SWA for smoother weight averaging. Dense stride-32 eval. +""" + +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +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", 42)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + 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)) + 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", 10)) + 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", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + 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)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.04)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 48)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +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: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +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, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_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) + + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern for pattern in + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,bigram.scale".split(",") + if pattern +) +FP16_KEEP_NAME_PATTERNS = ("tok_emb",) + +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 "bigram" in name: + return "bigram" + if ".attn." in name: + return "attn" + return "other" + +def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip_range).clamp_min(1e-12).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + +def quantize_int8_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + return quantize_intN_per_row(t, clip_range=127) + +def mixed_quantize(state_dict: dict[str, Tensor], int6_cats: set[str]): + 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() <= 8192: + 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 any(p in name for p in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if cat in int6_cats and t.ndim >= 1: + clip = 15 if cat == "mlp" else 31 # int5 for MLP, int6 for attn + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{5 if clip == 15 else 6}"} + else: + q, s = quantize_int8_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed(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[name] + 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 + + +# ----------------------------- +# 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): + def forward(self, x: Tensor) -> Tensor: + return F.linear(x, self.weight.to(x.dtype), self.bias.to(x.dtype) if self.bias is not None else None) + + +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(p in name for p 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): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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._seq_len_cached != seq_len or self._cos_cached.device != device: + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + 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: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + 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.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + return self.proj(y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + 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: + return self.proj(torch.relu(self.fc(x)).square()) + + +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 Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, + rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + 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)) + return x + + +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: float, tie_embeddings: bool, + tied_embed_init_std: float, logit_softcap: float, rope_base: float, + qk_gain_init: float, bigram_vocab_size: int = 0, bigram_dim: int = 128): + super().__init__() + self.tie_embeddings = tie_embeddings + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.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.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ 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._init_weights(tied_embed_init_std, num_layers) + + def _init_weights(self, tied_embed_init_std: float, num_layers: int) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=tied_embed_init_std) + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + 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, +) -> tuple[float, float]: + seq_len = 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 >= stride or ws == 0] + 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() + 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 = 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 = 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 + + +# ----------------------------- +# TRAINING +# ----------------------------- + +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 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + 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) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:tokens:{val_tokens.numel() - 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, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + ).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(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + scalar_params = [p for name, p in block_named_params + if p.ndim < 2 or any(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW(tok_params, betas=(args.beta1, args.beta2), + eps=args.adam_eps, weight_decay=args.weight_decay, fused=True) + optimizer_muon = Muon(matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.weight_decay) + 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.weight_decay, fused=True) + optimizers = [optimizer_tok, optimizer_muon, optimizer_scalar] + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params} world_size:{world_size} grad_accum:{grad_accum_steps}") + log0(f"layers:{args.num_layers} dim:{args.model_dim} heads:{args.num_heads} kv:{args.num_kv_heads}") + log0(f"mlp_mult:{args.mlp_mult} bigram:{args.bigram_vocab_size} seq_len:{args.train_seq_len}") + log0(f"batch_tokens:{args.train_batch_tokens} lr:{args.matrix_lr} wd:{args.weight_decay}") + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all(): + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup + if args.warmup_steps > 0: + initial_model_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for ws in range(args.warmup_steps): + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + wl = model(x, y) + (wl * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if ws + 1 == args.warmup_steps or (ws + 1) % 10 == 0: + log0(f"warmup_step:{ws + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, st in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(st) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # Training loop + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_sd = {n: t.detach().cpu().clone() for n, t in base_model.state_dict().items()} if args.ema_enabled else 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 step:{step}") + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + 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_mom = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_mom + 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_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + # SWA + if args.swa_enabled and scale < args.swa_start_frac 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 + + # EMA update + if ema_sd is not None: + with torch.no_grad(): + for n, p in base_model.state_dict().items(): + ema_sd[n].lerp_(p.cpu(), 1.0 - args.ema_decay) + + if args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None): + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_ms:.0f}ms step_avg:{approx_ms / step:.2f}ms") + + reached_cap = max_wallclock_ms is not None and approx_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + t_cap = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(t_cap, op=dist.ReduceOp.MAX) + reached_cap = bool(t_cap.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0(f"peak memory: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB") + + # Apply EMA or SWA + if ema_sd is not None: + base_model.load_state_dict(ema_sd, strict=True) + log0("Loaded EMA weights") + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + cur = base_model.state_dict() + avg = {name: (tensor / swa_count).to(dtype=cur[name].dtype) for name, tensor in swa_state.items()} + base_model.load_state_dict(avg, strict=True) + + # Save raw model + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + log0(f"Raw model: {os.path.getsize('final_model.pt')} bytes") + + # Magnitude pruning + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + threshold = torch.quantile(param.abs().float().flatten(), 0.05) + param.masked_fill_(param.abs() < threshold, 0.0) + + # Mixed int5/int6 quantization + compression + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize(sd_cpu, {"mlp", "attn", "bigram"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + qsize = os.path.getsize("final_model.int8.ptz") + csize = len(code.encode("utf-8")) + log0(f"Quantized model ({_COMPRESSOR}): {qsize} bytes") + log0(f"Code size: {csize} bytes") + log0(f"Total submission: {qsize + csize} bytes") + + # Roundtrip validation + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + blob = f.read() + if _COMPRESSOR == "zstd": + raw = zstandard.ZstdDecompressor().decompress(blob) + else: + raw = zlib.decompress(blob) + qs = torch.load(io.BytesIO(raw), map_location="cpu") + deq = dequantize_mixed(qs["w"], qs["m"], sd_cpu) + base_model.load_state_dict(deq, strict=True) + + # Sliding window eval + torch.cuda.synchronize() + t_eval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0(f"eval:sliding stride={args.eval_stride} batch={args.eval_batch_seqs}") + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_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.eval_batch_seqs, + ) + else: + q_val_loss, q_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, + ) + torch.cuda.synchronize() + log0(f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_eval):.0f}ms") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() + +==================================================================================================== +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:tokens:62021632 +model_params:24468561 world_size:8 grad_accum:1 +layers:10 dim:512 heads:8 kv:4 +mlp_mult:3.0 bigram:2048 seq_len:2048 +batch_tokens:786432 lr:0.02 wd:0.04 +warmup_step:10/20 +warmup_step:20/20 +step:1/20000 train_loss:6.9300 train_time:278ms step_avg:278.19ms +step:2/20000 train_loss:7.9847 train_time:405ms step_avg:202.42ms +step:3/20000 train_loss:7.6546 train_time:594ms step_avg:198.05ms +step:4/20000 train_loss:6.9167 train_time:720ms step_avg:180.08ms +step:5/20000 train_loss:6.7094 train_time:931ms step_avg:186.27ms +step:6/20000 train_loss:6.6178 train_time:1168ms step_avg:194.73ms +step:7/20000 train_loss:6.5318 train_time:1363ms step_avg:194.65ms +step:8/20000 train_loss:6.5571 train_time:1501ms step_avg:187.58ms +step:9/20000 train_loss:6.3505 train_time:1706ms step_avg:189.54ms +step:10/20000 train_loss:6.1252 train_time:1940ms step_avg:193.97ms +step:100/20000 train_loss:3.2359 train_time:20680ms step_avg:206.80ms +step:200/20000 train_loss:2.4328 train_time:40900ms step_avg:204.50ms +step:300/20000 train_loss:2.5825 train_time:50006ms step_avg:166.69ms +step:400/20000 train_loss:2.4432 train_time:59112ms step_avg:147.78ms +step:500/20000 train_loss:2.4190 train_time:68106ms step_avg:136.21ms +step:600/20000 train_loss:2.3560 train_time:77202ms step_avg:128.67ms +step:700/20000 train_loss:2.3645 train_time:86315ms step_avg:123.31ms +step:800/20000 train_loss:2.2558 train_time:95423ms step_avg:119.28ms +step:900/20000 train_loss:2.1340 train_time:104512ms step_avg:116.12ms +step:1000/20000 train_loss:2.2865 train_time:113503ms step_avg:113.50ms +step:1100/20000 train_loss:2.3330 train_time:122589ms step_avg:111.44ms +step:1200/20000 train_loss:2.3631 train_time:131741ms step_avg:109.78ms +step:1300/20000 train_loss:2.1104 train_time:140823ms step_avg:108.33ms +step:1400/20000 train_loss:2.1935 train_time:149900ms step_avg:107.07ms +step:1500/20000 train_loss:2.2314 train_time:158865ms step_avg:105.91ms +step:1600/20000 train_loss:2.0836 train_time:167943ms step_avg:104.96ms +step:1700/20000 train_loss:2.1532 train_time:177023ms step_avg:104.13ms +step:1800/20000 train_loss:2.1707 train_time:186094ms step_avg:103.39ms +step:1900/20000 train_loss:2.1390 train_time:195052ms step_avg:102.66ms +step:2000/20000 train_loss:2.0793 train_time:204127ms step_avg:102.06ms +step:2100/20000 train_loss:2.0561 train_time:213190ms step_avg:101.52ms +step:2200/20000 train_loss:2.1478 train_time:222255ms step_avg:101.02ms +step:2300/20000 train_loss:2.1154 train_time:231320ms step_avg:100.57ms +step:2400/20000 train_loss:2.0729 train_time:240278ms step_avg:100.12ms +step:2500/20000 train_loss:2.1765 train_time:249344ms step_avg:99.74ms +step:2600/20000 train_loss:2.1195 train_time:258408ms step_avg:99.39ms +step:2700/20000 train_loss:2.1098 train_time:267473ms step_avg:99.06ms +step:2800/20000 train_loss:2.1647 train_time:276543ms step_avg:98.77ms +step:2900/20000 train_loss:2.0320 train_time:285483ms step_avg:98.44ms +step:3000/20000 train_loss:2.1693 train_time:294526ms step_avg:98.18ms +step:3100/20000 train_loss:2.0472 train_time:303588ms step_avg:97.93ms +step:3200/20000 train_loss:2.1801 train_time:312648ms step_avg:97.70ms +step:3300/20000 train_loss:2.0784 train_time:321594ms step_avg:97.45ms +step:3400/20000 train_loss:2.0279 train_time:330646ms step_avg:97.25ms +step:3500/20000 train_loss:2.1795 train_time:339704ms step_avg:97.06ms +step:3600/20000 train_loss:2.0982 train_time:348764ms step_avg:96.88ms +step:3700/20000 train_loss:2.0933 train_time:357808ms step_avg:96.70ms +step:3800/20000 train_loss:2.0698 train_time:366752ms step_avg:96.51ms +step:3900/20000 train_loss:2.0723 train_time:375810ms step_avg:96.36ms +step:4000/20000 train_loss:1.9721 train_time:384865ms step_avg:96.22ms +step:4100/20000 train_loss:2.0134 train_time:393924ms step_avg:96.08ms +step:4200/20000 train_loss:2.1496 train_time:402986ms step_avg:95.95ms +step:4300/20000 train_loss:2.0538 train_time:411926ms step_avg:95.80ms +step:4400/20000 train_loss:2.0308 train_time:420977ms step_avg:95.68ms +step:4500/20000 train_loss:2.1194 train_time:430037ms step_avg:95.56ms +step:4600/20000 train_loss:1.8382 train_time:439090ms step_avg:95.45ms +step:4700/20000 train_loss:2.2290 train_time:448034ms step_avg:95.33ms +step:4800/20000 train_loss:2.4274 train_time:457092ms step_avg:95.23ms +step:4900/20000 train_loss:2.0456 train_time:466148ms step_avg:95.13ms +step:5000/20000 train_loss:2.0984 train_time:475206ms step_avg:95.04ms +step:5100/20000 train_loss:2.1240 train_time:484265ms step_avg:94.95ms +swa:start step:5150 +step:5200/20000 train_loss:2.0386 train_time:493277ms step_avg:94.86ms +step:5300/20000 train_loss:2.0022 train_time:502382ms step_avg:94.79ms +step:5400/20000 train_loss:2.0416 train_time:511484ms step_avg:94.72ms +step:5500/20000 train_loss:2.0117 train_time:520573ms step_avg:94.65ms +step:5600/20000 train_loss:1.9492 train_time:529686ms step_avg:94.59ms +step:5700/20000 train_loss:2.0053 train_time:538688ms step_avg:94.51ms +step:5800/20000 train_loss:1.9901 train_time:547788ms step_avg:94.45ms +step:5900/20000 train_loss:1.8960 train_time:556894ms step_avg:94.39ms +step:6000/20000 train_loss:1.9337 train_time:566000ms step_avg:94.33ms +step:6100/20000 train_loss:1.9099 train_time:575004ms step_avg:94.26ms +step:6200/20000 train_loss:1.9431 train_time:584125ms step_avg:94.21ms +step:6300/20000 train_loss:1.9404 train_time:593234ms step_avg:94.16ms +step:6374/20000 val_loss:1.9556 val_bpb:1.1582 train_time:600052ms step_avg:94.14ms +stopping_early: wallclock_cap train_time:600052ms step:6374 +peak memory: 18856 MiB +swa:applying averaged 25 checkpoints +Raw model: 96340267 bytes +Quantized model (zstd): 14855638 bytes +Code size: 46541 bytes +Total submission: 14902179 bytes +eval:sliding stride=32 batch=48 +final_int8_zlib_roundtrip val_loss:1.9422 val_bpb:1.1503 eval_time:333160ms +final_int8_zlib_roundtrip_exact val_loss:1.94223960 val_bpb:1.15030751 diff --git a/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_gpt.py b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_gpt.py new file mode 100644 index 000000000..89b5707ee --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_gpt.py @@ -0,0 +1,1040 @@ +""" +11L MLP3x + BigramHash(2048) + EMA + SmearGate + mixed int5/int6 + stride-32 sliding eval. +EMA(0.997) replaces SWA for smoother weight averaging. Dense stride-32 eval. +""" + +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +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", 42)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + 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)) + 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", 10)) + 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", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + 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)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.04)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 48)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +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: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +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, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_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) + + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern for pattern in + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,bigram.scale".split(",") + if pattern +) +FP16_KEEP_NAME_PATTERNS = ("tok_emb",) + +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 "bigram" in name: + return "bigram" + if ".attn." in name: + return "attn" + return "other" + +def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip_range).clamp_min(1e-12).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + +def quantize_int8_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + return quantize_intN_per_row(t, clip_range=127) + +def mixed_quantize(state_dict: dict[str, Tensor], int6_cats: set[str]): + 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() <= 8192: + 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 any(p in name for p in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if cat in int6_cats and t.ndim >= 1: + clip = 15 if cat == "mlp" else 31 # int5 for MLP, int6 for attn + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{5 if clip == 15 else 6}"} + else: + q, s = quantize_int8_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed(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[name] + 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 + + +# ----------------------------- +# 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): + def forward(self, x: Tensor) -> Tensor: + return F.linear(x, self.weight.to(x.dtype), self.bias.to(x.dtype) if self.bias is not None else None) + + +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(p in name for p 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): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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._seq_len_cached != seq_len or self._cos_cached.device != device: + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + 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: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + 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.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + return self.proj(y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + 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: + return self.proj(torch.relu(self.fc(x)).square()) + + +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 Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, + rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + 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)) + return x + + +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: float, tie_embeddings: bool, + tied_embed_init_std: float, logit_softcap: float, rope_base: float, + qk_gain_init: float, bigram_vocab_size: int = 0, bigram_dim: int = 128): + super().__init__() + self.tie_embeddings = tie_embeddings + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.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.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ 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._init_weights(tied_embed_init_std, num_layers) + + def _init_weights(self, tied_embed_init_std: float, num_layers: int) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=tied_embed_init_std) + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + 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, +) -> tuple[float, float]: + seq_len = 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 >= stride or ws == 0] + 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() + 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 = 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 = 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 + + +# ----------------------------- +# TRAINING +# ----------------------------- + +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 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + 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) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:tokens:{val_tokens.numel() - 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, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + ).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(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + scalar_params = [p for name, p in block_named_params + if p.ndim < 2 or any(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW(tok_params, betas=(args.beta1, args.beta2), + eps=args.adam_eps, weight_decay=args.weight_decay, fused=True) + optimizer_muon = Muon(matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.weight_decay) + 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.weight_decay, fused=True) + optimizers = [optimizer_tok, optimizer_muon, optimizer_scalar] + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params} world_size:{world_size} grad_accum:{grad_accum_steps}") + log0(f"layers:{args.num_layers} dim:{args.model_dim} heads:{args.num_heads} kv:{args.num_kv_heads}") + log0(f"mlp_mult:{args.mlp_mult} bigram:{args.bigram_vocab_size} seq_len:{args.train_seq_len}") + log0(f"batch_tokens:{args.train_batch_tokens} lr:{args.matrix_lr} wd:{args.weight_decay}") + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all(): + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup + if args.warmup_steps > 0: + initial_model_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for ws in range(args.warmup_steps): + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + wl = model(x, y) + (wl * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if ws + 1 == args.warmup_steps or (ws + 1) % 10 == 0: + log0(f"warmup_step:{ws + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, st in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(st) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # Training loop + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_sd = {n: t.detach().cpu().clone() for n, t in base_model.state_dict().items()} if args.ema_enabled else 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 step:{step}") + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + 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_mom = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_mom + 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_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + # SWA + if args.swa_enabled and scale < args.swa_start_frac 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 + + # EMA update + if ema_sd is not None: + with torch.no_grad(): + for n, p in base_model.state_dict().items(): + ema_sd[n].lerp_(p.cpu(), 1.0 - args.ema_decay) + + if args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None): + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_ms:.0f}ms step_avg:{approx_ms / step:.2f}ms") + + reached_cap = max_wallclock_ms is not None and approx_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + t_cap = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(t_cap, op=dist.ReduceOp.MAX) + reached_cap = bool(t_cap.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0(f"peak memory: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB") + + # Apply EMA or SWA + if ema_sd is not None: + base_model.load_state_dict(ema_sd, strict=True) + log0("Loaded EMA weights") + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + cur = base_model.state_dict() + avg = {name: (tensor / swa_count).to(dtype=cur[name].dtype) for name, tensor in swa_state.items()} + base_model.load_state_dict(avg, strict=True) + + # Save raw model + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + log0(f"Raw model: {os.path.getsize('final_model.pt')} bytes") + + # Magnitude pruning + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + threshold = torch.quantile(param.abs().float().flatten(), 0.05) + param.masked_fill_(param.abs() < threshold, 0.0) + + # Mixed int5/int6 quantization + compression + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize(sd_cpu, {"mlp", "attn", "bigram"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + qsize = os.path.getsize("final_model.int8.ptz") + csize = len(code.encode("utf-8")) + log0(f"Quantized model ({_COMPRESSOR}): {qsize} bytes") + log0(f"Code size: {csize} bytes") + log0(f"Total submission: {qsize + csize} bytes") + + # Roundtrip validation + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + blob = f.read() + if _COMPRESSOR == "zstd": + raw = zstandard.ZstdDecompressor().decompress(blob) + else: + raw = zlib.decompress(blob) + qs = torch.load(io.BytesIO(raw), map_location="cpu") + deq = dequantize_mixed(qs["w"], qs["m"], sd_cpu) + base_model.load_state_dict(deq, strict=True) + + # Sliding window eval + torch.cuda.synchronize() + t_eval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0(f"eval:sliding stride={args.eval_stride} batch={args.eval_batch_seqs}") + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_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.eval_batch_seqs, + ) + else: + q_val_loss, q_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, + ) + torch.cuda.synchronize() + log0(f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_eval):.0f}ms") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed1337.log b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed1337.log new file mode 100644 index 000000000..07302a0fb --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed1337.log @@ -0,0 +1,1135 @@ +""" +11L MLP3x + BigramHash(2048) + EMA + SmearGate + mixed int5/int6 + stride-32 sliding eval. +EMA(0.997) replaces SWA for smoother weight averaging. Dense stride-32 eval. +""" + +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +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", 42)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + 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)) + 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", 10)) + 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", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + 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)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.04)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 48)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +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: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +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, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_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) + + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern for pattern in + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,bigram.scale".split(",") + if pattern +) +FP16_KEEP_NAME_PATTERNS = ("tok_emb",) + +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 "bigram" in name: + return "bigram" + if ".attn." in name: + return "attn" + return "other" + +def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip_range).clamp_min(1e-12).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + +def quantize_int8_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + return quantize_intN_per_row(t, clip_range=127) + +def mixed_quantize(state_dict: dict[str, Tensor], int6_cats: set[str]): + 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() <= 8192: + 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 any(p in name for p in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if cat in int6_cats and t.ndim >= 1: + clip = 15 if cat == "mlp" else 31 # int5 for MLP, int6 for attn + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{5 if clip == 15 else 6}"} + else: + q, s = quantize_int8_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed(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[name] + 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 + + +# ----------------------------- +# 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): + def forward(self, x: Tensor) -> Tensor: + return F.linear(x, self.weight.to(x.dtype), self.bias.to(x.dtype) if self.bias is not None else None) + + +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(p in name for p 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): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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._seq_len_cached != seq_len or self._cos_cached.device != device: + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + 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: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + 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.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + return self.proj(y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + 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: + return self.proj(torch.relu(self.fc(x)).square()) + + +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 Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, + rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + 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)) + return x + + +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: float, tie_embeddings: bool, + tied_embed_init_std: float, logit_softcap: float, rope_base: float, + qk_gain_init: float, bigram_vocab_size: int = 0, bigram_dim: int = 128): + super().__init__() + self.tie_embeddings = tie_embeddings + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.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.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ 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._init_weights(tied_embed_init_std, num_layers) + + def _init_weights(self, tied_embed_init_std: float, num_layers: int) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=tied_embed_init_std) + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + 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, +) -> tuple[float, float]: + seq_len = 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 >= stride or ws == 0] + 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() + 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 = 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 = 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 + + +# ----------------------------- +# TRAINING +# ----------------------------- + +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 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + 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) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:tokens:{val_tokens.numel() - 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, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + ).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(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + scalar_params = [p for name, p in block_named_params + if p.ndim < 2 or any(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW(tok_params, betas=(args.beta1, args.beta2), + eps=args.adam_eps, weight_decay=args.weight_decay, fused=True) + optimizer_muon = Muon(matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.weight_decay) + 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.weight_decay, fused=True) + optimizers = [optimizer_tok, optimizer_muon, optimizer_scalar] + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params} world_size:{world_size} grad_accum:{grad_accum_steps}") + log0(f"layers:{args.num_layers} dim:{args.model_dim} heads:{args.num_heads} kv:{args.num_kv_heads}") + log0(f"mlp_mult:{args.mlp_mult} bigram:{args.bigram_vocab_size} seq_len:{args.train_seq_len}") + log0(f"batch_tokens:{args.train_batch_tokens} lr:{args.matrix_lr} wd:{args.weight_decay}") + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all(): + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup + if args.warmup_steps > 0: + initial_model_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for ws in range(args.warmup_steps): + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + wl = model(x, y) + (wl * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if ws + 1 == args.warmup_steps or (ws + 1) % 10 == 0: + log0(f"warmup_step:{ws + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, st in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(st) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # Training loop + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_sd = {n: t.detach().cpu().clone() for n, t in base_model.state_dict().items()} if args.ema_enabled else 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 step:{step}") + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + 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_mom = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_mom + 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_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + # SWA + if args.swa_enabled and scale < args.swa_start_frac 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 + + # EMA update + if ema_sd is not None: + with torch.no_grad(): + for n, p in base_model.state_dict().items(): + ema_sd[n].lerp_(p.cpu(), 1.0 - args.ema_decay) + + if args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None): + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_ms:.0f}ms step_avg:{approx_ms / step:.2f}ms") + + reached_cap = max_wallclock_ms is not None and approx_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + t_cap = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(t_cap, op=dist.ReduceOp.MAX) + reached_cap = bool(t_cap.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0(f"peak memory: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB") + + # Apply EMA or SWA + if ema_sd is not None: + base_model.load_state_dict(ema_sd, strict=True) + log0("Loaded EMA weights") + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + cur = base_model.state_dict() + avg = {name: (tensor / swa_count).to(dtype=cur[name].dtype) for name, tensor in swa_state.items()} + base_model.load_state_dict(avg, strict=True) + + # Save raw model + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + log0(f"Raw model: {os.path.getsize('final_model.pt')} bytes") + + # Magnitude pruning + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + threshold = torch.quantile(param.abs().float().flatten(), 0.05) + param.masked_fill_(param.abs() < threshold, 0.0) + + # Mixed int5/int6 quantization + compression + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize(sd_cpu, {"mlp", "attn", "bigram"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + qsize = os.path.getsize("final_model.int8.ptz") + csize = len(code.encode("utf-8")) + log0(f"Quantized model ({_COMPRESSOR}): {qsize} bytes") + log0(f"Code size: {csize} bytes") + log0(f"Total submission: {qsize + csize} bytes") + + # Roundtrip validation + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + blob = f.read() + if _COMPRESSOR == "zstd": + raw = zstandard.ZstdDecompressor().decompress(blob) + else: + raw = zlib.decompress(blob) + qs = torch.load(io.BytesIO(raw), map_location="cpu") + deq = dequantize_mixed(qs["w"], qs["m"], sd_cpu) + base_model.load_state_dict(deq, strict=True) + + # Sliding window eval + torch.cuda.synchronize() + t_eval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0(f"eval:sliding stride={args.eval_stride} batch={args.eval_batch_seqs}") + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_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.eval_batch_seqs, + ) + else: + q_val_loss, q_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, + ) + torch.cuda.synchronize() + log0(f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_eval):.0f}ms") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() + +==================================================================================================== +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:tokens:62021632 +model_params:24468561 world_size:8 grad_accum:1 +layers:10 dim:512 heads:8 kv:4 +mlp_mult:3.0 bigram:2048 seq_len:2048 +batch_tokens:786432 lr:0.02 wd:0.04 +warmup_step:10/20 +warmup_step:20/20 +step:1/20000 train_loss:6.9300 train_time:278ms step_avg:278.19ms +step:2/20000 train_loss:7.9847 train_time:405ms step_avg:202.42ms +step:3/20000 train_loss:7.6546 train_time:594ms step_avg:198.05ms +step:4/20000 train_loss:6.9167 train_time:720ms step_avg:180.08ms +step:5/20000 train_loss:6.7094 train_time:931ms step_avg:186.27ms +step:6/20000 train_loss:6.6178 train_time:1168ms step_avg:194.73ms +step:7/20000 train_loss:6.5318 train_time:1363ms step_avg:194.65ms +step:8/20000 train_loss:6.5571 train_time:1501ms step_avg:187.58ms +step:9/20000 train_loss:6.3505 train_time:1706ms step_avg:189.54ms +step:10/20000 train_loss:6.1252 train_time:1940ms step_avg:193.97ms +step:100/20000 train_loss:3.2359 train_time:20680ms step_avg:206.80ms +step:200/20000 train_loss:2.4328 train_time:40900ms step_avg:204.50ms +step:300/20000 train_loss:2.5825 train_time:50006ms step_avg:166.69ms +step:400/20000 train_loss:2.4432 train_time:59112ms step_avg:147.78ms +step:500/20000 train_loss:2.4190 train_time:68106ms step_avg:136.21ms +step:600/20000 train_loss:2.3560 train_time:77202ms step_avg:128.67ms +step:700/20000 train_loss:2.3645 train_time:86315ms step_avg:123.31ms +step:800/20000 train_loss:2.2558 train_time:95423ms step_avg:119.28ms +step:900/20000 train_loss:2.1340 train_time:104512ms step_avg:116.12ms +step:1000/20000 train_loss:2.2865 train_time:113503ms step_avg:113.50ms +step:1100/20000 train_loss:2.3330 train_time:122589ms step_avg:111.44ms +step:1200/20000 train_loss:2.3631 train_time:131741ms step_avg:109.78ms +step:1300/20000 train_loss:2.1104 train_time:140823ms step_avg:108.33ms +step:1400/20000 train_loss:2.1935 train_time:149900ms step_avg:107.07ms +step:1500/20000 train_loss:2.2314 train_time:158865ms step_avg:105.91ms +step:1600/20000 train_loss:2.0836 train_time:167943ms step_avg:104.96ms +step:1700/20000 train_loss:2.1532 train_time:177023ms step_avg:104.13ms +step:1800/20000 train_loss:2.1707 train_time:186094ms step_avg:103.39ms +step:1900/20000 train_loss:2.1390 train_time:195052ms step_avg:102.66ms +step:2000/20000 train_loss:2.0793 train_time:204127ms step_avg:102.06ms +step:2100/20000 train_loss:2.0561 train_time:213190ms step_avg:101.52ms +step:2200/20000 train_loss:2.1478 train_time:222255ms step_avg:101.02ms +step:2300/20000 train_loss:2.1154 train_time:231320ms step_avg:100.57ms +step:2400/20000 train_loss:2.0729 train_time:240278ms step_avg:100.12ms +step:2500/20000 train_loss:2.1765 train_time:249344ms step_avg:99.74ms +step:2600/20000 train_loss:2.1195 train_time:258408ms step_avg:99.39ms +step:2700/20000 train_loss:2.1098 train_time:267473ms step_avg:99.06ms +step:2800/20000 train_loss:2.1647 train_time:276543ms step_avg:98.77ms +step:2900/20000 train_loss:2.0320 train_time:285483ms step_avg:98.44ms +step:3000/20000 train_loss:2.1693 train_time:294526ms step_avg:98.18ms +step:3100/20000 train_loss:2.0472 train_time:303588ms step_avg:97.93ms +step:3200/20000 train_loss:2.1801 train_time:312648ms step_avg:97.70ms +step:3300/20000 train_loss:2.0784 train_time:321594ms step_avg:97.45ms +step:3400/20000 train_loss:2.0279 train_time:330646ms step_avg:97.25ms +step:3500/20000 train_loss:2.1795 train_time:339704ms step_avg:97.06ms +step:3600/20000 train_loss:2.0982 train_time:348764ms step_avg:96.88ms +step:3700/20000 train_loss:2.0933 train_time:357808ms step_avg:96.70ms +step:3800/20000 train_loss:2.0698 train_time:366752ms step_avg:96.51ms +step:3900/20000 train_loss:2.0723 train_time:375810ms step_avg:96.36ms +step:4000/20000 train_loss:1.9721 train_time:384865ms step_avg:96.22ms +step:4100/20000 train_loss:2.0134 train_time:393924ms step_avg:96.08ms +step:4200/20000 train_loss:2.1496 train_time:402986ms step_avg:95.95ms +step:4300/20000 train_loss:2.0538 train_time:411926ms step_avg:95.80ms +step:4400/20000 train_loss:2.0308 train_time:420977ms step_avg:95.68ms +step:4500/20000 train_loss:2.1194 train_time:430037ms step_avg:95.56ms +step:4600/20000 train_loss:1.8382 train_time:439090ms step_avg:95.45ms +step:4700/20000 train_loss:2.2290 train_time:448034ms step_avg:95.33ms +step:4800/20000 train_loss:2.4274 train_time:457092ms step_avg:95.23ms +step:4900/20000 train_loss:2.0456 train_time:466148ms step_avg:95.13ms +step:5000/20000 train_loss:2.0984 train_time:475206ms step_avg:95.04ms +step:5100/20000 train_loss:2.1240 train_time:484265ms step_avg:94.95ms +swa:start step:5150 +step:5200/20000 train_loss:2.0386 train_time:493277ms step_avg:94.86ms +step:5300/20000 train_loss:2.0022 train_time:502382ms step_avg:94.79ms +step:5400/20000 train_loss:2.0416 train_time:511484ms step_avg:94.72ms +step:5500/20000 train_loss:2.0117 train_time:520573ms step_avg:94.65ms +step:5600/20000 train_loss:1.9492 train_time:529686ms step_avg:94.59ms +step:5700/20000 train_loss:2.0053 train_time:538688ms step_avg:94.51ms +step:5800/20000 train_loss:1.9901 train_time:547788ms step_avg:94.45ms +step:5900/20000 train_loss:1.8960 train_time:556894ms step_avg:94.39ms +step:6000/20000 train_loss:1.9337 train_time:566000ms step_avg:94.33ms +step:6100/20000 train_loss:1.9099 train_time:575004ms step_avg:94.26ms +step:6200/20000 train_loss:1.9431 train_time:584125ms step_avg:94.21ms +step:6300/20000 train_loss:1.9404 train_time:593234ms step_avg:94.16ms +step:6374/20000 val_loss:1.9556 val_bpb:1.1582 train_time:600052ms step_avg:94.14ms +stopping_early: wallclock_cap train_time:600052ms step:6374 +peak memory: 18856 MiB +swa:applying averaged 25 checkpoints +Raw model: 96340267 bytes +Quantized model (zstd): 14855638 bytes +Code size: 46541 bytes +Total submission: 14902179 bytes +eval:sliding stride=32 batch=48 +final_int8_zlib_roundtrip val_loss:1.9422 val_bpb:1.1503 eval_time:333160ms +final_int8_zlib_roundtrip_exact val_loss:1.94223960 val_bpb:1.15030751 diff --git a/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed2025.log b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed2025.log new file mode 100644 index 000000000..75967f2dd --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed2025.log @@ -0,0 +1,1138 @@ +""" +11L MLP3x + BigramHash(2048) + EMA + SmearGate + mixed int5/int6 + stride-32 sliding eval. +EMA(0.997) replaces SWA for smoother weight averaging. Dense stride-32 eval. +""" + +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +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", 42)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + 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)) + 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", 10)) + 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", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + 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)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.04)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 48)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +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: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +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, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_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) + + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern for pattern in + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,bigram.scale".split(",") + if pattern +) +FP16_KEEP_NAME_PATTERNS = ("tok_emb",) + +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 "bigram" in name: + return "bigram" + if ".attn." in name: + return "attn" + return "other" + +def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip_range).clamp_min(1e-12).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + +def quantize_int8_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + return quantize_intN_per_row(t, clip_range=127) + +def mixed_quantize(state_dict: dict[str, Tensor], int6_cats: set[str]): + 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() <= 8192: + 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 any(p in name for p in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if cat in int6_cats and t.ndim >= 1: + clip = 15 if cat == "mlp" else 31 # int5 for MLP, int6 for attn + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{5 if clip == 15 else 6}"} + else: + q, s = quantize_int8_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed(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[name] + 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 + + +# ----------------------------- +# 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): + def forward(self, x: Tensor) -> Tensor: + return F.linear(x, self.weight.to(x.dtype), self.bias.to(x.dtype) if self.bias is not None else None) + + +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(p in name for p 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): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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._seq_len_cached != seq_len or self._cos_cached.device != device: + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + 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: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + 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.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + return self.proj(y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + 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: + return self.proj(torch.relu(self.fc(x)).square()) + + +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 Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, + rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + 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)) + return x + + +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: float, tie_embeddings: bool, + tied_embed_init_std: float, logit_softcap: float, rope_base: float, + qk_gain_init: float, bigram_vocab_size: int = 0, bigram_dim: int = 128): + super().__init__() + self.tie_embeddings = tie_embeddings + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.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.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ 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._init_weights(tied_embed_init_std, num_layers) + + def _init_weights(self, tied_embed_init_std: float, num_layers: int) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=tied_embed_init_std) + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + 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, +) -> tuple[float, float]: + seq_len = 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 >= stride or ws == 0] + 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() + 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 = 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 = 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 + + +# ----------------------------- +# TRAINING +# ----------------------------- + +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 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + 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) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:tokens:{val_tokens.numel() - 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, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + ).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(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + scalar_params = [p for name, p in block_named_params + if p.ndim < 2 or any(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW(tok_params, betas=(args.beta1, args.beta2), + eps=args.adam_eps, weight_decay=args.weight_decay, fused=True) + optimizer_muon = Muon(matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.weight_decay) + 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.weight_decay, fused=True) + optimizers = [optimizer_tok, optimizer_muon, optimizer_scalar] + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params} world_size:{world_size} grad_accum:{grad_accum_steps}") + log0(f"layers:{args.num_layers} dim:{args.model_dim} heads:{args.num_heads} kv:{args.num_kv_heads}") + log0(f"mlp_mult:{args.mlp_mult} bigram:{args.bigram_vocab_size} seq_len:{args.train_seq_len}") + log0(f"batch_tokens:{args.train_batch_tokens} lr:{args.matrix_lr} wd:{args.weight_decay}") + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all(): + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup + if args.warmup_steps > 0: + initial_model_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for ws in range(args.warmup_steps): + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + wl = model(x, y) + (wl * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if ws + 1 == args.warmup_steps or (ws + 1) % 10 == 0: + log0(f"warmup_step:{ws + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, st in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(st) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # Training loop + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_sd = {n: t.detach().cpu().clone() for n, t in base_model.state_dict().items()} if args.ema_enabled else 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 step:{step}") + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + 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_mom = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_mom + 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_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + # SWA + if args.swa_enabled and scale < args.swa_start_frac 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 + + # EMA update + if ema_sd is not None: + with torch.no_grad(): + for n, p in base_model.state_dict().items(): + ema_sd[n].lerp_(p.cpu(), 1.0 - args.ema_decay) + + if args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None): + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_ms:.0f}ms step_avg:{approx_ms / step:.2f}ms") + + reached_cap = max_wallclock_ms is not None and approx_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + t_cap = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(t_cap, op=dist.ReduceOp.MAX) + reached_cap = bool(t_cap.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0(f"peak memory: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB") + + # Apply EMA or SWA + if ema_sd is not None: + base_model.load_state_dict(ema_sd, strict=True) + log0("Loaded EMA weights") + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + cur = base_model.state_dict() + avg = {name: (tensor / swa_count).to(dtype=cur[name].dtype) for name, tensor in swa_state.items()} + base_model.load_state_dict(avg, strict=True) + + # Save raw model + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + log0(f"Raw model: {os.path.getsize('final_model.pt')} bytes") + + # Magnitude pruning + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + threshold = torch.quantile(param.abs().float().flatten(), 0.05) + param.masked_fill_(param.abs() < threshold, 0.0) + + # Mixed int5/int6 quantization + compression + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize(sd_cpu, {"mlp", "attn", "bigram"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + qsize = os.path.getsize("final_model.int8.ptz") + csize = len(code.encode("utf-8")) + log0(f"Quantized model ({_COMPRESSOR}): {qsize} bytes") + log0(f"Code size: {csize} bytes") + log0(f"Total submission: {qsize + csize} bytes") + + # Roundtrip validation + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + blob = f.read() + if _COMPRESSOR == "zstd": + raw = zstandard.ZstdDecompressor().decompress(blob) + else: + raw = zlib.decompress(blob) + qs = torch.load(io.BytesIO(raw), map_location="cpu") + deq = dequantize_mixed(qs["w"], qs["m"], sd_cpu) + base_model.load_state_dict(deq, strict=True) + + # Sliding window eval + torch.cuda.synchronize() + t_eval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0(f"eval:sliding stride={args.eval_stride} batch={args.eval_batch_seqs}") + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_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.eval_batch_seqs, + ) + else: + q_val_loss, q_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, + ) + torch.cuda.synchronize() + log0(f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_eval):.0f}ms") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() + +==================================================================================================== +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:tokens:62021632 +model_params:24468561 world_size:8 grad_accum:1 +layers:10 dim:512 heads:8 kv:4 +mlp_mult:3.0 bigram:2048 seq_len:2048 +batch_tokens:786432 lr:0.02 wd:0.04 +warmup_step:10/20 +warmup_step:20/20 +step:1/20000 train_loss:6.9314 train_time:148ms step_avg:147.76ms +step:2/20000 train_loss:7.9716 train_time:210ms step_avg:105.03ms +step:3/20000 train_loss:7.5769 train_time:298ms step_avg:99.39ms +step:4/20000 train_loss:7.0420 train_time:386ms step_avg:96.39ms +step:5/20000 train_loss:6.8574 train_time:474ms step_avg:94.72ms +step:6/20000 train_loss:6.7871 train_time:562ms step_avg:93.59ms +step:7/20000 train_loss:6.6956 train_time:649ms step_avg:92.73ms +step:8/20000 train_loss:6.5858 train_time:736ms step_avg:92.06ms +step:9/20000 train_loss:6.3353 train_time:824ms step_avg:91.59ms +step:10/20000 train_loss:6.0857 train_time:912ms step_avg:91.17ms +step:100/20000 train_loss:3.2457 train_time:8939ms step_avg:89.39ms +step:200/20000 train_loss:2.4397 train_time:18041ms step_avg:90.21ms +step:300/20000 train_loss:2.5757 train_time:27133ms step_avg:90.44ms +step:400/20000 train_loss:2.4373 train_time:36209ms step_avg:90.52ms +step:500/20000 train_loss:2.4171 train_time:45186ms step_avg:90.37ms +step:600/20000 train_loss:2.3490 train_time:54280ms step_avg:90.47ms +step:700/20000 train_loss:2.3652 train_time:63393ms step_avg:90.56ms +step:800/20000 train_loss:2.2488 train_time:72524ms step_avg:90.65ms +step:900/20000 train_loss:2.1420 train_time:81630ms step_avg:90.70ms +step:1000/20000 train_loss:2.2838 train_time:90645ms step_avg:90.65ms +step:1100/20000 train_loss:2.3329 train_time:99763ms step_avg:90.69ms +step:1200/20000 train_loss:2.3654 train_time:108868ms step_avg:90.72ms +step:1300/20000 train_loss:2.1110 train_time:117974ms step_avg:90.75ms +step:1400/20000 train_loss:2.1922 train_time:127078ms step_avg:90.77ms +step:1500/20000 train_loss:2.2316 train_time:136062ms step_avg:90.71ms +step:1600/20000 train_loss:2.0873 train_time:145157ms step_avg:90.72ms +step:1700/20000 train_loss:2.1545 train_time:154247ms step_avg:90.73ms +step:1800/20000 train_loss:2.1727 train_time:163335ms step_avg:90.74ms +step:1900/20000 train_loss:2.1376 train_time:172310ms step_avg:90.69ms +step:2000/20000 train_loss:2.0774 train_time:181395ms step_avg:90.70ms +step:2100/20000 train_loss:2.0565 train_time:190481ms step_avg:90.71ms +step:2200/20000 train_loss:2.1437 train_time:199561ms step_avg:90.71ms +step:2300/20000 train_loss:2.1205 train_time:208647ms step_avg:90.72ms +step:2400/20000 train_loss:2.0742 train_time:217613ms step_avg:90.67ms +step:2500/20000 train_loss:2.1756 train_time:226692ms step_avg:90.68ms +step:2600/20000 train_loss:2.1168 train_time:235766ms step_avg:90.68ms +step:2700/20000 train_loss:2.1104 train_time:244840ms step_avg:90.68ms +step:2800/20000 train_loss:2.1631 train_time:253913ms step_avg:90.68ms +step:2900/20000 train_loss:2.0333 train_time:262871ms step_avg:90.65ms +step:3000/20000 train_loss:2.1680 train_time:271945ms step_avg:90.65ms +step:3100/20000 train_loss:2.0474 train_time:281012ms step_avg:90.65ms +step:3200/20000 train_loss:2.1820 train_time:290079ms step_avg:90.65ms +step:3300/20000 train_loss:2.0836 train_time:299035ms step_avg:90.62ms +step:3400/20000 train_loss:2.0298 train_time:308099ms step_avg:90.62ms +step:3500/20000 train_loss:2.1897 train_time:317160ms step_avg:90.62ms +step:3600/20000 train_loss:2.1097 train_time:326228ms step_avg:90.62ms +step:3700/20000 train_loss:2.1066 train_time:335301ms step_avg:90.62ms +step:3800/20000 train_loss:2.0869 train_time:344251ms step_avg:90.59ms +step:3900/20000 train_loss:2.0871 train_time:353314ms step_avg:90.59ms +step:4000/20000 train_loss:1.9841 train_time:362372ms step_avg:90.59ms +step:4100/20000 train_loss:2.0259 train_time:371443ms step_avg:90.60ms +step:4200/20000 train_loss:2.1601 train_time:380508ms step_avg:90.60ms +step:4300/20000 train_loss:2.0642 train_time:389449ms step_avg:90.57ms +step:4400/20000 train_loss:2.0404 train_time:398510ms step_avg:90.57ms +step:4500/20000 train_loss:2.1279 train_time:407575ms step_avg:90.57ms +step:4600/20000 train_loss:1.8510 train_time:416629ms step_avg:90.57ms +step:4700/20000 train_loss:2.2403 train_time:425577ms step_avg:90.55ms +step:4800/20000 train_loss:2.4376 train_time:434638ms step_avg:90.55ms +step:4900/20000 train_loss:2.0570 train_time:443701ms step_avg:90.55ms +step:5000/20000 train_loss:2.1115 train_time:452767ms step_avg:90.55ms +step:5100/20000 train_loss:2.1298 train_time:461829ms step_avg:90.55ms +step:5200/20000 train_loss:2.0506 train_time:470779ms step_avg:90.53ms +step:5300/20000 train_loss:2.0131 train_time:479840ms step_avg:90.54ms +step:5400/20000 train_loss:2.0527 train_time:488900ms step_avg:90.54ms +swa:start step:5450 +step:5500/20000 train_loss:2.0216 train_time:498039ms step_avg:90.55ms +step:5600/20000 train_loss:1.9609 train_time:507160ms step_avg:90.56ms +step:5700/20000 train_loss:2.0180 train_time:516169ms step_avg:90.56ms +step:5800/20000 train_loss:1.9997 train_time:525278ms step_avg:90.57ms +step:5900/20000 train_loss:1.9077 train_time:534389ms step_avg:90.57ms +step:6000/20000 train_loss:1.9438 train_time:543510ms step_avg:90.58ms +step:6100/20000 train_loss:1.9227 train_time:552504ms step_avg:90.57ms +step:6200/20000 train_loss:1.9537 train_time:561615ms step_avg:90.58ms +step:6300/20000 train_loss:1.9503 train_time:570721ms step_avg:90.59ms +step:6400/20000 train_loss:2.0017 train_time:579832ms step_avg:90.60ms +step:6500/20000 train_loss:2.0815 train_time:588941ms step_avg:90.61ms +step:6600/20000 train_loss:1.8457 train_time:597940ms step_avg:90.60ms +step:6622/20000 val_loss:1.9534 val_bpb:1.1569 train_time:600075ms step_avg:90.62ms +stopping_early: wallclock_cap train_time:600075ms step:6622 +peak memory: 18856 MiB +swa:applying averaged 24 checkpoints +Raw model: 96340267 bytes +Quantized model (zstd): 14944394 bytes +Code size: 46541 bytes +Total submission: 14990935 bytes +eval:sliding stride=32 batch=48 +final_int8_zlib_roundtrip val_loss:1.9357 val_bpb:1.1464 eval_time:333117ms +final_int8_zlib_roundtrip_exact val_loss:1.93571149 val_bpb:1.14644118 diff --git a/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed42.log b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed42.log new file mode 100644 index 000000000..004ab5059 --- /dev/null +++ b/records/track_10min_16mb/2026-03-21_10L_MLP3x_BigramHash_SWA_Stride32/train_seed42.log @@ -0,0 +1,1138 @@ +""" +11L MLP3x + BigramHash(2048) + EMA + SmearGate + mixed int5/int6 + stride-32 sliding eval. +EMA(0.997) replaces SWA for smoother weight averaging. Dense stride-32 eval. +""" + +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +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", 42)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + 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)) + 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", 10)) + 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", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + 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)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.04)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 32)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 48)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +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: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +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, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_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) + + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern for pattern in + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,bigram.scale".split(",") + if pattern +) +FP16_KEEP_NAME_PATTERNS = ("tok_emb",) + +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 "bigram" in name: + return "bigram" + if ".attn." in name: + return "attn" + return "other" + +def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip_range).clamp_min(1e-12).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range + 1), clip_range).to(torch.int8) + return q, scale + +def quantize_int8_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + return quantize_intN_per_row(t, clip_range=127) + +def mixed_quantize(state_dict: dict[str, Tensor], int6_cats: set[str]): + 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() <= 8192: + 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 any(p in name for p in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if cat in int6_cats and t.ndim >= 1: + clip = 15 if cat == "mlp" else 31 # int5 for MLP, int6 for attn + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{5 if clip == 15 else 6}"} + else: + q, s = quantize_int8_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed(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[name] + 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 + + +# ----------------------------- +# 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): + def forward(self, x: Tensor) -> Tensor: + return F.linear(x, self.weight.to(x.dtype), self.bias.to(x.dtype) if self.bias is not None else None) + + +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(p in name for p 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): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + 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._seq_len_cached != seq_len or self._cos_cached.device != device: + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + 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: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + 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.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + return self.proj(y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + 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: + return self.proj(torch.relu(self.fc(x)).square()) + + +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 Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, + rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + 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)) + return x + + +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: float, tie_embeddings: bool, + tied_embed_init_std: float, logit_softcap: float, rope_base: float, + qk_gain_init: float, bigram_vocab_size: int = 0, bigram_dim: int = 128): + super().__init__() + self.tie_embeddings = tie_embeddings + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.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.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ 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._init_weights(tied_embed_init_std, num_layers) + + def _init_weights(self, tied_embed_init_std: float, num_layers: int) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=tied_embed_init_std) + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + 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] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + 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) + 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, +) -> tuple[float, float]: + seq_len = 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 >= stride or ws == 0] + 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() + 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 = 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 = 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 + + +# ----------------------------- +# TRAINING +# ----------------------------- + +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 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + 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) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:tokens:{val_tokens.numel() - 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, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + ).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(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + scalar_params = [p for name, p in block_named_params + if p.ndim < 2 or any(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW(tok_params, betas=(args.beta1, args.beta2), + eps=args.adam_eps, weight_decay=args.weight_decay, fused=True) + optimizer_muon = Muon(matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.weight_decay) + 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.weight_decay, fused=True) + optimizers = [optimizer_tok, optimizer_muon, optimizer_scalar] + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params} world_size:{world_size} grad_accum:{grad_accum_steps}") + log0(f"layers:{args.num_layers} dim:{args.model_dim} heads:{args.num_heads} kv:{args.num_kv_heads}") + log0(f"mlp_mult:{args.mlp_mult} bigram:{args.bigram_vocab_size} seq_len:{args.train_seq_len}") + log0(f"batch_tokens:{args.train_batch_tokens} lr:{args.matrix_lr} wd:{args.weight_decay}") + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all(): + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup + if args.warmup_steps > 0: + initial_model_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for ws in range(args.warmup_steps): + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + wl = model(x, y) + (wl * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if ws + 1 == args.warmup_steps or (ws + 1) % 10 == 0: + log0(f"warmup_step:{ws + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, st in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(st) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # Training loop + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_sd = {n: t.detach().cpu().clone() for n, t in base_model.state_dict().items()} if args.ema_enabled else 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 step:{step}") + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for ms in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = ms == 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): + 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_mom = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_mom + 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_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + # SWA + if args.swa_enabled and scale < args.swa_start_frac 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 + + # EMA update + if ema_sd is not None: + with torch.no_grad(): + for n, p in base_model.state_dict().items(): + ema_sd[n].lerp_(p.cpu(), 1.0 - args.ema_decay) + + if args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None): + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_ms:.0f}ms step_avg:{approx_ms / step:.2f}ms") + + reached_cap = max_wallclock_ms is not None and approx_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + t_cap = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(t_cap, op=dist.ReduceOp.MAX) + reached_cap = bool(t_cap.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0(f"peak memory: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB") + + # Apply EMA or SWA + if ema_sd is not None: + base_model.load_state_dict(ema_sd, strict=True) + log0("Loaded EMA weights") + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + cur = base_model.state_dict() + avg = {name: (tensor / swa_count).to(dtype=cur[name].dtype) for name, tensor in swa_state.items()} + base_model.load_state_dict(avg, strict=True) + + # Save raw model + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + log0(f"Raw model: {os.path.getsize('final_model.pt')} bytes") + + # Magnitude pruning + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + threshold = torch.quantile(param.abs().float().flatten(), 0.05) + param.masked_fill_(param.abs() < threshold, 0.0) + + # Mixed int5/int6 quantization + compression + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize(sd_cpu, {"mlp", "attn", "bigram"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + qsize = os.path.getsize("final_model.int8.ptz") + csize = len(code.encode("utf-8")) + log0(f"Quantized model ({_COMPRESSOR}): {qsize} bytes") + log0(f"Code size: {csize} bytes") + log0(f"Total submission: {qsize + csize} bytes") + + # Roundtrip validation + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + blob = f.read() + if _COMPRESSOR == "zstd": + raw = zstandard.ZstdDecompressor().decompress(blob) + else: + raw = zlib.decompress(blob) + qs = torch.load(io.BytesIO(raw), map_location="cpu") + deq = dequantize_mixed(qs["w"], qs["m"], sd_cpu) + base_model.load_state_dict(deq, strict=True) + + # Sliding window eval + torch.cuda.synchronize() + t_eval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0(f"eval:sliding stride={args.eval_stride} batch={args.eval_batch_seqs}") + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_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.eval_batch_seqs, + ) + else: + q_val_loss, q_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, + ) + torch.cuda.synchronize() + log0(f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_eval):.0f}ms") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() + +==================================================================================================== +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:tokens:62021632 +model_params:24468561 world_size:8 grad_accum:1 +layers:10 dim:512 heads:8 kv:4 +mlp_mult:3.0 bigram:2048 seq_len:2048 +batch_tokens:786432 lr:0.02 wd:0.04 +warmup_step:10/20 +warmup_step:20/20 +step:1/20000 train_loss:6.9326 train_time:204ms step_avg:204.44ms +step:2/20000 train_loss:7.9320 train_time:267ms step_avg:133.63ms +step:3/20000 train_loss:7.5438 train_time:355ms step_avg:118.22ms +step:4/20000 train_loss:6.9530 train_time:443ms step_avg:110.69ms +step:5/20000 train_loss:6.8135 train_time:530ms step_avg:106.08ms +step:6/20000 train_loss:6.7152 train_time:619ms step_avg:103.17ms +step:7/20000 train_loss:6.6518 train_time:707ms step_avg:100.94ms +step:8/20000 train_loss:6.6457 train_time:794ms step_avg:99.26ms +step:9/20000 train_loss:6.2991 train_time:882ms step_avg:97.96ms +step:10/20000 train_loss:6.0619 train_time:971ms step_avg:97.13ms +step:100/20000 train_loss:3.2199 train_time:8991ms step_avg:89.91ms +step:200/20000 train_loss:2.4313 train_time:18044ms step_avg:90.22ms +step:300/20000 train_loss:2.5774 train_time:27172ms step_avg:90.57ms +step:400/20000 train_loss:2.4403 train_time:36232ms step_avg:90.58ms +step:500/20000 train_loss:2.4225 train_time:45205ms step_avg:90.41ms +step:600/20000 train_loss:2.3493 train_time:54297ms step_avg:90.49ms +step:700/20000 train_loss:2.3655 train_time:63398ms step_avg:90.57ms +step:800/20000 train_loss:2.2518 train_time:72507ms step_avg:90.63ms +step:900/20000 train_loss:2.1435 train_time:81602ms step_avg:90.67ms +step:1000/20000 train_loss:2.2878 train_time:90608ms step_avg:90.61ms +step:1100/20000 train_loss:2.3325 train_time:99719ms step_avg:90.65ms +step:1200/20000 train_loss:2.3671 train_time:108826ms step_avg:90.69ms +step:1300/20000 train_loss:2.1136 train_time:117932ms step_avg:90.72ms +step:1400/20000 train_loss:2.2022 train_time:127025ms step_avg:90.73ms +step:1500/20000 train_loss:2.2341 train_time:136000ms step_avg:90.67ms +step:1600/20000 train_loss:2.0884 train_time:145105ms step_avg:90.69ms +step:1700/20000 train_loss:2.1588 train_time:154201ms step_avg:90.71ms +step:1800/20000 train_loss:2.1772 train_time:163273ms step_avg:90.71ms +step:1900/20000 train_loss:2.1417 train_time:172248ms step_avg:90.66ms +step:2000/20000 train_loss:2.0803 train_time:181325ms step_avg:90.66ms +step:2100/20000 train_loss:2.0598 train_time:190396ms step_avg:90.66ms +step:2200/20000 train_loss:2.1475 train_time:199467ms step_avg:90.67ms +step:2300/20000 train_loss:2.1206 train_time:208539ms step_avg:90.67ms +step:2400/20000 train_loss:2.0780 train_time:217496ms step_avg:90.62ms +step:2500/20000 train_loss:2.1835 train_time:226564ms step_avg:90.63ms +step:2600/20000 train_loss:2.1190 train_time:235619ms step_avg:90.62ms +step:2700/20000 train_loss:2.1122 train_time:244683ms step_avg:90.62ms +step:2800/20000 train_loss:2.1657 train_time:253739ms step_avg:90.62ms +step:2900/20000 train_loss:2.0352 train_time:262685ms step_avg:90.58ms +step:3000/20000 train_loss:2.1706 train_time:271741ms step_avg:90.58ms +step:3100/20000 train_loss:2.0501 train_time:280804ms step_avg:90.58ms +step:3200/20000 train_loss:2.1837 train_time:289858ms step_avg:90.58ms +step:3300/20000 train_loss:2.0836 train_time:298804ms step_avg:90.55ms +step:3400/20000 train_loss:2.0347 train_time:307860ms step_avg:90.55ms +step:3500/20000 train_loss:2.1934 train_time:316924ms step_avg:90.55ms +step:3600/20000 train_loss:2.1113 train_time:325981ms step_avg:90.55ms +step:3700/20000 train_loss:2.1073 train_time:335039ms step_avg:90.55ms +step:3800/20000 train_loss:2.0874 train_time:343984ms step_avg:90.52ms +step:3900/20000 train_loss:2.0917 train_time:353049ms step_avg:90.53ms +step:4000/20000 train_loss:1.9850 train_time:362105ms step_avg:90.53ms +step:4100/20000 train_loss:2.0279 train_time:371154ms step_avg:90.53ms +step:4200/20000 train_loss:2.1685 train_time:380219ms step_avg:90.53ms +step:4300/20000 train_loss:2.0671 train_time:389164ms step_avg:90.50ms +step:4400/20000 train_loss:2.0442 train_time:398218ms step_avg:90.50ms +step:4500/20000 train_loss:2.1309 train_time:407276ms step_avg:90.51ms +step:4600/20000 train_loss:1.8529 train_time:416324ms step_avg:90.51ms +step:4700/20000 train_loss:2.2418 train_time:425265ms step_avg:90.48ms +step:4800/20000 train_loss:2.4398 train_time:434324ms step_avg:90.48ms +step:4900/20000 train_loss:2.0582 train_time:443392ms step_avg:90.49ms +step:5000/20000 train_loss:2.1145 train_time:452450ms step_avg:90.49ms +step:5100/20000 train_loss:2.1342 train_time:461512ms step_avg:90.49ms +step:5200/20000 train_loss:2.0494 train_time:470463ms step_avg:90.47ms +step:5300/20000 train_loss:2.0145 train_time:479525ms step_avg:90.48ms +step:5400/20000 train_loss:2.0567 train_time:488581ms step_avg:90.48ms +swa:start step:5450 +step:5500/20000 train_loss:2.0268 train_time:497708ms step_avg:90.49ms +step:5600/20000 train_loss:1.9651 train_time:506814ms step_avg:90.50ms +step:5700/20000 train_loss:2.0194 train_time:515820ms step_avg:90.49ms +step:5800/20000 train_loss:2.0035 train_time:524949ms step_avg:90.51ms +step:5900/20000 train_loss:1.9123 train_time:534056ms step_avg:90.52ms +step:6000/20000 train_loss:1.9477 train_time:543171ms step_avg:90.53ms +step:6100/20000 train_loss:1.9215 train_time:552171ms step_avg:90.52ms +step:6200/20000 train_loss:1.9575 train_time:561275ms step_avg:90.53ms +step:6300/20000 train_loss:1.9507 train_time:570388ms step_avg:90.54ms +step:6400/20000 train_loss:2.0064 train_time:579498ms step_avg:90.55ms +step:6500/20000 train_loss:2.0854 train_time:588604ms step_avg:90.55ms +step:6600/20000 train_loss:1.8494 train_time:597590ms step_avg:90.54ms +step:6626/20000 val_loss:1.9560 val_bpb:1.1585 train_time:600077ms step_avg:90.56ms +stopping_early: wallclock_cap train_time:600077ms step:6626 +peak memory: 18856 MiB +swa:applying averaged 24 checkpoints +Raw model: 96340267 bytes +Quantized model (zstd): 14728517 bytes +Code size: 46541 bytes +Total submission: 14775058 bytes +eval:sliding stride=32 batch=48 +final_int8_zlib_roundtrip val_loss:1.9405 val_bpb:1.1493 eval_time:333333ms +final_int8_zlib_roundtrip_exact val_loss:1.94048786 val_bpb:1.14927002