diff --git a/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/56869d79-f695-4ef4-a7ee-2ff630965978.txt b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/56869d79-f695-4ef4-a7ee-2ff630965978.txt new file mode 100644 index 0000000000..f925fe4e15 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/56869d79-f695-4ef4-a7ee-2ff630965978.txt @@ -0,0 +1,1236 @@ +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 +import concurrent.futures +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 5000)) + 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", 2048)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 512)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 2.125)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_max_base = float(os.environ.get("ROPE_MAX_BASE", 8192.0)) + rope_min_base = float(os.environ.get("ROPE_MIN_BASE", 512.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.3)) + head_lr = float(os.environ.get("HEAD_LR", 0.04)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.125)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 256)) + 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", 1.0)) +@torch.compile(mode="max-autotune-no-cudagraphs", fullgraph=True) +def zeropower_via_newtonschulz5(G, steps=5, eps=1e-7): + X = torch.where(torch.isfinite(G), G, 0.0).to(torch.bfloat16) + norm = torch.linalg.matrix_norm(X) + eps + X = X / norm + transposed = X.size(0) > X.size(1) + if transposed: + X = X.T + a, b, c = 3.4445, -4.7750, 2.0315 + for _ in range(steps): + A = X @ X.T + B = torch.addmm(A, A, A, beta=b, alpha=c) + X = torch.addmm(X, B, X, beta=a, alpha=1.0) + return X.T if transposed else X +@torch.no_grad() +def get_hadamard_matrix(n, device): + p2 = 2**math.ceil(math.log2(n)) + H = torch.tensor([[1.0]], device=device) + while H.shape[0] < p2: + H = torch.cat([torch.cat([H, H], dim=1), torch.cat([H, -H], dim=1)], dim=0) + return H[:n, :n] / math.sqrt(p2) +@torch.no_grad() +def apply_zero_init(model, std=0.02): + for m in model.modules(): + if isinstance(m, nn.Embedding): + d_out, d_in = m.weight.shape + H = get_hadamard_matrix(max(d_out, d_in), m.weight.device) + m.weight.copy_(H[:d_out, :d_in]*std) + linears = [sub for sub in m.children() if isinstance(sub, nn.Linear)] + if linears: + for i, l in enumerate(linears): + d_out, d_in = l.weight.shape + if i == len(linears) - 1: + nn.init.normal_(l.weight, std=1e-5) + else: + H = get_hadamard_matrix(max(d_out, d_in), l.weight.device) + l.weight.copy_(H[:d_out, :d_in]) + if l.bias is not None: + nn.init.zeros_(l.bias) +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float = 0.95, backend_steps: int = 5, nesterov: bool = True): + defaults = dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov) + super().__init__(params, defaults) + self._is_initialized = False + def _init_state(self): + 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: + if "updates_flat" in group or not group["params"]: + continue + params = group["params"] + total_params = sum(p.numel() for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + group["updates_flat"] = updates_flat + group["param_views"] = [] + group["rank_params"] = [] + group["rank_param_views"] = [] + curr = 0 + for i, p in enumerate(params): + numel = p.numel() + view = updates_flat[curr : curr + numel].view_as(p) + group["param_views"].append(view) + if i % world_size == rank: + group["rank_params"].append(p) + group["rank_param_views"].append(view) + curr += numel + @torch.no_grad() + def step(self, closure=None): + loss = closure() if closure is not None else None + self._init_state() + for group in self.param_groups: + if not group["params"]: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + updates_flat = group["updates_flat"] + updates_flat.zero_() + for p, view in zip(group["rank_params"], group["rank_param_views"]): + if p.grad is None: + continue + 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) + u = g.add(buf, alpha=momentum) if nesterov else buf + original_shape = u.shape + if u.ndim > 2: + u = u.view(u.size(0), -1) + u_prepared = u.to(dtype=torch.bfloat16, memory_format=torch.contiguous_format, non_blocking=True) + g_ns = zeropower_via_newtonschulz5(u_prepared, steps=backend_steps) + g_ns.mul_(max(1, g_ns.size(0) / g_ns.size(1))**0.5) + view.copy_(g_ns.view(original_shape)) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + params = group["params"] + views = group["param_views"] + if updates_flat.dtype != params[0].dtype: + views = [v.to(dtype=p.dtype, non_blocking=True) for v, p in zip(views, params)] + torch._foreach_add_(params, views, alpha=-lr) + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = self.future_tokens.result() + self.pos = 0 + self.future_tokens = self._queue_next_file() + 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, self.world_size, self.device = rank, world_size, device + self.stream = TokenStream(pattern) + self.next_x, self.next_y = None, None + self.transfer_stream = torch.cuda.Stream(device) + def preload(self, global_tokens: int, seq_len: int, grad_accum_steps: int): + with torch.cuda.stream(self.transfer_stream): + 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].pin_memory().to( + device=self.device, dtype=torch.int32, non_blocking=True + ) + self.next_x = local[:-1].reshape(-1, seq_len) + self.next_y = local[1:].reshape(-1, seq_len).to(torch.long) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + if self.next_x is None: + self.preload(global_tokens, seq_len, grad_accum_steps) + torch.cuda.current_stream().wait_stream(self.transfer_stream) + x, y = self.next_x, self.next_y + self.preload(global_tokens, seq_len, grad_accum_steps) + return x, y +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) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, max_seq_len: int, p: float = 0.5, base: float = 10000.0): + super().__init__() + self.rotary_dim = (int(dim * p) // 8) * 8 + inv_freq = 1.0 / (base ** (torch.arange(0, self.rotary_dim, 2, dtype=torch.float32) / self.rotary_dim)) + t = torch.arange(max_seq_len, dtype=torch.float32) + freqs = torch.outer(t, inv_freq) + self.register_buffer("cos", freqs.cos().view(1, 1, max_seq_len, -1), persistent=False) + self.register_buffer("sin", freqs.sin().view(1, 1, max_seq_len, -1), persistent=False) + def forward(self, x: Tensor): + t = x.size(2) + return self.cos[:, :, :t, :].to(x.dtype), self.sin[:, :, :t, :].to(x.dtype) +def rotate_half(x: Tensor): + x1, x2 = x.chunk(2, dim=-1) + return torch.cat((-x2, x1), dim=-1) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + d = cos.shape[-1] * 2 + x_rop = x[..., :d] + x_pass = x[..., d:] + x_rop = x_rop.view(*x_rop.shape[:-1], -1, 2) + x0 = x_rop[..., 0] + x1 = x_rop[..., 1] + res = torch.stack([x0 * cos - x1 * sin, x1 * cos + x0 * sin], dim=-1) + return torch.cat([res.flatten(-2), x_pass], dim=-1) +class CausalSelfAttention(nn.Module): + def __init__(self, dim, num_heads, num_kv_heads, rope_base, qk_gain_init, seq_len=1024, use_rope=True, rope_proportion=0.5): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + self.kv_dim = self.num_kv_heads * self.head_dim + self.c_qkv = nn.Linear(dim, dim + 2 * self.kv_dim, bias=False) + self.proj = nn.Linear(dim, dim, bias=False) + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, max_seq_len=seq_len, p=rope_proportion, base=rope_base) + self.use_rope = use_rope + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + qkv = self.c_qkv(x) + q, k, v = qkv.split([dim, self.kv_dim, self.kv_dim], dim=-1) + q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = k.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = v.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),)) + if self.use_rope: + cos, sin = self.rotary(q) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.view(1, -1, 1, 1).to(q.dtype) + y = F.scaled_dot_product_attention( + q, k, v, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads) + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + super().__init__() + self.hidden_dim = int(mlp_mult * 2/3 * dim // 64) * 64 + self.c_fc = nn.Linear(dim, 2 * self.hidden_dim, bias=False) + self.c_proj = nn.Linear(self.hidden_dim, dim, bias=False) + def forward(self, x: Tensor) -> Tensor: + fused_x = self.c_fc(x) + gate, value = fused_x.chunk(2, dim=-1) + return self.c_proj(F.silu(gate) * value) +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, + seq_len: int=1024, + use_rope: bool=True, + rope_proportion: float=0.5, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, seq_len, use_rope, rope_proportion) + self.mlp = MLP(dim, mlp_mult) + def forward(self, x: Tensor) -> Tensor: + x = x + self.attn(self.attn_norm(x)) + x = x + self.mlp(self.mlp_norm(x)) + return x +def get_linear_progression_kv_heads(layer_idx, total_layers, num_heads): + min_kv = 1 + max_kv = num_heads + fraction = layer_idx / (total_layers - 1) + raw_kv = min_kv + (max_kv - min_kv) * fraction + valid_kvs = [i for i in range(1, num_heads + 1) if num_heads % i == 0] + kv_heads = min(valid_kvs, key=lambda x: abs(x - raw_kv)) + return int(max(min_kv, kv_heads)) +def get_rope_p_smooth(i: int, num_layers: int, p_min=0.25, p_max=0.75) -> float: + if num_layers <= 1: + return p_min + progress = i / (num_layers - 1) + scale = math.sin(progress * math.pi) + return p_min + (p_max - p_min) * scale +def get_rope_base_progression(layer_idx: int, total_layers: int, min_base: float, max_base: float) -> float: + if total_layers <= 1: + return max_base + fraction = layer_idx / (total_layers - 1) + return min_base * (max_base / min_base) ** fraction +def get_linear_progression_mlp_mult(layer_idx: int, total_layers: int, base_mult: int) -> float: + min_mult = float(base_mult) * 0.75 + max_mult = float(base_mult) * 1.25 + fraction = layer_idx / max(1, total_layers - 1) + return min_mult + (max_mult - min_mult) * fraction +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_max_base: float, + rope_min_base: float, + qk_gain_init: float, + seq_len: int=1024, + ): + super().__init__() + self.tie_embeddings = tie_embeddings + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.blocks = nn.ModuleList([ + Block( + model_dim, + num_heads, + get_linear_progression_kv_heads(i, num_layers, num_kv_heads), + get_linear_progression_mlp_mult(i, num_layers, mlp_mult), + get_rope_base_progression(i, num_layers, rope_min_base, rope_max_base), + qk_gain_init, + seq_len=seq_len, + use_rope=True, + rope_proportion=get_rope_p_smooth(i, num_layers) + ) for i in range(num_layers) + ]) + self.final_norm = RMSNorm() + self.logit_softcap = logit_softcap + self.lm_head = None if tie_embeddings else nn.Linear(model_dim, vocab_size, bias=False) + apply_zero_init(self, std=tied_embed_init_std) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + emb = self.tok_emb(input_ids) + x = F.rms_norm(emb, (emb.size(-1),)) + for block in self.blocks: + x = block(x) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits = F.linear(x, self.tok_emb.weight) + else: + logits = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits.float() / self.logit_softcap) + return F.cross_entropy(logits, targets) +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + torch.set_float32_matmul_precision('high') + import torch._inductor.config as inductor_config + inductor_config.fx_graph_cache = True + inductor_config.triton.unique_kernel_names = True + inductor_config.freezing = True + inductor_config.shape_padding = True + inductor_config.epilogue_fusion = True + 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) + @torch.compiler.disable + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + 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_max_base=args.rope_max_base, + rope_min_base=args.rope_min_base, + qk_gain_init=args.qk_gain_init, + seq_len=args.train_seq_len + ).to(device).bfloat16() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, fullgraph=True, mode="reduce-overhead") + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False, gradient_as_bucket_view=True) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + gpt_scalars = [p for name, p in base_model.named_parameters() if "skip" in name] + scalar_params.extend(gpt_scalars) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.11.0+cu130 +Fri May 1 06:22:08 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 43C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 37C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 35C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 43C P0 125W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 45C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 37C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 43C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 35C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:16318536 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.03 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:512 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:10/512 +warmup_step:20/512 +warmup_step:30/512 +warmup_step:40/512 +warmup_step:50/512 +warmup_step:60/512 +warmup_step:70/512 +warmup_step:80/512 +warmup_step:90/512 +warmup_step:100/512 +warmup_step:110/512 +warmup_step:120/512 +warmup_step:130/512 +warmup_step:140/512 +warmup_step:150/512 +warmup_step:160/512 +warmup_step:170/512 +warmup_step:180/512 +warmup_step:190/512 +warmup_step:200/512 +warmup_step:210/512 +warmup_step:220/512 +warmup_step:230/512 +warmup_step:240/512 +warmup_step:250/512 +warmup_step:260/512 +warmup_step:270/512 +warmup_step:280/512 +warmup_step:290/512 +warmup_step:300/512 +warmup_step:310/512 +warmup_step:320/512 +warmup_step:330/512 +warmup_step:340/512 +warmup_step:350/512 +warmup_step:360/512 +warmup_step:370/512 +warmup_step:380/512 +warmup_step:390/512 +warmup_step:400/512 +warmup_step:410/512 +warmup_step:420/512 +warmup_step:430/512 +warmup_step:440/512 +warmup_step:450/512 +warmup_step:460/512 +warmup_step:470/512 +warmup_step:480/512 +warmup_step:490/512 +warmup_step:500/512 +warmup_step:510/512 +warmup_step:512/512 +step:0/20000 val_loss:6.9326 val_bpb:4.1059 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9321 train_time:96ms step_avg:96.29ms +step:2/20000 train_loss:12.7577 train_time:108ms step_avg:53.86ms +step:3/20000 train_loss:9.7676 train_time:146ms step_avg:48.78ms +step:4/20000 train_loss:7.0821 train_time:183ms step_avg:45.79ms +step:5/20000 train_loss:6.3340 train_time:221ms step_avg:44.14ms +step:6/20000 train_loss:6.9930 train_time:258ms step_avg:43.00ms +step:7/20000 train_loss:6.4102 train_time:295ms step_avg:42.19ms +step:8/20000 train_loss:6.2008 train_time:333ms step_avg:41.62ms +step:9/20000 train_loss:5.8136 train_time:368ms step_avg:40.86ms +step:10/20000 train_loss:5.5661 train_time:404ms step_avg:40.42ms +step:100/20000 train_loss:3.3085 train_time:3832ms step_avg:38.32ms +step:200/20000 train_loss:2.8155 train_time:7645ms step_avg:38.23ms +step:300/20000 train_loss:2.4686 train_time:11462ms step_avg:38.21ms +step:400/20000 train_loss:2.3356 train_time:15294ms step_avg:38.23ms +step:500/20000 train_loss:2.4812 train_time:19125ms step_avg:38.25ms +step:600/20000 train_loss:2.5382 train_time:22957ms step_avg:38.26ms +step:700/20000 train_loss:2.4505 train_time:26789ms step_avg:38.27ms +step:800/20000 train_loss:2.3076 train_time:30627ms step_avg:38.28ms +step:900/20000 train_loss:2.3598 train_time:34462ms step_avg:38.29ms +step:1000/20000 train_loss:2.3954 train_time:38301ms step_avg:38.30ms +step:1100/20000 train_loss:2.2898 train_time:42136ms step_avg:38.31ms +step:1200/20000 train_loss:2.4160 train_time:45976ms step_avg:38.31ms +step:1300/20000 train_loss:2.3857 train_time:49816ms step_avg:38.32ms +step:1400/20000 train_loss:2.4578 train_time:53655ms step_avg:38.33ms +step:1500/20000 train_loss:2.2588 train_time:57498ms step_avg:38.33ms +step:1600/20000 train_loss:2.1302 train_time:61350ms step_avg:38.34ms +step:1700/20000 train_loss:2.2090 train_time:65194ms step_avg:38.35ms +step:1800/20000 train_loss:2.2405 train_time:69038ms step_avg:38.35ms +step:1900/20000 train_loss:2.2255 train_time:72889ms step_avg:38.36ms +step:2000/20000 train_loss:2.2935 train_time:76735ms step_avg:38.37ms +step:2100/20000 train_loss:2.3010 train_time:80584ms step_avg:38.37ms +step:2200/20000 train_loss:2.1174 train_time:84447ms step_avg:38.38ms +step:2300/20000 train_loss:2.4189 train_time:88291ms step_avg:38.39ms +step:2400/20000 train_loss:2.2397 train_time:92140ms step_avg:38.39ms +step:2500/20000 train_loss:2.1675 train_time:95981ms step_avg:38.39ms +step:2600/20000 train_loss:2.4472 train_time:99816ms step_avg:38.39ms +step:2700/20000 train_loss:2.1975 train_time:103656ms step_avg:38.39ms +step:2800/20000 train_loss:2.2769 train_time:107490ms step_avg:38.39ms +step:2900/20000 train_loss:2.2256 train_time:111335ms step_avg:38.39ms +step:3000/20000 train_loss:2.2668 train_time:115172ms step_avg:38.39ms +step:3100/20000 train_loss:2.2495 train_time:119006ms step_avg:38.39ms +step:3200/20000 train_loss:2.2295 train_time:122836ms step_avg:38.39ms +step:3300/20000 train_loss:2.2828 train_time:126673ms step_avg:38.39ms +step:3400/20000 train_loss:2.2037 train_time:130507ms step_avg:38.38ms +step:3500/20000 train_loss:2.2975 train_time:134337ms step_avg:38.38ms +step:3600/20000 train_loss:2.1646 train_time:138169ms step_avg:38.38ms +step:3700/20000 train_loss:2.1945 train_time:142005ms step_avg:38.38ms +step:3800/20000 train_loss:2.2685 train_time:145833ms step_avg:38.38ms +step:3900/20000 train_loss:2.0385 train_time:149664ms step_avg:38.38ms +step:4000/20000 train_loss:2.2095 train_time:153493ms step_avg:38.37ms +step:4100/20000 train_loss:2.2325 train_time:157327ms step_avg:38.37ms +step:4200/20000 train_loss:2.2192 train_time:161158ms step_avg:38.37ms +step:4300/20000 train_loss:2.0673 train_time:164984ms step_avg:38.37ms +step:4400/20000 train_loss:2.1623 train_time:168813ms step_avg:38.37ms +step:4500/20000 train_loss:2.3004 train_time:172642ms step_avg:38.36ms +step:4600/20000 train_loss:2.0217 train_time:176472ms step_avg:38.36ms +step:4700/20000 train_loss:2.3170 train_time:180297ms step_avg:38.36ms +step:4800/20000 train_loss:2.3109 train_time:184125ms step_avg:38.36ms +step:4900/20000 train_loss:2.2106 train_time:187953ms step_avg:38.36ms +step:5000/20000 train_loss:2.0821 train_time:191778ms step_avg:38.36ms +step:5000/20000 val_loss:2.2000 val_bpb:1.3030 train_time:191807ms step_avg:38.36ms +step:5100/20000 train_loss:2.0779 train_time:195601ms step_avg:38.35ms +step:5200/20000 train_loss:2.2210 train_time:199425ms step_avg:38.35ms +step:5300/20000 train_loss:2.2597 train_time:203248ms step_avg:38.35ms +step:5400/20000 train_loss:2.2354 train_time:207075ms step_avg:38.35ms +step:5500/20000 train_loss:2.1865 train_time:210897ms step_avg:38.34ms +step:5600/20000 train_loss:2.2334 train_time:214723ms step_avg:38.34ms +step:5700/20000 train_loss:2.2238 train_time:218545ms step_avg:38.34ms +step:5800/20000 train_loss:2.1977 train_time:222373ms step_avg:38.34ms +step:5900/20000 train_loss:2.1414 train_time:226194ms step_avg:38.34ms +step:6000/20000 train_loss:2.2692 train_time:230018ms step_avg:38.34ms +step:6100/20000 train_loss:2.1666 train_time:233843ms step_avg:38.33ms +step:6200/20000 train_loss:2.1356 train_time:237665ms step_avg:38.33ms +step:6300/20000 train_loss:2.0845 train_time:241489ms step_avg:38.33ms +step:6400/20000 train_loss:2.2149 train_time:245315ms step_avg:38.33ms +step:6500/20000 train_loss:2.1277 train_time:249139ms step_avg:38.33ms +step:6600/20000 train_loss:2.1770 train_time:252960ms step_avg:38.33ms +step:6700/20000 train_loss:2.2093 train_time:256789ms step_avg:38.33ms +step:6800/20000 train_loss:2.2402 train_time:260611ms step_avg:38.33ms +step:6900/20000 train_loss:2.1462 train_time:264433ms step_avg:38.32ms +step:7000/20000 train_loss:2.2669 train_time:268252ms step_avg:38.32ms +step:7100/20000 train_loss:2.1219 train_time:272076ms step_avg:38.32ms +step:7200/20000 train_loss:2.2491 train_time:275902ms step_avg:38.32ms +step:7300/20000 train_loss:2.1476 train_time:279726ms step_avg:38.32ms +step:7400/20000 train_loss:2.1668 train_time:283547ms step_avg:38.32ms +step:7500/20000 train_loss:2.1642 train_time:287369ms step_avg:38.32ms +step:7600/20000 train_loss:2.0476 train_time:291191ms step_avg:38.31ms +step:7700/20000 train_loss:2.1384 train_time:295013ms step_avg:38.31ms +step:7800/20000 train_loss:2.1988 train_time:298835ms step_avg:38.31ms +step:7900/20000 train_loss:2.1842 train_time:302657ms step_avg:38.31ms +step:8000/20000 train_loss:2.1684 train_time:306481ms step_avg:38.31ms +step:8100/20000 train_loss:2.1974 train_time:310303ms step_avg:38.31ms +step:8200/20000 train_loss:2.2376 train_time:314135ms step_avg:38.31ms +step:8300/20000 train_loss:2.1766 train_time:317962ms step_avg:38.31ms +step:8400/20000 train_loss:2.1880 train_time:321787ms step_avg:38.31ms +step:8500/20000 train_loss:2.1806 train_time:325605ms step_avg:38.31ms +step:8600/20000 train_loss:2.1877 train_time:329429ms step_avg:38.31ms +step:8700/20000 train_loss:2.0807 train_time:333247ms step_avg:38.30ms +step:8800/20000 train_loss:2.1613 train_time:337070ms step_avg:38.30ms +step:8900/20000 train_loss:2.2540 train_time:340886ms step_avg:38.30ms +step:9000/20000 train_loss:2.0775 train_time:344708ms step_avg:38.30ms +step:9100/20000 train_loss:2.3561 train_time:348527ms step_avg:38.30ms +step:9200/20000 train_loss:2.1404 train_time:352349ms step_avg:38.30ms +step:9300/20000 train_loss:2.1881 train_time:356167ms step_avg:38.30ms +step:9400/20000 train_loss:2.1932 train_time:359989ms step_avg:38.30ms +step:9500/20000 train_loss:2.3075 train_time:363807ms step_avg:38.30ms +step:9600/20000 train_loss:2.1974 train_time:367631ms step_avg:38.29ms +step:9700/20000 train_loss:2.1641 train_time:371446ms step_avg:38.29ms +step:9800/20000 train_loss:2.1332 train_time:375266ms step_avg:38.29ms +step:9900/20000 train_loss:2.1992 train_time:379078ms step_avg:38.29ms +step:10000/20000 train_loss:2.1724 train_time:382904ms step_avg:38.29ms +step:10000/20000 val_loss:2.1669 val_bpb:1.2834 train_time:382933ms step_avg:38.29ms +step:10100/20000 train_loss:2.1487 train_time:386718ms step_avg:38.29ms +step:10200/20000 train_loss:2.1234 train_time:390539ms step_avg:38.29ms +step:10300/20000 train_loss:2.2396 train_time:394413ms step_avg:38.29ms +step:10400/20000 train_loss:2.1579 train_time:398257ms step_avg:38.29ms +step:10500/20000 train_loss:2.0369 train_time:402076ms step_avg:38.29ms +step:10600/20000 train_loss:2.0318 train_time:405895ms step_avg:38.29ms +step:10700/20000 train_loss:2.1239 train_time:409714ms step_avg:38.29ms +step:10800/20000 train_loss:2.2393 train_time:413529ms step_avg:38.29ms +step:10900/20000 train_loss:2.2075 train_time:417349ms step_avg:38.29ms +step:11000/20000 train_loss:2.1734 train_time:421168ms step_avg:38.29ms +step:11100/20000 train_loss:2.1196 train_time:424996ms step_avg:38.29ms +step:11200/20000 train_loss:2.1242 train_time:428815ms step_avg:38.29ms +step:11300/20000 train_loss:2.0554 train_time:432636ms step_avg:38.29ms +step:11400/20000 train_loss:2.1112 train_time:436455ms step_avg:38.29ms +step:11500/20000 train_loss:2.1653 train_time:440278ms step_avg:38.29ms +step:11600/20000 train_loss:2.1158 train_time:444098ms step_avg:38.28ms +step:11700/20000 train_loss:2.2628 train_time:447919ms step_avg:38.28ms +step:11800/20000 train_loss:2.1500 train_time:451741ms step_avg:38.28ms +step:11900/20000 train_loss:2.1074 train_time:455568ms step_avg:38.28ms +step:12000/20000 train_loss:2.1282 train_time:459388ms step_avg:38.28ms +step:12100/20000 train_loss:2.1586 train_time:463213ms step_avg:38.28ms +step:12200/20000 train_loss:2.2711 train_time:467030ms step_avg:38.28ms +step:12300/20000 train_loss:2.1612 train_time:470852ms step_avg:38.28ms +step:12400/20000 train_loss:1.9223 train_time:474675ms step_avg:38.28ms +step:12500/20000 train_loss:2.4385 train_time:478492ms step_avg:38.28ms +step:12600/20000 train_loss:2.1458 train_time:482311ms step_avg:38.28ms +step:12700/20000 train_loss:2.1435 train_time:486131ms step_avg:38.28ms +step:12800/20000 train_loss:2.1655 train_time:489949ms step_avg:38.28ms +step:12900/20000 train_loss:2.2066 train_time:493769ms step_avg:38.28ms +step:13000/20000 train_loss:2.2481 train_time:497594ms step_avg:38.28ms +step:13100/20000 train_loss:2.0823 train_time:501411ms step_avg:38.28ms +step:13200/20000 train_loss:2.2604 train_time:505241ms step_avg:38.28ms +step:13300/20000 train_loss:2.1139 train_time:509060ms step_avg:38.28ms +step:13400/20000 train_loss:2.1364 train_time:512883ms step_avg:38.27ms +step:13500/20000 train_loss:2.0993 train_time:516702ms step_avg:38.27ms +step:13600/20000 train_loss:2.0107 train_time:520525ms step_avg:38.27ms +step:13700/20000 train_loss:2.1780 train_time:524342ms step_avg:38.27ms +step:13800/20000 train_loss:2.0868 train_time:528172ms step_avg:38.27ms +step:13900/20000 train_loss:2.1781 train_time:532013ms step_avg:38.27ms +step:14000/20000 train_loss:2.1436 train_time:535841ms step_avg:38.27ms +step:14100/20000 train_loss:2.1131 train_time:539663ms step_avg:38.27ms +step:14200/20000 train_loss:2.2242 train_time:543489ms step_avg:38.27ms +step:14300/20000 train_loss:2.1785 train_time:547308ms step_avg:38.27ms +step:14400/20000 train_loss:2.1152 train_time:551130ms step_avg:38.27ms +step:14500/20000 train_loss:2.0416 train_time:554954ms step_avg:38.27ms +step:14600/20000 train_loss:2.1680 train_time:558772ms step_avg:38.27ms +step:14700/20000 train_loss:2.1258 train_time:562594ms step_avg:38.27ms +step:14800/20000 train_loss:1.9416 train_time:566414ms step_avg:38.27ms +step:14900/20000 train_loss:1.8653 train_time:570235ms step_avg:38.27ms +step:15000/20000 train_loss:2.0555 train_time:574054ms step_avg:38.27ms +step:15000/20000 val_loss:2.1115 val_bpb:1.2506 train_time:574083ms step_avg:38.27ms +step:15100/20000 train_loss:2.1918 train_time:577878ms step_avg:38.27ms +step:15200/20000 train_loss:2.1684 train_time:581696ms step_avg:38.27ms +step:15300/20000 train_loss:2.1991 train_time:585519ms step_avg:38.27ms +step:15400/20000 train_loss:2.0722 train_time:589338ms step_avg:38.27ms +step:15500/20000 train_loss:2.0847 train_time:593162ms step_avg:38.27ms +step:15600/20000 train_loss:2.0858 train_time:596980ms step_avg:38.27ms +step:15680/20000 val_loss:2.1038 val_bpb:1.2460 train_time:600065ms step_avg:38.27ms +stopping_early: wallclock_cap train_time:600065ms step:15680/20000 +peak memory allocated: 8612 MiB reserved: 8974 MiB +Serialized model: 32656041 bytes +Code size: 39067 bytes +Total submission size: 32695108 bytes +Serialized model int8+zlib: 15664859 bytes (payload:16379168 raw_torch:16404714 payload_ratio:1.99x) +Total submission size int8+zlib: 15703926 bytes +final_int8_zlib_roundtrip val_loss:2.1142 val_bpb:1.2521 eval_time:1392ms +final_int8_zlib_roundtrip_exact val_loss:2.11416887 val_bpb:1.25212989 diff --git a/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/README.md b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/README.md new file mode 100644 index 0000000000..e90d0680f2 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/README.md @@ -0,0 +1,52 @@ +# Constrained by Time +*(Not in the traditional sense)* + +Between schoolwork and other commitments, this was definitely an on-and-off type of project. +- Prior to this, I have never trained any language models of any kind, let alone in a speedrunning context. +- I was planning on full-scale testing but had to wait until 4/30/2026 for a sufficient compute grant for full H100 testing. With only a day to really experiment, there wasn't much time for me to validate anything or find very good ideas. + - Smaller-scale testing was done on Google Colab and the RTX 5090s provided by RunPod, but these were slow and it was hard to really know what worked when moving so slowly. + +There were many (frankly) strange implementations that I tried experimenting with with the help of AI. +- Extreme depth/low dim +- Low depth/high dim +- Rolling data inside each batch (give the model one more chance at "harder" data) +- Dynamic loss scaling +- You'll find more by going through [my repo's](https://github.com/AlstonTang/parameter-golf) commit history. + +Although this probably isn't going to shatter any records (at all), I do hope that this at least shines an interesting light at some potential ideas that may be integrated into future GPT training/speedruns. +- Even if this ultimately doesn't go very far alone, it would be nice if some of the ideas in this implementation were explored. Perhaps there's too much going on in this implementation, and that together they're conflicting each other. Or perhaps's it's merely a hyperparameter configuration away from getting solid results. + +My focus ended up not being on Test-Time Training (TTT) or any significant implementation-specifc optimizations (e.g. fp8 training). Rather, my focus was on the underlying architecture itself, and really (trying towards) pushing the limits of what a conventional transformer can do. +- I may continue experiments even after this competition, since research isn't just one and done! I may take some ideas from my implementation and iteratively add it to new language models I may train in the future as I both learn more about LLMs and advanced deep learning in general. +- Some commits may show attempts implementation-specific optimizations (e.g. my attempt at fp8 training), but these usually either failed or led to training instability during experimentation. + +## ZerO initalization +I was interested in this paper: https://arxiv.org/pdf/2110.12661 +- Zhao et. al. describes how performance of deep networks can be both better and more reproducable through a more deterministic initialization method involving Hadamard/Identity-like matrices + +The actual usage of this initialization does involve some level of non-determinism (and likely somewhat deviates from the original algorithm), but it's less pronounced than fully random initalization. + +## Progression of Various Model Hyperparameters +Throughout each layer, I utilized a progression of KV head count, rope proportion, and MLP expansion factor, all of which increase as the layer's depth within the model increases. The rationale is as follows: +1. Earlier layers most likely focus on nearby context and shouldn't worry about long-range dependencies. +2. As tokens go further into the model, more information is going to be needed. + +Whether these should all be scaled linearly, geometrically, inversely, or something else is a question for another day. + +## So, could it work? +Well, maybe if there was more time (to experiment + training time) and a better implementation. +- Of course, hyperparameter choice also plays a role. However, I did not have too much time to really test anything. + +## What would I have done if I had more time? +1. Tuning hyperparameters. +2. More experimentation with other strange hypotheses. +3. Optimizing the implementation. + +If I wasn't constrained by the constraints, I would also test it on larger-scale models (e.g. 100 million+ parameters). +- Perhaps the model was too small to really realize the potential gains of my proposed implementation. +- Going beyong 16 MB would be nice to see if my ideas could potentially fly! + +## Why just one Result? +Time... +- If I had more time, I would submit more results (and probably more refined ones). +- The result should (hopefully) be reproducable across seeds due to the more deterministic [ZerO intialization](#zero-initalization) being used. diff --git a/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/requirements.txt b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/requirements.txt new file mode 100644 index 0000000000..911b0e52f0 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/requirements.txt @@ -0,0 +1,10 @@ +numpy +tqdm +torch +huggingface-hub +kernels +setuptools +typing-extensions==4.15.0 +datasets +tiktoken +sentencepiece \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/submission.json b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/submission.json new file mode 100644 index 0000000000..35be36e82c --- /dev/null +++ b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/submission.json @@ -0,0 +1,11 @@ +{ + "track": "non_record_16mb", + "date": "2026-04-30", + "name": "ZerO Initalization + Progressive KV, RoPE proportion and base, and MLP mult", + "author": "Alston Tang", + "github_id": "AlstonTang", + "val_bpb": 1.25212989, + "val_loss": 2.11416887, + "bytes_total": 15703926 + } + \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/train_gpt.py b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/train_gpt.py new file mode 100644 index 0000000000..09bcd6fcb5 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-30_ZerO _init+progressive_kv_rope_and_mlp_mult/train_gpt.py @@ -0,0 +1,941 @@ +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 +import concurrent.futures +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 5000)) + 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", 2048)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 512)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 2.125)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_max_base = float(os.environ.get("ROPE_MAX_BASE", 8192.0)) + rope_min_base = float(os.environ.get("ROPE_MIN_BASE", 512.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.3)) + head_lr = float(os.environ.get("HEAD_LR", 0.04)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.125)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 256)) + 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", 1.0)) +@torch.compile(mode="max-autotune-no-cudagraphs", fullgraph=True) +def zeropower_via_newtonschulz5(G, steps=5, eps=1e-7): + X = torch.where(torch.isfinite(G), G, 0.0).to(torch.bfloat16) + norm = torch.linalg.matrix_norm(X) + eps + X = X / norm + transposed = X.size(0) > X.size(1) + if transposed: + X = X.T + a, b, c = 3.4445, -4.7750, 2.0315 + for _ in range(steps): + A = X @ X.T + B = torch.addmm(A, A, A, beta=b, alpha=c) + X = torch.addmm(X, B, X, beta=a, alpha=1.0) + return X.T if transposed else X +@torch.no_grad() +def get_hadamard_matrix(n, device): + p2 = 2**math.ceil(math.log2(n)) + H = torch.tensor([[1.0]], device=device) + while H.shape[0] < p2: + H = torch.cat([torch.cat([H, H], dim=1), torch.cat([H, -H], dim=1)], dim=0) + return H[:n, :n] / math.sqrt(p2) +@torch.no_grad() +def apply_zero_init(model, std=0.02): + for m in model.modules(): + if isinstance(m, nn.Embedding): + d_out, d_in = m.weight.shape + H = get_hadamard_matrix(max(d_out, d_in), m.weight.device) + m.weight.copy_(H[:d_out, :d_in]*std) + linears = [sub for sub in m.children() if isinstance(sub, nn.Linear)] + if linears: + for i, l in enumerate(linears): + d_out, d_in = l.weight.shape + if i == len(linears) - 1: + nn.init.normal_(l.weight, std=1e-5) + else: + H = get_hadamard_matrix(max(d_out, d_in), l.weight.device) + l.weight.copy_(H[:d_out, :d_in]) + if l.bias is not None: + nn.init.zeros_(l.bias) +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float = 0.95, backend_steps: int = 5, nesterov: bool = True): + defaults = dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov) + super().__init__(params, defaults) + self._is_initialized = False + def _init_state(self): + 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: + if "updates_flat" in group or not group["params"]: + continue + params = group["params"] + total_params = sum(p.numel() for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + group["updates_flat"] = updates_flat + group["param_views"] = [] + group["rank_params"] = [] + group["rank_param_views"] = [] + curr = 0 + for i, p in enumerate(params): + numel = p.numel() + view = updates_flat[curr : curr + numel].view_as(p) + group["param_views"].append(view) + if i % world_size == rank: + group["rank_params"].append(p) + group["rank_param_views"].append(view) + curr += numel + @torch.no_grad() + def step(self, closure=None): + loss = closure() if closure is not None else None + self._init_state() + for group in self.param_groups: + if not group["params"]: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + updates_flat = group["updates_flat"] + updates_flat.zero_() + for p, view in zip(group["rank_params"], group["rank_param_views"]): + if p.grad is None: + continue + 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) + u = g.add(buf, alpha=momentum) if nesterov else buf + original_shape = u.shape + if u.ndim > 2: + u = u.view(u.size(0), -1) + u_prepared = u.to(dtype=torch.bfloat16, memory_format=torch.contiguous_format, non_blocking=True) + g_ns = zeropower_via_newtonschulz5(u_prepared, steps=backend_steps) + g_ns.mul_(max(1, g_ns.size(0) / g_ns.size(1))**0.5) + view.copy_(g_ns.view(original_shape)) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + params = group["params"] + views = group["param_views"] + if updates_flat.dtype != params[0].dtype: + views = [v.to(dtype=p.dtype, non_blocking=True) for v, p in zip(views, params)] + torch._foreach_add_(params, views, alpha=-lr) + return loss +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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) +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = self.future_tokens.result() + self.pos = 0 + self.future_tokens = self._queue_next_file() + 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, self.world_size, self.device = rank, world_size, device + self.stream = TokenStream(pattern) + self.next_x, self.next_y = None, None + self.transfer_stream = torch.cuda.Stream(device) + def preload(self, global_tokens: int, seq_len: int, grad_accum_steps: int): + with torch.cuda.stream(self.transfer_stream): + 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].pin_memory().to( + device=self.device, dtype=torch.int32, non_blocking=True + ) + self.next_x = local[:-1].reshape(-1, seq_len) + self.next_y = local[1:].reshape(-1, seq_len).to(torch.long) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + if self.next_x is None: + self.preload(global_tokens, seq_len, grad_accum_steps) + torch.cuda.current_stream().wait_stream(self.transfer_stream) + x, y = self.next_x, self.next_y + self.preload(global_tokens, seq_len, grad_accum_steps) + return x, y +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) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, max_seq_len: int, p: float = 0.5, base: float = 10000.0): + super().__init__() + self.rotary_dim = (int(dim * p) // 8) * 8 + inv_freq = 1.0 / (base ** (torch.arange(0, self.rotary_dim, 2, dtype=torch.float32) / self.rotary_dim)) + t = torch.arange(max_seq_len, dtype=torch.float32) + freqs = torch.outer(t, inv_freq) + self.register_buffer("cos", freqs.cos().view(1, 1, max_seq_len, -1), persistent=False) + self.register_buffer("sin", freqs.sin().view(1, 1, max_seq_len, -1), persistent=False) + def forward(self, x: Tensor): + t = x.size(2) + return self.cos[:, :, :t, :].to(x.dtype), self.sin[:, :, :t, :].to(x.dtype) +def rotate_half(x: Tensor): + x1, x2 = x.chunk(2, dim=-1) + return torch.cat((-x2, x1), dim=-1) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + d = cos.shape[-1] * 2 + x_rop = x[..., :d] + x_pass = x[..., d:] + x_rop = x_rop.view(*x_rop.shape[:-1], -1, 2) + x0 = x_rop[..., 0] + x1 = x_rop[..., 1] + res = torch.stack([x0 * cos - x1 * sin, x1 * cos + x0 * sin], dim=-1) + return torch.cat([res.flatten(-2), x_pass], dim=-1) +class CausalSelfAttention(nn.Module): + def __init__(self, dim, num_heads, num_kv_heads, rope_base, qk_gain_init, seq_len=1024, use_rope=True, rope_proportion=0.5): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + self.kv_dim = self.num_kv_heads * self.head_dim + self.c_qkv = nn.Linear(dim, dim + 2 * self.kv_dim, bias=False) + self.proj = nn.Linear(dim, dim, bias=False) + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, max_seq_len=seq_len, p=rope_proportion, base=rope_base) + self.use_rope = use_rope + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + qkv = self.c_qkv(x) + q, k, v = qkv.split([dim, self.kv_dim, self.kv_dim], dim=-1) + q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = k.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = v.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),)) + if self.use_rope: + cos, sin = self.rotary(q) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.view(1, -1, 1, 1).to(q.dtype) + y = F.scaled_dot_product_attention( + q, k, v, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads) + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + super().__init__() + self.hidden_dim = int(mlp_mult * 2/3 * dim // 64) * 64 + self.c_fc = nn.Linear(dim, 2 * self.hidden_dim, bias=False) + self.c_proj = nn.Linear(self.hidden_dim, dim, bias=False) + def forward(self, x: Tensor) -> Tensor: + fused_x = self.c_fc(x) + gate, value = fused_x.chunk(2, dim=-1) + return self.c_proj(F.silu(gate) * value) +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, + seq_len: int=1024, + use_rope: bool=True, + rope_proportion: float=0.5, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, seq_len, use_rope, rope_proportion) + self.mlp = MLP(dim, mlp_mult) + def forward(self, x: Tensor) -> Tensor: + x = x + self.attn(self.attn_norm(x)) + x = x + self.mlp(self.mlp_norm(x)) + return x +def get_linear_progression_kv_heads(layer_idx, total_layers, num_heads): + min_kv = 1 + max_kv = num_heads + fraction = layer_idx / (total_layers - 1) + raw_kv = min_kv + (max_kv - min_kv) * fraction + valid_kvs = [i for i in range(1, num_heads + 1) if num_heads % i == 0] + kv_heads = min(valid_kvs, key=lambda x: abs(x - raw_kv)) + return int(max(min_kv, kv_heads)) +def get_rope_p_smooth(i: int, num_layers: int, p_min=0.25, p_max=0.75) -> float: + if num_layers <= 1: + return p_min + progress = i / (num_layers - 1) + scale = math.sin(progress * math.pi) + return p_min + (p_max - p_min) * scale +def get_rope_base_progression(layer_idx: int, total_layers: int, min_base: float, max_base: float) -> float: + if total_layers <= 1: + return max_base + fraction = layer_idx / (total_layers - 1) + return min_base * (max_base / min_base) ** fraction +def get_linear_progression_mlp_mult(layer_idx: int, total_layers: int, base_mult: int) -> float: + min_mult = float(base_mult) * 0.75 + max_mult = float(base_mult) * 1.25 + fraction = layer_idx / max(1, total_layers - 1) + return min_mult + (max_mult - min_mult) * fraction +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_max_base: float, + rope_min_base: float, + qk_gain_init: float, + seq_len: int=1024, + ): + super().__init__() + self.tie_embeddings = tie_embeddings + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.blocks = nn.ModuleList([ + Block( + model_dim, + num_heads, + get_linear_progression_kv_heads(i, num_layers, num_kv_heads), + get_linear_progression_mlp_mult(i, num_layers, mlp_mult), + get_rope_base_progression(i, num_layers, rope_min_base, rope_max_base), + qk_gain_init, + seq_len=seq_len, + use_rope=True, + rope_proportion=get_rope_p_smooth(i, num_layers) + ) for i in range(num_layers) + ]) + self.final_norm = RMSNorm() + self.logit_softcap = logit_softcap + self.lm_head = None if tie_embeddings else nn.Linear(model_dim, vocab_size, bias=False) + apply_zero_init(self, std=tied_embed_init_std) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + emb = self.tok_emb(input_ids) + x = F.rms_norm(emb, (emb.size(-1),)) + for block in self.blocks: + x = block(x) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits = F.linear(x, self.tok_emb.weight) + else: + logits = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits.float() / self.logit_softcap) + return F.cross_entropy(logits, targets) +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + torch.set_float32_matmul_precision('high') + import torch._inductor.config as inductor_config + inductor_config.fx_graph_cache = True + inductor_config.triton.unique_kernel_names = True + inductor_config.freezing = True + inductor_config.shape_padding = True + inductor_config.epilogue_fusion = True + 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) + @torch.compiler.disable + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + 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_max_base=args.rope_max_base, + rope_min_base=args.rope_min_base, + qk_gain_init=args.qk_gain_init, + seq_len=args.train_seq_len + ).to(device).bfloat16() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, fullgraph=True, mode="reduce-overhead") + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False, gradient_as_bucket_view=True) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + gpt_scalars = [p for name, p in base_model.named_parameters() if "skip" in name] + scalar_params.extend(gpt_scalars) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/README.md b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/README.md new file mode 100644 index 0000000000..cf6e8abf15 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/README.md @@ -0,0 +1,66 @@ +# Redoing ZerO and A Follow Up To PR 2104 +The original [writeup](https://github.com/AlstonTang/parameter-golf/blob/fef7edc8c96ce169f31754f8deae1334a76e0fba/records/track_non_record_16mb/2026-04-30_ZerO%20_init%2Bprogressive_kv_rope_and_mlp_mult/README.md) in [PR 2104](https://github.com/openai/parameter-golf/pull/2104) was written semi-hastily in order to at least submit something in time within the deadline. However, after a bit more experimentation, there were some surprising results! +- The ZerO implementation from generative AI used in PR 2104 was flawed. +- If you just want to see results, either view logs or see [here](#the-surprise) for a high-level summary. + +In this submission I dump various training logs, plus the final train_gpt.py used in the final run. +- Keep in mind that each training log has a copy of train_gpt.py used during the run. + +I do not explore whether progression of various hyperparameters within each layer will have worked. +- for more details + implementation, see the [pull request](https://github.com/openai/parameter-golf/pull/2104). + +## Correcting ZerO +Previously, I had generative AI try to implement the [ZerO](https://arxiv.org/pdf/2110.12661) initialization function. However, due to issues in prompting and visibility of the document (a URL was sent instead of the underlying PDF document), the model most likely hallucinated the implementation. +- It did take inspiration from the usage of Hadamard matrices, though this could also have been a hallucination or perhaps was not fully memorized from the training corpus that the LLM I prompted went through. +- However, due to various other ideas being tested concurrently, validation of this implementation was largely overlooked during the challenge. + - Admittedly, it isn't a really good idea to test too many things at once. + +However, post-challenge, I wanted to see what ZerO could really do given a fresh start. I had previously suspected that the implementation was perhaps somewhat off due to non-determinisism being used, but I hadn't really thought of generating a fresh implementation until now. + +When feeding generative AI the actual PDF, it was able to more accurately generate a valid implementation of ZerO. +- The implementation used in this submission does include usage beyond transformers (e.g. Convolutions). +- The implementation also takes into account the existing transformer implementation in the original train_gpt.py. + +Note that, although I tried to set $W_k$ and $W_v$ as zeroed matrices as described in the paper, since the projection right after attention was also zeroed out due to the existing implementation, the model trained very poorly. +- Hence, there is still a slight deviation from the canonical ZerO implementation. + +## Reimplementation +The implementation is based on the original train_gpt.py, not the submitted one in PR 2104. +- This means that we can more accurately see whether or not ZerO works and is doing its thing instead of it being potentially masked by the other hypotheses concurrently tested within the submitted train_gpt.py in PR 2104. + +## So was ZerO the Bottleneck? +It would seem so, at least for GPT speedrunning. Although more testing is needed, when reading through section 4.1, you will see that the standard initialization actually beats out ZerO with a small number of layers. +- Perhaps ZerO may perform better with more layers and/or more training time. + +Interesting things to note are that: +1. Experimentation of ZerO within the paper was primarily focused on Convolutional Neural Networks (specifically ResNet). +2. Much of the paper focuses on the math (and proofs) instead of emperical testing. This could explain why the initialization makes the model slightly worse in practice despite sounding better on paper. + +I do have a [few plans](#future-work) to see if ZerO really is a bottleneck in a more realistic setting beyond speedrunning. + +## Experimentation +Without the constant knowledge of having to get something submitted, this went surprisingly smoothly +- Less focus on tuning hyperparameters to get something in one day, more focus on getting things right. + +With a much more accurate and deterministic intialization, it could be likely that one training run is all that is needed when testing hypotheses moving forward. + +Initially, there was an invalid implementation due to the naming conventions used in the train_gpt.py file. This was resolved in later training runs, but for convenience, the following list shows the names of log(s) containing invalid implementations: +- e31e596e-e21a-4c48-829c-78233c992cc8.txt +- fe927335-4827-415a-b543-8b5d2706de4c.txt + +These are still included so that you can kind of see the experiemntation process. All other logs contain the correct implementation of ZerO. + +## The surprise +As it turns out, due to the fundamental nature of ZerO, we actually get a much better compression result. + +Within run_logs/3c25e790-2f8a-4c36-881c-67066cf1e465.txt, although the model contained 17,059,912 parameters, the serialized model only reaches a size of 13,695,416 bytes. Although this could be partly attributed to a higher bpb (suggesting that the model learned less within the 10 minutes of training time), I strongly suspect that this is due to the low-rank learning trajectory that the model goes through within ZerO initialization. Because of this, the model is naturally more compressible. + +Adding one more layer (so that the model has 10 layers, 8 query heads, 4 kv heads, and a dim of 512 with mlp_mult of 2), and running it for the full 20,000 steps (see run_logs/24d3334a-0ddf-45ca-8be8-9dbd470f8866.txt), we get a final bpb of ~1.2494 with the submission size still only taking up 15,221,665 despite a parameter count of 18,897,488. + +I initially thought that ZerO would at least yield better loss. Although it didn't, the improvements in compression were very surprising, and I'm interested in further increasing parameter efficiency with ZerO. + +## Future Work +My plans for this consist of the following: +1. Testing ZerO with more step counts +2. Investigating how to best alter hyperparameters using ZerO initialization +3. Using ZerO with much larger (and deeper!) models. diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/requirements.txt b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/requirements.txt new file mode 100644 index 0000000000..911b0e52f0 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/requirements.txt @@ -0,0 +1,10 @@ +numpy +tqdm +torch +huggingface-hub +kernels +setuptools +typing-extensions==4.15.0 +datasets +tiktoken +sentencepiece \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/24d3334a-0ddf-45ca-8be8-9dbd470f8866.txt b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/24d3334a-0ddf-45ca-8be8-9dbd470f8866.txt new file mode 100644 index 0000000000..fe4e1f95c0 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/24d3334a-0ddf-45ca-8be8-9dbd470f8866.txt @@ -0,0 +1,1476 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +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 + +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 +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 500)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + 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 = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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 +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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 +# ----------------------------- + +def apply_zero_init(model: nn.Module) -> nn.Module: + """ + Applies ZerO initialization to a given PyTorch nn.Module in-place. + + This function implements: + - Algorithm 1 (Linear/FFN Layers using Identity, Partial Identity, and Hadamard) + - Algorithm 2 (Convolution Layers) + + Args: + model (nn.Module): The PyTorch model to initialize. + + Returns: + nn.Module: The model with initialized weights. + """ + + @torch.no_grad() + def _zero_init_tensor(tensor: torch.Tensor, out_f: int, in_f: int): + """Core logic corresponding to Algorithm 1 in the paper.""" + tensor.zero_() + + if out_f <= in_f: + # P_l == Q_l: Identity mapping + # P_l < Q_l: Partial identity (Propagates first P_l dimensions) + tensor.copy_(torch.eye(out_f, in_f, dtype=tensor.dtype, device=tensor.device)) + else: + # P_l > Q_l: Hadamard mapping (e.g., expanding MLPs) + m = math.ceil(math.log2(out_f)) + m = max(m, 1) # Ensure at least m=1 + + # Recursively generate Hadamard matrix H_m + H = torch.tensor([[1.0]], dtype=tensor.dtype, device=tensor.device) + for _ in range(m): + H1 = torch.cat([H, H], dim=1) + H2 = torch.cat([H, -H], dim=1) + H = torch.cat([H1, H2], dim=0) + + # Normalization factor defined in the paper: c = 2^{-(m-1)/2} + c = 2.0 ** (-(m - 1) / 2.0) + + # Apply scaled Hadamard top-left submatrix (I* H_m I*) + tensor.copy_(c * H[:out_f, :in_f]) + + # Iterate through and parse all network submodules + for name, module in model.named_modules(): + + # 1. Respect the script's architectural _zero_init flags (e.g., `proj` & `lm_head`). + # This achieves the paper's goal of dynamical isometry (Identity residual pass) + # and avoids the gradient deadlock. + if getattr(module, '_zero_init', False): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.zero_() + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + continue + + # 2. Linear Layers (c_q, c_k, c_v, fc) + # c_q (512->512) becomes Identity. + # c_k/c_v (512->256) become Partial Identities. + # fc (512->1024) becomes a Hadamard transform. + if isinstance(module, nn.Linear): + _zero_init_tensor(module.weight.data, module.out_features, module.in_features) + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 3. Convolutional Layers (Implementation of Algorithm 2) + elif isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)): + out_c, in_c = module.out_channels, module.in_channels + module.weight.data.zero_() + + # Find center index, taking n <- floor(k / 2) for each dimension + centers = tuple(k // 2 for k in module.kernel_size) + + # Create an out_c x in_c block utilizing the base ZerO logic + block = torch.zeros(out_c, in_c, dtype=module.weight.dtype, device=module.weight.device) + _zero_init_tensor(block, out_c, in_c) + + # Place block in the center slice of the kernel + if len(centers) == 1: + module.weight.data[:, :, centers[0]] = block + elif len(centers) == 2: + module.weight.data[:, :, centers[0], centers[1]] = block + elif len(centers) == 3: + module.weight.data[:, :, centers[0], centers[1], centers[2]] = block + + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 4. Normalization Layers (Standard practice is Scale=1, Bias=0) + elif isinstance(module, (nn.LayerNorm, nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.fill_(1.0) + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + + return model + +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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + 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._sin_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__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.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), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + 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: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + 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: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + 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") + + +# ----------------------------- +# 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, fullgraph=True) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + base_model = apply_zero_init(base_model) + 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 + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Fri May 1 17:33:22 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 39C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 32C P0 115W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 32C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 38C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 40C P0 126W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 33C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 38C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 31C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:500 max_wallclock_seconds:0.000 +seed:1337 +warmup_step:10/500 +warmup_step:20/500 +warmup_step:30/500 +warmup_step:40/500 +warmup_step:50/500 +warmup_step:60/500 +warmup_step:70/500 +warmup_step:80/500 +warmup_step:90/500 +warmup_step:100/500 +warmup_step:110/500 +warmup_step:120/500 +warmup_step:130/500 +warmup_step:140/500 +warmup_step:150/500 +warmup_step:160/500 +warmup_step:170/500 +warmup_step:180/500 +warmup_step:190/500 +warmup_step:200/500 +warmup_step:210/500 +warmup_step:220/500 +warmup_step:230/500 +warmup_step:240/500 +warmup_step:250/500 +warmup_step:260/500 +warmup_step:270/500 +warmup_step:280/500 +warmup_step:290/500 +warmup_step:300/500 +warmup_step:310/500 +warmup_step:320/500 +warmup_step:330/500 +warmup_step:340/500 +warmup_step:350/500 +warmup_step:360/500 +warmup_step:370/500 +warmup_step:380/500 +warmup_step:390/500 +warmup_step:400/500 +warmup_step:410/500 +warmup_step:420/500 +warmup_step:430/500 +warmup_step:440/500 +warmup_step:450/500 +warmup_step:460/500 +warmup_step:470/500 +warmup_step:480/500 +warmup_step:490/500 +warmup_step:500/500 +step:0/20000 val_loss:6.9363 val_bpb:4.1080 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9355 train_time:32ms step_avg:32.15ms +step:2/20000 train_loss:15.7807 train_time:80ms step_avg:40.10ms +step:3/20000 train_loss:7.3458 train_time:128ms step_avg:42.72ms +step:4/20000 train_loss:7.2066 train_time:176ms step_avg:43.97ms +step:5/20000 train_loss:7.2272 train_time:224ms step_avg:44.72ms +step:6/20000 train_loss:7.3731 train_time:271ms step_avg:45.23ms +step:7/20000 train_loss:6.4855 train_time:319ms step_avg:45.61ms +step:8/20000 train_loss:6.1886 train_time:367ms step_avg:45.90ms +step:9/20000 train_loss:5.9110 train_time:415ms step_avg:46.13ms +step:10/20000 train_loss:5.7178 train_time:463ms step_avg:46.30ms +step:200/20000 train_loss:2.9555 train_time:9534ms step_avg:47.67ms +step:400/20000 train_loss:2.4261 train_time:19087ms step_avg:47.72ms +step:600/20000 train_loss:2.5914 train_time:28659ms step_avg:47.76ms +step:800/20000 train_loss:2.3420 train_time:38249ms step_avg:47.81ms +step:1000/20000 train_loss:2.4166 train_time:47846ms step_avg:47.85ms +step:1000/20000 val_loss:2.3827 val_bpb:1.4112 train_time:47869ms step_avg:47.87ms +step:1200/20000 train_loss:2.4324 train_time:57449ms step_avg:47.87ms +step:1400/20000 train_loss:2.4740 train_time:67057ms step_avg:47.90ms +step:1600/20000 train_loss:2.1459 train_time:76658ms step_avg:47.91ms +step:1800/20000 train_loss:2.2476 train_time:86252ms step_avg:47.92ms +step:2000/20000 train_loss:2.3005 train_time:95845ms step_avg:47.92ms +step:2000/20000 val_loss:2.2831 val_bpb:1.3522 train_time:95869ms step_avg:47.93ms +step:2200/20000 train_loss:2.1276 train_time:105434ms step_avg:47.92ms +step:2400/20000 train_loss:2.2487 train_time:115017ms step_avg:47.92ms +step:2600/20000 train_loss:2.4535 train_time:124600ms step_avg:47.92ms +step:2800/20000 train_loss:2.2854 train_time:134172ms step_avg:47.92ms +step:3000/20000 train_loss:2.2749 train_time:143733ms step_avg:47.91ms +step:3000/20000 val_loss:2.2428 val_bpb:1.3283 train_time:143757ms step_avg:47.92ms +step:3200/20000 train_loss:2.2357 train_time:153292ms step_avg:47.90ms +step:3400/20000 train_loss:2.2073 train_time:162849ms step_avg:47.90ms +step:3600/20000 train_loss:2.1720 train_time:172407ms step_avg:47.89ms +step:3800/20000 train_loss:2.2673 train_time:181960ms step_avg:47.88ms +step:4000/20000 train_loss:2.2074 train_time:191513ms step_avg:47.88ms +step:4000/20000 val_loss:2.2175 val_bpb:1.3133 train_time:191537ms step_avg:47.88ms +step:4200/20000 train_loss:2.2220 train_time:201101ms step_avg:47.88ms +step:4400/20000 train_loss:2.1612 train_time:210646ms step_avg:47.87ms +step:4600/20000 train_loss:2.0193 train_time:220198ms step_avg:47.87ms +step:4800/20000 train_loss:2.3137 train_time:229750ms step_avg:47.86ms +step:5000/20000 train_loss:2.0826 train_time:239293ms step_avg:47.86ms +step:5000/20000 val_loss:2.2017 val_bpb:1.3040 train_time:239317ms step_avg:47.86ms +step:5200/20000 train_loss:2.2194 train_time:248831ms step_avg:47.85ms +step:5400/20000 train_loss:2.2337 train_time:258370ms step_avg:47.85ms +step:5600/20000 train_loss:2.2370 train_time:267904ms step_avg:47.84ms +step:5800/20000 train_loss:2.1881 train_time:277438ms step_avg:47.83ms +step:6000/20000 train_loss:2.2698 train_time:286976ms step_avg:47.83ms +step:6000/20000 val_loss:2.1927 val_bpb:1.2986 train_time:287000ms step_avg:47.83ms +step:6200/20000 train_loss:2.1367 train_time:296509ms step_avg:47.82ms +step:6400/20000 train_loss:2.2112 train_time:306038ms step_avg:47.82ms +step:6600/20000 train_loss:2.1733 train_time:315569ms step_avg:47.81ms +step:6800/20000 train_loss:2.2418 train_time:325104ms step_avg:47.81ms +step:7000/20000 train_loss:2.2779 train_time:334637ms step_avg:47.81ms +step:7000/20000 val_loss:2.1807 val_bpb:1.2915 train_time:334660ms step_avg:47.81ms +step:7200/20000 train_loss:2.2507 train_time:344211ms step_avg:47.81ms +step:7400/20000 train_loss:2.1676 train_time:353742ms step_avg:47.80ms +step:7600/20000 train_loss:2.0482 train_time:363268ms step_avg:47.80ms +step:7800/20000 train_loss:2.1976 train_time:372790ms step_avg:47.79ms +step:8000/20000 train_loss:2.1654 train_time:382312ms step_avg:47.79ms +step:8000/20000 val_loss:2.1717 val_bpb:1.2862 train_time:382336ms step_avg:47.79ms +step:8200/20000 train_loss:2.2361 train_time:391848ms step_avg:47.79ms +step:8400/20000 train_loss:2.1826 train_time:401416ms step_avg:47.79ms +step:8600/20000 train_loss:2.1872 train_time:410941ms step_avg:47.78ms +step:8800/20000 train_loss:2.1560 train_time:420468ms step_avg:47.78ms +step:9000/20000 train_loss:2.0747 train_time:429995ms step_avg:47.78ms +step:9000/20000 val_loss:2.1668 val_bpb:1.2833 train_time:430018ms step_avg:47.78ms +step:9200/20000 train_loss:2.1297 train_time:439526ms step_avg:47.77ms +step:9400/20000 train_loss:2.1854 train_time:449052ms step_avg:47.77ms +step:9600/20000 train_loss:2.1940 train_time:458577ms step_avg:47.77ms +step:9800/20000 train_loss:2.1239 train_time:468102ms step_avg:47.77ms +step:10000/20000 train_loss:2.1634 train_time:477629ms step_avg:47.76ms +step:10000/20000 val_loss:2.1610 val_bpb:1.2799 train_time:477652ms step_avg:47.77ms +step:10200/20000 train_loss:2.1188 train_time:487160ms step_avg:47.76ms +step:10400/20000 train_loss:2.1489 train_time:496693ms step_avg:47.76ms +step:10600/20000 train_loss:2.0260 train_time:506220ms step_avg:47.76ms +step:10800/20000 train_loss:2.2336 train_time:515746ms step_avg:47.75ms +step:11000/20000 train_loss:2.1634 train_time:525274ms step_avg:47.75ms +step:11000/20000 val_loss:2.1551 val_bpb:1.2764 train_time:525298ms step_avg:47.75ms +step:11200/20000 train_loss:2.1209 train_time:534796ms step_avg:47.75ms +step:11400/20000 train_loss:2.1049 train_time:544318ms step_avg:47.75ms +step:11600/20000 train_loss:2.1093 train_time:553842ms step_avg:47.74ms +step:11800/20000 train_loss:2.1475 train_time:563367ms step_avg:47.74ms +step:12000/20000 train_loss:2.1222 train_time:572897ms step_avg:47.74ms +step:12000/20000 val_loss:2.1500 val_bpb:1.2734 train_time:572921ms step_avg:47.74ms +step:12200/20000 train_loss:2.2634 train_time:582423ms step_avg:47.74ms +step:12400/20000 train_loss:1.9120 train_time:591994ms step_avg:47.74ms +step:12600/20000 train_loss:2.1398 train_time:601521ms step_avg:47.74ms +step:12800/20000 train_loss:2.1643 train_time:611055ms step_avg:47.74ms +step:13000/20000 train_loss:2.2421 train_time:620585ms step_avg:47.74ms +step:13000/20000 val_loss:2.1498 val_bpb:1.2732 train_time:620609ms step_avg:47.74ms +step:13200/20000 train_loss:2.2502 train_time:630114ms step_avg:47.74ms +step:13400/20000 train_loss:2.1292 train_time:639639ms step_avg:47.73ms +step:13600/20000 train_loss:2.0075 train_time:649167ms step_avg:47.73ms +step:13800/20000 train_loss:2.0835 train_time:658698ms step_avg:47.73ms +step:14000/20000 train_loss:2.1477 train_time:668223ms step_avg:47.73ms +step:14000/20000 val_loss:2.1431 val_bpb:1.2693 train_time:668246ms step_avg:47.73ms +step:14200/20000 train_loss:2.2300 train_time:677747ms step_avg:47.73ms +step:14400/20000 train_loss:2.1275 train_time:687272ms step_avg:47.73ms +step:14600/20000 train_loss:2.1833 train_time:696795ms step_avg:47.73ms +step:14800/20000 train_loss:1.9672 train_time:706320ms step_avg:47.72ms +step:15000/20000 train_loss:2.0831 train_time:715860ms step_avg:47.72ms +step:15000/20000 val_loss:2.1385 val_bpb:1.2666 train_time:715884ms step_avg:47.73ms +step:15200/20000 train_loss:2.1946 train_time:725389ms step_avg:47.72ms +step:15400/20000 train_loss:2.0976 train_time:734914ms step_avg:47.72ms +step:15600/20000 train_loss:2.1188 train_time:744440ms step_avg:47.72ms +step:15800/20000 train_loss:1.9666 train_time:753967ms step_avg:47.72ms +step:16000/20000 train_loss:2.1792 train_time:763496ms step_avg:47.72ms +step:16000/20000 val_loss:2.1364 val_bpb:1.2653 train_time:763520ms step_avg:47.72ms +step:16200/20000 train_loss:2.0572 train_time:773024ms step_avg:47.72ms +step:16400/20000 train_loss:2.0803 train_time:782549ms step_avg:47.72ms +step:16600/20000 train_loss:2.0230 train_time:792111ms step_avg:47.72ms +step:16800/20000 train_loss:2.2437 train_time:801632ms step_avg:47.72ms +step:17000/20000 train_loss:2.1546 train_time:811162ms step_avg:47.72ms +step:17000/20000 val_loss:2.1339 val_bpb:1.2638 train_time:811185ms step_avg:47.72ms +step:17200/20000 train_loss:2.1488 train_time:820691ms step_avg:47.71ms +step:17400/20000 train_loss:2.0464 train_time:830216ms step_avg:47.71ms +step:17600/20000 train_loss:2.1303 train_time:839742ms step_avg:47.71ms +step:17800/20000 train_loss:2.2072 train_time:849266ms step_avg:47.71ms +step:18000/20000 train_loss:2.1204 train_time:858792ms step_avg:47.71ms +step:18000/20000 val_loss:2.1331 val_bpb:1.2633 train_time:858815ms step_avg:47.71ms +step:18200/20000 train_loss:2.3465 train_time:868315ms step_avg:47.71ms +step:18400/20000 train_loss:2.1184 train_time:877843ms step_avg:47.71ms +step:18600/20000 train_loss:2.1484 train_time:887374ms step_avg:47.71ms +step:18800/20000 train_loss:2.1997 train_time:896902ms step_avg:47.71ms +step:19000/20000 train_loss:2.1245 train_time:906423ms step_avg:47.71ms +step:19000/20000 val_loss:2.1240 val_bpb:1.2580 train_time:906447ms step_avg:47.71ms +step:19200/20000 train_loss:1.9684 train_time:916010ms step_avg:47.71ms +step:19400/20000 train_loss:2.1891 train_time:925577ms step_avg:47.71ms +step:19600/20000 train_loss:2.2643 train_time:935103ms step_avg:47.71ms +step:19800/20000 train_loss:1.9416 train_time:944630ms step_avg:47.71ms +step:20000/20000 train_loss:2.1303 train_time:954159ms step_avg:47.71ms +step:20000/20000 val_loss:2.0938 val_bpb:1.2400 train_time:954183ms step_avg:47.71ms +peak memory allocated: 11263 MiB reserved: 11320 MiB +Serialized model: 74578915 bytes +Code size: 52042 bytes +Total submission size: 74630957 bytes +Serialized model int8+zlib: 15169623 bytes (payload:19030336 raw_torch:19080377 payload_ratio:3.92x) +Total submission size int8+zlib: 15221665 bytes +final_int8_zlib_roundtrip val_loss:2.1095 val_bpb:1.2494 eval_time:1566ms +final_int8_zlib_roundtrip_exact val_loss:2.10952960 val_bpb:1.24938225 diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/3c25e790-2f8a-4c36-881c-67066cf1e465.txt b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/3c25e790-2f8a-4c36-881c-67066cf1e465.txt new file mode 100644 index 0000000000..37e760da50 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/3c25e790-2f8a-4c36-881c-67066cf1e465.txt @@ -0,0 +1,1399 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +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 + +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 +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 100)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.05)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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 +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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 +# ----------------------------- + +def apply_zero_init(model: nn.Module) -> nn.Module: + """ + Applies ZerO initialization to a given PyTorch nn.Module in-place. + + This function implements: + - Algorithm 1 (Linear/FFN Layers using Identity, Partial Identity, and Hadamard) + - Algorithm 2 (Convolution Layers) + + Args: + model (nn.Module): The PyTorch model to initialize. + + Returns: + nn.Module: The model with initialized weights. + """ + + @torch.no_grad() + def _zero_init_tensor(tensor: torch.Tensor, out_f: int, in_f: int): + """Core logic corresponding to Algorithm 1 in the paper.""" + tensor.zero_() + + if out_f <= in_f: + # P_l == Q_l: Identity mapping + # P_l < Q_l: Partial identity (Propagates first P_l dimensions) + tensor.copy_(torch.eye(out_f, in_f, dtype=tensor.dtype, device=tensor.device)) + else: + # P_l > Q_l: Hadamard mapping (e.g., expanding MLPs) + m = math.ceil(math.log2(out_f)) + m = max(m, 1) # Ensure at least m=1 + + # Recursively generate Hadamard matrix H_m + H = torch.tensor([[1.0]], dtype=tensor.dtype, device=tensor.device) + for _ in range(m): + H1 = torch.cat([H, H], dim=1) + H2 = torch.cat([H, -H], dim=1) + H = torch.cat([H1, H2], dim=0) + + # Normalization factor defined in the paper: c = 2^{-(m-1)/2} + c = 2.0 ** (-(m - 1) / 2.0) + + # Apply scaled Hadamard top-left submatrix (I* H_m I*) + tensor.copy_(c * H[:out_f, :in_f]) + + # Iterate through and parse all network submodules + for name, module in model.named_modules(): + + # 1. Respect the script's architectural _zero_init flags (e.g., `proj` & `lm_head`). + # This achieves the paper's goal of dynamical isometry (Identity residual pass) + # and avoids the gradient deadlock. + if getattr(module, '_zero_init', False): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.zero_() + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + continue + + # 2. Linear Layers (c_q, c_k, c_v, fc) + # c_q (512->512) becomes Identity. + # c_k/c_v (512->256) become Partial Identities. + # fc (512->1024) becomes a Hadamard transform. + if isinstance(module, nn.Linear): + _zero_init_tensor(module.weight.data, module.out_features, module.in_features) + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 3. Convolutional Layers (Implementation of Algorithm 2) + elif isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)): + out_c, in_c = module.out_channels, module.in_channels + module.weight.data.zero_() + + # Find center index, taking n <- floor(k / 2) for each dimension + centers = tuple(k // 2 for k in module.kernel_size) + + # Create an out_c x in_c block utilizing the base ZerO logic + block = torch.zeros(out_c, in_c, dtype=module.weight.dtype, device=module.weight.device) + _zero_init_tensor(block, out_c, in_c) + + # Place block in the center slice of the kernel + if len(centers) == 1: + module.weight.data[:, :, centers[0]] = block + elif len(centers) == 2: + module.weight.data[:, :, centers[0], centers[1]] = block + elif len(centers) == 3: + module.weight.data[:, :, centers[0], centers[1], centers[2]] = block + + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 4. Normalization Layers (Standard practice is Scale=1, Bias=0) + elif isinstance(module, (nn.LayerNorm, nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.fill_(1.0) + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + + return model + +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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + 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._sin_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__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.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), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + 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: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + 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: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + 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") + + +# ----------------------------- +# 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, fullgraph=True) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + base_model = apply_zero_init(base_model) + 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 + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Fri May 1 16:39:39 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 39C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 33C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 33C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 38C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 42C P0 129W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 34C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 39C P0 121W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 32C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059912 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.05 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:100 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:10/100 +warmup_step:20/100 +warmup_step:30/100 +warmup_step:40/100 +warmup_step:50/100 +warmup_step:60/100 +warmup_step:70/100 +warmup_step:80/100 +warmup_step:90/100 +warmup_step:100/100 +step:0/20000 val_loss:6.9357 val_bpb:4.1077 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9370 train_time:27ms step_avg:26.58ms +step:2/20000 train_loss:15.6862 train_time:73ms step_avg:36.37ms +step:3/20000 train_loss:7.2524 train_time:118ms step_avg:39.23ms +step:4/20000 train_loss:7.3808 train_time:164ms step_avg:41.02ms +step:5/20000 train_loss:7.5702 train_time:208ms step_avg:41.53ms +step:6/20000 train_loss:7.7805 train_time:267ms step_avg:44.49ms +step:7/20000 train_loss:6.6300 train_time:302ms step_avg:43.21ms +step:8/20000 train_loss:6.1898 train_time:345ms step_avg:43.10ms +step:9/20000 train_loss:5.9347 train_time:387ms step_avg:42.96ms +step:10/20000 train_loss:5.6995 train_time:428ms step_avg:42.79ms +step:200/20000 train_loss:2.9527 train_time:8780ms step_avg:43.90ms +step:400/20000 train_loss:2.4409 train_time:17552ms step_avg:43.88ms +step:600/20000 train_loss:2.6121 train_time:26339ms step_avg:43.90ms +step:800/20000 train_loss:2.3574 train_time:35129ms step_avg:43.91ms +step:1000/20000 train_loss:2.4316 train_time:43937ms step_avg:43.94ms +step:1000/20000 val_loss:2.3994 val_bpb:1.4211 train_time:43950ms step_avg:43.95ms +step:1200/20000 train_loss:2.4514 train_time:52775ms step_avg:43.98ms +step:1400/20000 train_loss:2.4901 train_time:61598ms step_avg:44.00ms +step:1600/20000 train_loss:2.1602 train_time:70441ms step_avg:44.03ms +step:1800/20000 train_loss:2.2684 train_time:79269ms step_avg:44.04ms +step:2000/20000 train_loss:2.3156 train_time:88094ms step_avg:44.05ms +step:2000/20000 val_loss:2.3011 val_bpb:1.3628 train_time:88109ms step_avg:44.05ms +step:2200/20000 train_loss:2.1455 train_time:96924ms step_avg:44.06ms +step:2400/20000 train_loss:2.2675 train_time:105742ms step_avg:44.06ms +step:2600/20000 train_loss:2.4753 train_time:114561ms step_avg:44.06ms +step:2800/20000 train_loss:2.3080 train_time:123367ms step_avg:44.06ms +step:3000/20000 train_loss:2.2924 train_time:132173ms step_avg:44.06ms +step:3000/20000 val_loss:2.2622 val_bpb:1.3398 train_time:132200ms step_avg:44.07ms +step:3200/20000 train_loss:2.2553 train_time:140984ms step_avg:44.06ms +step:3400/20000 train_loss:2.2284 train_time:149782ms step_avg:44.05ms +step:3600/20000 train_loss:2.1919 train_time:158602ms step_avg:44.06ms +step:3800/20000 train_loss:2.2899 train_time:167394ms step_avg:44.05ms +step:4000/20000 train_loss:2.2306 train_time:176184ms step_avg:44.05ms +step:4000/20000 val_loss:2.2379 val_bpb:1.3254 train_time:176204ms step_avg:44.05ms +step:4200/20000 train_loss:2.2384 train_time:185050ms step_avg:44.06ms +step:4400/20000 train_loss:2.1830 train_time:193836ms step_avg:44.05ms +step:4600/20000 train_loss:2.0421 train_time:202628ms step_avg:44.05ms +step:4800/20000 train_loss:2.3297 train_time:211418ms step_avg:44.05ms +step:5000/20000 train_loss:2.0996 train_time:220196ms step_avg:44.04ms +step:5000/20000 val_loss:2.2224 val_bpb:1.3163 train_time:220210ms step_avg:44.04ms +step:5200/20000 train_loss:2.2382 train_time:228983ms step_avg:44.04ms +step:5400/20000 train_loss:2.2570 train_time:237772ms step_avg:44.03ms +step:5600/20000 train_loss:2.2551 train_time:246556ms step_avg:44.03ms +step:5800/20000 train_loss:2.2178 train_time:255344ms step_avg:44.02ms +step:6000/20000 train_loss:2.2837 train_time:264127ms step_avg:44.02ms +step:6000/20000 val_loss:2.2127 val_bpb:1.3105 train_time:264141ms step_avg:44.02ms +step:6200/20000 train_loss:2.1591 train_time:272909ms step_avg:44.02ms +step:6400/20000 train_loss:2.2351 train_time:281692ms step_avg:44.01ms +step:6600/20000 train_loss:2.1986 train_time:290473ms step_avg:44.01ms +step:6800/20000 train_loss:2.2593 train_time:299250ms step_avg:44.01ms +step:7000/20000 train_loss:2.2935 train_time:308036ms step_avg:44.01ms +step:7000/20000 val_loss:2.2025 val_bpb:1.3044 train_time:308051ms step_avg:44.01ms +step:7200/20000 train_loss:2.2722 train_time:316819ms step_avg:44.00ms +step:7400/20000 train_loss:2.1888 train_time:325594ms step_avg:44.00ms +step:7600/20000 train_loss:2.0719 train_time:334388ms step_avg:44.00ms +step:7800/20000 train_loss:2.2174 train_time:343173ms step_avg:44.00ms +step:8000/20000 train_loss:2.1836 train_time:351961ms step_avg:44.00ms +step:8000/20000 val_loss:2.1938 val_bpb:1.2993 train_time:351974ms step_avg:44.00ms +step:8200/20000 train_loss:2.2516 train_time:360751ms step_avg:43.99ms +step:8400/20000 train_loss:2.2039 train_time:369595ms step_avg:44.00ms +step:8600/20000 train_loss:2.2087 train_time:378377ms step_avg:44.00ms +step:8800/20000 train_loss:2.1797 train_time:387173ms step_avg:44.00ms +step:9000/20000 train_loss:2.0962 train_time:395957ms step_avg:44.00ms +step:9000/20000 val_loss:2.1889 val_bpb:1.2964 train_time:395971ms step_avg:44.00ms +step:9200/20000 train_loss:2.1523 train_time:404751ms step_avg:43.99ms +step:9400/20000 train_loss:2.2119 train_time:413533ms step_avg:43.99ms +step:9600/20000 train_loss:2.2201 train_time:422325ms step_avg:43.99ms +step:9800/20000 train_loss:2.1550 train_time:431106ms step_avg:43.99ms +step:10000/20000 train_loss:2.1850 train_time:439893ms step_avg:43.99ms +step:10000/20000 val_loss:2.1846 val_bpb:1.2938 train_time:439908ms step_avg:43.99ms +step:10200/20000 train_loss:2.1424 train_time:448683ms step_avg:43.99ms +step:10400/20000 train_loss:2.1723 train_time:457466ms step_avg:43.99ms +step:10600/20000 train_loss:2.0463 train_time:466256ms step_avg:43.99ms +step:10800/20000 train_loss:2.2569 train_time:475038ms step_avg:43.98ms +step:11000/20000 train_loss:2.1925 train_time:483830ms step_avg:43.98ms +step:11000/20000 val_loss:2.1781 val_bpb:1.2900 train_time:483844ms step_avg:43.99ms +step:11200/20000 train_loss:2.1436 train_time:492619ms step_avg:43.98ms +step:11400/20000 train_loss:2.1248 train_time:501403ms step_avg:43.98ms +step:11600/20000 train_loss:2.1365 train_time:510193ms step_avg:43.98ms +step:11800/20000 train_loss:2.1706 train_time:518971ms step_avg:43.98ms +step:12000/20000 train_loss:2.1453 train_time:527759ms step_avg:43.98ms +step:12000/20000 val_loss:2.1730 val_bpb:1.2870 train_time:527774ms step_avg:43.98ms +step:12200/20000 train_loss:2.2864 train_time:536547ms step_avg:43.98ms +step:12400/20000 train_loss:1.9287 train_time:545394ms step_avg:43.98ms +step:12600/20000 train_loss:2.1581 train_time:554180ms step_avg:43.98ms +step:12800/20000 train_loss:2.1741 train_time:562952ms step_avg:43.98ms +step:13000/20000 train_loss:2.2488 train_time:571722ms step_avg:43.98ms +step:13000/20000 val_loss:2.1538 val_bpb:1.2756 train_time:571747ms step_avg:43.98ms +step:13200/20000 train_loss:2.2517 train_time:580590ms step_avg:43.98ms +step:13400/20000 train_loss:2.1198 train_time:589373ms step_avg:43.98ms +step:13600/20000 train_loss:1.9812 train_time:598159ms step_avg:43.98ms +step:13643/20000 val_loss:2.1313 val_bpb:1.2623 train_time:600049ms step_avg:43.98ms +stopping_early: wallclock_cap train_time:600049ms step:13643/20000 +peak memory allocated: 10119 MiB reserved: 10438 MiB +Serialized model: 67224983 bytes +Code size: 52041 bytes +Total submission size: 67277024 bytes +Serialized model int8+zlib: 13695416 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x) +Total submission size int8+zlib: 13747457 bytes +final_int8_zlib_roundtrip val_loss:2.1480 val_bpb:1.2722 eval_time:1414ms +final_int8_zlib_roundtrip_exact val_loss:2.14799565 val_bpb:1.27216402 diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/8c0bf12c-29d7-4dad-8cb7-6fa04b89b309.txt b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/8c0bf12c-29d7-4dad-8cb7-6fa04b89b309.txt new file mode 100644 index 0000000000..4044fd2419 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/8c0bf12c-29d7-4dad-8cb7-6fa04b89b309.txt @@ -0,0 +1,1431 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +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 + +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 +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 500)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + 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 = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.05)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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 +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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 +# ----------------------------- + +def apply_zero_init(model: nn.Module) -> nn.Module: + """ + Applies ZerO initialization to a given PyTorch nn.Module in-place. + + This function implements: + - Algorithm 1 (Linear/FFN Layers using Identity, Partial Identity, and Hadamard) + - Algorithm 2 (Convolution Layers) + + Args: + model (nn.Module): The PyTorch model to initialize. + + Returns: + nn.Module: The model with initialized weights. + """ + + @torch.no_grad() + def _zero_init_tensor(tensor: torch.Tensor, out_f: int, in_f: int): + """Core logic corresponding to Algorithm 1 in the paper.""" + tensor.zero_() + + if out_f <= in_f: + # P_l == Q_l: Identity mapping + # P_l < Q_l: Partial identity (Propagates first P_l dimensions) + tensor.copy_(torch.eye(out_f, in_f, dtype=tensor.dtype, device=tensor.device)) + else: + # P_l > Q_l: Hadamard mapping (e.g., expanding MLPs) + m = math.ceil(math.log2(out_f)) + m = max(m, 1) # Ensure at least m=1 + + # Recursively generate Hadamard matrix H_m + H = torch.tensor([[1.0]], dtype=tensor.dtype, device=tensor.device) + for _ in range(m): + H1 = torch.cat([H, H], dim=1) + H2 = torch.cat([H, -H], dim=1) + H = torch.cat([H1, H2], dim=0) + + # Normalization factor defined in the paper: c = 2^{-(m-1)/2} + c = 2.0 ** (-(m - 1) / 2.0) + + # Apply scaled Hadamard top-left submatrix (I* H_m I*) + tensor.copy_(c * H[:out_f, :in_f]) + + # Iterate through and parse all network submodules + for name, module in model.named_modules(): + + # 1. Respect the script's architectural _zero_init flags (e.g., `proj` & `lm_head`). + # This achieves the paper's goal of dynamical isometry (Identity residual pass) + # and avoids the gradient deadlock. + if getattr(module, '_zero_init', False): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.zero_() + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + continue + + # 2. Linear Layers (c_q, c_k, c_v, fc) + # c_q (512->512) becomes Identity. + # c_k/c_v (512->256) become Partial Identities. + # fc (512->1024) becomes a Hadamard transform. + if isinstance(module, nn.Linear): + _zero_init_tensor(module.weight.data, module.out_features, module.in_features) + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 3. Convolutional Layers (Implementation of Algorithm 2) + elif isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)): + out_c, in_c = module.out_channels, module.in_channels + module.weight.data.zero_() + + # Find center index, taking n <- floor(k / 2) for each dimension + centers = tuple(k // 2 for k in module.kernel_size) + + # Create an out_c x in_c block utilizing the base ZerO logic + block = torch.zeros(out_c, in_c, dtype=module.weight.dtype, device=module.weight.device) + _zero_init_tensor(block, out_c, in_c) + + # Place block in the center slice of the kernel + if len(centers) == 1: + module.weight.data[:, :, centers[0]] = block + elif len(centers) == 2: + module.weight.data[:, :, centers[0], centers[1]] = block + elif len(centers) == 3: + module.weight.data[:, :, centers[0], centers[1], centers[2]] = block + + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 4. Normalization Layers (Standard practice is Scale=1, Bias=0) + elif isinstance(module, (nn.LayerNorm, nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.fill_(1.0) + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + + return model + +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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + 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._sin_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__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.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), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + 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: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + 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: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + 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") + + +# ----------------------------- +# 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, fullgraph=True) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + base_model = apply_zero_init(base_model) + 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 + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Fri May 1 16:59:34 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 39C P0 122W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 33C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 33C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 39C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 42C P0 127W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 33C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 39C P0 120W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 32C P0 115W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.05 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:500 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:10/500 +warmup_step:20/500 +warmup_step:30/500 +warmup_step:40/500 +warmup_step:50/500 +warmup_step:60/500 +warmup_step:70/500 +warmup_step:80/500 +warmup_step:90/500 +warmup_step:100/500 +warmup_step:110/500 +warmup_step:120/500 +warmup_step:130/500 +warmup_step:140/500 +warmup_step:150/500 +warmup_step:160/500 +warmup_step:170/500 +warmup_step:180/500 +warmup_step:190/500 +warmup_step:200/500 +warmup_step:210/500 +warmup_step:220/500 +warmup_step:230/500 +warmup_step:240/500 +warmup_step:250/500 +warmup_step:260/500 +warmup_step:270/500 +warmup_step:280/500 +warmup_step:290/500 +warmup_step:300/500 +warmup_step:310/500 +warmup_step:320/500 +warmup_step:330/500 +warmup_step:340/500 +warmup_step:350/500 +warmup_step:360/500 +warmup_step:370/500 +warmup_step:380/500 +warmup_step:390/500 +warmup_step:400/500 +warmup_step:410/500 +warmup_step:420/500 +warmup_step:430/500 +warmup_step:440/500 +warmup_step:450/500 +warmup_step:460/500 +warmup_step:470/500 +warmup_step:480/500 +warmup_step:490/500 +warmup_step:500/500 +step:0/20000 val_loss:6.9363 val_bpb:4.1080 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9355 train_time:34ms step_avg:33.81ms +step:2/20000 train_loss:15.4054 train_time:88ms step_avg:43.76ms +step:3/20000 train_loss:7.0580 train_time:135ms step_avg:45.11ms +step:4/20000 train_loss:7.5619 train_time:184ms step_avg:45.92ms +step:5/20000 train_loss:7.6829 train_time:232ms step_avg:46.32ms +step:6/20000 train_loss:7.7755 train_time:279ms step_avg:46.55ms +step:7/20000 train_loss:6.6480 train_time:327ms step_avg:46.78ms +step:8/20000 train_loss:6.1307 train_time:376ms step_avg:47.03ms +step:9/20000 train_loss:5.8335 train_time:425ms step_avg:47.21ms +step:10/20000 train_loss:5.6066 train_time:472ms step_avg:47.25ms +step:200/20000 train_loss:2.9260 train_time:9718ms step_avg:48.59ms +step:400/20000 train_loss:2.4167 train_time:19466ms step_avg:48.66ms +step:600/20000 train_loss:2.5982 train_time:29229ms step_avg:48.71ms +step:800/20000 train_loss:2.3428 train_time:38994ms step_avg:48.74ms +step:1000/20000 train_loss:2.4160 train_time:48764ms step_avg:48.76ms +step:1000/20000 val_loss:2.3810 val_bpb:1.4101 train_time:48780ms step_avg:48.78ms +step:1200/20000 train_loss:2.4330 train_time:58537ms step_avg:48.78ms +step:1400/20000 train_loss:2.4768 train_time:68306ms step_avg:48.79ms +step:1600/20000 train_loss:2.1452 train_time:78069ms step_avg:48.79ms +step:1800/20000 train_loss:2.2518 train_time:87833ms step_avg:48.80ms +step:2000/20000 train_loss:2.2958 train_time:97588ms step_avg:48.79ms +step:2000/20000 val_loss:2.2838 val_bpb:1.3526 train_time:97603ms step_avg:48.80ms +step:2200/20000 train_loss:2.1265 train_time:107348ms step_avg:48.79ms +step:2400/20000 train_loss:2.2490 train_time:117089ms step_avg:48.79ms +step:2600/20000 train_loss:2.4605 train_time:126824ms step_avg:48.78ms +step:2800/20000 train_loss:2.2921 train_time:136549ms step_avg:48.77ms +step:3000/20000 train_loss:2.2760 train_time:146271ms step_avg:48.76ms +step:3000/20000 val_loss:2.2444 val_bpb:1.3293 train_time:146287ms step_avg:48.76ms +step:3200/20000 train_loss:2.2376 train_time:155990ms step_avg:48.75ms +step:3400/20000 train_loss:2.2064 train_time:165700ms step_avg:48.74ms +step:3600/20000 train_loss:2.1746 train_time:175408ms step_avg:48.72ms +step:3800/20000 train_loss:2.2689 train_time:185099ms step_avg:48.71ms +step:4000/20000 train_loss:2.2102 train_time:194790ms step_avg:48.70ms +step:4000/20000 val_loss:2.2196 val_bpb:1.3146 train_time:194806ms step_avg:48.70ms +step:4200/20000 train_loss:2.2205 train_time:204545ms step_avg:48.70ms +step:4400/20000 train_loss:2.1643 train_time:214231ms step_avg:48.69ms +step:4600/20000 train_loss:2.0250 train_time:223919ms step_avg:48.68ms +step:4800/20000 train_loss:2.3136 train_time:233603ms step_avg:48.67ms +step:5000/20000 train_loss:2.0813 train_time:243284ms step_avg:48.66ms +step:5000/20000 val_loss:2.2034 val_bpb:1.3050 train_time:243300ms step_avg:48.66ms +step:5200/20000 train_loss:2.2222 train_time:252970ms step_avg:48.65ms +step:5400/20000 train_loss:2.2378 train_time:262653ms step_avg:48.64ms +step:5600/20000 train_loss:2.2314 train_time:272333ms step_avg:48.63ms +step:5800/20000 train_loss:2.1918 train_time:282020ms step_avg:48.62ms +step:6000/20000 train_loss:2.2623 train_time:291701ms step_avg:48.62ms +step:6000/20000 val_loss:2.1940 val_bpb:1.2994 train_time:291716ms step_avg:48.62ms +step:6200/20000 train_loss:2.1408 train_time:301380ms step_avg:48.61ms +step:6400/20000 train_loss:2.2189 train_time:311065ms step_avg:48.60ms +step:6600/20000 train_loss:2.1810 train_time:320736ms step_avg:48.60ms +step:6800/20000 train_loss:2.2474 train_time:330420ms step_avg:48.59ms +step:7000/20000 train_loss:2.2785 train_time:340102ms step_avg:48.59ms +step:7000/20000 val_loss:2.1836 val_bpb:1.2933 train_time:340118ms step_avg:48.59ms +step:7200/20000 train_loss:2.2525 train_time:349777ms step_avg:48.58ms +step:7400/20000 train_loss:2.1674 train_time:359453ms step_avg:48.57ms +step:7600/20000 train_loss:2.0531 train_time:369126ms step_avg:48.57ms +step:7800/20000 train_loss:2.1992 train_time:378797ms step_avg:48.56ms +step:8000/20000 train_loss:2.1641 train_time:388473ms step_avg:48.56ms +step:8000/20000 val_loss:2.1747 val_bpb:1.2880 train_time:388489ms step_avg:48.56ms +step:8200/20000 train_loss:2.2369 train_time:398161ms step_avg:48.56ms +step:8400/20000 train_loss:2.1847 train_time:407900ms step_avg:48.56ms +step:8600/20000 train_loss:2.1873 train_time:417583ms step_avg:48.56ms +step:8800/20000 train_loss:2.1570 train_time:427272ms step_avg:48.55ms +step:9000/20000 train_loss:2.0783 train_time:436952ms step_avg:48.55ms +step:9000/20000 val_loss:2.1697 val_bpb:1.2850 train_time:436968ms step_avg:48.55ms +step:9200/20000 train_loss:2.1343 train_time:446634ms step_avg:48.55ms +step:9400/20000 train_loss:2.1895 train_time:456322ms step_avg:48.54ms +step:9600/20000 train_loss:2.2026 train_time:466001ms step_avg:48.54ms +step:9800/20000 train_loss:2.1330 train_time:475675ms step_avg:48.54ms +step:10000/20000 train_loss:2.1662 train_time:485356ms step_avg:48.54ms +step:10000/20000 val_loss:2.1644 val_bpb:1.2819 train_time:485372ms step_avg:48.54ms +step:10200/20000 train_loss:2.1213 train_time:495028ms step_avg:48.53ms +step:10400/20000 train_loss:2.1505 train_time:504709ms step_avg:48.53ms +step:10600/20000 train_loss:2.0258 train_time:514399ms step_avg:48.53ms +step:10800/20000 train_loss:2.2354 train_time:524090ms step_avg:48.53ms +step:11000/20000 train_loss:2.1705 train_time:533770ms step_avg:48.52ms +step:11000/20000 val_loss:2.1580 val_bpb:1.2781 train_time:533786ms step_avg:48.53ms +step:11200/20000 train_loss:2.1251 train_time:543447ms step_avg:48.52ms +step:11400/20000 train_loss:2.1018 train_time:553123ms step_avg:48.52ms +step:11600/20000 train_loss:2.1047 train_time:562918ms step_avg:48.53ms +step:11800/20000 train_loss:2.1291 train_time:572608ms step_avg:48.53ms +step:12000/20000 train_loss:2.0992 train_time:582295ms step_avg:48.52ms +step:12000/20000 val_loss:2.1257 val_bpb:1.2589 train_time:582311ms step_avg:48.53ms +step:12200/20000 train_loss:2.2375 train_time:591987ms step_avg:48.52ms +step:12366/20000 val_loss:2.1141 val_bpb:1.2521 train_time:600048ms step_avg:48.52ms +stopping_early: wallclock_cap train_time:600048ms step:12366/20000 +peak memory allocated: 11263 MiB reserved: 11320 MiB +Serialized model: 74578915 bytes +Code size: 52042 bytes +Total submission size: 74630957 bytes +Serialized model int8+zlib: 15165177 bytes (payload:19030336 raw_torch:19080377 payload_ratio:3.92x) +Total submission size int8+zlib: 15217219 bytes +final_int8_zlib_roundtrip val_loss:2.1298 val_bpb:1.2614 eval_time:1561ms +final_int8_zlib_roundtrip_exact val_loss:2.12975066 val_bpb:1.26135830 diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/bf9a8898-d4e8-40dd-8786-320b140eb700.txt b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/bf9a8898-d4e8-40dd-8786-320b140eb700.txt new file mode 100644 index 0000000000..3822203d52 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/bf9a8898-d4e8-40dd-8786-320b140eb700.txt @@ -0,0 +1,1431 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +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 + +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 +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 500)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + 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 = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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 +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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 +# ----------------------------- + +def apply_zero_init(model: nn.Module) -> nn.Module: + """ + Applies ZerO initialization to a given PyTorch nn.Module in-place. + + This function implements: + - Algorithm 1 (Linear/FFN Layers using Identity, Partial Identity, and Hadamard) + - Algorithm 2 (Convolution Layers) + + Args: + model (nn.Module): The PyTorch model to initialize. + + Returns: + nn.Module: The model with initialized weights. + """ + + @torch.no_grad() + def _zero_init_tensor(tensor: torch.Tensor, out_f: int, in_f: int): + """Core logic corresponding to Algorithm 1 in the paper.""" + tensor.zero_() + + if out_f <= in_f: + # P_l == Q_l: Identity mapping + # P_l < Q_l: Partial identity (Propagates first P_l dimensions) + tensor.copy_(torch.eye(out_f, in_f, dtype=tensor.dtype, device=tensor.device)) + else: + # P_l > Q_l: Hadamard mapping (e.g., expanding MLPs) + m = math.ceil(math.log2(out_f)) + m = max(m, 1) # Ensure at least m=1 + + # Recursively generate Hadamard matrix H_m + H = torch.tensor([[1.0]], dtype=tensor.dtype, device=tensor.device) + for _ in range(m): + H1 = torch.cat([H, H], dim=1) + H2 = torch.cat([H, -H], dim=1) + H = torch.cat([H1, H2], dim=0) + + # Normalization factor defined in the paper: c = 2^{-(m-1)/2} + c = 2.0 ** (-(m - 1) / 2.0) + + # Apply scaled Hadamard top-left submatrix (I* H_m I*) + tensor.copy_(c * H[:out_f, :in_f]) + + # Iterate through and parse all network submodules + for name, module in model.named_modules(): + + # 1. Respect the script's architectural _zero_init flags (e.g., `proj` & `lm_head`). + # This achieves the paper's goal of dynamical isometry (Identity residual pass) + # and avoids the gradient deadlock. + if getattr(module, '_zero_init', False): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.zero_() + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + continue + + # 2. Linear Layers (c_q, c_k, c_v, fc) + # c_q (512->512) becomes Identity. + # c_k/c_v (512->256) become Partial Identities. + # fc (512->1024) becomes a Hadamard transform. + if isinstance(module, nn.Linear): + _zero_init_tensor(module.weight.data, module.out_features, module.in_features) + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 3. Convolutional Layers (Implementation of Algorithm 2) + elif isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)): + out_c, in_c = module.out_channels, module.in_channels + module.weight.data.zero_() + + # Find center index, taking n <- floor(k / 2) for each dimension + centers = tuple(k // 2 for k in module.kernel_size) + + # Create an out_c x in_c block utilizing the base ZerO logic + block = torch.zeros(out_c, in_c, dtype=module.weight.dtype, device=module.weight.device) + _zero_init_tensor(block, out_c, in_c) + + # Place block in the center slice of the kernel + if len(centers) == 1: + module.weight.data[:, :, centers[0]] = block + elif len(centers) == 2: + module.weight.data[:, :, centers[0], centers[1]] = block + elif len(centers) == 3: + module.weight.data[:, :, centers[0], centers[1], centers[2]] = block + + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 4. Normalization Layers (Standard practice is Scale=1, Bias=0) + elif isinstance(module, (nn.LayerNorm, nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.fill_(1.0) + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + + return model + +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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + 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._sin_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__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.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), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + 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: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + 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: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + 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") + + +# ----------------------------- +# 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, fullgraph=True) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + base_model = apply_zero_init(base_model) + 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 + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Fri May 1 17:21:24 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 43C P0 123W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 34C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 34C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 41C P0 120W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 44C P0 130W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 34C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 42C P0 123W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 33C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:500 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:10/500 +warmup_step:20/500 +warmup_step:30/500 +warmup_step:40/500 +warmup_step:50/500 +warmup_step:60/500 +warmup_step:70/500 +warmup_step:80/500 +warmup_step:90/500 +warmup_step:100/500 +warmup_step:110/500 +warmup_step:120/500 +warmup_step:130/500 +warmup_step:140/500 +warmup_step:150/500 +warmup_step:160/500 +warmup_step:170/500 +warmup_step:180/500 +warmup_step:190/500 +warmup_step:200/500 +warmup_step:210/500 +warmup_step:220/500 +warmup_step:230/500 +warmup_step:240/500 +warmup_step:250/500 +warmup_step:260/500 +warmup_step:270/500 +warmup_step:280/500 +warmup_step:290/500 +warmup_step:300/500 +warmup_step:310/500 +warmup_step:320/500 +warmup_step:330/500 +warmup_step:340/500 +warmup_step:350/500 +warmup_step:360/500 +warmup_step:370/500 +warmup_step:380/500 +warmup_step:390/500 +warmup_step:400/500 +warmup_step:410/500 +warmup_step:420/500 +warmup_step:430/500 +warmup_step:440/500 +warmup_step:450/500 +warmup_step:460/500 +warmup_step:470/500 +warmup_step:480/500 +warmup_step:490/500 +warmup_step:500/500 +step:0/20000 val_loss:6.9363 val_bpb:4.1080 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9355 train_time:34ms step_avg:33.69ms +step:2/20000 train_loss:15.7807 train_time:86ms step_avg:42.82ms +step:3/20000 train_loss:7.3458 train_time:134ms step_avg:44.55ms +step:4/20000 train_loss:7.2066 train_time:183ms step_avg:45.85ms +step:5/20000 train_loss:7.2275 train_time:230ms step_avg:46.08ms +step:6/20000 train_loss:7.3732 train_time:278ms step_avg:46.37ms +step:7/20000 train_loss:6.4856 train_time:326ms step_avg:46.59ms +step:8/20000 train_loss:6.1885 train_time:375ms step_avg:46.82ms +step:9/20000 train_loss:5.9111 train_time:423ms step_avg:46.98ms +step:10/20000 train_loss:5.7179 train_time:471ms step_avg:47.09ms +step:200/20000 train_loss:2.9473 train_time:9698ms step_avg:48.49ms +step:400/20000 train_loss:2.4203 train_time:19401ms step_avg:48.50ms +step:600/20000 train_loss:2.5958 train_time:29130ms step_avg:48.55ms +step:800/20000 train_loss:2.3421 train_time:38857ms step_avg:48.57ms +step:1000/20000 train_loss:2.4133 train_time:48584ms step_avg:48.58ms +step:1000/20000 val_loss:2.3795 val_bpb:1.4093 train_time:48601ms step_avg:48.60ms +step:1200/20000 train_loss:2.4334 train_time:58329ms step_avg:48.61ms +step:1400/20000 train_loss:2.4710 train_time:68065ms step_avg:48.62ms +step:1600/20000 train_loss:2.1405 train_time:77806ms step_avg:48.63ms +step:1800/20000 train_loss:2.2448 train_time:87536ms step_avg:48.63ms +step:2000/20000 train_loss:2.2988 train_time:97267ms step_avg:48.63ms +step:2000/20000 val_loss:2.2813 val_bpb:1.3511 train_time:97284ms step_avg:48.64ms +step:2200/20000 train_loss:2.1245 train_time:106998ms step_avg:48.64ms +step:2400/20000 train_loss:2.2451 train_time:116722ms step_avg:48.63ms +step:2600/20000 train_loss:2.4558 train_time:126440ms step_avg:48.63ms +step:2800/20000 train_loss:2.2855 train_time:136153ms step_avg:48.63ms +step:3000/20000 train_loss:2.2724 train_time:145860ms step_avg:48.62ms +step:3000/20000 val_loss:2.2407 val_bpb:1.3271 train_time:145876ms step_avg:48.63ms +step:3200/20000 train_loss:2.2333 train_time:155570ms step_avg:48.62ms +step:3400/20000 train_loss:2.2067 train_time:165270ms step_avg:48.61ms +step:3600/20000 train_loss:2.1651 train_time:174973ms step_avg:48.60ms +step:3800/20000 train_loss:2.2628 train_time:184671ms step_avg:48.60ms +step:4000/20000 train_loss:2.2087 train_time:194366ms step_avg:48.59ms +step:4000/20000 val_loss:2.2175 val_bpb:1.3133 train_time:194382ms step_avg:48.60ms +step:4200/20000 train_loss:2.2175 train_time:204136ms step_avg:48.60ms +step:4400/20000 train_loss:2.1617 train_time:213826ms step_avg:48.60ms +step:4600/20000 train_loss:2.0184 train_time:223526ms step_avg:48.59ms +step:4800/20000 train_loss:2.3111 train_time:233218ms step_avg:48.59ms +step:5000/20000 train_loss:2.0792 train_time:242920ms step_avg:48.58ms +step:5000/20000 val_loss:2.2005 val_bpb:1.3033 train_time:242944ms step_avg:48.59ms +step:5200/20000 train_loss:2.2193 train_time:252632ms step_avg:48.58ms +step:5400/20000 train_loss:2.2357 train_time:262335ms step_avg:48.58ms +step:5600/20000 train_loss:2.2336 train_time:272027ms step_avg:48.58ms +step:5800/20000 train_loss:2.1926 train_time:281718ms step_avg:48.57ms +step:6000/20000 train_loss:2.2634 train_time:291419ms step_avg:48.57ms +step:6000/20000 val_loss:2.1906 val_bpb:1.2974 train_time:291435ms step_avg:48.57ms +step:6200/20000 train_loss:2.1317 train_time:301112ms step_avg:48.57ms +step:6400/20000 train_loss:2.2120 train_time:310805ms step_avg:48.56ms +step:6600/20000 train_loss:2.1746 train_time:320490ms step_avg:48.56ms +step:6800/20000 train_loss:2.2376 train_time:330175ms step_avg:48.56ms +step:7000/20000 train_loss:2.2767 train_time:339874ms step_avg:48.55ms +step:7000/20000 val_loss:2.1799 val_bpb:1.2911 train_time:339889ms step_avg:48.56ms +step:7200/20000 train_loss:2.2503 train_time:349600ms step_avg:48.56ms +step:7400/20000 train_loss:2.1664 train_time:359291ms step_avg:48.55ms +step:7600/20000 train_loss:2.0471 train_time:368976ms step_avg:48.55ms +step:7800/20000 train_loss:2.1940 train_time:378661ms step_avg:48.55ms +step:8000/20000 train_loss:2.1614 train_time:388350ms step_avg:48.54ms +step:8000/20000 val_loss:2.1705 val_bpb:1.2855 train_time:388367ms step_avg:48.55ms +step:8200/20000 train_loss:2.2335 train_time:398040ms step_avg:48.54ms +step:8400/20000 train_loss:2.1793 train_time:407787ms step_avg:48.55ms +step:8600/20000 train_loss:2.1869 train_time:417469ms step_avg:48.54ms +step:8800/20000 train_loss:2.1529 train_time:427166ms step_avg:48.54ms +step:9000/20000 train_loss:2.0746 train_time:436845ms step_avg:48.54ms +step:9000/20000 val_loss:2.1663 val_bpb:1.2830 train_time:436862ms step_avg:48.54ms +step:9200/20000 train_loss:2.1343 train_time:446535ms step_avg:48.54ms +step:9400/20000 train_loss:2.1866 train_time:456312ms step_avg:48.54ms +step:9600/20000 train_loss:2.1989 train_time:466006ms step_avg:48.54ms +step:9800/20000 train_loss:2.1228 train_time:475685ms step_avg:48.54ms +step:10000/20000 train_loss:2.1626 train_time:485374ms step_avg:48.54ms +step:10000/20000 val_loss:2.1603 val_bpb:1.2794 train_time:485390ms step_avg:48.54ms +step:10200/20000 train_loss:2.1160 train_time:495062ms step_avg:48.54ms +step:10400/20000 train_loss:2.1504 train_time:504743ms step_avg:48.53ms +step:10600/20000 train_loss:2.0270 train_time:514434ms step_avg:48.53ms +step:10800/20000 train_loss:2.2328 train_time:524140ms step_avg:48.53ms +step:11000/20000 train_loss:2.1689 train_time:533835ms step_avg:48.53ms +step:11000/20000 val_loss:2.1548 val_bpb:1.2762 train_time:533852ms step_avg:48.53ms +step:11200/20000 train_loss:2.1158 train_time:543529ms step_avg:48.53ms +step:11400/20000 train_loss:2.0994 train_time:553225ms step_avg:48.53ms +step:11600/20000 train_loss:2.0988 train_time:562917ms step_avg:48.53ms +step:11800/20000 train_loss:2.1232 train_time:572599ms step_avg:48.53ms +step:12000/20000 train_loss:2.0914 train_time:582281ms step_avg:48.52ms +step:12000/20000 val_loss:2.1221 val_bpb:1.2569 train_time:582305ms step_avg:48.53ms +step:12200/20000 train_loss:2.2333 train_time:591982ms step_avg:48.52ms +step:12366/20000 val_loss:2.1106 val_bpb:1.2500 train_time:600049ms step_avg:48.52ms +stopping_early: wallclock_cap train_time:600049ms step:12366/20000 +peak memory allocated: 11263 MiB reserved: 11320 MiB +Serialized model: 74578915 bytes +Code size: 52042 bytes +Total submission size: 74630957 bytes +Serialized model int8+zlib: 15164905 bytes (payload:19030336 raw_torch:19080377 payload_ratio:3.92x) +Total submission size int8+zlib: 15216947 bytes +final_int8_zlib_roundtrip val_loss:2.1255 val_bpb:1.2588 eval_time:1570ms +final_int8_zlib_roundtrip_exact val_loss:2.12551013 val_bpb:1.25884682 diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/e31e596e-e21a-4c48-829c-78233c992cc8.txt b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/e31e596e-e21a-4c48-829c-78233c992cc8.txt new file mode 100644 index 0000000000..2e48843808 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/e31e596e-e21a-4c48-829c-78233c992cc8.txt @@ -0,0 +1,1455 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +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 + +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 +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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 +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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 +# ----------------------------- + +def apply_zero_init(model: nn.Module) -> nn.Module: + """ + Applies ZerO initialization to a given PyTorch nn.Module in-place. + + This function implements: + - Algorithm 1 (Linear/FFN Layers) + - Algorithm 2 (Convolution Layers) + - Section 4.1 Transformer specific initialization (W_Q=Identity, W_K/W_V=Zeros) + + Args: + model (nn.Module): The PyTorch model to initialize. + + Returns: + nn.Module: The model with initialized weights. + """ + + @torch.no_grad() + def _zero_init_tensor(tensor: torch.Tensor, out_f: int, in_f: int): + """Core logic corresponding to Algorithm 1 in the paper.""" + tensor.zero_() + + if out_f <= in_f: + # P_l == Q_l: Identity mapping + # P_l < Q_l: Partial identity (Propagates first P_l dimensions) + # torch.eye intrinsically handles rectangular partial-identities. + tensor.copy_(torch.eye(out_f, in_f, dtype=tensor.dtype, device=tensor.device)) + else: + # P_l > Q_l: Hadamard mapping + m = math.ceil(math.log2(out_f)) + m = max(m, 1) # Ensure at least m=1 + + # Recursively generate Hadamard matrix H_m + H = torch.tensor([[1.0]], dtype=tensor.dtype, device=tensor.device) + for _ in range(m): + H1 = torch.cat([H, H], dim=1) + H2 = torch.cat([H, -H], dim=1) + H = torch.cat([H1, H2], dim=0) + + # Normalization factor defined in the paper: c = 2^{-(m-1)/2} + c = 2.0 ** (-(m - 1) / 2.0) + + # Apply scaled Hadamard top-left submatrix (I* H_m I*) + tensor.copy_(c * H[:out_f, :in_f]) + + @torch.no_grad() + def _init_mha_zero(module: nn.MultiheadAttention): + """Initializes PyTorch's native MultiheadAttention according to Sec 4.1""" + embed_dim = module.embed_dim + + # in_proj_weight groups Q, K, and V + if module.in_proj_weight is not None: + module.in_proj_weight.zero_() + # W_Q is identity + module.in_proj_weight[:embed_dim, :].copy_( + torch.eye(embed_dim, dtype=module.in_proj_weight.dtype, device=module.in_proj_weight.device) + ) + # W_K, W_V remain zero (as handled by .zero_()) + else: + # If instantiated with separate weights + if getattr(module, 'q_proj_weight', None) is not None: + module.q_proj_weight.copy_( + torch.eye(embed_dim, dtype=module.q_proj_weight.dtype, device=module.q_proj_weight.device) + ) + if getattr(module, 'k_proj_weight', None) is not None: + module.k_proj_weight.zero_() + if getattr(module, 'v_proj_weight', None) is not None: + module.v_proj_weight.zero_() + + if getattr(module, 'in_proj_bias', None) is not None: + module.in_proj_bias.zero_() + + # The output projection relies on the standard algorithm (P_l = Q_l) -> Identity + if getattr(module, 'out_proj', None) is not None and getattr(module.out_proj, 'weight', None) is not None: + _zero_init_tensor(module.out_proj.weight, module.out_proj.out_features, module.out_proj.in_features) + if getattr(module.out_proj, 'bias', None) is not None: + module.out_proj.bias.zero_() + + @torch.no_grad() + def _init_conv_zero(module): + """Initializes Convolutions according to Algorithm 2""" + out_c, in_c = module.out_channels, module.in_channels + module.weight.zero_() + + # Find center index, taking n <- floor(k / 2) for each dimension + centers = tuple(k // 2 for k in module.kernel_size) + + # Create an out_c x in_c block utilizing the base ZerO logic + block = torch.zeros(out_c, in_c, dtype=module.weight.dtype, device=module.weight.device) + _zero_init_tensor(block, out_c, in_c) + + # Place block in the center slice of the kernel + if len(centers) == 1: # Conv1D + module.weight[:, :, centers[0]] = block + elif len(centers) == 2: # Conv2D + module.weight[:, :, centers[0], centers[1]] = block + elif len(centers) == 3: # Conv3D + module.weight[:, :, centers[0], centers[1], centers[2]] = block + + if getattr(module, 'bias', None) is not None: + module.bias.zero_() + + # Iterate through and parse all network submodules + for name, module in model.named_modules(): + + # 1. Transformers MultiheadAttention (PyTorch Native MHA) + if isinstance(module, nn.MultiheadAttention): + _init_mha_zero(module) + + # 2. Linear Layers (Applies to Feed-Forward Networks or HuggingFace Custom QKV Projections) + elif isinstance(module, nn.Linear): + name_lower = name.split('.')[-1].lower() + + # W_K, W_V at zero (Fallback heuristics for custom attention modules) + if any(k in name_lower for k in ['k_proj', 'v_proj', 'key', 'value']): + module.weight.data.zero_() + + # W_Q as identity (Fallback heuristics for custom attention modules) + elif any(q in name_lower for q in ['q_proj', 'query']): + module.weight.data.copy_( + torch.eye(module.out_features, module.in_features, dtype=module.weight.dtype, device=module.weight.device) + ) + + # Generic linear layer (e.g., Dimension expanding/shrinking FFN) + else: + _zero_init_tensor(module.weight, module.out_features, module.in_features) + + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 3. Convolutional Layers + elif isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)): + _init_conv_zero(module) + + # 4. Normalization Layers (Standard practice is Scale=1, Bias=0) + elif isinstance(module, (nn.LayerNorm, nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): + if getattr(module, 'weight', None) is not None: + module.weight.data.fill_(1.0) + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + return model + +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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + 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._sin_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__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.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), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + 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: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + 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: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + 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") + + +# ----------------------------- +# 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 + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + base_model = apply_zero_init(base_model) + 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 + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Fri May 1 15:54:01 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 33C P0 119W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 31C P0 115W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 31C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 32C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 35C P0 122W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 32C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 33C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 30C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059912 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9315 val_bpb:4.1053 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9313 train_time:35ms step_avg:34.95ms +step:2/20000 train_loss:11.5793 train_time:74ms step_avg:36.76ms +step:3/20000 train_loss:7.3305 train_time:118ms step_avg:39.18ms +step:4/20000 train_loss:6.4159 train_time:160ms step_avg:40.09ms +step:5/20000 train_loss:6.0377 train_time:205ms step_avg:41.09ms +step:6/20000 train_loss:6.2271 train_time:249ms step_avg:41.46ms +step:7/20000 train_loss:5.9458 train_time:293ms step_avg:41.86ms +step:8/20000 train_loss:5.8528 train_time:336ms step_avg:42.02ms +step:9/20000 train_loss:5.7863 train_time:380ms step_avg:42.23ms +step:10/20000 train_loss:5.7175 train_time:424ms step_avg:42.35ms +step:200/20000 train_loss:3.0837 train_time:8740ms step_avg:43.70ms +step:400/20000 train_loss:2.4340 train_time:17910ms step_avg:44.77ms +step:600/20000 train_loss:2.5968 train_time:26686ms step_avg:44.48ms +step:800/20000 train_loss:2.3361 train_time:35450ms step_avg:44.31ms +step:1000/20000 train_loss:2.4013 train_time:44257ms step_avg:44.26ms +step:1000/20000 val_loss:2.3684 val_bpb:1.4027 train_time:44272ms step_avg:44.27ms +step:1200/20000 train_loss:2.4126 train_time:53068ms step_avg:44.22ms +step:1400/20000 train_loss:2.4552 train_time:61880ms step_avg:44.20ms +step:1600/20000 train_loss:2.1209 train_time:70707ms step_avg:44.19ms +step:1800/20000 train_loss:2.2201 train_time:79530ms step_avg:44.18ms +step:2000/20000 train_loss:2.2712 train_time:88349ms step_avg:44.17ms +step:2000/20000 val_loss:2.2565 val_bpb:1.3364 train_time:88364ms step_avg:44.18ms +step:2200/20000 train_loss:2.0940 train_time:97181ms step_avg:44.17ms +step:2400/20000 train_loss:2.2174 train_time:105993ms step_avg:44.16ms +step:2600/20000 train_loss:2.4347 train_time:114808ms step_avg:44.16ms +step:2800/20000 train_loss:2.2522 train_time:123633ms step_avg:44.15ms +step:3000/20000 train_loss:2.2455 train_time:132437ms step_avg:44.15ms +step:3000/20000 val_loss:2.2117 val_bpb:1.3099 train_time:132451ms step_avg:44.15ms +step:3200/20000 train_loss:2.2056 train_time:141233ms step_avg:44.14ms +step:3400/20000 train_loss:2.1791 train_time:150025ms step_avg:44.13ms +step:3600/20000 train_loss:2.1333 train_time:158824ms step_avg:44.12ms +step:3800/20000 train_loss:2.2386 train_time:167613ms step_avg:44.11ms +step:4000/20000 train_loss:2.1768 train_time:176396ms step_avg:44.10ms +step:4000/20000 val_loss:2.1844 val_bpb:1.2937 train_time:176410ms step_avg:44.10ms +step:4200/20000 train_loss:2.1867 train_time:185285ms step_avg:44.12ms +step:4400/20000 train_loss:2.1291 train_time:194065ms step_avg:44.11ms +step:4600/20000 train_loss:1.9870 train_time:202854ms step_avg:44.10ms +step:4800/20000 train_loss:2.2776 train_time:211640ms step_avg:44.09ms +step:5000/20000 train_loss:2.0421 train_time:220421ms step_avg:44.08ms +step:5000/20000 val_loss:2.1662 val_bpb:1.2829 train_time:220434ms step_avg:44.09ms +step:5200/20000 train_loss:2.1840 train_time:229207ms step_avg:44.08ms +step:5400/20000 train_loss:2.1990 train_time:237993ms step_avg:44.07ms +step:5600/20000 train_loss:2.1972 train_time:246769ms step_avg:44.07ms +step:5800/20000 train_loss:2.1559 train_time:255545ms step_avg:44.06ms +step:6000/20000 train_loss:2.2323 train_time:264331ms step_avg:44.06ms +step:6000/20000 val_loss:2.1560 val_bpb:1.2769 train_time:264344ms step_avg:44.06ms +step:6200/20000 train_loss:2.0976 train_time:273199ms step_avg:44.06ms +step:6400/20000 train_loss:2.1702 train_time:281981ms step_avg:44.06ms +step:6600/20000 train_loss:2.1342 train_time:290758ms step_avg:44.05ms +step:6800/20000 train_loss:2.2053 train_time:299542ms step_avg:44.05ms +step:7000/20000 train_loss:2.2348 train_time:308319ms step_avg:44.05ms +step:7000/20000 val_loss:2.1433 val_bpb:1.2694 train_time:308334ms step_avg:44.05ms +step:7200/20000 train_loss:2.2099 train_time:317103ms step_avg:44.04ms +step:7400/20000 train_loss:2.1284 train_time:325890ms step_avg:44.04ms +step:7600/20000 train_loss:2.0098 train_time:334673ms step_avg:44.04ms +step:7800/20000 train_loss:2.1552 train_time:343460ms step_avg:44.03ms +step:8000/20000 train_loss:2.1270 train_time:352237ms step_avg:44.03ms +step:8000/20000 val_loss:2.1333 val_bpb:1.2634 train_time:352251ms step_avg:44.03ms +step:8200/20000 train_loss:2.1960 train_time:361006ms step_avg:44.03ms +step:8400/20000 train_loss:2.1513 train_time:369845ms step_avg:44.03ms +step:8600/20000 train_loss:2.1514 train_time:378621ms step_avg:44.03ms +step:8800/20000 train_loss:2.1110 train_time:387420ms step_avg:44.03ms +step:9000/20000 train_loss:2.0373 train_time:396203ms step_avg:44.02ms +step:9000/20000 val_loss:2.1278 val_bpb:1.2602 train_time:396217ms step_avg:44.02ms +step:9200/20000 train_loss:2.0951 train_time:404984ms step_avg:44.02ms +step:9400/20000 train_loss:2.1393 train_time:413762ms step_avg:44.02ms +step:9600/20000 train_loss:2.1571 train_time:422546ms step_avg:44.02ms +step:9800/20000 train_loss:2.0810 train_time:431318ms step_avg:44.01ms +step:10000/20000 train_loss:2.1207 train_time:440096ms step_avg:44.01ms +step:10000/20000 val_loss:2.1216 val_bpb:1.2565 train_time:440110ms step_avg:44.01ms +step:10200/20000 train_loss:2.0766 train_time:448869ms step_avg:44.01ms +step:10400/20000 train_loss:2.1118 train_time:457642ms step_avg:44.00ms +step:10600/20000 train_loss:1.9861 train_time:466421ms step_avg:44.00ms +step:10800/20000 train_loss:2.1940 train_time:475206ms step_avg:44.00ms +step:11000/20000 train_loss:2.1234 train_time:483991ms step_avg:44.00ms +step:11000/20000 val_loss:2.1153 val_bpb:1.2528 train_time:484005ms step_avg:44.00ms +step:11200/20000 train_loss:2.0787 train_time:492763ms step_avg:44.00ms +step:11400/20000 train_loss:2.0635 train_time:501539ms step_avg:43.99ms +step:11600/20000 train_loss:2.0741 train_time:510320ms step_avg:43.99ms +step:11800/20000 train_loss:2.1060 train_time:519094ms step_avg:43.99ms +step:12000/20000 train_loss:2.0810 train_time:527876ms step_avg:43.99ms +step:12000/20000 val_loss:2.1095 val_bpb:1.2494 train_time:527890ms step_avg:43.99ms +step:12200/20000 train_loss:2.2265 train_time:536662ms step_avg:43.99ms +step:12400/20000 train_loss:1.8727 train_time:545506ms step_avg:43.99ms +step:12600/20000 train_loss:2.0998 train_time:554286ms step_avg:43.99ms +step:12800/20000 train_loss:2.1115 train_time:563070ms step_avg:43.99ms +step:13000/20000 train_loss:2.1838 train_time:571849ms step_avg:43.99ms +step:13000/20000 val_loss:2.0895 val_bpb:1.2375 train_time:571862ms step_avg:43.99ms +step:13200/20000 train_loss:2.1871 train_time:580628ms step_avg:43.99ms +step:13400/20000 train_loss:2.0566 train_time:589407ms step_avg:43.99ms +step:13600/20000 train_loss:1.9241 train_time:598195ms step_avg:43.98ms +step:13642/20000 val_loss:2.0669 val_bpb:1.2242 train_time:600038ms step_avg:43.98ms +stopping_early: wallclock_cap train_time:600038ms step:13642/20000 +peak memory allocated: 10119 MiB reserved: 10294 MiB +Serialized model: 67224983 bytes +Code size: 54138 bytes +Total submission size: 67279121 bytes +Serialized model int8+zlib: 15585390 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x) +Total submission size int8+zlib: 15639528 bytes +final_int8_zlib_roundtrip val_loss:2.0810 val_bpb:1.2325 eval_time:1410ms +final_int8_zlib_roundtrip_exact val_loss:2.08096715 val_bpb:1.23246596 diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/fe927335-4827-415a-b543-8b5d2706de4c.txt b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/fe927335-4827-415a-b543-8b5d2706de4c.txt new file mode 100644 index 0000000000..262271b0eb --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/run_logs/fe927335-4827-415a-b543-8b5d2706de4c.txt @@ -0,0 +1,1445 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +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 + +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 +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 100)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.05)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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 +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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 +# ----------------------------- + +def apply_zero_init(model: nn.Module) -> nn.Module: + """ + Applies ZerO initialization to a given PyTorch nn.Module in-place. + + This function implements: + - Algorithm 1 (Linear/FFN Layers) + - Algorithm 2 (Convolution Layers) + - Section 4.1 Transformer specific initialization (W_Q=Identity, W_K/W_V=Zeros) + + Args: + model (nn.Module): The PyTorch model to initialize. + + Returns: + nn.Module: The model with initialized weights. + """ + + @torch.no_grad() + def _zero_init_tensor(tensor: torch.Tensor, out_f: int, in_f: int): + """Core logic corresponding to Algorithm 1 in the paper.""" + tensor.zero_() + + if out_f <= in_f: + # P_l == Q_l: Identity mapping + # P_l < Q_l: Partial identity (Propagates first P_l dimensions) + # torch.eye intrinsically handles rectangular partial-identities. + tensor.copy_(torch.eye(out_f, in_f, dtype=tensor.dtype, device=tensor.device)) + else: + # P_l > Q_l: Hadamard mapping + m = math.ceil(math.log2(out_f)) + m = max(m, 1) # Ensure at least m=1 + + # Recursively generate Hadamard matrix H_m + H = torch.tensor([[1.0]], dtype=tensor.dtype, device=tensor.device) + for _ in range(m): + H1 = torch.cat([H, H], dim=1) + H2 = torch.cat([H, -H], dim=1) + H = torch.cat([H1, H2], dim=0) + + # Normalization factor defined in the paper: c = 2^{-(m-1)/2} + c = 2.0 ** (-(m - 1) / 2.0) + + # Apply scaled Hadamard top-left submatrix (I* H_m I*) + tensor.copy_(c * H[:out_f, :in_f]) + + @torch.no_grad() + def _init_mha_zero(module: nn.MultiheadAttention): + """Initializes PyTorch's native MultiheadAttention according to Sec 4.1""" + embed_dim = module.embed_dim + + # in_proj_weight groups Q, K, and V + if module.in_proj_weight is not None: + module.in_proj_weight.zero_() + # W_Q is identity + module.in_proj_weight[:embed_dim, :].copy_( + torch.eye(embed_dim, dtype=module.in_proj_weight.dtype, device=module.in_proj_weight.device) + ) + # W_K, W_V remain zero (as handled by .zero_()) + else: + # If instantiated with separate weights + if getattr(module, 'q_proj_weight', None) is not None: + module.q_proj_weight.copy_( + torch.eye(embed_dim, dtype=module.q_proj_weight.dtype, device=module.q_proj_weight.device) + ) + if getattr(module, 'k_proj_weight', None) is not None: + module.k_proj_weight.zero_() + if getattr(module, 'v_proj_weight', None) is not None: + module.v_proj_weight.zero_() + + if getattr(module, 'in_proj_bias', None) is not None: + module.in_proj_bias.zero_() + + # The output projection relies on the standard algorithm (P_l = Q_l) -> Identity + if getattr(module, 'out_proj', None) is not None and getattr(module.out_proj, 'weight', None) is not None: + _zero_init_tensor(module.out_proj.weight, module.out_proj.out_features, module.out_proj.in_features) + if getattr(module.out_proj, 'bias', None) is not None: + module.out_proj.bias.zero_() + + @torch.no_grad() + def _init_conv_zero(module): + """Initializes Convolutions according to Algorithm 2""" + out_c, in_c = module.out_channels, module.in_channels + module.weight.zero_() + + # Find center index, taking n <- floor(k / 2) for each dimension + centers = tuple(k // 2 for k in module.kernel_size) + + # Create an out_c x in_c block utilizing the base ZerO logic + block = torch.zeros(out_c, in_c, dtype=module.weight.dtype, device=module.weight.device) + _zero_init_tensor(block, out_c, in_c) + + # Place block in the center slice of the kernel + if len(centers) == 1: # Conv1D + module.weight[:, :, centers[0]] = block + elif len(centers) == 2: # Conv2D + module.weight[:, :, centers[0], centers[1]] = block + elif len(centers) == 3: # Conv3D + module.weight[:, :, centers[0], centers[1], centers[2]] = block + + if getattr(module, 'bias', None) is not None: + module.bias.zero_() + + # Iterate through and parse all network submodules + for name, module in model.named_modules(): + + # 1. Transformers MultiheadAttention (PyTorch Native MHA) + if isinstance(module, nn.MultiheadAttention): + _init_mha_zero(module) + + # 2. Linear Layers (Applies to Feed-Forward Networks or HuggingFace Custom QKV Projections) + elif isinstance(module, nn.Linear): + name_lower = name.split('.')[-1].lower() + + # W_K, W_V at zero (Fallback heuristics for custom attention modules) + if any(k in name_lower for k in ['k_proj', 'v_proj', 'key', 'value']): + module.weight.data.zero_() + + # W_Q as identity (Fallback heuristics for custom attention modules) + elif any(q in name_lower for q in ['q_proj', 'query']): + module.weight.data.copy_( + torch.eye(module.out_features, module.in_features, dtype=module.weight.dtype, device=module.weight.device) + ) + + # Generic linear layer (e.g., Dimension expanding/shrinking FFN) + else: + _zero_init_tensor(module.weight, module.out_features, module.in_features) + + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 3. Convolutional Layers + elif isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)): + _init_conv_zero(module) + + # 4. Normalization Layers (Standard practice is Scale=1, Bias=0) + elif isinstance(module, (nn.LayerNorm, nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): + if getattr(module, 'weight', None) is not None: + module.weight.data.fill_(1.0) + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + return model + +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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + 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._sin_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__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.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), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + 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: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + 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: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + 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") + + +# ----------------------------- +# 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, fullgraph=True) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + base_model = apply_zero_init(base_model) + 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 + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Fri May 1 16:17:15 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 34C P0 120W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 31C P0 117W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 31C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 33C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 36C P0 123W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 32C P0 118W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 34C P0 119W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 31C P0 116W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059912 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.05 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:100 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:10/100 +warmup_step:20/100 +warmup_step:30/100 +warmup_step:40/100 +warmup_step:50/100 +warmup_step:60/100 +warmup_step:70/100 +warmup_step:80/100 +warmup_step:90/100 +warmup_step:100/100 +step:0/20000 val_loss:6.9315 val_bpb:4.1053 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9313 train_time:32ms step_avg:31.75ms +step:2/20000 train_loss:11.5735 train_time:78ms step_avg:38.84ms +step:3/20000 train_loss:7.1929 train_time:123ms step_avg:40.89ms +step:4/20000 train_loss:6.3270 train_time:167ms step_avg:41.81ms +step:5/20000 train_loss:6.0549 train_time:211ms step_avg:42.24ms +step:6/20000 train_loss:6.1655 train_time:256ms step_avg:42.66ms +step:7/20000 train_loss:5.8802 train_time:299ms step_avg:42.66ms +step:8/20000 train_loss:5.8191 train_time:342ms step_avg:42.74ms +step:9/20000 train_loss:5.7510 train_time:384ms step_avg:42.64ms +step:10/20000 train_loss:5.6883 train_time:429ms step_avg:42.88ms +step:200/20000 train_loss:3.0501 train_time:8729ms step_avg:43.64ms +step:400/20000 train_loss:2.4298 train_time:17491ms step_avg:43.73ms +step:600/20000 train_loss:2.5824 train_time:26245ms step_avg:43.74ms +step:800/20000 train_loss:2.3211 train_time:35007ms step_avg:43.76ms +step:1000/20000 train_loss:2.3946 train_time:43784ms step_avg:43.78ms +step:1000/20000 val_loss:2.3563 val_bpb:1.3955 train_time:43800ms step_avg:43.80ms +step:1200/20000 train_loss:2.4059 train_time:52566ms step_avg:43.80ms +step:1400/20000 train_loss:2.4449 train_time:61368ms step_avg:43.83ms +step:1600/20000 train_loss:2.1165 train_time:70173ms step_avg:43.86ms +step:1800/20000 train_loss:2.2146 train_time:78960ms step_avg:43.87ms +step:2000/20000 train_loss:2.2682 train_time:87756ms step_avg:43.88ms +step:2000/20000 val_loss:2.2519 val_bpb:1.3337 train_time:87770ms step_avg:43.89ms +step:2200/20000 train_loss:2.0919 train_time:96553ms step_avg:43.89ms +step:2400/20000 train_loss:2.2119 train_time:105340ms step_avg:43.89ms +step:2600/20000 train_loss:2.4323 train_time:114118ms step_avg:43.89ms +step:2800/20000 train_loss:2.2503 train_time:122918ms step_avg:43.90ms +step:3000/20000 train_loss:2.2446 train_time:131762ms step_avg:43.92ms +step:3000/20000 val_loss:2.2075 val_bpb:1.3074 train_time:131777ms step_avg:43.93ms +step:3200/20000 train_loss:2.2037 train_time:140550ms step_avg:43.92ms +step:3400/20000 train_loss:2.1692 train_time:149329ms step_avg:43.92ms +step:3600/20000 train_loss:2.1296 train_time:158110ms step_avg:43.92ms +step:3800/20000 train_loss:2.2317 train_time:166864ms step_avg:43.91ms +step:4000/20000 train_loss:2.1715 train_time:175643ms step_avg:43.91ms +step:4000/20000 val_loss:2.1805 val_bpb:1.2914 train_time:175658ms step_avg:43.91ms +step:4200/20000 train_loss:2.1877 train_time:184467ms step_avg:43.92ms +step:4400/20000 train_loss:2.1236 train_time:193243ms step_avg:43.92ms +step:4600/20000 train_loss:1.9859 train_time:202015ms step_avg:43.92ms +step:4800/20000 train_loss:2.2733 train_time:210787ms step_avg:43.91ms +step:5000/20000 train_loss:2.0361 train_time:219552ms step_avg:43.91ms +step:5000/20000 val_loss:2.1631 val_bpb:1.2811 train_time:219567ms step_avg:43.91ms +step:5200/20000 train_loss:2.1834 train_time:228313ms step_avg:43.91ms +step:5400/20000 train_loss:2.1937 train_time:237082ms step_avg:43.90ms +step:5600/20000 train_loss:2.1897 train_time:245839ms step_avg:43.90ms +step:5800/20000 train_loss:2.1550 train_time:254602ms step_avg:43.90ms +step:6000/20000 train_loss:2.2332 train_time:263360ms step_avg:43.89ms +step:6000/20000 val_loss:2.1526 val_bpb:1.2749 train_time:263375ms step_avg:43.90ms +step:6200/20000 train_loss:2.0989 train_time:272117ms step_avg:43.89ms +step:6400/20000 train_loss:2.1725 train_time:280883ms step_avg:43.89ms +step:6600/20000 train_loss:2.1283 train_time:289631ms step_avg:43.88ms +step:6800/20000 train_loss:2.2035 train_time:298379ms step_avg:43.88ms +step:7000/20000 train_loss:2.2323 train_time:307142ms step_avg:43.88ms +step:7000/20000 val_loss:2.1415 val_bpb:1.2683 train_time:307157ms step_avg:43.88ms +step:7200/20000 train_loss:2.2079 train_time:315894ms step_avg:43.87ms +step:7400/20000 train_loss:2.1283 train_time:324655ms step_avg:43.87ms +step:7600/20000 train_loss:2.0099 train_time:333407ms step_avg:43.87ms +step:7800/20000 train_loss:2.1548 train_time:342160ms step_avg:43.87ms +step:8000/20000 train_loss:2.1226 train_time:350916ms step_avg:43.86ms +step:8000/20000 val_loss:2.1314 val_bpb:1.2623 train_time:350930ms step_avg:43.87ms +step:8200/20000 train_loss:2.1968 train_time:359678ms step_avg:43.86ms +step:8400/20000 train_loss:2.1435 train_time:368495ms step_avg:43.87ms +step:8600/20000 train_loss:2.1418 train_time:377246ms step_avg:43.87ms +step:8800/20000 train_loss:2.1074 train_time:386013ms step_avg:43.87ms +step:9000/20000 train_loss:2.0360 train_time:394770ms step_avg:43.86ms +step:9000/20000 val_loss:2.1262 val_bpb:1.2593 train_time:394784ms step_avg:43.86ms +step:9200/20000 train_loss:2.0950 train_time:403535ms step_avg:43.86ms +step:9400/20000 train_loss:2.1434 train_time:412289ms step_avg:43.86ms +step:9600/20000 train_loss:2.1575 train_time:421036ms step_avg:43.86ms +step:9800/20000 train_loss:2.0780 train_time:429798ms step_avg:43.86ms +step:10000/20000 train_loss:2.1219 train_time:438556ms step_avg:43.86ms +step:10000/20000 val_loss:2.1198 val_bpb:1.2555 train_time:438582ms step_avg:43.86ms +step:10200/20000 train_loss:2.0745 train_time:447326ms step_avg:43.86ms +step:10400/20000 train_loss:2.1079 train_time:456077ms step_avg:43.85ms +step:10600/20000 train_loss:1.9842 train_time:464835ms step_avg:43.85ms +step:10800/20000 train_loss:2.1929 train_time:473598ms step_avg:43.85ms +step:11000/20000 train_loss:2.1213 train_time:482368ms step_avg:43.85ms +step:11000/20000 val_loss:2.1136 val_bpb:1.2518 train_time:482383ms step_avg:43.85ms +step:11200/20000 train_loss:2.0772 train_time:491132ms step_avg:43.85ms +step:11400/20000 train_loss:2.0655 train_time:499891ms step_avg:43.85ms +step:11600/20000 train_loss:2.0719 train_time:508649ms step_avg:43.85ms +step:11800/20000 train_loss:2.0984 train_time:517410ms step_avg:43.85ms +step:12000/20000 train_loss:2.0816 train_time:526173ms step_avg:43.85ms +step:12000/20000 val_loss:2.1085 val_bpb:1.2488 train_time:526188ms step_avg:43.85ms +step:12200/20000 train_loss:2.2273 train_time:534932ms step_avg:43.85ms +step:12400/20000 train_loss:1.8696 train_time:543766ms step_avg:43.85ms +step:12600/20000 train_loss:2.0964 train_time:552523ms step_avg:43.85ms +step:12800/20000 train_loss:2.1153 train_time:561272ms step_avg:43.85ms +step:13000/20000 train_loss:2.1838 train_time:570022ms step_avg:43.85ms +step:13000/20000 val_loss:2.0897 val_bpb:1.2376 train_time:570037ms step_avg:43.85ms +step:13200/20000 train_loss:2.1885 train_time:578784ms step_avg:43.85ms +step:13400/20000 train_loss:2.0564 train_time:587538ms step_avg:43.85ms +step:13600/20000 train_loss:1.9180 train_time:596292ms step_avg:43.85ms +step:13686/20000 val_loss:2.0654 val_bpb:1.2233 train_time:600051ms step_avg:43.84ms +stopping_early: wallclock_cap train_time:600051ms step:13686/20000 +peak memory allocated: 10119 MiB reserved: 10438 MiB +Serialized model: 67224983 bytes +Code size: 54155 bytes +Total submission size: 67279138 bytes +Serialized model int8+zlib: 15645913 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x) +Total submission size int8+zlib: 15700068 bytes +final_int8_zlib_roundtrip val_loss:2.0792 val_bpb:1.2314 eval_time:1407ms +final_int8_zlib_roundtrip_exact val_loss:2.07922175 val_bpb:1.23143224 diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/submission.json b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/submission.json new file mode 100644 index 0000000000..279d5ba326 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/submission.json @@ -0,0 +1,11 @@ +{ + "track": "non_record_16mb", + "date": "2026-05-01", + "name": "Non-record: Corrected ZerO initialization + Follow up to PR 2104", + "author": "Alston Tang", + "github_id": "AlstonTang", + "val_bpb": 1.24938225, + "val_loss": 2.10952960, + "bytes_total": 15221665 + } + \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/train_gpt.py b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/train_gpt.py new file mode 100644 index 0000000000..1e111ea556 --- /dev/null +++ b/records/track_non_record_16mb/2026-05-01_Follow_up_to_PR_2104/train_gpt.py @@ -0,0 +1,1223 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +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 + +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 +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 500)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + 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 = int(os.environ.get("MLP_MULT", 2)) + 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)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + 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 +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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 +# ----------------------------- + +def apply_zero_init(model: nn.Module) -> nn.Module: + """ + Applies ZerO initialization to a given PyTorch nn.Module in-place. + + This function implements: + - Algorithm 1 (Linear/FFN Layers using Identity, Partial Identity, and Hadamard) + - Algorithm 2 (Convolution Layers) + + Args: + model (nn.Module): The PyTorch model to initialize. + + Returns: + nn.Module: The model with initialized weights. + """ + + @torch.no_grad() + def _zero_init_tensor(tensor: torch.Tensor, out_f: int, in_f: int): + """Core logic corresponding to Algorithm 1 in the paper.""" + tensor.zero_() + + if out_f <= in_f: + # P_l == Q_l: Identity mapping + # P_l < Q_l: Partial identity (Propagates first P_l dimensions) + tensor.copy_(torch.eye(out_f, in_f, dtype=tensor.dtype, device=tensor.device)) + else: + # P_l > Q_l: Hadamard mapping (e.g., expanding MLPs) + m = math.ceil(math.log2(out_f)) + m = max(m, 1) # Ensure at least m=1 + + # Recursively generate Hadamard matrix H_m + H = torch.tensor([[1.0]], dtype=tensor.dtype, device=tensor.device) + for _ in range(m): + H1 = torch.cat([H, H], dim=1) + H2 = torch.cat([H, -H], dim=1) + H = torch.cat([H1, H2], dim=0) + + # Normalization factor defined in the paper: c = 2^{-(m-1)/2} + c = 2.0 ** (-(m - 1) / 2.0) + + # Apply scaled Hadamard top-left submatrix (I* H_m I*) + tensor.copy_(c * H[:out_f, :in_f]) + + # Iterate through and parse all network submodules + for name, module in model.named_modules(): + + # 1. Respect the script's architectural _zero_init flags (e.g., `proj` & `lm_head`). + # This achieves the paper's goal of dynamical isometry (Identity residual pass) + # and avoids the gradient deadlock. + if getattr(module, '_zero_init', False): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.zero_() + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + continue + + # 2. Linear Layers (c_q, c_k, c_v, fc) + # c_q (512->512) becomes Identity. + # c_k/c_v (512->256) become Partial Identities. + # fc (512->1024) becomes a Hadamard transform. + if isinstance(module, nn.Linear): + _zero_init_tensor(module.weight.data, module.out_features, module.in_features) + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 3. Convolutional Layers (Implementation of Algorithm 2) + elif isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)): + out_c, in_c = module.out_channels, module.in_channels + module.weight.data.zero_() + + # Find center index, taking n <- floor(k / 2) for each dimension + centers = tuple(k // 2 for k in module.kernel_size) + + # Create an out_c x in_c block utilizing the base ZerO logic + block = torch.zeros(out_c, in_c, dtype=module.weight.dtype, device=module.weight.device) + _zero_init_tensor(block, out_c, in_c) + + # Place block in the center slice of the kernel + if len(centers) == 1: + module.weight.data[:, :, centers[0]] = block + elif len(centers) == 2: + module.weight.data[:, :, centers[0], centers[1]] = block + elif len(centers) == 3: + module.weight.data[:, :, centers[0], centers[1], centers[2]] = block + + if getattr(module, 'bias', None) is not None: + module.bias.data.zero_() + + # 4. Normalization Layers (Standard practice is Scale=1, Bias=0) + elif isinstance(module, (nn.LayerNorm, nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): + if hasattr(module, 'weight') and module.weight is not None: + module.weight.data.fill_(1.0) + if hasattr(module, 'bias') and module.bias is not None: + module.bias.data.zero_() + + return model + +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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + 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._sin_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__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.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), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + 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: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + 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: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + 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") + + +# ----------------------------- +# 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, fullgraph=True) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + 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, + ).to(device).bfloat16() + base_model = apply_zero_init(base_model) + 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 + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + 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_qeval):.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()