diff --git a/records/track_10min_16mb/2026-04-01_21L_PRP/README.md b/records/track_10min_16mb/2026-04-01_21L_PRP/README.md new file mode 100644 index 0000000000..b9f9e23028 --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_21L_PRP/README.md @@ -0,0 +1,109 @@ +# 21L + PRP SwiGLU + Shared Attention + sp8192 + 2048 context length + +While this run demonstrates end-to-end results, the current int8+zlib artifact (17.7 MB) exceeds the 16 MB track limit. Furthermore, due to local hardware constraints (Nvidia RTX 3080), the model processed only 2.5–5% of the total target tokens required for a full convergence run (defined as a <10-minute training window on an 8xH100 cluster) + +This variant moves away from the default 1k-vocab, 1024-context, 9-layer starter and spends the budget on a deeper 8k-vocab model with parameter sharing, PRP-based MLPs, and a longer-horizon LR schedule. + +### Changes from Baseline + +**1. 8K tokenizer + longer context** + +The script switches from `fineweb10B_sp1024` to `fineweb10B_sp8192`, uses the 8k SentencePiece tokenizer, and doubles context length to `TRAIN_SEQ_LEN=2048`. The default model shape becomes 21 layers at width 512 with 8 query heads and 2 KV heads. + +**2. PRP SwiGLU MLP instead of the baseline relu^2 MLP** + +The dense MLP is replaced with a SwiGLU block built from 'ParametrizedRandomProjection' (https://arxiv.org/pdf/2512.13480). Each PRP layer keeps a fixed random projection buffer and only trains low-dimensional controls: input scaling, output scaling, output bias, and a rank-32 LoRA update. The MLP projects once to `2 * hidden`, then splits into gate and value halves. A higher MLP multiplier is used: 4 instead of 2. + +**3. Pairwise shared attention in the transformer body** + +Layer 1 and the final layer remain unique. Interior layers share attention modules in adjacent pairs: `(2,3)`, `(4,5)`, `(6,7)`, and so on. Each block still keeps its own norms, residual mixing, scales, and MLP, so depth increases without paying for 21 fully independent attention stacks. + +In the logged run, that produces 21 transformer blocks but only 12 unique attention modules. + +**4. Separate optimizer path for PRP vector controls** + +The script adds `unique_named_parameters` and splits PRP vector parameters (`.alpha`, `.weight`, `.bias`) into their own Adam group with `PRP_LR`. Matrix-shaped parameters still use Muon, while the remaining scalar and control tensors stay on Adam. + +**5. Three-phase cosine LR schedule** + +The default warmdown schedule is replaced with a longer-run schedule: +- cosine warmup from `LR_INIT_SCALE` to 1.0 +- cosine flash drop from 1.0 to `LR_MAIN_SCALE` +- powered cosine tail from `LR_MAIN_SCALE` to `LR_MIN_SCALE`, shaped by `LR_GAMMA` + +This will be adjusted for the full schedule run. + +**6. Shared-aware int8 export** + +The post-training export remains per-row int8 for 2D tensors and fp16/fp32 passthrough for small or control tensors, but it now deduplicates repeated storage before counting payload bytes. That matters once attention modules or embeddings are shared or tied. + +### Default Configuration in This Record + +```bash +DATA_PATH=./data/datasets/fineweb10B_sp8192 +TOKENIZER_PATH=./data/tokenizers/fineweb_8192_bpe.model + +VOCAB_SIZE=8192 +NUM_LAYERS=21 +MODEL_DIM=512 +NUM_HEADS=8 +NUM_KV_HEADS=2 +MLP_MULT=4 +TIE_EMBEDDINGS=1 + +TRAIN_SEQ_LEN=2048 +TRAIN_BATCH_TOKENS=65536 +VAL_BATCH_SIZE=65536 +VAL_FRAC=0.25 + +ITERATIONS=4000 +WARMUP_STEPS=2 +LR_MAIN_SCALE=0.5 +LR_MIN_SCALE=0.05 +LR_INIT_SCALE=0.0 +LR_GAMMA=0.8 + +TIED_EMBED_LR=0.03 +MATRIX_LR=0.03 +SCALAR_LR=0.03 +PRP_LR=0.06 +MUON_MOMENTUM=0.95 +``` + +Notes: +- `LR_WARMUP_STEPS` defaults to `int(0.01 * ITERATIONS)`. +- `LR_DROP_STEPS` defaults to `int(0.2 * ITERATIONS)`. +- The file lives under the 10-minute track, but the current defaults are closer to a longer research run than a finalized 10-minute submission. + +### Actual Results + +| Metric | Value | +|--------|-------| +| Final pre-quant val_bpb | 1.3527 | +| Final pre-quant val_loss | 3.4657 | +| int8+zlib roundtrip val_bpb | 1.3540 | +| int8+zlib roundtrip val_loss | 3.4690 | +| Quantization gap | +0.0013 BPB | +| Total params | 17,171,040 | +| Peak memory allocated | 7,329 MiB | +| Serialized model | 50,783,797 bytes | +| int8+zlib artifact | 17,603,010 bytes | +| Total submission size (artifact + code) | 17,658,504 bytes | + + +### Validation Trajectory + +| Step | val_loss | val_bpb | +|------|----------|---------| +| 0 | 9.0102 | 3.5169 | +| 400 | 4.2170 | 1.6460 | +| 800 | 3.9027 | 1.5233 | +| 1200 | 3.7963 | 1.4818 | +| 1600 | 3.7088 | 1.4476 | +| 2000 | 3.6501 | 1.4247 | +| 2400 | 3.6022 | 1.4060 | +| 2800 | 3.5515 | 1.3862 | +| 3200 | 3.5092 | 1.3697 | +| 3600 | 3.4804 | 1.3585 | +| 4000 | 3.4657 | 1.3527 | + diff --git a/records/track_10min_16mb/2026-04-01_21L_PRP/prp.py b/records/track_10min_16mb/2026-04-01_21L_PRP/prp.py new file mode 100644 index 0000000000..e456d34c24 --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_21L_PRP/prp.py @@ -0,0 +1,97 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +import math + + +class ParametrizedRandomProjection(nn.Module): + _layer_counter = 0 + def __init__( + self, + in_features, + out_features, + projection_type="random", + lora_rank: int = 32, + lora_alpha: float = 1.0, + ): + super().__init__() + self.in_features = in_features + self.out_features = out_features + self.projection_type = projection_type + self._layer_id = ParametrizedRandomProjection._layer_counter + ParametrizedRandomProjection._layer_counter += 1 + self.seed = hash((in_features, out_features, projection_type, self._layer_id)) % (2 ** 32) + + self.alpha = nn.Parameter(torch.empty(in_features)) + self.weight = nn.Parameter(torch.empty(out_features)) + self.bias = nn.Parameter(torch.empty(out_features)) + + std_in = 1 / math.sqrt(in_features) + nn.init.normal_(self.alpha, mean=1.0, std=std_in) + nn.init.normal_(self.weight, mean=1.0, std=std_in) + nn.init.zeros_(self.bias) + + # ========== Fixed random projection (non-learnable) ========== + g = torch.Generator(device="cpu") + g.manual_seed(self.seed) + std_dev = 1 / math.sqrt(self.in_features) + if self.projection_type == "random": + proj = torch.randn(self.in_features, self.out_features, generator=g) * std_dev + elif self.projection_type == "sparse": + val_scale = math.sqrt(3.0 / self.in_features) + u = torch.rand(self.in_features, self.out_features, generator=g) + proj = torch.zeros(self.in_features, self.out_features) + proj[u < 1 / 6] = -val_scale + proj[u >= 5 / 6] = val_scale + elif self.projection_type == "orthogonal": + proj = torch.randn(self.in_features, self.out_features, generator=g) * std_dev + nn.init.orthogonal_(proj) + elif self.projection_type == "uniform": + proj = (torch.rand(self.in_features, self.out_features, generator=g) * 2 - 1) * std_dev * math.sqrt(3) + else: + raise ValueError("Unsupported projection_type") + + # keep both names for compatibility + self.register_buffer("proj", proj, persistent=False) + self.register_buffer("fixed_proj", proj, persistent=False) + + # ========== LoRA low-rank adaptation ========== + self.lora_rank = lora_rank + self.lora_alpha = lora_alpha + self.lora_A = nn.Parameter(torch.empty(self.in_features, lora_rank)) + self.lora_B = nn.Parameter(torch.empty(lora_rank, self.out_features)) + nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5)) + nn.init.zeros_(self.lora_B) + self.lora_scale = float(self.lora_alpha) / max(1, self.lora_rank) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + # Move proj to correct device/dtype + proj = self.proj.to(dtype=x.dtype, device=x.device) + + # Element-wise scaling (no per-input bias) + x_scaled = x * self.alpha.unsqueeze(0) + + # Fixed projection path + out_fixed = x_scaled @ proj + out_fixed = out_fixed * self.weight.unsqueeze(0) + self.bias.unsqueeze(0) + + # LoRA adaptation as separate path + lora_out = (x @ self.lora_A) @ self.lora_B + lora_out = lora_out * self.lora_scale + + return out_fixed + lora_out + + def count_trainable_params(self) -> int: + return sum(p.numel() for p in self.parameters() if p.requires_grad) + + def __repr__(self): + return ( + f"{self.__class__.__name__}(" + f"in_features={self.in_features}, " + f"out_features={self.out_features}, " + f"projection_type='{self.projection_type}', " + f"lora_rank={self.lora_rank}, " + f"lora_alpha={self.lora_alpha}, " + f"trainable_params={self.count_trainable_params():,})" + ) + diff --git a/records/track_10min_16mb/2026-04-01_21L_PRP/submission.json b/records/track_10min_16mb/2026-04-01_21L_PRP/submission.json new file mode 100644 index 0000000000..c2f22a1616 --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_21L_PRP/submission.json @@ -0,0 +1,8 @@ +{ + "author": "UniversalComputingResearch", + "name": "21L + PRP SwiGLU + Shared Attention + sp8192 + 2048 context length", + "github": "https://github.com/UniversalComputingResearch", + "val_bpb": 1.3540, + "description": "Custom architecture experiment for Parameter Golf.", + "status": "Awaiting compute credits" +} diff --git a/records/track_10min_16mb/2026-04-01_21L_PRP/tokenizer_getter.py b/records/track_10min_16mb/2026-04-01_21L_PRP/tokenizer_getter.py new file mode 100644 index 0000000000..29b43a53a7 --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_21L_PRP/tokenizer_getter.py @@ -0,0 +1,24 @@ +import os +from huggingface_hub import hf_hub_download + +repo_id = "sproos/parameter-golf-tokenizers" + +# Define your custom paths (matching train_gpt.py expectations) +tok_dir = "./data/tokenizers/" +data_dir = "./data/datasets/fineweb10B_sp8192/" + +# 1. Download Tokenizer files (8k version) +tokenizer_files = ["tokenizers/fineweb_8192_bpe.model", "tokenizers/fineweb_8192_bpe.vocab"] +for file in tokenizer_files: + print(f"Downloading {file}...") + hf_hub_download(repo_id=repo_id, filename=file, local_dir=tok_dir) + +# 2. Download Dataset shards (Train 00-79 + Val) — 8192 vocab version +dataset_subdir = "datasets/fineweb10B_sp8192" +dataset_files = [f"{dataset_subdir}/fineweb_train_{i:06d}.bin" for i in range(80)] +dataset_files.append(f"{dataset_subdir}/fineweb_val_000000.bin") + +for file in dataset_files: + print(f"Downloading {file}...") + hf_hub_download(repo_id=repo_id, filename=file, local_dir=data_dir) + diff --git a/records/track_10min_16mb/2026-04-01_21L_PRP/train.log b/records/track_10min_16mb/2026-04-01_21L_PRP/train.log new file mode 100644 index 0000000000..ea423b3d5a --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_21L_PRP/train.log @@ -0,0 +1,1614 @@ +""" +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 +from prp import ParametrizedRandomProjection + + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp8192") + 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_8192_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", 65_536)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 400)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + # Fraction of the validation split to use (0 < VAL_FRAC <= 1.0) + val_frac = float(os.environ.get("VAL_FRAC", 0.25)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 4000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 2)) + # LR scheduling knobs (training warmup, flash drop, main phase, and gamma curvature) + lr_warmup_steps = int(os.environ.get("LR_WARMUP_STEPS", iterations * 0.01)) + lr_drop_steps = int(os.environ.get("LR_DROP_STEPS", iterations * 0.2)) + # Fractional scales relative to each param-group's base_lr + lr_main_scale = float(os.environ.get("LR_MAIN_SCALE", 0.5)) + lr_min_scale = float(os.environ.get("LR_MIN_SCALE", 0.05)) + lr_init_scale = float(os.environ.get("LR_INIT_SCALE", 0.0)) + lr_gamma = float(os.environ.get("LR_GAMMA", 0.8)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 65_536)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 3600.0 * 2)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 8192)) + num_layers = int(os.environ.get("NUM_LAYERS", 21)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 2)) + 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", 4)) + 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", 0.4)) # not used since tied + head_lr = float(os.environ.get("HEAD_LR", 0.02)) # not used since tied + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.004)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.03)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.03)) + prp_lr = float(os.environ.get("PRP_LR", 0.06)) + 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", lr_warmup_steps + lr_drop_steps)) + 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, + ) + + # Track already-seen storage pointers to handle shared/tied parameters + seen_ptr: dict[int, str] = {} + 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) + + ptr = tensor.data_ptr() + first_name = seen_ptr.get(ptr) + if first_name is not None: + # Reuse already-computed quantization/passthrough for this storage + if first_name in quantized: + quantized[name] = quantized[first_name] + scales[name] = scales[first_name] + dtypes[name] = dtypes[first_name] + if first_name in qmeta: + qmeta[name] = qmeta[first_name] + elif first_name in passthrough: + passthrough[name] = passthrough[first_name] + else: + passthrough[name] = t + # Do not double-count payload bytes; they were counted at first occurrence. + continue + + seen_ptr[ptr] = name + + 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) + # 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 + +def print_model_param_summary(model: nn.Module, logger: callable = print, top_n: int | None = None, show_percent: bool = True) -> None: + from collections import defaultdict + + total = sum(int(p.numel()) for p in model.parameters()) + trainable = sum(int(p.numel()) for p in model.parameters() if p.requires_grad) + dtype_bytes = sum(int(p.element_size()) * int(p.numel()) for p in model.parameters()) + + per_module: dict[str, int] = defaultdict(int) + for name, p in model.named_parameters(): + module_name = ".".join(name.split(".")[:-1]) or "" + per_module[module_name] += int(p.numel()) + + items = sorted(per_module.items(), key=lambda x: x[1], reverse=True) + logger(f"Total params: {total:,} Trainable: {trainable:,} Memory: {dtype_bytes/1e6:.2f} MB") + + display = items if top_n is None else items[:top_n] + for mod, cnt in display: + pct = (cnt / total * 100.0) if total else 0.0 + if show_percent: + logger(f"{mod:50s} {cnt:12,d} {pct:6.2f}%") + else: + logger(f"{mod:50s} {cnt:12,d}") + +# ----------------------------- +# 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 +# ----------------------------- + +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() + + +def unique_named_parameters(module: nn.Module): + """Yield (name, param) for the module, skipping duplicate Parameter identities. + + Useful when modules share Parameters (tied weights) so consumers can operate + on a canonical set of parameters. + """ + seen_ids = set() + for name, p in module.named_parameters(): + if id(p) in seen_ids: + continue + seen_ids.add(id(p)) + yield name, p + + +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, + tie_even_blocks: bool = True, + ): + 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): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # SwiGLU: hidden = int(2/3 * mlp_mult * dim) for parameter match, but here keep mlp_mult * dim for simplicity + hidden = mlp_mult * dim + # Project once to 2*hidden, then split for SwiGLU + self.fc = ParametrizedRandomProjection( + in_features=dim, + out_features=2 * hidden, + projection_type="random", + ) + self.final = ParametrizedRandomProjection( + in_features=hidden, + out_features=dim, + projection_type="random", + ) + self.final._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + # SwiGLU: (x @ W1) * SiLU(x @ W2) @ W3, but project once to 2*hidden then split + x_proj = self.fc(x) + hidden = x_proj.shape[-1] // 2 + x1, x2 = x_proj[..., :hidden], x_proj[..., hidden:] + gated = x1 * torch.nn.functional.silu(x2) + return self.final(gated) + + +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, + tie_even_blocks: bool = True, + ): + 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() + if tie_even_blocks: + # New sharing logic: + # - Keep layer 1 and the final layer unique (not shared). + # - For the body (layers 2..num_layers-1) create shared Blocks per pair: + # (2,3), (4,5), (6,7), ... Each pair reuses the same Block instance. + # If there is an unpaired body layer (odd count), it will still be + # handled by the pairing math and remain unique if it's the final layer. + shared_attn_map: dict[int, nn.Module] = {} + for idx in range(1, num_layers + 1): + if idx == 1 or idx == num_layers: + # first and last layers are unique + self.blocks.append( + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + ) + else: + # compute pair start index for this body layer (1-based) + pair_start = 2 + ((idx - 2) // 2) * 2 + if pair_start not in shared_attn_map: + # Create a shared attention module for this pair + shared_attn_map[pair_start] = CausalSelfAttention( + model_dim, + num_heads, + num_kv_heads, + rope_base, + qk_gain_init, + ) + # Each block gets its own MLP, but shares attention + block = Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + block.attn = shared_attn_map[pair_start] + self.blocks.append(block) + else: + 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) + # Optionally reduce validation set size to speed up experiments. + val_frac = float(os.environ.get("VAL_FRAC", getattr(args, "val_frac", 1.0))) + if not (0.0 < val_frac <= 1.0): + raise ValueError(f"VAL_FRAC must be in (0, 1], got {val_frac}") + if val_frac < 1.0: + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + keep = int(math.ceil(total_seqs * val_frac)) * args.train_seq_len + 1 + val_tokens = val_tokens[:keep] + 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() + 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 + + # Print per-layer parameter summary to the training log + try: + print_model_param_summary(base_model, logger=log0) + except Exception: + # Best-effort summary; don't fail training if logging fails + print_model_param_summary(base_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 + # Use unique_named_parameters to avoid duplicate Parameter identities when blocks are shared + block_named_params = list(unique_named_parameters(base_model.blocks)) + PRP_VECTOR_PATTERNS = (".alpha", ".weight", ".bias") + def _is_prp_vector(name: str, p) -> bool: + return p.ndim == 1 and ("mlp.fc." in name or "mlp.final." in name) and any(name.endswith(pat) for pat in PRP_VECTOR_PATTERNS) + 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) + ] + prp_vector_params = [ + p + for name, p in block_named_params + if _is_prp_vector(name, p) + ] + prp_vector_ids = {id(p) for p in prp_vector_params} + 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)) + and id(p) not in prp_vector_ids + ] + 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, + ) + optimizer_prp = torch.optim.Adam( + [{"params": prp_vector_params, "lr": args.prp_lr, "base_lr": args.prp_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar, optimizer_prp] + 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} prp_lr:{args.prp_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: + + # Phase 1: cosine warmup from lr_init_scale -> 1.0 over lr_warmup_steps + if args.lr_warmup_steps > 0 and step < args.lr_warmup_steps: + t = step + cos_frac = 0.5 * (1.0 - math.cos(math.pi * t / max(args.lr_warmup_steps, 1))) + return args.lr_init_scale + (1.0 - args.lr_init_scale) * cos_frac + + # Phase 2: flash drop (cosine) from 1.0 -> lr_main_scale over lr_drop_steps + if args.lr_drop_steps > 0 and step < args.lr_warmup_steps + args.lr_drop_steps: + t_drop = step - args.lr_warmup_steps + # cosine from 1 -> 0 + cos = 0.5 * (1.0 + math.cos(math.pi * t_drop / max(args.lr_drop_steps, 1))) + return args.lr_main_scale + (1.0 - args.lr_main_scale) * cos + + # Phase 3: main powered-cosine anneal from lr_main_scale -> lr_min_scale with gamma curvature + main_total = args.iterations - args.lr_warmup_steps - args.lr_drop_steps + if main_total <= 0: + return args.lr_min_scale + t_main = step - args.lr_warmup_steps - args.lr_drop_steps + t_main = max(0, min(t_main, main_total)) + cosine_decay = 0.5 * (1.0 + math.cos(math.pi * t_main / max(main_total, 1))) + # Apply gamma curvature to stretch/shape the tail + return args.lr_min_scale + (args.lr_main_scale - args.lr_min_scale) * (cosine_decay ** args.lr_gamma) + + # 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 <= 100 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 3.11.2 (main, Apr 28 2025, 14:11:48) [GCC 12.2.0] +Running PyTorch 2.7.1+cu126 +Wed Apr 1 15:00:32 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 595.54 Driver Version: 595.79 CUDA Version: 13.2 | ++-----------------------------------------+------------------------+----------------------+ +| 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 GeForce RTX 3080 On | 00000000:01:00.0 On | N/A | +| 0% 55C P0 114W / 345W | 1123MiB / 10240MiB | 0% Default | +| | | N/A | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 485 G /Xwayland N/A | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_8192_bpe.model +train_loader:dataset:fineweb10B_sp8192 train_shards:5 +val_loader:shards pattern=./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin tokens:10137600 +Total params: 17,171,040 Trainable: 17,171,040 Memory: 50.66 MB +tok_emb 4,194,304 24.43% +blocks.0.attn.c_q 262,144 1.53% +blocks.0.attn.proj 262,144 1.53% +blocks.1.attn.c_q 262,144 1.53% +blocks.1.attn.proj 262,144 1.53% +blocks.3.attn.c_q 262,144 1.53% +blocks.3.attn.proj 262,144 1.53% +blocks.5.attn.c_q 262,144 1.53% +blocks.5.attn.proj 262,144 1.53% +blocks.7.attn.c_q 262,144 1.53% +blocks.7.attn.proj 262,144 1.53% +blocks.9.attn.c_q 262,144 1.53% +blocks.9.attn.proj 262,144 1.53% +blocks.11.attn.c_q 262,144 1.53% +blocks.11.attn.proj 262,144 1.53% +blocks.13.attn.c_q 262,144 1.53% +blocks.13.attn.proj 262,144 1.53% +blocks.15.attn.c_q 262,144 1.53% +blocks.15.attn.proj 262,144 1.53% +blocks.17.attn.c_q 262,144 1.53% +blocks.17.attn.proj 262,144 1.53% +blocks.19.attn.c_q 262,144 1.53% +blocks.19.attn.proj 262,144 1.53% +blocks.20.attn.c_q 262,144 1.53% +blocks.20.attn.proj 262,144 1.53% +blocks.0.mlp.fc 156,160 0.91% +blocks.1.mlp.fc 156,160 0.91% +blocks.2.mlp.fc 156,160 0.91% +blocks.3.mlp.fc 156,160 0.91% +blocks.4.mlp.fc 156,160 0.91% +blocks.5.mlp.fc 156,160 0.91% +blocks.6.mlp.fc 156,160 0.91% +blocks.7.mlp.fc 156,160 0.91% +blocks.8.mlp.fc 156,160 0.91% +blocks.9.mlp.fc 156,160 0.91% +blocks.10.mlp.fc 156,160 0.91% +blocks.11.mlp.fc 156,160 0.91% +blocks.12.mlp.fc 156,160 0.91% +blocks.13.mlp.fc 156,160 0.91% +blocks.14.mlp.fc 156,160 0.91% +blocks.15.mlp.fc 156,160 0.91% +blocks.16.mlp.fc 156,160 0.91% +blocks.17.mlp.fc 156,160 0.91% +blocks.18.mlp.fc 156,160 0.91% +blocks.19.mlp.fc 156,160 0.91% +blocks.20.mlp.fc 156,160 0.91% +blocks.0.mlp.final 84,992 0.49% +blocks.1.mlp.final 84,992 0.49% +blocks.2.mlp.final 84,992 0.49% +blocks.3.mlp.final 84,992 0.49% +blocks.4.mlp.final 84,992 0.49% +blocks.5.mlp.final 84,992 0.49% +blocks.6.mlp.final 84,992 0.49% +blocks.7.mlp.final 84,992 0.49% +blocks.8.mlp.final 84,992 0.49% +blocks.9.mlp.final 84,992 0.49% +blocks.10.mlp.final 84,992 0.49% +blocks.11.mlp.final 84,992 0.49% +blocks.12.mlp.final 84,992 0.49% +blocks.13.mlp.final 84,992 0.49% +blocks.14.mlp.final 84,992 0.49% +blocks.15.mlp.final 84,992 0.49% +blocks.16.mlp.final 84,992 0.49% +blocks.17.mlp.final 84,992 0.49% +blocks.18.mlp.final 84,992 0.49% +blocks.19.mlp.final 84,992 0.49% +blocks.20.mlp.final 84,992 0.49% +blocks.0.attn.c_k 65,536 0.38% +blocks.0.attn.c_v 65,536 0.38% +blocks.1.attn.c_k 65,536 0.38% +blocks.1.attn.c_v 65,536 0.38% +blocks.3.attn.c_k 65,536 0.38% +blocks.3.attn.c_v 65,536 0.38% +blocks.5.attn.c_k 65,536 0.38% +blocks.5.attn.c_v 65,536 0.38% +blocks.7.attn.c_k 65,536 0.38% +blocks.7.attn.c_v 65,536 0.38% +blocks.9.attn.c_k 65,536 0.38% +blocks.9.attn.c_v 65,536 0.38% +blocks.11.attn.c_k 65,536 0.38% +blocks.11.attn.c_v 65,536 0.38% +blocks.13.attn.c_k 65,536 0.38% +blocks.13.attn.c_v 65,536 0.38% +blocks.15.attn.c_k 65,536 0.38% +blocks.15.attn.c_v 65,536 0.38% +blocks.17.attn.c_k 65,536 0.38% +blocks.17.attn.c_v 65,536 0.38% +blocks.19.attn.c_k 65,536 0.38% +blocks.19.attn.c_v 65,536 0.38% +blocks.20.attn.c_k 65,536 0.38% +blocks.20.attn.c_v 65,536 0.38% + 5,120 0.03% +blocks.0 2,048 0.01% +blocks.1 2,048 0.01% +blocks.2 2,048 0.01% +blocks.3 2,048 0.01% +blocks.4 2,048 0.01% +blocks.5 2,048 0.01% +blocks.6 2,048 0.01% +blocks.7 2,048 0.01% +blocks.8 2,048 0.01% +blocks.9 2,048 0.01% +blocks.10 2,048 0.01% +blocks.11 2,048 0.01% +blocks.12 2,048 0.01% +blocks.13 2,048 0.01% +blocks.14 2,048 0.01% +blocks.15 2,048 0.01% +blocks.16 2,048 0.01% +blocks.17 2,048 0.01% +blocks.18 2,048 0.01% +blocks.19 2,048 0.01% +blocks.20 2,048 0.01% +blocks.0.attn 8 0.00% +blocks.1.attn 8 0.00% +blocks.3.attn 8 0.00% +blocks.5.attn 8 0.00% +blocks.7.attn 8 0.00% +blocks.9.attn 8 0.00% +blocks.11.attn 8 0.00% +blocks.13.attn 8 0.00% +blocks.15.attn 8 0.00% +blocks.17.attn 8 0.00% +blocks.19.attn 8 0.00% +blocks.20.attn 8 0.00% +model_params:17171040 +world_size:1 grad_accum_steps:8 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:2 +tie_embeddings:True embed_lr:0.03 head_lr:0.0 matrix_lr:0.03 scalar_lr:0.03 prp_lr:0.06 +train_batch_tokens:65536 train_seq_len:2048 iterations:4000 warmup_steps:2 max_wallclock_seconds:7200.000 +seed:1337 +warmup_step:1/2 +warmup_step:2/2 +step:0/4000 val_loss:9.0102 val_bpb:3.5169 train_time:0ms step_avg:0.01ms +step:1/4000 train_loss:9.0107 train_time:1466ms step_avg:1466.26ms +step:2/4000 train_loss:9.0100 train_time:2908ms step_avg:1453.99ms +step:3/4000 train_loss:9.0048 train_time:4397ms step_avg:1465.59ms +step:4/4000 train_loss:8.9833 train_time:5883ms step_avg:1470.83ms +step:5/4000 train_loss:8.9408 train_time:7344ms step_avg:1468.77ms +step:6/4000 train_loss:8.8739 train_time:8864ms step_avg:1477.28ms +step:7/4000 train_loss:8.7573 train_time:10274ms step_avg:1467.77ms +step:8/4000 train_loss:8.6094 train_time:11700ms step_avg:1462.56ms +step:9/4000 train_loss:8.3667 train_time:13116ms step_avg:1457.36ms +step:10/4000 train_loss:8.0158 train_time:14567ms step_avg:1456.74ms +step:11/4000 train_loss:7.5980 train_time:16032ms step_avg:1457.49ms +step:12/4000 train_loss:7.1966 train_time:17424ms step_avg:1451.98ms +step:13/4000 train_loss:6.9445 train_time:18820ms step_avg:1447.73ms +step:14/4000 train_loss:6.8814 train_time:20204ms step_avg:1443.11ms +step:15/4000 train_loss:6.8540 train_time:21586ms step_avg:1439.07ms +step:16/4000 train_loss:6.7408 train_time:22995ms step_avg:1437.17ms +step:17/4000 train_loss:6.7347 train_time:24516ms step_avg:1442.14ms +step:18/4000 train_loss:6.5704 train_time:26032ms step_avg:1446.21ms +step:19/4000 train_loss:6.4021 train_time:27554ms step_avg:1450.22ms +step:20/4000 train_loss:6.4025 train_time:29110ms step_avg:1455.49ms +step:21/4000 train_loss:6.0681 train_time:30646ms step_avg:1459.36ms +step:22/4000 train_loss:6.1421 train_time:32183ms step_avg:1462.88ms +step:23/4000 train_loss:6.2051 train_time:33724ms step_avg:1466.26ms +step:24/4000 train_loss:6.1865 train_time:35261ms step_avg:1469.20ms +step:25/4000 train_loss:6.1297 train_time:36829ms step_avg:1473.16ms +step:26/4000 train_loss:6.0523 train_time:38339ms step_avg:1474.58ms +step:27/4000 train_loss:6.1787 train_time:39883ms step_avg:1477.14ms +step:28/4000 train_loss:6.0469 train_time:41436ms step_avg:1479.84ms +step:29/4000 train_loss:5.8544 train_time:42973ms step_avg:1481.82ms +step:30/4000 train_loss:5.9607 train_time:44565ms step_avg:1485.49ms +step:31/4000 train_loss:6.0748 train_time:46140ms step_avg:1488.39ms +step:32/4000 train_loss:6.0144 train_time:47678ms step_avg:1489.94ms +step:33/4000 train_loss:6.0363 train_time:49190ms step_avg:1490.61ms +step:34/4000 train_loss:6.0068 train_time:50703ms step_avg:1491.27ms +step:35/4000 train_loss:6.0546 train_time:52215ms step_avg:1491.87ms +step:36/4000 train_loss:5.8725 train_time:53726ms step_avg:1492.38ms +step:37/4000 train_loss:5.9328 train_time:55250ms step_avg:1493.23ms +step:38/4000 train_loss:5.9411 train_time:56763ms step_avg:1493.76ms +step:39/4000 train_loss:5.8241 train_time:58272ms step_avg:1494.16ms +step:40/4000 train_loss:5.8473 train_time:59783ms step_avg:1494.57ms +step:41/4000 train_loss:5.8413 train_time:61296ms step_avg:1495.03ms +step:42/4000 train_loss:5.8743 train_time:62829ms step_avg:1495.92ms +step:43/4000 train_loss:5.8594 train_time:64403ms step_avg:1497.75ms +step:44/4000 train_loss:5.7509 train_time:65953ms step_avg:1498.93ms +step:45/4000 train_loss:5.7855 train_time:67520ms step_avg:1500.44ms +step:46/4000 train_loss:5.7237 train_time:69091ms step_avg:1501.98ms +step:47/4000 train_loss:5.7650 train_time:70679ms step_avg:1503.81ms +step:48/4000 train_loss:5.7486 train_time:72271ms step_avg:1505.65ms +step:49/4000 train_loss:5.5558 train_time:73871ms step_avg:1507.56ms +step:50/4000 train_loss:5.6160 train_time:75488ms step_avg:1509.77ms +step:51/4000 train_loss:5.7767 train_time:77171ms step_avg:1513.15ms +step:52/4000 train_loss:5.7962 train_time:78763ms step_avg:1514.68ms +step:53/4000 train_loss:5.6287 train_time:80342ms step_avg:1515.89ms +step:54/4000 train_loss:5.6450 train_time:81895ms step_avg:1516.57ms +step:55/4000 train_loss:5.7057 train_time:83473ms step_avg:1517.69ms +step:56/4000 train_loss:5.6428 train_time:85032ms step_avg:1518.43ms +step:57/4000 train_loss:5.6108 train_time:86578ms step_avg:1518.92ms +step:58/4000 train_loss:5.6387 train_time:88109ms step_avg:1519.12ms +step:59/4000 train_loss:5.5729 train_time:89683ms step_avg:1520.06ms +step:60/4000 train_loss:5.6263 train_time:91195ms step_avg:1519.92ms +step:61/4000 train_loss:5.6270 train_time:92661ms step_avg:1519.03ms +step:62/4000 train_loss:5.6577 train_time:94156ms step_avg:1518.65ms +step:63/4000 train_loss:5.5811 train_time:95657ms step_avg:1518.37ms +step:64/4000 train_loss:5.5810 train_time:97171ms step_avg:1518.30ms +step:65/4000 train_loss:5.5652 train_time:98735ms step_avg:1519.00ms +step:66/4000 train_loss:5.5476 train_time:100257ms step_avg:1519.05ms +step:67/4000 train_loss:5.4965 train_time:101764ms step_avg:1518.86ms +step:68/4000 train_loss:5.6307 train_time:103263ms step_avg:1518.57ms +step:69/4000 train_loss:5.4078 train_time:104741ms step_avg:1517.99ms +step:70/4000 train_loss:5.6589 train_time:106239ms step_avg:1517.71ms +step:71/4000 train_loss:5.4794 train_time:107704ms step_avg:1516.96ms +step:72/4000 train_loss:5.4162 train_time:109152ms step_avg:1516.00ms +step:73/4000 train_loss:5.5552 train_time:110588ms step_avg:1514.90ms +step:74/4000 train_loss:5.4504 train_time:112018ms step_avg:1513.76ms +step:75/4000 train_loss:5.4800 train_time:113450ms step_avg:1512.67ms +step:76/4000 train_loss:5.4039 train_time:114881ms step_avg:1511.60ms +step:77/4000 train_loss:5.4704 train_time:116313ms step_avg:1510.55ms +step:78/4000 train_loss:5.4211 train_time:117744ms step_avg:1509.54ms +step:79/4000 train_loss:5.3627 train_time:119201ms step_avg:1508.87ms +step:80/4000 train_loss:5.3976 train_time:120652ms step_avg:1508.15ms +step:81/4000 train_loss:5.3854 train_time:122088ms step_avg:1507.26ms +step:82/4000 train_loss:5.4266 train_time:123519ms step_avg:1506.33ms +step:83/4000 train_loss:5.3513 train_time:124951ms step_avg:1505.43ms +step:84/4000 train_loss:5.3415 train_time:126383ms step_avg:1504.56ms +step:85/4000 train_loss:5.2543 train_time:127816ms step_avg:1503.72ms +step:86/4000 train_loss:5.3205 train_time:129250ms step_avg:1502.91ms +step:87/4000 train_loss:5.2840 train_time:130683ms step_avg:1502.10ms +step:88/4000 train_loss:5.3121 train_time:132124ms step_avg:1501.41ms +step:89/4000 train_loss:5.3281 train_time:133592ms step_avg:1501.03ms +step:90/4000 train_loss:5.2674 train_time:135027ms step_avg:1500.30ms +step:91/4000 train_loss:5.3901 train_time:136488ms step_avg:1499.87ms +step:92/4000 train_loss:5.2806 train_time:137941ms step_avg:1499.35ms +step:93/4000 train_loss:5.2922 train_time:139423ms step_avg:1499.17ms +step:94/4000 train_loss:5.3340 train_time:140861ms step_avg:1498.52ms +step:95/4000 train_loss:5.2743 train_time:142295ms step_avg:1497.85ms +step:96/4000 train_loss:5.3632 train_time:143724ms step_avg:1497.12ms +step:97/4000 train_loss:5.3033 train_time:145159ms step_avg:1496.49ms +step:98/4000 train_loss:5.2685 train_time:146603ms step_avg:1495.95ms +step:99/4000 train_loss:5.1979 train_time:148106ms step_avg:1496.02ms +step:100/4000 train_loss:5.2947 train_time:149625ms step_avg:1496.25ms +step:200/4000 train_loss:4.9269 train_time:291547ms step_avg:1457.74ms +step:300/4000 train_loss:4.4978 train_time:439279ms step_avg:1464.26ms +step:400/4000 train_loss:4.3156 train_time:584198ms step_avg:1460.50ms +step:400/4000 val_loss:4.2170 val_bpb:1.6460 train_time:584199ms step_avg:1460.50ms +step:500/4000 train_loss:3.9304 train_time:733498ms step_avg:1467.00ms +step:600/4000 train_loss:3.9694 train_time:879590ms step_avg:1465.98ms +step:700/4000 train_loss:3.8861 train_time:1018792ms step_avg:1455.42ms +step:800/4000 train_loss:4.0357 train_time:1157404ms step_avg:1446.75ms +step:800/4000 val_loss:3.9027 val_bpb:1.5233 train_time:1157404ms step_avg:1446.76ms +step:900/4000 train_loss:3.8725 train_time:1296749ms step_avg:1440.83ms +step:1000/4000 train_loss:3.9312 train_time:1437619ms step_avg:1437.62ms +step:1100/4000 train_loss:3.7792 train_time:1580835ms step_avg:1437.12ms +step:1200/4000 train_loss:3.8756 train_time:1723458ms step_avg:1436.21ms +step:1200/4000 val_loss:3.7963 val_bpb:1.4818 train_time:1723458ms step_avg:1436.22ms +step:1300/4000 train_loss:3.7214 train_time:1862973ms step_avg:1433.06ms +step:1400/4000 train_loss:3.8289 train_time:2002155ms step_avg:1430.11ms +step:1500/4000 train_loss:3.7423 train_time:2141555ms step_avg:1427.70ms +step:1600/4000 train_loss:3.7806 train_time:2281502ms step_avg:1425.94ms +step:1600/4000 val_loss:3.7088 val_bpb:1.4476 train_time:2281502ms step_avg:1425.94ms +step:1700/4000 train_loss:3.6855 train_time:2419864ms step_avg:1423.45ms +step:1800/4000 train_loss:3.5586 train_time:2558430ms step_avg:1421.35ms +step:1900/4000 train_loss:3.5357 train_time:2697557ms step_avg:1419.77ms +step:2000/4000 train_loss:3.6078 train_time:2836787ms step_avg:1418.39ms +step:2000/4000 val_loss:3.6501 val_bpb:1.4247 train_time:2836787ms step_avg:1418.39ms +step:2100/4000 train_loss:3.6979 train_time:2975888ms step_avg:1417.09ms +step:2200/4000 train_loss:3.6231 train_time:3116042ms step_avg:1416.38ms +step:2300/4000 train_loss:3.6264 train_time:3254754ms step_avg:1415.11ms +step:2400/4000 train_loss:3.6386 train_time:3393411ms step_avg:1413.92ms +step:2400/4000 val_loss:3.6022 val_bpb:1.4060 train_time:3393412ms step_avg:1413.92ms +step:2500/4000 train_loss:3.5390 train_time:3532154ms step_avg:1412.86ms +step:2600/4000 train_loss:3.7658 train_time:3670836ms step_avg:1411.86ms +step:2700/4000 train_loss:3.6511 train_time:3808804ms step_avg:1410.67ms +step:2800/4000 train_loss:3.6159 train_time:3946749ms step_avg:1409.55ms +step:2800/4000 val_loss:3.5515 val_bpb:1.3862 train_time:3946750ms step_avg:1409.55ms +step:2900/4000 train_loss:3.5555 train_time:4084771ms step_avg:1408.54ms +step:3000/4000 train_loss:3.5243 train_time:4222931ms step_avg:1407.64ms +step:3100/4000 train_loss:3.2796 train_time:4362351ms step_avg:1407.21ms +step:3200/4000 train_loss:3.5057 train_time:4501608ms step_avg:1406.75ms +step:3200/4000 val_loss:3.5092 val_bpb:1.3697 train_time:4501609ms step_avg:1406.75ms +step:3300/4000 train_loss:3.5309 train_time:4641024ms step_avg:1406.37ms +step:3400/4000 train_loss:3.5036 train_time:4779888ms step_avg:1405.85ms +step:3500/4000 train_loss:3.3632 train_time:4920019ms step_avg:1405.72ms +step:3600/4000 train_loss:3.4954 train_time:5067954ms step_avg:1407.77ms +step:3600/4000 val_loss:3.4804 val_bpb:1.3585 train_time:5067955ms step_avg:1407.77ms +step:3700/4000 train_loss:3.2293 train_time:5212689ms step_avg:1408.83ms +step:3800/4000 train_loss:3.3355 train_time:5352113ms step_avg:1408.45ms +step:3900/4000 train_loss:3.5224 train_time:5493014ms step_avg:1408.47ms +step:4000/4000 train_loss:3.5189 train_time:5632309ms step_avg:1408.08ms +step:4000/4000 val_loss:3.4657 val_bpb:1.3527 train_time:5632309ms step_avg:1408.08ms +peak memory allocated: 7329 MiB reserved: 8056 MiB +Serialized model: 50783797 bytes +Code size: 55494 bytes +Total submission size: 50839291 bytes +Serialized model int8+zlib: 17603010 bytes (payload:21242560 raw_torch:21364933 payload_ratio:3.50x) +Total submission size int8+zlib: 17658504 bytes +final_int8_zlib_roundtrip val_loss:3.4690 val_bpb:1.3540 eval_time:64514ms +final_int8_zlib_roundtrip_exact val_loss:3.46896188 val_bpb:1.35401763 diff --git a/records/track_10min_16mb/2026-04-01_21L_PRP/train_gpt.py b/records/track_10min_16mb/2026-04-01_21L_PRP/train_gpt.py new file mode 100644 index 0000000000..cb832ea870 --- /dev/null +++ b/records/track_10min_16mb/2026-04-01_21L_PRP/train_gpt.py @@ -0,0 +1,1292 @@ +""" +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 +from prp import ParametrizedRandomProjection + + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp8192") + 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_8192_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)) + # Fraction of the validation split to use (0 < VAL_FRAC <= 1.0) + val_frac = float(os.environ.get("VAL_FRAC", 1.0)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 10_000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 5)) + # LR scheduling knobs (training warmup, flash drop, main phase, and gamma curvature) + lr_warmup_steps = int(os.environ.get("LR_WARMUP_STEPS", iterations * 0.01 )) + lr_drop_steps = int(os.environ.get("LR_DROP_STEPS", iterations * 0.25)) + # Fractional scales relative to each param-group's base_lr + lr_main_scale = float(os.environ.get("LR_MAIN_SCALE", 0.5)) + lr_min_scale = float(os.environ.get("LR_MIN_SCALE", 0.05)) + lr_init_scale = float(os.environ.get("LR_INIT_SCALE", 0.0)) + lr_gamma = float(os.environ.get("LR_GAMMA", 0.8)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 4800)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 8192)) + num_layers = int(os.environ.get("NUM_LAYERS", 21)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 2)) + 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", 4)) + 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", 0.4)) # not used since tied + head_lr = float(os.environ.get("HEAD_LR", 0.02)) # not used since tied + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.004)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.03)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.03)) + prp_lr = float(os.environ.get("PRP_LR", 0.06)) + 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", lr_warmup_steps + lr_drop_steps)) + 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, + ) + + # Track already-seen storage pointers to handle shared/tied parameters + seen_ptr: dict[int, str] = {} + 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) + + ptr = tensor.data_ptr() + first_name = seen_ptr.get(ptr) + if first_name is not None: + # Reuse already-computed quantization/passthrough for this storage + if first_name in quantized: + quantized[name] = quantized[first_name] + scales[name] = scales[first_name] + dtypes[name] = dtypes[first_name] + if first_name in qmeta: + qmeta[name] = qmeta[first_name] + elif first_name in passthrough: + passthrough[name] = passthrough[first_name] + else: + passthrough[name] = t + # Do not double-count payload bytes; they were counted at first occurrence. + continue + + seen_ptr[ptr] = name + + 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) + # 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 + +def print_model_param_summary(model: nn.Module, logger: callable = print, top_n: int | None = None, show_percent: bool = True) -> None: + from collections import defaultdict + + total = sum(int(p.numel()) for p in model.parameters()) + trainable = sum(int(p.numel()) for p in model.parameters() if p.requires_grad) + dtype_bytes = sum(int(p.element_size()) * int(p.numel()) for p in model.parameters()) + + per_module: dict[str, int] = defaultdict(int) + for name, p in model.named_parameters(): + module_name = ".".join(name.split(".")[:-1]) or "" + per_module[module_name] += int(p.numel()) + + items = sorted(per_module.items(), key=lambda x: x[1], reverse=True) + logger(f"Total params: {total:,} Trainable: {trainable:,} Memory: {dtype_bytes/1e6:.2f} MB") + + display = items if top_n is None else items[:top_n] + for mod, cnt in display: + pct = (cnt / total * 100.0) if total else 0.0 + if show_percent: + logger(f"{mod:50s} {cnt:12,d} {pct:6.2f}%") + else: + logger(f"{mod:50s} {cnt:12,d}") + +# ----------------------------- +# 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 +# ----------------------------- + +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() + + +def unique_named_parameters(module: nn.Module): + """Yield (name, param) for the module, skipping duplicate Parameter identities. + + Useful when modules share Parameters (tied weights) so consumers can operate + on a canonical set of parameters. + """ + seen_ids = set() + for name, p in module.named_parameters(): + if id(p) in seen_ids: + continue + seen_ids.add(id(p)) + yield name, p + + +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, + tie_even_blocks: bool = True, + ): + 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): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # SwiGLU: hidden = int(2/3 * mlp_mult * dim) for parameter match, but here keep mlp_mult * dim for simplicity + hidden = mlp_mult * dim + # Project once to 2*hidden, then split for SwiGLU + self.fc = ParametrizedRandomProjection( + in_features=dim, + out_features=2 * hidden, + projection_type="random", + ) + self.final = ParametrizedRandomProjection( + in_features=hidden, + out_features=dim, + projection_type="random", + ) + self.final._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + # SwiGLU: (x @ W1) * SiLU(x @ W2) @ W3, but project once to 2*hidden then split + x_proj = self.fc(x) + hidden = x_proj.shape[-1] // 2 + x1, x2 = x_proj[..., :hidden], x_proj[..., hidden:] + gated = x1 * torch.nn.functional.silu(x2) + return self.final(gated) + + +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, + tie_even_blocks: bool = True, + ): + 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() + if tie_even_blocks: + # New sharing logic: + # - Keep layer 1 and the final layer unique (not shared). + # - For the body (layers 2..num_layers-1) create shared Blocks per pair: + # (2,3), (4,5), (6,7), ... Each pair reuses the same Block instance. + # If there is an unpaired body layer (odd count), it will still be + # handled by the pairing math and remain unique if it's the final layer. + shared_attn_map: dict[int, nn.Module] = {} + for idx in range(1, num_layers + 1): + if idx == 1 or idx == num_layers: + # first and last layers are unique + self.blocks.append( + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + ) + else: + # compute pair start index for this body layer (1-based) + pair_start = 2 + ((idx - 2) // 2) * 2 + if pair_start not in shared_attn_map: + # Create a shared attention module for this pair + shared_attn_map[pair_start] = CausalSelfAttention( + model_dim, + num_heads, + num_kv_heads, + rope_base, + qk_gain_init, + ) + # Each block gets its own MLP, but shares attention + block = Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + block.attn = shared_attn_map[pair_start] + self.blocks.append(block) + else: + 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) + # Optionally reduce validation set size to speed up experiments. + val_frac = float(os.environ.get("VAL_FRAC", getattr(args, "val_frac", 1.0))) + if not (0.0 < val_frac <= 1.0): + raise ValueError(f"VAL_FRAC must be in (0, 1], got {val_frac}") + if val_frac < 1.0: + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + keep = int(math.ceil(total_seqs * val_frac)) * args.train_seq_len + 1 + val_tokens = val_tokens[:keep] + 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() + 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 + + # Print per-layer parameter summary to the training log + try: + print_model_param_summary(base_model, logger=log0) + except Exception: + # Best-effort summary; don't fail training if logging fails + print_model_param_summary(base_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 + # Use unique_named_parameters to avoid duplicate Parameter identities when blocks are shared + block_named_params = list(unique_named_parameters(base_model.blocks)) + PRP_VECTOR_PATTERNS = (".alpha", ".weight", ".bias") + def _is_prp_vector(name: str, p) -> bool: + return p.ndim == 1 and ("mlp.fc." in name or "mlp.final." in name) and any(name.endswith(pat) for pat in PRP_VECTOR_PATTERNS) + 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) + ] + prp_vector_params = [ + p + for name, p in block_named_params + if _is_prp_vector(name, p) + ] + prp_vector_ids = {id(p) for p in prp_vector_params} + 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)) + and id(p) not in prp_vector_ids + ] + 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, + ) + optimizer_prp = torch.optim.Adam( + [{"params": prp_vector_params, "lr": args.prp_lr, "base_lr": args.prp_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar, optimizer_prp] + 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} prp_lr:{args.prp_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: + + # Phase 1: cosine warmup from lr_init_scale -> 1.0 over lr_warmup_steps + if args.lr_warmup_steps > 0 and step < args.lr_warmup_steps: + t = step + cos_frac = 0.5 * (1.0 - math.cos(math.pi * t / max(args.lr_warmup_steps, 1))) + return args.lr_init_scale + (1.0 - args.lr_init_scale) * cos_frac + + # Phase 2: flash drop (cosine) from 1.0 -> lr_main_scale over lr_drop_steps + if args.lr_drop_steps > 0 and step < args.lr_warmup_steps + args.lr_drop_steps: + t_drop = step - args.lr_warmup_steps + # cosine from 1 -> 0 + cos = 0.5 * (1.0 + math.cos(math.pi * t_drop / max(args.lr_drop_steps, 1))) + return args.lr_main_scale + (1.0 - args.lr_main_scale) * cos + + # Phase 3: main powered-cosine anneal from lr_main_scale -> lr_min_scale with gamma curvature + main_total = args.iterations - args.lr_warmup_steps - args.lr_drop_steps + if main_total <= 0: + return args.lr_min_scale + t_main = step - args.lr_warmup_steps - args.lr_drop_steps + t_main = max(0, min(t_main, main_total)) + cosine_decay = 0.5 * (1.0 + math.cos(math.pi * t_main / max(main_total, 1))) + # Apply gamma curvature to stretch/shape the tail + return args.lr_min_scale + (args.lr_main_scale - args.lr_min_scale) * (cosine_decay ** args.lr_gamma) + + # 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 <= 100 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()