From 1c4a491273b0256562f82b7bcd72615fac8cb723 Mon Sep 17 00:00:00 2001 From: Robby Sneiderman Date: Sun, 22 Mar 2026 10:59:38 -0500 Subject: [PATCH] Add EBLS Learned Sharing submission (10min/16MB track) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Empirical Bayes Layer Sharing: 3 shared blocks × 3 virtual layers with per-virtual-layer LoRA deviations gated by learned shrinkage gammas. Val BPB: 1.3441 (post-quant) / 1.2105 (pre-quant) Artifact: 16,224,826 bytes | 8×H100 SXM, 4572 steps, 10 min Co-Authored-By: Claude Opus 4.6 --- .../2026-03-22_EBLS_LearnedSharing/README.md | 33 + .../submission.json | 11 + .../2026-03-22_EBLS_LearnedSharing/train.log | 1544 +++++++++++++++++ .../train_gpt.py | 1393 +++++++++++++++ 4 files changed, 2981 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/README.md create mode 100644 records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/submission.json create mode 100644 records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train.log create mode 100644 records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train_gpt.py diff --git a/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/README.md b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/README.md new file mode 100644 index 000000000..2b18a9cc2 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/README.md @@ -0,0 +1,33 @@ +# EBLS Learned Sharing + +**Track**: 10 min / 16 MB +**Val BPB (post-quant)**: 1.3441 +**Val BPB (pre-quant)**: 1.2105 +**Artifact size**: 16,224,826 bytes +**Date**: 2026-03-22 + +## Approach + +Empirical Bayes Layer Sharing (EBLS): 3 shared transformer blocks, each applied 3× for 9 effective layers. Per-virtual-layer rank-8 LoRA deviations gated by learned shrinkage factors γ_i = σ(logit_i). Shrinkage regularization encourages weight sharing unless deviation helps. + +## Architecture + +- **Dimension**: 1024, **Heads**: 16Q / 4KV (GQA) +- **Layers**: 3 shared blocks × 3 = 9 virtual layers +- **LoRA rank**: 8 (attention + MLP) +- **MLP**: 3× expansion with ReLU² +- **Features**: SmearGate, BigramHash(10240), U-Net skips +- **Optimizer**: Muon (WD=0.04) + Adam (LoRA, embeddings, scalars) +- **Quantization**: Int6 STE QAT + zstd-22 + +## Key Finding + +The model discovers optimal sharing automatically: +- MLP gammas → 0.0000 across all virtual layers (fully shared) +- Attention gammas → 0.0035 for layer 0, ~0 otherwise (minimal specialization) + +## Reproduce + +```bash +bash eval/eval.sh +``` diff --git a/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/submission.json b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/submission.json new file mode 100644 index 000000000..17261a3b9 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/submission.json @@ -0,0 +1,11 @@ +{ + "author": "Robby Sneiderman", + "github_id": "Robby955", + "name": "EBLS Learned Sharing", + "blurb": "Empirical Bayes Layer Sharing: 3 shared blocks × 3 virtual layers with per-virtual-layer LoRA deviations gated by learned shrinkage gammas. Model discovers MLP weights should be fully shared (gamma→0) while attention needs minimal specialization. 1024-dim, int6+zstd-22, Muon+Adam.", + "date": "2026-03-22T00:00:00Z", + "val_loss": 2.2694, + "val_bpb": 1.3441, + "bytes_total": 16224826, + "bytes_code": 62684 +} diff --git a/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train.log b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train.log new file mode 100644 index 000000000..701f7ea8c --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train.log @@ -0,0 +1,1544 @@ +""" +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: `train_gpt.py` and `train_gpt_mlx.py` must never be 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 zstandard as zstd +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 1024)) + num_heads = int(os.environ.get("NUM_HEADS", 16)) + mlp_mult = int(os.environ.get("MLP_MULT", 3)) + 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)) + + # BigramHash + SmearGate parameters. + bigram_buckets = int(os.environ.get("BIGRAM_BUCKETS", 10240)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + # EBLS (Empirical Bayes Layer Sharing) parameters. + lora_rank = int(os.environ.get("LORA_RANK", 8)) + shrinkage_lambda = float(os.environ.get("SHRINKAGE_LAMBDA", 0.01)) + num_shared_blocks = int(os.environ.get("NUM_SHARED_BLOCKS", 3)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + muon_weight_decay = float(os.environ.get("MUON_WEIGHT_DECAY", 0.04)) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + +# ----------------------------- +# 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, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov, weight_decay=weight_decay), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group["weight_decay"] + + # Decoupled weight decay (applied before update) + if wd > 0: + for p in params: + p.data.mul_(1.0 - lr * wd) + + 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) + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int = 64, + batch_size: int = 64, +) -> tuple[float, float]: + """Sliding window eval: overlapping windows with stride, score only last `stride` tokens.""" + seq_len = args.train_seq_len + total = val_tokens.numel() - 1 + max_start = total - seq_len + all_starts = list(range(0, max_start + 1, stride)) + my_starts = all_starts[rank::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) + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_starts), batch_size): + batch_starts = my_starts[bi:bi + batch_size] + bsz = len(batch_starts) + x_batch = torch.stack([val_tokens[s:s + seq_len] for s in batch_starts]).to(device=device, dtype=torch.int64) + y_batch = torch.stack([val_tokens[s + 1:s + seq_len + 1] for s in batch_starts]).to(device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = base_model.get_logits(x_batch) # (bsz, seq_len, vocab) + score_logits = logits[:, -stride:, :].reshape(-1, logits.size(-1)) + score_targets = y_batch[:, -stride:].reshape(-1) + losses = F.cross_entropy(score_logits.float(), score_targets, reduction='none') + val_loss_sum += losses.to(torch.float64).sum() + val_token_count += float(score_targets.numel()) + prev_ids = x_batch[:, -stride:].reshape(-1) + tgt_ids = score_targets + 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() + base_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,attn_shrinkage_logits,mlp_shrinkage_logits,gate_logit", + ).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 +# Int6 quantization: [-31, 31] range packed into int8 storage +INT6_RANGE = 31 + +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]: + """Quantize to int6 range [-31, 31] stored in int8 containers.""" + t32 = t.float() + qr = INT6_RANGE + if t32.ndim == 2: + clip_abs = t32.abs().amax(dim=1) + scale = (clip_abs / qr).clamp_min(1.0 / qr) + q = torch.clamp(torch.round(t32 / scale[:, None]), -qr, qr).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(t32.abs().max().item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / qr if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(t32 / scale), -qr, qr).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +def fake_quantize_int6(w: Tensor) -> Tensor: + """Fake int6 quantization with straight-through estimator for QAT.""" + scale = w.float().abs().amax(dim=-1, keepdim=True).clamp(min=1e-8) / 31.0 + w_q = (w.float() / scale).round().clamp(-31, 31) * scale + return w + (w_q - w).detach() # STE: forward uses quantized, backward uses original + +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + # During training, applies fake int6 quantization (STE) to close the quantization gap. + def forward(self, x: Tensor) -> Tensor: + w = self.weight + if self.training: + w = fake_quantize_int6(w) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + if self.num_kv_heads != self.num_heads: + rep = self.num_heads // self.num_kv_heads + k = k[:, :, None, :, :].expand(bsz, self.num_kv_heads, rep, seqlen, self.head_dim).reshape(bsz, self.num_heads, seqlen, self.head_dim) + v = v[:, :, None, :, :].expand(bsz, self.num_kv_heads, rep, seqlen, self.head_dim).reshape(bsz, self.num_heads, seqlen, self.head_dim) + y = F.scaled_dot_product_attention(q, k, v, attn_mask=None, is_causal=True) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class SmearGate(nn.Module): + """Per-dimension gate blending current token with previous token embedding.""" + def __init__(self, dim: int, init_logit: float = 3.0): + super().__init__() + # sigmoid(3.0) ≈ 0.95 → mostly keep current token + self.gate_logit = nn.Parameter(torch.full((dim,), init_logit, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + gate = torch.sigmoid(self.gate_logit).to(x.dtype) + # Shift right: prev token embedding for position i is x at position i-1 + x_prev = F.pad(x[:, :-1, :], (0, 0, 1, 0)) # zero-pad first position + return gate * x + (1 - gate) * x_prev + + +class BigramHash(nn.Module): + """Hash-based bigram embedding: maps (prev_token, cur_token) pairs to learned vectors.""" + def __init__(self, num_buckets: int, embed_dim: int, model_dim: int): + super().__init__() + self.num_buckets = num_buckets + self.embed = nn.Embedding(num_buckets, embed_dim) + self.proj = CastedLinear(embed_dim, model_dim, bias=False) + nn.init.normal_(self.embed.weight, std=0.01) + nn.init.zeros_(self.proj.weight) + + def forward(self, input_ids: Tensor) -> Tensor: + # Hash bigrams: prev_id * large_prime + cur_id, mod num_buckets + prev_ids = F.pad(input_ids[:, :-1], (1, 0)) # zero for first position + bigram_hash = ((prev_ids.long() * 104729 + input_ids.long()) % self.num_buckets).long() + return self.proj(self.embed(bigram_hash)) + + +class EBLSBlock(nn.Module): + """Transformer block with Empirical Bayes Layer Sharing. + + Shared base attention + MLP weights are reused across virtual layers. + Per-virtual-layer LoRA deviations provide specialization, gated by + learned shrinkage factors gamma_i = sigmoid(logit_i). + """ + + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + lora_rank: int, + num_virtual_layers: int, + ): + 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.num_virtual_layers = num_virtual_layers + self.lora_rank = lora_rank + # Per-virtual-layer scales and residual mixing (indexed by virtual_layer_idx). + # Each virtual layer gets its own gating, matching the baseline's per-layer independence. + self.attn_scales = nn.Parameter(torch.ones(num_virtual_layers, dim, dtype=torch.float32)) + self.mlp_scales = nn.Parameter(torch.ones(num_virtual_layers, dim, dtype=torch.float32)) + self.resid_mixes = nn.Parameter( + torch.stack([torch.stack((torch.ones(dim), torch.zeros(dim))) for _ in range(num_virtual_layers)]).float() + ) + # Stacked LoRA tensors for torch.compile compatibility (indexed by virtual_layer_idx). + # A initialized with small random values, B initialized to zero → deviation starts at zero. + self.attn_lora_A = nn.Parameter(torch.randn(num_virtual_layers, dim, lora_rank) * (1.0 / lora_rank)) + self.attn_lora_B = nn.Parameter(torch.zeros(num_virtual_layers, lora_rank, dim)) + self.mlp_lora_A = nn.Parameter(torch.randn(num_virtual_layers, dim, lora_rank) * (1.0 / lora_rank)) + self.mlp_lora_B = nn.Parameter(torch.zeros(num_virtual_layers, lora_rank, dim)) + # Granular shrinkage: separate gammas for attention vs MLP per virtual layer. + # sigmoid(-2.0) ≈ 0.12, so layers start mostly tied. + self.attn_shrinkage_logits = nn.Parameter(torch.full((num_virtual_layers,), -2.0)) + self.mlp_shrinkage_logits = nn.Parameter(torch.full((num_virtual_layers,), -2.0)) + + def forward(self, x: Tensor, x0: Tensor, virtual_layer_idx: int) -> Tensor: + gamma_attn = torch.sigmoid(self.attn_shrinkage_logits[virtual_layer_idx]) + gamma_mlp = torch.sigmoid(self.mlp_shrinkage_logits[virtual_layer_idx]) + mix = self.resid_mixes[virtual_layer_idx].to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + # Shared attention + LoRA deviation + normed = self.attn_norm(x) + attn_out = self.attn(normed) + lora_attn = normed @ self.attn_lora_A[virtual_layer_idx].to(x.dtype) @ self.attn_lora_B[virtual_layer_idx].to(x.dtype) + attn_out = attn_out + gamma_attn.to(x.dtype) * lora_attn + x = x + self.attn_scales[virtual_layer_idx].to(dtype=x.dtype)[None, None, :] * attn_out + # Shared MLP + LoRA deviation + normed_mlp = self.mlp_norm(x) + mlp_out = self.mlp(normed_mlp) + lora_mlp = normed_mlp @ self.mlp_lora_A[virtual_layer_idx].to(x.dtype) @ self.mlp_lora_B[virtual_layer_idx].to(x.dtype) + mlp_out = mlp_out + gamma_mlp.to(x.dtype) * lora_mlp + x = x + self.mlp_scales[virtual_layer_idx].to(dtype=x.dtype)[None, None, :] * mlp_out + 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, + lora_rank: int = 8, + num_shared_blocks: int = 3, + bigram_buckets: int = 10240, + bigram_dim: int = 128, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + if num_layers % num_shared_blocks != 0: + raise ValueError(f"num_layers ({num_layers}) must be divisible by num_shared_blocks ({num_shared_blocks})") + 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.smear_gate = SmearGate(model_dim) + self.bigram_hash = BigramHash(bigram_buckets, bigram_dim, model_dim) + # EBLS: shared blocks with virtual layer schedule + self.num_shared_blocks = num_shared_blocks + self.virtual_layers_per_block = num_layers // num_shared_blocks + num_effective_layers = num_shared_blocks * self.virtual_layers_per_block + self.num_encoder_layers = num_effective_layers // 2 + self.num_decoder_layers = num_effective_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.shared_blocks = nn.ModuleList( + [ + EBLSBlock( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + lora_rank, + self.virtual_layers_per_block, + ) + for _ in range(num_shared_blocks) + ] + ) + # Pre-build virtual layer schedule: (block_idx, virtual_idx) tuples + self.schedule = tuple( + (block_idx, v) + for block_idx in range(num_shared_blocks) + for v in range(self.virtual_layers_per_block) + ) + 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 _run_layers(self, input_ids: Tensor) -> Tensor: + """Shared encoder-decoder forward, returns final hidden states.""" + x = self.tok_emb(input_ids) + x = x + self.bigram_hash(input_ids) + x = self.smear_gate(x) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + block_idx, v_idx = self.schedule[i] + x = self.shared_blocks[block_idx](x, x0, v_idx) + 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() + block_idx, v_idx = self.schedule[self.num_encoder_layers + i] + x = self.shared_blocks[block_idx](x, x0, v_idx) + return self.final_norm(x) + + def _get_logits(self, hidden: Tensor) -> Tensor: + """Project hidden states to vocabulary logits with softcap.""" + flat = hidden.reshape(-1, hidden.size(-1)) + if self.tie_embeddings: + logits_proj = F.linear(flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(flat) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + hidden = self._run_layers(input_ids) + logits = self._get_logits(hidden) + return F.cross_entropy(logits.float(), target_ids.reshape(-1), reduction="mean") + + @torch.no_grad() + def get_logits(self, input_ids: Tensor) -> Tensor: + """Return full logit tensor (batch, seq_len, vocab_size) for inference.""" + hidden = self._run_layers(input_ids) + bsz, seq_len, _ = hidden.shape + logits = self._get_logits(hidden) + return logits.reshape(bsz, seq_len, -1) + + @torch.no_grad() + def generate(self, input_ids: Tensor, max_new_tokens: int = 128, temperature: float = 0.8, top_k: int = 50) -> Tensor: + """Autoregressive generation from a prompt.""" + ids = input_ids.clone() + for _ in range(max_new_tokens): + context = ids[:, -1024:] # Limit to seq_len window + logits = self.get_logits(context)[:, -1, :] / max(temperature, 1e-6) + if top_k > 0: + v, _ = torch.topk(logits, min(top_k, logits.size(-1))) + logits[logits < v[:, [-1]]] = float("-inf") + probs = F.softmax(logits.float(), dim=-1) + next_id = torch.multinomial(probs, num_samples=1) + ids = torch.cat([ids, next_id], dim=1) + return ids + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + if not int(os.environ.get("SKIP_COMPILE", "0")): + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + lora_rank=args.lora_rank, + num_shared_blocks=args.num_shared_blocks, + bigram_buckets=args.bigram_buckets, + bigram_dim=args.bigram_dim, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # Keep LoRA params in fp32 for optimizer quality (same pattern as CastedLinear). + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if "lora" in name and param.dtype != torch.float32: + param.data = param.data.float() + if int(os.environ.get("SKIP_COMPILE", "0")): + compiled_model = base_model + else: + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in shared blocks use MATRIX_LR via Muon (excludes LoRA) + # - everything else (scalars, LoRA 3D tensors, shrinkage logits) uses SCALAR_LR via Adam + block_named_params = list(base_model.shared_blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + # BigramHash proj is a 2D CastedLinear → include in Muon + matrix_params.append(base_model.bigram_hash.proj.weight) + matrix_param_ids = {id(p) for p in matrix_params} + scalar_params = [p for _, p in block_named_params if id(p) not in matrix_param_ids] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + # SmearGate gate_logit → scalar Adam + scalar_params.append(base_model.smear_gate.gate_logit) + # BigramHash embed → scalar Adam (embedding, not a Muon matrix) + scalar_params.append(base_model.bigram_hash.embed.weight) + 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, + weight_decay=args.muon_weight_decay, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + n_shared = sum(p.numel() for b in base_model.shared_blocks for n, p in b.named_parameters() if "lora" not in n and "shrinkage" not in n) + n_lora = sum(p.numel() for p in base_model.parameters() if p.ndim == 3) + log0(f"model_params:{n_params} (shared:{n_shared} lora:{n_lora} other:{n_params - n_shared - n_lora})") + log0(f"ebls: num_shared_blocks:{args.num_shared_blocks} virtual_layers_per_block:{base_model.virtual_layers_per_block} lora_rank:{args.lora_rank} shrinkage_lambda:{args.shrinkage_lambda}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + # SWA: accumulate weight averages during late warmdown + swa_state: dict[str, Tensor] = {} + swa_count = 0 + + 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, + ) + attn_gammas = [ + torch.sigmoid(block.attn_shrinkage_logits[v]).item() + for block in base_model.shared_blocks + for v in range(block.num_virtual_layers) + ] + mlp_gammas = [ + torch.sigmoid(block.mlp_shrinkage_logits[v]).item() + for block in base_model.shared_blocks + for v in range(block.num_virtual_layers) + ] + 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" + ) + log0(f"attn_gammas: {[f'{g:.4f}' for g in attn_gammas]}") + log0(f"mlp_gammas: {[f'{g:.4f}' for g in mlp_gammas]}") + 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) + # Shrinkage regularization: penalize deviation from shared weights. + if args.shrinkage_lambda > 0: + shrink_reg = torch.sigmoid(torch.cat([ + block.attn_shrinkage_logits for block in base_model.shared_blocks + ] + [ + block.mlp_shrinkage_logits for block in base_model.shared_blocks + ])).sum() + loss = loss + args.shrinkage_lambda * shrink_reg + 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 + + # SWA: accumulate during late warmdown + if scale < 1.0 and scale <= args.swa_start_frac and args.swa_every > 0 and step % args.swa_every == 0: + with torch.no_grad(): + for name, param in base_model.state_dict().items(): + if name not in swa_state: + swa_state[name] = param.detach().cpu().clone().float() + else: + swa_state[name] += param.detach().cpu().float() + swa_count += 1 + + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # Apply SWA averaged weights if collected + if swa_count > 0: + log0(f"swa: applying averaged weights from {swa_count} checkpoints") + avg_state = {name: (t / swa_count) for name, t in swa_state.items()} + # Cast back to original dtypes + orig_state = base_model.state_dict() + for name in avg_state: + avg_state[name] = avg_state[name].to(dtype=orig_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + del swa_state, avg_state + + # ----------------------------- + # 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() + cctx = zstd.ZstdCompressor(level=22) + quant_blob = cctx.compress(quant_raw) + 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 int6+zstd: {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 int6+zstd: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + dctx = zstd.ZstdDecompressor() + quant_state = torch.load(io.BytesIO(dctx.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + # Standard eval for comparison + 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_int6_zstd_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" + ) + # Sliding window eval (stride=64) — skip if SKIP_COMPILE set (dev mode) + if not int(os.environ.get("SKIP_COMPILE", "0")): + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_size=64, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms stride:64" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + else: + log0("sliding_window_eval: skipped (SKIP_COMPILE/dev mode)") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.11.10 (main, Sep 7 2024, 18:35:41) [GCC 11.4.0] +Running PyTorch 2.4.1+cu124 +Sun Mar 22 15:25:47 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 35C P0 123W / 700W | 4336MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 32C P0 121W / 700W | 4384MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 32C P0 120W / 700W | 4384MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 35C P0 124W / 700W | 4384MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 36C P0 119W / 700W | 4384MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 34C P0 122W / 700W | 4384MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 35C P0 119W / 700W | 4384MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 32C P0 117W / 700W | 4144MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| 0 N/A N/A 16246 C /usr/bin/python 4326MiB | +| 1 N/A N/A 16247 C /usr/bin/python 4374MiB | +| 2 N/A N/A 16248 C /usr/bin/python 4374MiB | +| 3 N/A N/A 16249 C /usr/bin/python 4374MiB | +| 4 N/A N/A 16250 C /usr/bin/python 4374MiB | +| 5 N/A N/A 16251 C /usr/bin/python 4374MiB | +| 6 N/A N/A 16252 C /usr/bin/python 4374MiB | +| 7 N/A N/A 16253 C /usr/bin/python 4134MiB | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:29566018 (shared:26775600 lora:313344 other:2477074) +ebls: num_shared_blocks:3 virtual_layers_per_block:3 lora_rank:8 shrinkage_lambda:0.01 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:16 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.02 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:7.0292 val_bpb:4.1631 train_time:0ms step_avg:0.02ms +attn_gammas: ['0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192'] +mlp_gammas: ['0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192', '0.1192'] +step:1/20000 train_loss:7.0784 train_time:168ms step_avg:167.54ms +step:2/20000 train_loss:21.8592 train_time:211ms step_avg:105.56ms +step:3/20000 train_loss:9.6747 train_time:343ms step_avg:114.46ms +step:4/20000 train_loss:9.0667 train_time:474ms step_avg:118.38ms +step:5/20000 train_loss:7.2876 train_time:603ms step_avg:120.70ms +step:6/20000 train_loss:7.4041 train_time:734ms step_avg:122.31ms +step:7/20000 train_loss:6.5683 train_time:864ms step_avg:123.41ms +step:8/20000 train_loss:6.3801 train_time:994ms step_avg:124.26ms +step:9/20000 train_loss:6.3831 train_time:1231ms step_avg:136.81ms +step:10/20000 train_loss:6.3478 train_time:1362ms step_avg:136.16ms +step:200/20000 train_loss:3.1368 train_time:26095ms step_avg:130.47ms +step:400/20000 train_loss:2.3893 train_time:52218ms step_avg:130.54ms +step:600/20000 train_loss:2.5535 train_time:78414ms step_avg:130.69ms +step:800/20000 train_loss:2.2928 train_time:104577ms step_avg:130.72ms +step:1000/20000 train_loss:2.3725 train_time:130684ms step_avg:130.68ms +step:1000/20000 val_loss:2.3315 val_bpb:1.3808 train_time:130778ms step_avg:130.78ms +attn_gammas: ['0.0112', '0.0014', '0.0004', '0.0005', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +mlp_gammas: ['0.0007', '0.0002', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +step:1200/20000 train_loss:2.3817 train_time:156723ms step_avg:130.60ms +step:1400/20000 train_loss:2.4335 train_time:182804ms step_avg:130.57ms +step:1600/20000 train_loss:2.0983 train_time:208785ms step_avg:130.49ms +step:1800/20000 train_loss:2.1880 train_time:234749ms step_avg:130.42ms +step:2000/20000 train_loss:2.2353 train_time:260696ms step_avg:130.35ms +step:2000/20000 val_loss:2.2179 val_bpb:1.3135 train_time:260790ms step_avg:130.39ms +attn_gammas: ['0.0089', '0.0017', '0.0018', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +mlp_gammas: ['0.0015', '0.0001', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +step:2200/20000 train_loss:2.0487 train_time:286638ms step_avg:130.29ms +step:2400/20000 train_loss:2.1738 train_time:312538ms step_avg:130.22ms +step:2600/20000 train_loss:2.3831 train_time:338428ms step_avg:130.16ms +step:2800/20000 train_loss:2.2095 train_time:364313ms step_avg:130.11ms +step:3000/20000 train_loss:2.1976 train_time:390218ms step_avg:130.07ms +step:3000/20000 val_loss:2.1625 val_bpb:1.2808 train_time:390310ms step_avg:130.10ms +attn_gammas: ['0.0054', '0.0015', '0.0016', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +mlp_gammas: ['0.0015', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +step:3200/20000 train_loss:2.1607 train_time:416083ms step_avg:130.03ms +step:3400/20000 train_loss:2.1324 train_time:441972ms step_avg:129.99ms +step:3600/20000 train_loss:2.0781 train_time:467830ms step_avg:129.95ms +step:3800/20000 train_loss:2.1638 train_time:493701ms step_avg:129.92ms +step:4000/20000 train_loss:2.0892 train_time:519567ms step_avg:129.89ms +step:4000/20000 val_loss:2.0986 val_bpb:1.2429 train_time:519661ms step_avg:129.92ms +attn_gammas: ['0.0040', '0.0015', '0.0014', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +mlp_gammas: ['0.0015', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +step:4200/20000 train_loss:2.0916 train_time:548564ms step_avg:130.61ms +step:4400/20000 train_loss:2.0042 train_time:577788ms step_avg:131.32ms +step:4572/20000 val_loss:2.0440 val_bpb:1.2105 train_time:600032ms step_avg:131.24ms +attn_gammas: ['0.0035', '0.0013', '0.0012', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +mlp_gammas: ['0.0012', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000', '0.0000'] +stopping_early: wallclock_cap train_time:600032ms step:4572/20000 +peak memory allocated: 29769 MiB reserved: 30418 MiB +swa: applying averaged weights from 9 checkpoints +Serialized model: 113564827 bytes +Code size: 62684 bytes +Total submission size: 113627511 bytes +Serialized model int6+zstd: 16162142 bytes (payload:30051592 raw_torch:30074488 payload_ratio:3.78x) +Total submission size int6+zstd: 16224826 bytes +final_int6_zstd_roundtrip val_loss:2.2694 val_bpb:1.3441 eval_time:4147ms diff --git a/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train_gpt.py b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train_gpt.py new file mode 100644 index 000000000..c39aa56c3 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_EBLS_LearnedSharing/train_gpt.py @@ -0,0 +1,1393 @@ +""" +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: `train_gpt.py` and `train_gpt_mlx.py` must never be 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 zstandard as zstd +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 1024)) + num_heads = int(os.environ.get("NUM_HEADS", 16)) + mlp_mult = int(os.environ.get("MLP_MULT", 3)) + 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)) + + # BigramHash + SmearGate parameters. + bigram_buckets = int(os.environ.get("BIGRAM_BUCKETS", 10240)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + # EBLS (Empirical Bayes Layer Sharing) parameters. + lora_rank = int(os.environ.get("LORA_RANK", 8)) + shrinkage_lambda = float(os.environ.get("SHRINKAGE_LAMBDA", 0.01)) + num_shared_blocks = int(os.environ.get("NUM_SHARED_BLOCKS", 3)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + muon_weight_decay = float(os.environ.get("MUON_WEIGHT_DECAY", 0.04)) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + +# ----------------------------- +# 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, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov, weight_decay=weight_decay), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group["weight_decay"] + + # Decoupled weight decay (applied before update) + if wd > 0: + for p in params: + p.data.mul_(1.0 - lr * wd) + + 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) + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int = 64, + batch_size: int = 64, +) -> tuple[float, float]: + """Sliding window eval: overlapping windows with stride, score only last `stride` tokens.""" + seq_len = args.train_seq_len + total = val_tokens.numel() - 1 + max_start = total - seq_len + all_starts = list(range(0, max_start + 1, stride)) + my_starts = all_starts[rank::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) + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_starts), batch_size): + batch_starts = my_starts[bi:bi + batch_size] + bsz = len(batch_starts) + x_batch = torch.stack([val_tokens[s:s + seq_len] for s in batch_starts]).to(device=device, dtype=torch.int64) + y_batch = torch.stack([val_tokens[s + 1:s + seq_len + 1] for s in batch_starts]).to(device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = base_model.get_logits(x_batch) # (bsz, seq_len, vocab) + score_logits = logits[:, -stride:, :].reshape(-1, logits.size(-1)) + score_targets = y_batch[:, -stride:].reshape(-1) + losses = F.cross_entropy(score_logits.float(), score_targets, reduction='none') + val_loss_sum += losses.to(torch.float64).sum() + val_token_count += float(score_targets.numel()) + prev_ids = x_batch[:, -stride:].reshape(-1) + tgt_ids = score_targets + 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() + base_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,attn_shrinkage_logits,mlp_shrinkage_logits,gate_logit", + ).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 +# Int6 quantization: [-31, 31] range packed into int8 storage +INT6_RANGE = 31 + +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]: + """Quantize to int6 range [-31, 31] stored in int8 containers.""" + t32 = t.float() + qr = INT6_RANGE + if t32.ndim == 2: + clip_abs = t32.abs().amax(dim=1) + scale = (clip_abs / qr).clamp_min(1.0 / qr) + q = torch.clamp(torch.round(t32 / scale[:, None]), -qr, qr).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(t32.abs().max().item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / qr if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(t32 / scale), -qr, qr).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +def fake_quantize_int6(w: Tensor) -> Tensor: + """Fake int6 quantization with straight-through estimator for QAT.""" + scale = w.float().abs().amax(dim=-1, keepdim=True).clamp(min=1e-8) / 31.0 + w_q = (w.float() / scale).round().clamp(-31, 31) * scale + return w + (w_q - w).detach() # STE: forward uses quantized, backward uses original + +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + # During training, applies fake int6 quantization (STE) to close the quantization gap. + def forward(self, x: Tensor) -> Tensor: + w = self.weight + if self.training: + w = fake_quantize_int6(w) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + if self.num_kv_heads != self.num_heads: + rep = self.num_heads // self.num_kv_heads + k = k[:, :, None, :, :].expand(bsz, self.num_kv_heads, rep, seqlen, self.head_dim).reshape(bsz, self.num_heads, seqlen, self.head_dim) + v = v[:, :, None, :, :].expand(bsz, self.num_kv_heads, rep, seqlen, self.head_dim).reshape(bsz, self.num_heads, seqlen, self.head_dim) + y = F.scaled_dot_product_attention(q, k, v, attn_mask=None, is_causal=True) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class SmearGate(nn.Module): + """Per-dimension gate blending current token with previous token embedding.""" + def __init__(self, dim: int, init_logit: float = 3.0): + super().__init__() + # sigmoid(3.0) ≈ 0.95 → mostly keep current token + self.gate_logit = nn.Parameter(torch.full((dim,), init_logit, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + gate = torch.sigmoid(self.gate_logit).to(x.dtype) + # Shift right: prev token embedding for position i is x at position i-1 + x_prev = F.pad(x[:, :-1, :], (0, 0, 1, 0)) # zero-pad first position + return gate * x + (1 - gate) * x_prev + + +class BigramHash(nn.Module): + """Hash-based bigram embedding: maps (prev_token, cur_token) pairs to learned vectors.""" + def __init__(self, num_buckets: int, embed_dim: int, model_dim: int): + super().__init__() + self.num_buckets = num_buckets + self.embed = nn.Embedding(num_buckets, embed_dim) + self.proj = CastedLinear(embed_dim, model_dim, bias=False) + nn.init.normal_(self.embed.weight, std=0.01) + nn.init.zeros_(self.proj.weight) + + def forward(self, input_ids: Tensor) -> Tensor: + # Hash bigrams: prev_id * large_prime + cur_id, mod num_buckets + prev_ids = F.pad(input_ids[:, :-1], (1, 0)) # zero for first position + bigram_hash = ((prev_ids.long() * 104729 + input_ids.long()) % self.num_buckets).long() + return self.proj(self.embed(bigram_hash)) + + +class EBLSBlock(nn.Module): + """Transformer block with Empirical Bayes Layer Sharing. + + Shared base attention + MLP weights are reused across virtual layers. + Per-virtual-layer LoRA deviations provide specialization, gated by + learned shrinkage factors gamma_i = sigmoid(logit_i). + """ + + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + lora_rank: int, + num_virtual_layers: int, + ): + 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.num_virtual_layers = num_virtual_layers + self.lora_rank = lora_rank + # Per-virtual-layer scales and residual mixing (indexed by virtual_layer_idx). + # Each virtual layer gets its own gating, matching the baseline's per-layer independence. + self.attn_scales = nn.Parameter(torch.ones(num_virtual_layers, dim, dtype=torch.float32)) + self.mlp_scales = nn.Parameter(torch.ones(num_virtual_layers, dim, dtype=torch.float32)) + self.resid_mixes = nn.Parameter( + torch.stack([torch.stack((torch.ones(dim), torch.zeros(dim))) for _ in range(num_virtual_layers)]).float() + ) + # Stacked LoRA tensors for torch.compile compatibility (indexed by virtual_layer_idx). + # A initialized with small random values, B initialized to zero → deviation starts at zero. + self.attn_lora_A = nn.Parameter(torch.randn(num_virtual_layers, dim, lora_rank) * (1.0 / lora_rank)) + self.attn_lora_B = nn.Parameter(torch.zeros(num_virtual_layers, lora_rank, dim)) + self.mlp_lora_A = nn.Parameter(torch.randn(num_virtual_layers, dim, lora_rank) * (1.0 / lora_rank)) + self.mlp_lora_B = nn.Parameter(torch.zeros(num_virtual_layers, lora_rank, dim)) + # Granular shrinkage: separate gammas for attention vs MLP per virtual layer. + # sigmoid(-2.0) ≈ 0.12, so layers start mostly tied. + self.attn_shrinkage_logits = nn.Parameter(torch.full((num_virtual_layers,), -2.0)) + self.mlp_shrinkage_logits = nn.Parameter(torch.full((num_virtual_layers,), -2.0)) + + def forward(self, x: Tensor, x0: Tensor, virtual_layer_idx: int) -> Tensor: + gamma_attn = torch.sigmoid(self.attn_shrinkage_logits[virtual_layer_idx]) + gamma_mlp = torch.sigmoid(self.mlp_shrinkage_logits[virtual_layer_idx]) + mix = self.resid_mixes[virtual_layer_idx].to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + # Shared attention + LoRA deviation + normed = self.attn_norm(x) + attn_out = self.attn(normed) + lora_attn = normed @ self.attn_lora_A[virtual_layer_idx].to(x.dtype) @ self.attn_lora_B[virtual_layer_idx].to(x.dtype) + attn_out = attn_out + gamma_attn.to(x.dtype) * lora_attn + x = x + self.attn_scales[virtual_layer_idx].to(dtype=x.dtype)[None, None, :] * attn_out + # Shared MLP + LoRA deviation + normed_mlp = self.mlp_norm(x) + mlp_out = self.mlp(normed_mlp) + lora_mlp = normed_mlp @ self.mlp_lora_A[virtual_layer_idx].to(x.dtype) @ self.mlp_lora_B[virtual_layer_idx].to(x.dtype) + mlp_out = mlp_out + gamma_mlp.to(x.dtype) * lora_mlp + x = x + self.mlp_scales[virtual_layer_idx].to(dtype=x.dtype)[None, None, :] * mlp_out + 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, + lora_rank: int = 8, + num_shared_blocks: int = 3, + bigram_buckets: int = 10240, + bigram_dim: int = 128, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + if num_layers % num_shared_blocks != 0: + raise ValueError(f"num_layers ({num_layers}) must be divisible by num_shared_blocks ({num_shared_blocks})") + 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.smear_gate = SmearGate(model_dim) + self.bigram_hash = BigramHash(bigram_buckets, bigram_dim, model_dim) + # EBLS: shared blocks with virtual layer schedule + self.num_shared_blocks = num_shared_blocks + self.virtual_layers_per_block = num_layers // num_shared_blocks + num_effective_layers = num_shared_blocks * self.virtual_layers_per_block + self.num_encoder_layers = num_effective_layers // 2 + self.num_decoder_layers = num_effective_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.shared_blocks = nn.ModuleList( + [ + EBLSBlock( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + lora_rank, + self.virtual_layers_per_block, + ) + for _ in range(num_shared_blocks) + ] + ) + # Pre-build virtual layer schedule: (block_idx, virtual_idx) tuples + self.schedule = tuple( + (block_idx, v) + for block_idx in range(num_shared_blocks) + for v in range(self.virtual_layers_per_block) + ) + 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 _run_layers(self, input_ids: Tensor) -> Tensor: + """Shared encoder-decoder forward, returns final hidden states.""" + x = self.tok_emb(input_ids) + x = x + self.bigram_hash(input_ids) + x = self.smear_gate(x) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + block_idx, v_idx = self.schedule[i] + x = self.shared_blocks[block_idx](x, x0, v_idx) + 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() + block_idx, v_idx = self.schedule[self.num_encoder_layers + i] + x = self.shared_blocks[block_idx](x, x0, v_idx) + return self.final_norm(x) + + def _get_logits(self, hidden: Tensor) -> Tensor: + """Project hidden states to vocabulary logits with softcap.""" + flat = hidden.reshape(-1, hidden.size(-1)) + if self.tie_embeddings: + logits_proj = F.linear(flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(flat) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + hidden = self._run_layers(input_ids) + logits = self._get_logits(hidden) + return F.cross_entropy(logits.float(), target_ids.reshape(-1), reduction="mean") + + @torch.no_grad() + def get_logits(self, input_ids: Tensor) -> Tensor: + """Return full logit tensor (batch, seq_len, vocab_size) for inference.""" + hidden = self._run_layers(input_ids) + bsz, seq_len, _ = hidden.shape + logits = self._get_logits(hidden) + return logits.reshape(bsz, seq_len, -1) + + @torch.no_grad() + def generate(self, input_ids: Tensor, max_new_tokens: int = 128, temperature: float = 0.8, top_k: int = 50) -> Tensor: + """Autoregressive generation from a prompt.""" + ids = input_ids.clone() + for _ in range(max_new_tokens): + context = ids[:, -1024:] # Limit to seq_len window + logits = self.get_logits(context)[:, -1, :] / max(temperature, 1e-6) + if top_k > 0: + v, _ = torch.topk(logits, min(top_k, logits.size(-1))) + logits[logits < v[:, [-1]]] = float("-inf") + probs = F.softmax(logits.float(), dim=-1) + next_id = torch.multinomial(probs, num_samples=1) + ids = torch.cat([ids, next_id], dim=1) + return ids + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + if not int(os.environ.get("SKIP_COMPILE", "0")): + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + lora_rank=args.lora_rank, + num_shared_blocks=args.num_shared_blocks, + bigram_buckets=args.bigram_buckets, + bigram_dim=args.bigram_dim, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + # Keep LoRA params in fp32 for optimizer quality (same pattern as CastedLinear). + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if "lora" in name and param.dtype != torch.float32: + param.data = param.data.float() + if int(os.environ.get("SKIP_COMPILE", "0")): + compiled_model = base_model + else: + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in shared blocks use MATRIX_LR via Muon (excludes LoRA) + # - everything else (scalars, LoRA 3D tensors, shrinkage logits) uses SCALAR_LR via Adam + block_named_params = list(base_model.shared_blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + # BigramHash proj is a 2D CastedLinear → include in Muon + matrix_params.append(base_model.bigram_hash.proj.weight) + matrix_param_ids = {id(p) for p in matrix_params} + scalar_params = [p for _, p in block_named_params if id(p) not in matrix_param_ids] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + # SmearGate gate_logit → scalar Adam + scalar_params.append(base_model.smear_gate.gate_logit) + # BigramHash embed → scalar Adam (embedding, not a Muon matrix) + scalar_params.append(base_model.bigram_hash.embed.weight) + 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, + weight_decay=args.muon_weight_decay, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + n_shared = sum(p.numel() for b in base_model.shared_blocks for n, p in b.named_parameters() if "lora" not in n and "shrinkage" not in n) + n_lora = sum(p.numel() for p in base_model.parameters() if p.ndim == 3) + log0(f"model_params:{n_params} (shared:{n_shared} lora:{n_lora} other:{n_params - n_shared - n_lora})") + log0(f"ebls: num_shared_blocks:{args.num_shared_blocks} virtual_layers_per_block:{base_model.virtual_layers_per_block} lora_rank:{args.lora_rank} shrinkage_lambda:{args.shrinkage_lambda}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + # SWA: accumulate weight averages during late warmdown + swa_state: dict[str, Tensor] = {} + swa_count = 0 + + 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, + ) + attn_gammas = [ + torch.sigmoid(block.attn_shrinkage_logits[v]).item() + for block in base_model.shared_blocks + for v in range(block.num_virtual_layers) + ] + mlp_gammas = [ + torch.sigmoid(block.mlp_shrinkage_logits[v]).item() + for block in base_model.shared_blocks + for v in range(block.num_virtual_layers) + ] + 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" + ) + log0(f"attn_gammas: {[f'{g:.4f}' for g in attn_gammas]}") + log0(f"mlp_gammas: {[f'{g:.4f}' for g in mlp_gammas]}") + 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) + # Shrinkage regularization: penalize deviation from shared weights. + if args.shrinkage_lambda > 0: + shrink_reg = torch.sigmoid(torch.cat([ + block.attn_shrinkage_logits for block in base_model.shared_blocks + ] + [ + block.mlp_shrinkage_logits for block in base_model.shared_blocks + ])).sum() + loss = loss + args.shrinkage_lambda * shrink_reg + 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 + + # SWA: accumulate during late warmdown + if scale < 1.0 and scale <= args.swa_start_frac and args.swa_every > 0 and step % args.swa_every == 0: + with torch.no_grad(): + for name, param in base_model.state_dict().items(): + if name not in swa_state: + swa_state[name] = param.detach().cpu().clone().float() + else: + swa_state[name] += param.detach().cpu().float() + swa_count += 1 + + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # Apply SWA averaged weights if collected + if swa_count > 0: + log0(f"swa: applying averaged weights from {swa_count} checkpoints") + avg_state = {name: (t / swa_count) for name, t in swa_state.items()} + # Cast back to original dtypes + orig_state = base_model.state_dict() + for name in avg_state: + avg_state[name] = avg_state[name].to(dtype=orig_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + del swa_state, avg_state + + # ----------------------------- + # 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() + cctx = zstd.ZstdCompressor(level=22) + quant_blob = cctx.compress(quant_raw) + 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 int6+zstd: {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 int6+zstd: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + dctx = zstd.ZstdDecompressor() + quant_state = torch.load(io.BytesIO(dctx.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + # Standard eval for comparison + 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_int6_zstd_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" + ) + # Sliding window eval (stride=64) — skip if SKIP_COMPILE set (dev mode) + if not int(os.environ.get("SKIP_COMPILE", "0")): + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_size=64, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms stride:64" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + else: + log0("sliding_window_eval: skipped (SKIP_COMPILE/dev mode)") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main()