diff --git a/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/README.md b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/README.md new file mode 100644 index 000000000..dc2a9aada --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/README.md @@ -0,0 +1,157 @@ +# Hybrid Spiking Neural Networks (SNNs) MLP + +**val_bpb: 1.2982** | **15.78 MB** | 8×H100 SXM + +A contest-friendly hybrid SNN submission built from the `train_gpt.py` baseline: keep dense GQA attention and the original training/eval/compression pipeline, but replace the standard feed-forward block with a small multi-step leaky integrate-and-fire (LIF-style) spiking MLP. + +Reference :https://arxiv.org/pdf/2203.14679 + +## Results (8×H100 80GB SXM) + +| Run | Val loss | **Val bpb** | Serialized model | int8+zlib model | Total submission | +|-----|----------|-------------|------------------|-----------------|------------------| +| SNN baseline | 2.1919 | **1.2982** | 67,233,157 | 15,723,303 | **15,776,086** | + +### Exact export log + +- Serialized model: **67,233,157 bytes** +- Code size: **52,783 bytes** +- Total submission size: **67,285,940 bytes** +- Serialized model int8+zlib: **15,723,303 bytes** + - payload: 17,179,020 bytes + - raw_torch: 17,231,715 bytes + - payload_ratio: 3.91x +- Total submission size int8+zlib: **15,776,086 bytes** +- final_int8_zlib_roundtrip val_loss: **2.19192126** +- final_int8_zlib_roundtrip val_bpb: **1.29817924** +- eval_time: **1435 ms** + +## Key Innovation: Replace MLP with a Multi-Step Spiking MLP + +The baseline keeps the attention path dense and only swaps the MLP inside each Transformer block. + +```python +# Standard baseline MLP +x = torch.relu(self.fc(x)) +out = self.proj(x.square()) +``` + +```python +# This submission: multi-step LIF-style spiking MLP +cur = self.fc(x) +mem = torch.zeros_like(cur) +spike_sum = torch.zeros_like(cur) +for _ in range(self.snn_steps): + mem = decay * mem + cur / self.snn_steps + over = mem - thresh + spike_soft = torch.sigmoid(grad_scale * over) + spike_hard = (over > 0).to(dtype=x.dtype) + spike = spike_hard + spike_soft - spike_soft.detach() + mem = mem - spike_hard * thresh + spike_sum = spike_sum + spike +rate = spike_sum / self.snn_steps +out = self.proj(rate * self.spike_out_scale) +``` + +The spiking pathway introduces: +- **multi-step membrane integration** instead of a one-shot activation +- **thresholded firing** instead of continuous hidden activations +- **surrogate-gradient training** via a sigmoid straight-through estimator +- **spike-rate regularization** during training + +This makes the FFN a small dynamical system rather than a static pointwise nonlinearity. + +## Why this is interesting + +This is **not** a fully spiking language model. It is a **hybrid Transformer + SNN-MLP** design: + +- embeddings, attention, residual path, and logits remain standard dense LM components +- only the feed-forward block is replaced by a spiking mechanism +- the original Parameter Golf training and export path stays intact + +That makes the experiment meaningful for the contest setting because it isolates one question: + +> Can spike neural network achieves good performance in a tiny language model under a strict size budget? + +## Training Architecture + +Baseline model shape from `train_gpt.py`: + +| Component | Setting | +|-----------|---------| +| Layers | 9 | +| Width | 512 | +| Attention | 8 heads, 4 KV heads (GQA) | +| Sequence length | 1024 | +| Vocab size | 1024 | +| MLP | **2× multi-step spiking MLP** | +| Embeddings | Tied | +| Position encoding | RoPE | +| Norm | RMSNorm | +| Residual structure | Encoder/decoder-style skip reuse | +| Logit stabilization | tanh softcap | +| Quantization/export | int8 + zlib | + +### Spiking hyperparameters + +| Parameter | Value | +|-----------|-------| +| `USE_SNN_MLP` | 1 | +| `SNN_STEPS` | 2 | +| `SNN_DECAY` | 0.8 | +| `SNN_THRESH_INIT` | 1.0 | +| `SNN_GRAD_SCALE` | 4.0 | +| `SNN_OUT_SCALE_INIT` | 1.0 | +| `SNN_RATE_LOSS` | 1e-4 | +| `SNN_RATE_TARGET` | 0.15 | + +## Optimizer Setup + +The script keeps the baseline split-optimizer recipe: + +- **Adam** for token embeddings +- **Muon** for matrix-shaped parameters +- **Adam** for scalar/vector parameters +- optional tied-embedding head path from the baseline remains unchanged + +This is important because the submission changes the model architecture without rewriting the overall training system. + +## Run Command + +```bash +RUN_ID=snn_baseline \ +DATA_PATH=./data/datasets/fineweb10B_sp1024 \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 \ +USE_SNN_MLP=1 \ +SNN_STEPS=2 \ +SNN_DECAY=0.8 \ +SNN_THRESH_INIT=1.0 \ +SNN_GRAD_SCALE=4.0 \ +SNN_OUT_SCALE_INIT=1.0 \ +SNN_RATE_LOSS=1e-4 \ +SNN_RATE_TARGET=0.15 \ +VAL_LOSS_EVERY=1000 \ +TRAIN_LOG_EVERY=200 \ +SEED=1337 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## Practical Takeaway + +This submission is best viewed as a **contest-friendly spiking experiment**, not as a claim that SNNs are already superior to standard dense LLMs on GPUs. + +What it demonstrates: +- a hybrid SNN block can be dropped into the baseline with minimal code changes +- the resulting model still trains end-to-end with the standard PyTorch + Muon pipeline +- the exported artifact remains under the **16 MB** challenge limit after int8+zlib compression + +## Limitations + +- This version projects only the **average spike rate** back to model dimension, which is a fairly aggressive information bottleneck. +- It is likely more interesting as an architectural experiment than as the strongest contest-optimized submission. + + +## Credits + +- **SNN adaptation**: `train_gpt_snn.py`, replacing the block MLP with a multi-step LIF-style spiking pathway while preserving the original contest pipeline diff --git a/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/submission.json b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/submission.json new file mode 100644 index 000000000..46d873b48 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/submission.json @@ -0,0 +1,18 @@ +{ + "author": "Junxuan Huang", + "github_id": "tsbiosky", + "name": "Hybrid Spiking Neural Networks MLP", + "blurb": "Non-record submission: hybrid spiking Transformer with a multi-step spiking MLP ", + "date": "2026-03-24T22:15:00Z", + "track": "non-record-16mb", + "val_loss": 2.19192126, + "val_bpb": 1.29817924, + "pre_quant_val_loss": null, + "pre_quant_val_bpb": null, + "step_stop": 3773, + "wallclock_seconds": 600, + "bytes_total": 15776086, + "bytes_model_int8_zlib": 15723303, + "bytes_code": 52783, + "gpu": "8xH100" + } \ No newline at end of file diff --git a/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/train_gpt.py b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/train_gpt.py new file mode 100644 index 000000000..3c7e8c2a8 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/train_gpt.py @@ -0,0 +1,1238 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + use_snn_mlp = bool(int(os.environ.get("USE_SNN_MLP", "1"))) + snn_steps = int(os.environ.get("SNN_STEPS", 2)) + snn_decay = float(os.environ.get("SNN_DECAY", 0.8)) + snn_thresh_init = float(os.environ.get("SNN_THRESH_INIT", 1.0)) + snn_grad_scale = float(os.environ.get("SNN_GRAD_SCALE", 4.0)) + snn_out_scale_init = float(os.environ.get("SNN_OUT_SCALE_INIT", 1.0)) + snn_rate_loss = float(os.environ.get("SNN_RATE_LOSS", 1e-4)) + snn_rate_target = float(os.environ.get("SNN_RATE_TARGET", 0.15)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Scale correction from Muon reference implementations. + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,decay_logit,spike_thresh,spike_out_scale", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +def _logit(p: float) -> float: + p = min(max(p, 1e-4), 1.0 - 1e-4) + return math.log(p / (1.0 - p)) + + +class SpikingMLP(nn.Module): + # Contest-friendly hybrid SNN block: keep attention dense, replace the MLP with + # a small multi-step leaky integrate-and-fire pathway using a sigmoid STE. + def __init__( + self, + dim: int, + mlp_mult: int, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + 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 + self.snn_steps = max(int(snn_steps), 1) + self.snn_grad_scale = float(snn_grad_scale) + self.decay_logit = nn.Parameter(torch.tensor(_logit(snn_decay), dtype=torch.float32)) + self.spike_thresh = nn.Parameter(torch.tensor(float(snn_thresh_init), dtype=torch.float32)) + self.spike_out_scale = nn.Parameter(torch.tensor(float(snn_out_scale_init), dtype=torch.float32)) + + def forward(self, x: Tensor) -> tuple[Tensor, Tensor]: + cur = self.fc(x) + mem = torch.zeros_like(cur) + spike_sum = torch.zeros_like(cur) + decay = self.decay_logit.sigmoid().to(dtype=x.dtype) + thresh = self.spike_thresh.to(dtype=x.dtype).clamp_min(0.05) + grad_scale = torch.tensor(self.snn_grad_scale, dtype=x.dtype, device=x.device) + for _ in range(self.snn_steps): + mem = decay * mem + cur / self.snn_steps + over = mem - thresh + spike_soft = torch.sigmoid(grad_scale * over) + spike_hard = (over > 0).to(dtype=x.dtype) + spike = spike_hard + spike_soft - spike_soft.detach() + mem = mem - spike_hard * thresh + spike_sum = spike_sum + spike + rate = spike_sum / self.snn_steps + out = self.proj(rate * self.spike_out_scale.to(dtype=x.dtype)) + spike_rate = rate.mean() + return out, spike_rate + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + use_snn_mlp: bool, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.use_snn_mlp = use_snn_mlp + self.mlp = ( + SpikingMLP(dim, mlp_mult, snn_steps, snn_decay, snn_thresh_init, snn_grad_scale, snn_out_scale_init) + if use_snn_mlp + else MLP(dim, mlp_mult) + ) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + if self.use_snn_mlp: + mlp_out, spike_rate = self.mlp(self.mlp_norm(x)) + else: + mlp_out = self.mlp(self.mlp_norm(x)) + spike_rate = x.new_zeros(()) + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * mlp_out + return x, spike_rate + + +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, + use_snn_mlp: bool, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + use_snn_mlp, + snn_steps, + snn_decay, + snn_thresh_init, + snn_grad_scale, + snn_out_scale_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor, return_aux: bool = False): + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + spike_rate_sum = x.new_zeros(()) + + # First half stores skips; second half reuses them in reverse order. + for i in range(self.num_encoder_layers): + x, block_spike_rate = self.blocks[i](x, x0) + spike_rate_sum = spike_rate_sum + block_spike_rate + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, block_spike_rate = self.blocks[self.num_encoder_layers + i](x, x0) + spike_rate_sum = spike_rate_sum + block_spike_rate + + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + ce_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if return_aux: + mean_spike_rate = spike_rate_sum / max(len(self.blocks), 1) + return ce_loss, mean_spike_rate + return ce_loss + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + use_snn_mlp=args.use_snn_mlp, + snn_steps=args.snn_steps, + snn_decay=args.snn_decay, + snn_thresh_init=args.snn_thresh_init, + snn_grad_scale=args.snn_grad_scale, + snn_out_scale_init=args.snn_out_scale_init, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0( + f"use_snn_mlp:{args.use_snn_mlp} snn_steps:{args.snn_steps} snn_decay:{args.snn_decay} " + f"snn_thresh_init:{args.snn_thresh_init} snn_rate_loss:{args.snn_rate_loss} " + f"snn_rate_target:{args.snn_rate_target}" + ) + 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, warmup_spike_rate = model(x, y, True) + if args.use_snn_mlp and args.snn_rate_loss > 0: + warmup_loss = warmup_loss + args.snn_rate_loss * (warmup_spike_rate - args.snn_rate_target).square() + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + spike_rate = 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, batch_spike_rate = model(x, y, True) + if args.use_snn_mlp and args.snn_rate_loss > 0: + loss = loss + args.snn_rate_loss * (batch_spike_rate - args.snn_rate_target).square() + train_loss += loss.detach() + spike_rate += batch_spike_rate.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + spike_rate /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"spike_rate:{spike_rate.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/train_log.txt b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/train_log.txt new file mode 100644 index 000000000..562072509 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-24_Hybrid_SNN_MLP/train_log.txt @@ -0,0 +1,3135 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 2)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + use_snn_mlp = bool(int(os.environ.get("USE_SNN_MLP", "1"))) + snn_steps = int(os.environ.get("SNN_STEPS", 1)) + snn_decay = float(os.environ.get("SNN_DECAY", 0.8)) + snn_thresh_init = float(os.environ.get("SNN_THRESH_INIT", 1.0)) + snn_grad_scale = float(os.environ.get("SNN_GRAD_SCALE", 4.0)) + snn_out_scale_init = float(os.environ.get("SNN_OUT_SCALE_INIT", 0.5)) + snn_rate_loss = float(os.environ.get("SNN_RATE_LOSS", 1e-4)) + snn_rate_target = float(os.environ.get("SNN_RATE_TARGET", 0.15)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Scale correction from Muon reference implementations. + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,decay_logit,spike_thresh,spike_out_scale", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +def _logit(p: float) -> float: + p = min(max(p, 1e-4), 1.0 - 1e-4) + return math.log(p / (1.0 - p)) + + +class SpikingMLP(nn.Module): + # Hybrid spike-gated relu^2 MLP: keep the strong dense pathway from the baseline + # and let a small LIF-style spike process gate the hidden activations. This keeps + # parameter count unchanged, improves contest practicality on GPUs, and still lets + # us test whether thresholded temporal dynamics help in the tiny-LM regime. + def __init__( + self, + dim: int, + mlp_mult: int, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + 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 + self.snn_steps = max(int(snn_steps), 1) + self.snn_grad_scale = float(snn_grad_scale) + self.decay_logit = nn.Parameter(torch.tensor(_logit(snn_decay), dtype=torch.float32)) + self.spike_thresh = nn.Parameter(torch.tensor(float(snn_thresh_init), dtype=torch.float32)) + self.spike_out_scale = nn.Parameter(torch.tensor(float(snn_out_scale_init), dtype=torch.float32)) + + def forward(self, x: Tensor) -> tuple[Tensor, Tensor]: + cur = self.fc(x) + dense = torch.relu(cur).square() + mem = torch.zeros_like(cur) + spike_sum = torch.zeros_like(cur) + decay = self.decay_logit.sigmoid().to(dtype=x.dtype) + thresh = self.spike_thresh.to(dtype=x.dtype).clamp_min(0.05) + grad_scale = torch.tensor(self.snn_grad_scale, dtype=x.dtype, device=x.device) + for _ in range(self.snn_steps): + mem = decay * mem + cur / self.snn_steps + over = mem - thresh + spike_soft = torch.sigmoid(grad_scale * over) + spike_hard = (over > 0).to(dtype=x.dtype) + spike = spike_hard + spike_soft - spike_soft.detach() + mem = mem - spike_hard * thresh + spike_sum = spike_sum + spike + rate = spike_sum / self.snn_steps + gate = 1.0 + self.spike_out_scale.to(dtype=x.dtype) * rate + out = self.proj(dense * gate) + spike_rate = rate.mean() + return out, spike_rate + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + use_snn_mlp: bool, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.use_snn_mlp = use_snn_mlp + self.mlp = ( + SpikingMLP(dim, mlp_mult, snn_steps, snn_decay, snn_thresh_init, snn_grad_scale, snn_out_scale_init) + if use_snn_mlp + else MLP(dim, mlp_mult) + ) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + if self.use_snn_mlp: + mlp_out, spike_rate = self.mlp(self.mlp_norm(x)) + else: + mlp_out = self.mlp(self.mlp_norm(x)) + spike_rate = x.new_zeros(()) + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * mlp_out + return x, spike_rate + + +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, + use_snn_mlp: bool, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + use_snn_mlp, + snn_steps, + snn_decay, + snn_thresh_init, + snn_grad_scale, + snn_out_scale_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor, return_aux: bool = False): + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + spike_rate_sum = x.new_zeros(()) + + # First half stores skips; second half reuses them in reverse order. + for i in range(self.num_encoder_layers): + x, block_spike_rate = self.blocks[i](x, x0) + spike_rate_sum = spike_rate_sum + block_spike_rate + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, block_spike_rate = self.blocks[self.num_encoder_layers + i](x, x0) + spike_rate_sum = spike_rate_sum + block_spike_rate + + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + ce_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if return_aux: + mean_spike_rate = spike_rate_sum / max(len(self.blocks), 1) + return ce_loss, mean_spike_rate + return ce_loss + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + use_snn_mlp=args.use_snn_mlp, + snn_steps=args.snn_steps, + snn_decay=args.snn_decay, + snn_thresh_init=args.snn_thresh_init, + snn_grad_scale=args.snn_grad_scale, + snn_out_scale_init=args.snn_out_scale_init, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0( + f"use_snn_mlp:{args.use_snn_mlp} snn_steps:{args.snn_steps} snn_decay:{args.snn_decay} " + f"snn_thresh_init:{args.snn_thresh_init} snn_rate_loss:{args.snn_rate_loss} " + f"snn_rate_target:{args.snn_rate_target}" + ) + 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, warmup_spike_rate = model(x, y, True) + if args.use_snn_mlp and args.snn_rate_loss > 0: + warmup_loss = warmup_loss + args.snn_rate_loss * (warmup_spike_rate - args.snn_rate_target).square() + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + spike_rate = 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, batch_spike_rate = model(x, y, True) + if args.use_snn_mlp and args.snn_rate_loss > 0: + loss = loss + args.snn_rate_loss * (batch_spike_rate - args.snn_rate_target).square() + train_loss += loss.detach() + spike_rate += batch_spike_rate.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + spike_rate /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"spike_rate:{spike_rate.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Tue Mar 24 21:55:39 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 | +| N/A 40C P0 152W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 35C P0 154W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 34C P0 152W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 37C P0 146W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 39C P0 150W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 34C P0 144W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 37C P0 150W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 33C P0 155W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:19292793 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:2 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +use_snn_mlp:True snn_steps:2 snn_decay:0.8 snn_thresh_init:1.0 snn_rate_loss:0.0001 snn_rate_target:0.15 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9365 val_bpb:4.1082 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9362 spike_rate:0.0136 train_time:49ms step_avg:48.68ms +step:2/20000 train_loss:16.5290 spike_rate:0.0204 train_time:99ms step_avg:49.32ms +step:3/20000 train_loss:8.3664 spike_rate:0.0297 train_time:184ms step_avg:61.39ms +step:4/20000 train_loss:6.5806 spike_rate:0.0400 train_time:244ms step_avg:61.01ms +step:5/20000 train_loss:6.6631 spike_rate:0.0498 train_time:301ms step_avg:60.11ms +step:6/20000 train_loss:7.4230 spike_rate:0.0537 train_time:382ms step_avg:63.68ms +step:7/20000 train_loss:6.3644 spike_rate:0.0564 train_time:442ms step_avg:63.08ms +step:8/20000 train_loss:6.1691 spike_rate:0.0586 train_time:502ms step_avg:62.80ms +step:9/20000 train_loss:6.0625 spike_rate:0.0618 train_time:561ms step_avg:62.36ms +step:10/20000 train_loss:5.9247 spike_rate:0.0669 train_time:620ms step_avg:61.98ms +step:50/20000 train_loss:4.0335 spike_rate:0.1406 train_time:2768ms step_avg:55.37ms +step:100/20000 train_loss:3.4417 spike_rate:0.1060 train_time:5462ms step_avg:54.62ms +step:150/20000 train_loss:3.1033 spike_rate:0.0688 train_time:8158ms step_avg:54.39ms +step:200/20000 train_loss:2.8798 spike_rate:0.0515 train_time:13177ms step_avg:65.89ms +step:200/20000 val_loss:2.8694 val_bpb:1.6994 train_time:13200ms step_avg:66.00ms +step:250/20000 train_loss:2.7660 spike_rate:0.0408 train_time:15878ms step_avg:63.51ms +step:300/20000 train_loss:2.4861 spike_rate:0.0347 train_time:18585ms step_avg:61.95ms +step:350/20000 train_loss:2.6753 spike_rate:0.0305 train_time:21292ms step_avg:60.83ms +step:400/20000 train_loss:2.3475 spike_rate:0.0277 train_time:26429ms step_avg:66.07ms +step:400/20000 val_loss:2.5557 val_bpb:1.5136 train_time:26451ms step_avg:66.13ms +step:450/20000 train_loss:2.5001 spike_rate:0.0225 train_time:29132ms step_avg:64.74ms +step:500/20000 train_loss:2.4872 spike_rate:0.0203 train_time:31833ms step_avg:63.67ms +step:550/20000 train_loss:2.3919 spike_rate:0.0168 train_time:34535ms step_avg:62.79ms +step:600/20000 train_loss:2.5327 spike_rate:0.0152 train_time:39519ms step_avg:65.87ms +step:600/20000 val_loss:2.4384 val_bpb:1.4441 train_time:39577ms step_avg:65.96ms +step:650/20000 train_loss:2.3768 spike_rate:0.0140 train_time:42259ms step_avg:65.01ms +step:700/20000 train_loss:2.4259 spike_rate:0.0129 train_time:44962ms step_avg:64.23ms +step:750/20000 train_loss:2.2644 spike_rate:0.0120 train_time:47672ms step_avg:63.56ms +step:800/20000 train_loss:2.2810 spike_rate:0.0110 train_time:53364ms step_avg:66.70ms +step:800/20000 val_loss:2.3697 val_bpb:1.4035 train_time:53386ms step_avg:66.73ms +step:850/20000 train_loss:2.7100 spike_rate:0.0101 train_time:57168ms step_avg:67.26ms +step:900/20000 train_loss:2.3260 spike_rate:0.0103 train_time:59868ms step_avg:66.52ms +step:950/20000 train_loss:2.3843 spike_rate:0.0095 train_time:62568ms step_avg:65.86ms +step:1000/20000 train_loss:2.3614 spike_rate:0.0093 train_time:68316ms step_avg:68.32ms +step:1000/20000 val_loss:2.3200 val_bpb:1.3740 train_time:68339ms step_avg:68.34ms +step:1050/20000 train_loss:2.4697 spike_rate:0.0093 train_time:71016ms step_avg:67.63ms +step:1100/20000 train_loss:2.2440 spike_rate:0.0095 train_time:73728ms step_avg:67.03ms +step:1150/20000 train_loss:2.2345 spike_rate:0.0092 train_time:79126ms step_avg:68.81ms +step:1200/20000 train_loss:2.3732 spike_rate:0.0090 train_time:81814ms step_avg:68.18ms +step:1200/20000 val_loss:2.2879 val_bpb:1.3550 train_time:81839ms step_avg:68.20ms +step:1250/20000 train_loss:2.1908 spike_rate:0.0087 train_time:84515ms step_avg:67.61ms +step:1300/20000 train_loss:2.3440 spike_rate:0.0084 train_time:87215ms step_avg:67.09ms +step:1350/20000 train_loss:2.2579 spike_rate:0.0084 train_time:91462ms step_avg:67.75ms +step:1400/20000 train_loss:2.4140 spike_rate:0.0082 train_time:94181ms step_avg:67.27ms +step:1400/20000 val_loss:2.2645 val_bpb:1.3412 train_time:94201ms step_avg:67.29ms +step:1450/20000 train_loss:2.2201 spike_rate:0.0080 train_time:97981ms step_avg:67.57ms +step:1500/20000 train_loss:2.2045 spike_rate:0.0082 train_time:100681ms step_avg:67.12ms +step:1550/20000 train_loss:2.1404 spike_rate:0.0081 train_time:106427ms step_avg:68.66ms +step:1600/20000 train_loss:2.0827 spike_rate:0.0081 train_time:109136ms step_avg:68.21ms +step:1600/20000 val_loss:2.2492 val_bpb:1.3321 train_time:109166ms step_avg:68.23ms +step:1650/20000 train_loss:2.2162 spike_rate:0.0079 train_time:111841ms step_avg:67.78ms +step:1700/20000 train_loss:2.1568 spike_rate:0.0079 train_time:114541ms step_avg:67.38ms +step:1750/20000 train_loss:2.2324 spike_rate:0.0075 train_time:120659ms step_avg:68.95ms +step:1800/20000 train_loss:2.1820 spike_rate:0.0078 train_time:123358ms step_avg:68.53ms +step:1800/20000 val_loss:2.2329 val_bpb:1.3224 train_time:123380ms step_avg:68.54ms +step:1850/20000 train_loss:2.2866 spike_rate:0.0074 train_time:127146ms step_avg:68.73ms +step:1900/20000 train_loss:2.1717 spike_rate:0.0079 train_time:129844ms step_avg:68.34ms +step:1950/20000 train_loss:2.1943 spike_rate:0.0076 train_time:136398ms step_avg:69.95ms +step:2000/20000 train_loss:2.2381 spike_rate:0.0078 train_time:139132ms step_avg:69.57ms +step:2000/20000 val_loss:2.2182 val_bpb:1.3138 train_time:139155ms step_avg:69.58ms +step:2050/20000 train_loss:2.2327 spike_rate:0.0075 train_time:141833ms step_avg:69.19ms +step:2100/20000 train_loss:2.2460 spike_rate:0.0074 train_time:147571ms step_avg:70.27ms +step:2150/20000 train_loss:2.1704 spike_rate:0.0076 train_time:150254ms step_avg:69.89ms +step:2200/20000 train_loss:2.0542 spike_rate:0.0076 train_time:152950ms step_avg:69.52ms +step:2200/20000 val_loss:2.2090 val_bpb:1.3083 train_time:152974ms step_avg:69.53ms +step:2250/20000 train_loss:2.1443 spike_rate:0.0073 train_time:155652ms step_avg:69.18ms +step:2300/20000 train_loss:2.3634 spike_rate:0.0074 train_time:162309ms step_avg:70.57ms +step:2350/20000 train_loss:2.1803 spike_rate:0.0073 train_time:165002ms step_avg:70.21ms +step:2400/20000 train_loss:2.1832 spike_rate:0.0072 train_time:167693ms step_avg:69.87ms +step:2400/20000 val_loss:2.1982 val_bpb:1.3019 train_time:167719ms step_avg:69.88ms +step:2450/20000 train_loss:2.1821 spike_rate:0.0071 train_time:171430ms step_avg:69.97ms +step:2500/20000 train_loss:2.1029 spike_rate:0.0073 train_time:177222ms step_avg:70.89ms +step:2550/20000 train_loss:2.1185 spike_rate:0.0074 train_time:179914ms step_avg:70.55ms +step:2600/20000 train_loss:2.3942 spike_rate:0.0071 train_time:182607ms step_avg:70.23ms +step:2600/20000 val_loss:2.2005 val_bpb:1.3032 train_time:182630ms step_avg:70.24ms +step:2650/20000 train_loss:2.2219 spike_rate:0.0072 train_time:185302ms step_avg:69.93ms +step:2700/20000 train_loss:2.1337 spike_rate:0.0073 train_time:191688ms step_avg:71.00ms +step:2750/20000 train_loss:2.3433 spike_rate:0.0070 train_time:194381ms step_avg:70.68ms +step:2800/20000 train_loss:2.2122 spike_rate:0.0071 train_time:197067ms step_avg:70.38ms +step:2800/20000 val_loss:2.1827 val_bpb:1.2927 train_time:197089ms step_avg:70.39ms +step:2850/20000 train_loss:2.1655 spike_rate:0.0072 train_time:199770ms step_avg:70.09ms +step:2900/20000 train_loss:2.1620 spike_rate:0.0072 train_time:205612ms step_avg:70.90ms +step:2950/20000 train_loss:2.2210 spike_rate:0.0072 train_time:208304ms step_avg:70.61ms +step:3000/20000 train_loss:2.2076 spike_rate:0.0070 train_time:210995ms step_avg:70.33ms +step:3000/20000 val_loss:2.1757 val_bpb:1.2886 train_time:211017ms step_avg:70.34ms +step:3050/20000 train_loss:2.1513 spike_rate:0.0065 train_time:213690ms step_avg:70.06ms +step:3100/20000 train_loss:2.1926 spike_rate:0.0068 train_time:218514ms step_avg:70.49ms +step:3150/20000 train_loss:2.1433 spike_rate:0.0064 train_time:221219ms step_avg:70.23ms +step:3200/20000 train_loss:2.1719 spike_rate:0.0064 train_time:223926ms step_avg:69.98ms +step:3200/20000 val_loss:2.1693 val_bpb:1.2848 train_time:223960ms step_avg:69.99ms +step:3250/20000 train_loss:2.0674 spike_rate:0.0065 train_time:229177ms step_avg:70.52ms +step:3300/20000 train_loss:2.2219 spike_rate:0.0063 train_time:231861ms step_avg:70.26ms +step:3350/20000 train_loss:2.0701 spike_rate:0.0063 train_time:234552ms step_avg:70.02ms +step:3400/20000 train_loss:2.1373 spike_rate:0.0063 train_time:237242ms step_avg:69.78ms +step:3400/20000 val_loss:2.1670 val_bpb:1.2834 train_time:237264ms step_avg:69.78ms +step:3450/20000 train_loss:2.0904 spike_rate:0.0062 train_time:242399ms step_avg:70.26ms +step:3500/20000 train_loss:2.2286 spike_rate:0.0062 train_time:245093ms step_avg:70.03ms +step:3550/20000 train_loss:2.3677 spike_rate:0.0062 train_time:247780ms step_avg:69.80ms +step:3600/20000 train_loss:2.0987 spike_rate:0.0062 train_time:250468ms step_avg:69.57ms +step:3600/20000 val_loss:2.1584 val_bpb:1.2783 train_time:250497ms step_avg:69.58ms +step:3650/20000 train_loss:2.1982 spike_rate:0.0060 train_time:256147ms step_avg:70.18ms +step:3700/20000 train_loss:2.1289 spike_rate:0.0062 train_time:258831ms step_avg:69.95ms +step:3750/20000 train_loss:2.1227 spike_rate:0.0060 train_time:261529ms step_avg:69.74ms +step:3800/20000 train_loss:2.2028 spike_rate:0.0062 train_time:264221ms step_avg:69.53ms +step:3800/20000 val_loss:2.1542 val_bpb:1.2759 train_time:264258ms step_avg:69.54ms +step:3850/20000 train_loss:2.1579 spike_rate:0.0062 train_time:270056ms step_avg:70.14ms +step:3900/20000 train_loss:1.9720 spike_rate:0.0061 train_time:272739ms step_avg:69.93ms +step:3950/20000 train_loss:2.1093 spike_rate:0.0063 train_time:275419ms step_avg:69.73ms +step:4000/20000 train_loss:2.1448 spike_rate:0.0062 train_time:278104ms step_avg:69.53ms +step:4000/20000 val_loss:2.1498 val_bpb:1.2732 train_time:278133ms step_avg:69.53ms +step:4050/20000 train_loss:2.0804 spike_rate:0.0064 train_time:282764ms step_avg:69.82ms +step:4100/20000 train_loss:2.1650 spike_rate:0.0060 train_time:285463ms step_avg:69.63ms +step:4150/20000 train_loss:2.3035 spike_rate:0.0059 train_time:288161ms step_avg:69.44ms +step:4200/20000 train_loss:2.1570 spike_rate:0.0060 train_time:293901ms step_avg:69.98ms +step:4200/20000 val_loss:2.1459 val_bpb:1.2709 train_time:293924ms step_avg:69.98ms +step:4250/20000 train_loss:2.1068 spike_rate:0.0062 train_time:296588ms step_avg:69.79ms +step:4300/20000 train_loss:2.0073 spike_rate:0.0060 train_time:299273ms step_avg:69.60ms +step:4350/20000 train_loss:2.1867 spike_rate:0.0061 train_time:301955ms step_avg:69.41ms +step:4400/20000 train_loss:2.0920 spike_rate:0.0062 train_time:308067ms step_avg:70.02ms +step:4400/20000 val_loss:2.1450 val_bpb:1.2704 train_time:308095ms step_avg:70.02ms +step:4450/20000 train_loss:2.0451 spike_rate:0.0059 train_time:311883ms step_avg:70.09ms +step:4500/20000 train_loss:2.2371 spike_rate:0.0060 train_time:314561ms step_avg:69.90ms +step:4550/20000 train_loss:2.0356 spike_rate:0.0060 train_time:317241ms step_avg:69.72ms +step:4600/20000 train_loss:1.9496 spike_rate:0.0063 train_time:324596ms step_avg:70.56ms +step:4600/20000 val_loss:2.1402 val_bpb:1.2676 train_time:324619ms step_avg:70.57ms +step:4650/20000 train_loss:2.0547 spike_rate:0.0060 train_time:327276ms step_avg:70.38ms +step:4700/20000 train_loss:2.2466 spike_rate:0.0060 train_time:329958ms step_avg:70.20ms +step:4750/20000 train_loss:1.9540 spike_rate:0.0063 train_time:332639ms step_avg:70.03ms +step:4800/20000 train_loss:2.2465 spike_rate:0.0058 train_time:337950ms step_avg:70.41ms +step:4800/20000 val_loss:2.1368 val_bpb:1.2655 train_time:337974ms step_avg:70.41ms +step:4850/20000 train_loss:2.1344 spike_rate:0.0061 train_time:341675ms step_avg:70.45ms +step:4900/20000 train_loss:2.1400 spike_rate:0.0060 train_time:344357ms step_avg:70.28ms +step:4950/20000 train_loss:2.3241 spike_rate:0.0058 train_time:347048ms step_avg:70.11ms +step:5000/20000 train_loss:2.0077 spike_rate:0.0062 train_time:353110ms step_avg:70.62ms +step:5000/20000 val_loss:2.1326 val_bpb:1.2630 train_time:353132ms step_avg:70.63ms +step:5050/20000 train_loss:2.1828 spike_rate:0.0061 train_time:355791ms step_avg:70.45ms +step:5100/20000 train_loss:2.0041 spike_rate:0.0062 train_time:358490ms step_avg:70.29ms +step:5150/20000 train_loss:2.2596 spike_rate:0.0060 train_time:363922ms step_avg:70.66ms +step:5200/20000 train_loss:2.1561 spike_rate:0.0060 train_time:366595ms step_avg:70.50ms +step:5200/20000 val_loss:2.1322 val_bpb:1.2628 train_time:366635ms step_avg:70.51ms +step:5250/20000 train_loss:2.1015 spike_rate:0.0060 train_time:370462ms step_avg:70.56ms +step:5300/20000 train_loss:2.1880 spike_rate:0.0060 train_time:373144ms step_avg:70.40ms +step:5350/20000 train_loss:2.1165 spike_rate:0.0060 train_time:378668ms step_avg:70.78ms +step:5400/20000 train_loss:2.1632 spike_rate:0.0060 train_time:381365ms step_avg:70.62ms +step:5400/20000 val_loss:2.1275 val_bpb:1.2600 train_time:381406ms step_avg:70.63ms +step:5450/20000 train_loss:2.1778 spike_rate:0.0061 train_time:384070ms step_avg:70.47ms +step:5500/20000 train_loss:2.1136 spike_rate:0.0063 train_time:386755ms step_avg:70.32ms +step:5550/20000 train_loss:2.0864 spike_rate:0.0061 train_time:392082ms step_avg:70.65ms +step:5600/20000 train_loss:2.1636 spike_rate:0.0060 train_time:394764ms step_avg:70.49ms +step:5600/20000 val_loss:2.1276 val_bpb:1.2601 train_time:394790ms step_avg:70.50ms +step:5650/20000 train_loss:2.0353 spike_rate:0.0060 train_time:397450ms step_avg:70.35ms +step:5700/20000 train_loss:2.1562 spike_rate:0.0060 train_time:400134ms step_avg:70.20ms +step:5750/20000 train_loss:2.1972 spike_rate:0.0060 train_time:405477ms step_avg:70.52ms +step:5800/20000 train_loss:2.1191 spike_rate:0.0061 train_time:408151ms step_avg:70.37ms +step:5800/20000 val_loss:2.1252 val_bpb:1.2587 train_time:408174ms step_avg:70.37ms +step:5850/20000 train_loss:2.1604 spike_rate:0.0061 train_time:410833ms step_avg:70.23ms +step:5900/20000 train_loss:2.0724 spike_rate:0.0060 train_time:413515ms step_avg:70.09ms +step:5950/20000 train_loss:2.1142 spike_rate:0.0061 train_time:419044ms step_avg:70.43ms +step:6000/20000 train_loss:2.1994 spike_rate:0.0058 train_time:421724ms step_avg:70.29ms +step:6000/20000 val_loss:2.1219 val_bpb:1.2567 train_time:421747ms step_avg:70.29ms +step:6050/20000 train_loss:2.1045 spike_rate:0.0059 train_time:424406ms step_avg:70.15ms +step:6100/20000 train_loss:2.0984 spike_rate:0.0060 train_time:427093ms step_avg:70.02ms +step:6150/20000 train_loss:2.0788 spike_rate:0.0061 train_time:432814ms step_avg:70.38ms +step:6200/20000 train_loss:2.0683 spike_rate:0.0060 train_time:435497ms step_avg:70.24ms +step:6200/20000 val_loss:2.1206 val_bpb:1.2559 train_time:435519ms step_avg:70.25ms +step:6250/20000 train_loss:2.1327 spike_rate:0.0059 train_time:438178ms step_avg:70.11ms +step:6300/20000 train_loss:2.0156 spike_rate:0.0060 train_time:443293ms step_avg:70.36ms +step:6350/20000 train_loss:1.9978 spike_rate:0.0059 train_time:445968ms step_avg:70.23ms +step:6400/20000 train_loss:2.1378 spike_rate:0.0060 train_time:448649ms step_avg:70.10ms +step:6400/20000 val_loss:2.1173 val_bpb:1.2540 train_time:448671ms step_avg:70.10ms +step:6450/20000 train_loss:2.0565 spike_rate:0.0061 train_time:451332ms step_avg:69.97ms +step:6500/20000 train_loss:2.0606 spike_rate:0.0061 train_time:456983ms step_avg:70.31ms +step:6550/20000 train_loss:2.1902 spike_rate:0.0060 train_time:459663ms step_avg:70.18ms +step:6600/20000 train_loss:2.1017 spike_rate:0.0061 train_time:462346ms step_avg:70.05ms +step:6600/20000 val_loss:2.1138 val_bpb:1.2519 train_time:462368ms step_avg:70.06ms +step:6650/20000 train_loss:2.2697 spike_rate:0.0056 train_time:465026ms step_avg:69.93ms +step:6700/20000 train_loss:2.1362 spike_rate:0.0059 train_time:469636ms step_avg:70.09ms +step:6750/20000 train_loss:2.3042 spike_rate:0.0059 train_time:472311ms step_avg:69.97ms +step:6800/20000 train_loss:2.1706 spike_rate:0.0060 train_time:474990ms step_avg:69.85ms +step:6800/20000 val_loss:2.1122 val_bpb:1.2510 train_time:475012ms step_avg:69.85ms +step:6850/20000 train_loss:1.9995 spike_rate:0.0061 train_time:477671ms step_avg:69.73ms +step:6900/20000 train_loss:2.0745 spike_rate:0.0062 train_time:482888ms step_avg:69.98ms +step:6950/20000 train_loss:2.1538 spike_rate:0.0060 train_time:485569ms step_avg:69.87ms +step:7000/20000 train_loss:2.2045 spike_rate:0.0060 train_time:488249ms step_avg:69.75ms +step:7000/20000 val_loss:2.1102 val_bpb:1.2498 train_time:488270ms step_avg:69.75ms +step:7050/20000 train_loss:2.2310 spike_rate:0.0060 train_time:490930ms step_avg:69.64ms +step:7100/20000 train_loss:2.0494 spike_rate:0.0060 train_time:495772ms step_avg:69.83ms +step:7150/20000 train_loss:2.1321 spike_rate:0.0060 train_time:498451ms step_avg:69.71ms +step:7200/20000 train_loss:2.1760 spike_rate:0.0061 train_time:501126ms step_avg:69.60ms +step:7200/20000 val_loss:2.1110 val_bpb:1.2502 train_time:501149ms step_avg:69.60ms +step:7250/20000 train_loss:2.0815 spike_rate:0.0061 train_time:505887ms step_avg:69.78ms +step:7300/20000 train_loss:2.0681 spike_rate:0.0058 train_time:508558ms step_avg:69.67ms +step:7350/20000 train_loss:2.1600 spike_rate:0.0060 train_time:511249ms step_avg:69.56ms +step:7400/20000 train_loss:2.0971 spike_rate:0.0060 train_time:513952ms step_avg:69.45ms +step:7400/20000 val_loss:2.1069 val_bpb:1.2478 train_time:513990ms step_avg:69.46ms +step:7450/20000 train_loss:2.0901 spike_rate:0.0061 train_time:519552ms step_avg:69.74ms +step:7500/20000 train_loss:2.0859 spike_rate:0.0060 train_time:522231ms step_avg:69.63ms +step:7550/20000 train_loss:2.1397 spike_rate:0.0056 train_time:524921ms step_avg:69.53ms +step:7600/20000 train_loss:1.9690 spike_rate:0.0059 train_time:527624ms step_avg:69.42ms +step:7600/20000 val_loss:2.1007 val_bpb:1.2442 train_time:527661ms step_avg:69.43ms +step:7650/20000 train_loss:2.2596 spike_rate:0.0057 train_time:533455ms step_avg:69.73ms +step:7700/20000 train_loss:2.0608 spike_rate:0.0053 train_time:536134ms step_avg:69.63ms +step:7750/20000 train_loss:2.0771 spike_rate:0.0053 train_time:538813ms step_avg:69.52ms +step:7800/20000 train_loss:2.1166 spike_rate:0.0051 train_time:541506ms step_avg:69.42ms +step:7800/20000 val_loss:2.0899 val_bpb:1.2378 train_time:541536ms step_avg:69.43ms +step:7850/20000 train_loss:1.9615 spike_rate:0.0051 train_time:546414ms step_avg:69.61ms +step:7900/20000 train_loss:2.0919 spike_rate:0.0053 train_time:549101ms step_avg:69.51ms +step:7950/20000 train_loss:2.0570 spike_rate:0.0051 train_time:551792ms step_avg:69.41ms +step:8000/20000 train_loss:2.0739 spike_rate:0.0050 train_time:554481ms step_avg:69.31ms +step:8000/20000 val_loss:2.0808 val_bpb:1.2324 train_time:554504ms step_avg:69.31ms +step:8050/20000 train_loss:2.0389 spike_rate:0.0051 train_time:562248ms step_avg:69.84ms +step:8100/20000 train_loss:2.1053 spike_rate:0.0050 train_time:564932ms step_avg:69.74ms +step:8150/20000 train_loss:2.2133 spike_rate:0.0051 train_time:567637ms step_avg:69.65ms +step:8200/20000 train_loss:2.1415 spike_rate:0.0050 train_time:570325ms step_avg:69.55ms +step:8200/20000 val_loss:2.0711 val_bpb:1.2266 train_time:570362ms step_avg:69.56ms +step:8250/20000 train_loss:2.0979 spike_rate:0.0051 train_time:576025ms step_avg:69.82ms +step:8300/20000 train_loss:2.0687 spike_rate:0.0050 train_time:578705ms step_avg:69.72ms +step:8350/20000 train_loss:2.1741 spike_rate:0.0049 train_time:581396ms step_avg:69.63ms +step:8400/20000 train_loss:2.0849 spike_rate:0.0049 train_time:586766ms step_avg:69.85ms +step:8400/20000 val_loss:2.0623 val_bpb:1.2214 train_time:586787ms step_avg:69.86ms +step:8450/20000 train_loss:2.1732 spike_rate:0.0052 train_time:589444ms step_avg:69.76ms +step:8500/20000 train_loss:2.0656 spike_rate:0.0050 train_time:592126ms step_avg:69.66ms +step:8550/20000 train_loss:2.1379 spike_rate:0.0049 train_time:594805ms step_avg:69.57ms +step:8583/20000 val_loss:2.0563 val_bpb:1.2179 train_time:599922ms step_avg:69.90ms +stopping_early: wallclock_cap train_time:599922ms step:8583/20000 +peak memory allocated: 14926 MiB reserved: 15058 MiB +Serialized model: 76173625 bytes +Code size: 53034 bytes +Total submission size: 76226659 bytes +Serialized model int8+zlib: 19196077 bytes (payload:20868580 raw_torch:20926339 payload_ratio:3.65x) +Total submission size int8+zlib: 19249111 bytes +final_int8_zlib_roundtrip val_loss:2.0685 val_bpb:1.2251 eval_time:1712ms +final_int8_zlib_roundtrip_exact val_loss:2.06848443 val_bpb:1.22507299 +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + use_snn_mlp = bool(int(os.environ.get("USE_SNN_MLP", "1"))) + snn_steps = int(os.environ.get("SNN_STEPS", 2)) + snn_decay = float(os.environ.get("SNN_DECAY", 0.8)) + snn_thresh_init = float(os.environ.get("SNN_THRESH_INIT", 1.0)) + snn_grad_scale = float(os.environ.get("SNN_GRAD_SCALE", 4.0)) + snn_out_scale_init = float(os.environ.get("SNN_OUT_SCALE_INIT", 1.0)) + snn_rate_loss = float(os.environ.get("SNN_RATE_LOSS", 1e-4)) + snn_rate_target = float(os.environ.get("SNN_RATE_TARGET", 0.15)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Scale correction from Muon reference implementations. + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,decay_logit,spike_thresh,spike_out_scale", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +def _logit(p: float) -> float: + p = min(max(p, 1e-4), 1.0 - 1e-4) + return math.log(p / (1.0 - p)) + + +class SpikingMLP(nn.Module): + # Contest-friendly hybrid SNN block: keep attention dense, replace the MLP with + # a small multi-step leaky integrate-and-fire pathway using a sigmoid STE. + def __init__( + self, + dim: int, + mlp_mult: int, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + 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 + self.snn_steps = max(int(snn_steps), 1) + self.snn_grad_scale = float(snn_grad_scale) + self.decay_logit = nn.Parameter(torch.tensor(_logit(snn_decay), dtype=torch.float32)) + self.spike_thresh = nn.Parameter(torch.tensor(float(snn_thresh_init), dtype=torch.float32)) + self.spike_out_scale = nn.Parameter(torch.tensor(float(snn_out_scale_init), dtype=torch.float32)) + + def forward(self, x: Tensor) -> tuple[Tensor, Tensor]: + cur = self.fc(x) + mem = torch.zeros_like(cur) + spike_sum = torch.zeros_like(cur) + decay = self.decay_logit.sigmoid().to(dtype=x.dtype) + thresh = self.spike_thresh.to(dtype=x.dtype).clamp_min(0.05) + grad_scale = torch.tensor(self.snn_grad_scale, dtype=x.dtype, device=x.device) + for _ in range(self.snn_steps): + mem = decay * mem + cur / self.snn_steps + over = mem - thresh + spike_soft = torch.sigmoid(grad_scale * over) + spike_hard = (over > 0).to(dtype=x.dtype) + spike = spike_hard + spike_soft - spike_soft.detach() + mem = mem - spike_hard * thresh + spike_sum = spike_sum + spike + rate = spike_sum / self.snn_steps + out = self.proj(rate * self.spike_out_scale.to(dtype=x.dtype)) + spike_rate = rate.mean() + return out, spike_rate + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + use_snn_mlp: bool, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.use_snn_mlp = use_snn_mlp + self.mlp = ( + SpikingMLP(dim, mlp_mult, snn_steps, snn_decay, snn_thresh_init, snn_grad_scale, snn_out_scale_init) + if use_snn_mlp + else MLP(dim, mlp_mult) + ) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + if self.use_snn_mlp: + mlp_out, spike_rate = self.mlp(self.mlp_norm(x)) + else: + mlp_out = self.mlp(self.mlp_norm(x)) + spike_rate = x.new_zeros(()) + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * mlp_out + return x, spike_rate + + +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, + use_snn_mlp: bool, + snn_steps: int, + snn_decay: float, + snn_thresh_init: float, + snn_grad_scale: float, + snn_out_scale_init: float, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + use_snn_mlp, + snn_steps, + snn_decay, + snn_thresh_init, + snn_grad_scale, + snn_out_scale_init, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward(self, input_ids: Tensor, target_ids: Tensor, return_aux: bool = False): + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + spike_rate_sum = x.new_zeros(()) + + # First half stores skips; second half reuses them in reverse order. + for i in range(self.num_encoder_layers): + x, block_spike_rate = self.blocks[i](x, x0) + spike_rate_sum = spike_rate_sum + block_spike_rate + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, block_spike_rate = self.blocks[self.num_encoder_layers + i](x, x0) + spike_rate_sum = spike_rate_sum + block_spike_rate + + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + ce_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if return_aux: + mean_spike_rate = spike_rate_sum / max(len(self.blocks), 1) + return ce_loss, mean_spike_rate + return ce_loss + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + use_snn_mlp=args.use_snn_mlp, + snn_steps=args.snn_steps, + snn_decay=args.snn_decay, + snn_thresh_init=args.snn_thresh_init, + snn_grad_scale=args.snn_grad_scale, + snn_out_scale_init=args.snn_out_scale_init, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.Adam( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0( + f"use_snn_mlp:{args.use_snn_mlp} snn_steps:{args.snn_steps} snn_decay:{args.snn_decay} " + f"snn_thresh_init:{args.snn_thresh_init} snn_rate_loss:{args.snn_rate_loss} " + f"snn_rate_target:{args.snn_rate_target}" + ) + 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, warmup_spike_rate = model(x, y, True) + if args.use_snn_mlp and args.snn_rate_loss > 0: + warmup_loss = warmup_loss + args.snn_rate_loss * (warmup_spike_rate - args.snn_rate_target).square() + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + spike_rate = 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, batch_spike_rate = model(x, y, True) + if args.use_snn_mlp and args.snn_rate_loss > 0: + loss = loss + args.snn_rate_loss * (batch_spike_rate - args.snn_rate_target).square() + train_loss += loss.detach() + spike_rate += batch_spike_rate.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + spike_rate /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"spike_rate:{spike_rate.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() + +==================================================================================================== +Running Python 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] +Running PyTorch 2.9.1+cu128 +Tue Mar 24 22:10:50 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 42C P0 151W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 | +| N/A 35C P0 153W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 | +| N/A 34C P0 152W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 | +| N/A 39C P0 147W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 | +| N/A 40C P0 149W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 | +| N/A 34C P0 144W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 | +| N/A 39C P0 152W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 | +| N/A 33C P0 152W / 700W | 1521MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17059939 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +use_snn_mlp:True snn_steps:2 snn_decay:0.8 snn_thresh_init:1.0 snn_rate_loss:0.0001 snn_rate_target:0.15 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9357 val_bpb:4.1077 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9370 spike_rate:0.0135 train_time:43ms step_avg:43.15ms +step:2/20000 train_loss:17.0278 spike_rate:0.0203 train_time:86ms step_avg:43.16ms +step:3/20000 train_loss:9.0828 spike_rate:0.0277 train_time:139ms step_avg:46.20ms +step:4/20000 train_loss:6.6325 spike_rate:0.0364 train_time:188ms step_avg:47.11ms +step:5/20000 train_loss:6.4899 spike_rate:0.0464 train_time:241ms step_avg:48.11ms +step:6/20000 train_loss:7.3360 spike_rate:0.0537 train_time:292ms step_avg:48.72ms +step:7/20000 train_loss:6.3036 spike_rate:0.0608 train_time:344ms step_avg:49.14ms +step:8/20000 train_loss:6.1525 spike_rate:0.0659 train_time:399ms step_avg:49.87ms +step:9/20000 train_loss:6.0778 spike_rate:0.0708 train_time:449ms step_avg:49.92ms +step:10/20000 train_loss:5.9979 spike_rate:0.0747 train_time:501ms step_avg:50.14ms +step:50/20000 train_loss:4.1893 spike_rate:0.1621 train_time:2321ms step_avg:46.42ms +step:100/20000 train_loss:3.4330 spike_rate:0.1426 train_time:4624ms step_avg:46.24ms +step:150/20000 train_loss:3.0477 spike_rate:0.1289 train_time:6922ms step_avg:46.14ms +step:200/20000 train_loss:2.8435 spike_rate:0.1147 train_time:12101ms step_avg:60.51ms +step:200/20000 val_loss:2.8336 val_bpb:1.6782 train_time:12114ms step_avg:60.57ms +step:250/20000 train_loss:2.7546 spike_rate:0.1050 train_time:15507ms step_avg:62.03ms +step:300/20000 train_loss:2.5080 spike_rate:0.0967 train_time:17815ms step_avg:59.38ms +step:350/20000 train_loss:2.6994 spike_rate:0.0918 train_time:20132ms step_avg:57.52ms +step:400/20000 train_loss:2.3989 spike_rate:0.0879 train_time:25045ms step_avg:62.61ms +step:400/20000 val_loss:2.6032 val_bpb:1.5418 train_time:25065ms step_avg:62.66ms +step:450/20000 train_loss:2.5483 spike_rate:0.0811 train_time:27346ms step_avg:60.77ms +step:500/20000 train_loss:2.5451 spike_rate:0.0791 train_time:29658ms step_avg:59.32ms +step:550/20000 train_loss:2.4540 spike_rate:0.0742 train_time:31962ms step_avg:58.11ms +step:600/20000 train_loss:2.6060 spike_rate:0.0708 train_time:37291ms step_avg:62.15ms +step:600/20000 val_loss:2.5103 val_bpb:1.4868 train_time:37311ms step_avg:62.18ms +step:650/20000 train_loss:2.4548 spike_rate:0.0718 train_time:39598ms step_avg:60.92ms +step:700/20000 train_loss:2.5138 spike_rate:0.0693 train_time:41902ms step_avg:59.86ms +step:750/20000 train_loss:2.3485 spike_rate:0.0684 train_time:44208ms step_avg:58.94ms +step:800/20000 train_loss:2.3724 spike_rate:0.0659 train_time:49289ms step_avg:61.61ms +step:800/20000 val_loss:2.4575 val_bpb:1.4555 train_time:49309ms step_avg:61.64ms +step:850/20000 train_loss:2.7934 spike_rate:0.0613 train_time:52698ms step_avg:62.00ms +step:900/20000 train_loss:2.4254 spike_rate:0.0640 train_time:55006ms step_avg:61.12ms +step:950/20000 train_loss:2.4846 spike_rate:0.0620 train_time:57321ms step_avg:60.34ms +step:1000/20000 train_loss:2.4557 spike_rate:0.0608 train_time:63359ms step_avg:63.36ms +step:1000/20000 val_loss:2.4196 val_bpb:1.4330 train_time:63381ms step_avg:63.38ms +step:1050/20000 train_loss:2.5695 spike_rate:0.0630 train_time:65666ms step_avg:62.54ms +step:1100/20000 train_loss:2.3514 spike_rate:0.0625 train_time:67972ms step_avg:61.79ms +step:1150/20000 train_loss:2.3421 spike_rate:0.0596 train_time:72955ms step_avg:63.44ms +step:1200/20000 train_loss:2.4777 spike_rate:0.0603 train_time:75244ms step_avg:62.70ms +step:1200/20000 val_loss:2.3953 val_bpb:1.4186 train_time:75265ms step_avg:62.72ms +step:1250/20000 train_loss:2.2980 spike_rate:0.0598 train_time:78685ms step_avg:62.95ms +step:1300/20000 train_loss:2.4522 spike_rate:0.0581 train_time:80988ms step_avg:62.30ms +step:1350/20000 train_loss:2.3681 spike_rate:0.0576 train_time:85337ms step_avg:63.21ms +step:1400/20000 train_loss:2.5235 spike_rate:0.0576 train_time:87640ms step_avg:62.60ms +step:1400/20000 val_loss:2.3792 val_bpb:1.4091 train_time:87661ms step_avg:62.61ms +step:1450/20000 train_loss:2.3357 spike_rate:0.0559 train_time:89945ms step_avg:62.03ms +step:1500/20000 train_loss:2.3332 spike_rate:0.0576 train_time:92266ms step_avg:61.51ms +step:1550/20000 train_loss:2.2511 spike_rate:0.0581 train_time:97578ms step_avg:62.95ms +step:1600/20000 train_loss:2.1939 spike_rate:0.0574 train_time:99882ms step_avg:62.43ms +step:1600/20000 val_loss:2.3641 val_bpb:1.4002 train_time:99903ms step_avg:62.44ms +step:1650/20000 train_loss:2.3309 spike_rate:0.0564 train_time:104236ms step_avg:63.17ms +step:1700/20000 train_loss:2.2679 spike_rate:0.0574 train_time:106540ms step_avg:62.67ms +step:1750/20000 train_loss:2.3443 spike_rate:0.0554 train_time:112495ms step_avg:64.28ms +step:1800/20000 train_loss:2.3024 spike_rate:0.0581 train_time:114795ms step_avg:63.78ms +step:1800/20000 val_loss:2.3520 val_bpb:1.3930 train_time:114818ms step_avg:63.79ms +step:1850/20000 train_loss:2.4035 spike_rate:0.0554 train_time:118253ms step_avg:63.92ms +step:1900/20000 train_loss:2.2979 spike_rate:0.0574 train_time:120558ms step_avg:63.45ms +step:1950/20000 train_loss:2.3089 spike_rate:0.0564 train_time:126608ms step_avg:64.93ms +step:2000/20000 train_loss:2.3533 spike_rate:0.0564 train_time:128910ms step_avg:64.45ms +step:2000/20000 val_loss:2.3380 val_bpb:1.3847 train_time:128930ms step_avg:64.47ms +step:2050/20000 train_loss:2.3540 spike_rate:0.0564 train_time:132315ms step_avg:64.54ms +step:2100/20000 train_loss:2.3644 spike_rate:0.0549 train_time:138006ms step_avg:65.72ms +step:2150/20000 train_loss:2.2833 spike_rate:0.0574 train_time:140315ms step_avg:65.26ms +step:2200/20000 train_loss:2.1883 spike_rate:0.0554 train_time:142635ms step_avg:64.83ms +step:2200/20000 val_loss:2.3280 val_bpb:1.3788 train_time:142658ms step_avg:64.84ms +step:2250/20000 train_loss:2.2629 spike_rate:0.0552 train_time:144942ms step_avg:64.42ms +step:2300/20000 train_loss:2.4747 spike_rate:0.0554 train_time:150319ms step_avg:65.36ms +step:2350/20000 train_loss:2.2980 spike_rate:0.0557 train_time:152621ms step_avg:64.95ms +step:2400/20000 train_loss:2.2990 spike_rate:0.0564 train_time:154923ms step_avg:64.55ms +step:2400/20000 val_loss:2.3182 val_bpb:1.3730 train_time:154950ms step_avg:64.56ms +step:2450/20000 train_loss:2.3027 spike_rate:0.0547 train_time:159385ms step_avg:65.06ms +step:2500/20000 train_loss:2.2293 spike_rate:0.0547 train_time:164241ms step_avg:65.70ms +step:2550/20000 train_loss:2.2346 spike_rate:0.0557 train_time:166548ms step_avg:65.31ms +step:2600/20000 train_loss:2.5052 spike_rate:0.0557 train_time:168859ms step_avg:64.95ms +step:2600/20000 val_loss:2.3185 val_bpb:1.3732 train_time:168879ms step_avg:64.95ms +step:2650/20000 train_loss:2.3461 spike_rate:0.0559 train_time:172296ms step_avg:65.02ms +step:2700/20000 train_loss:2.2510 spike_rate:0.0544 train_time:179350ms step_avg:66.43ms +step:2750/20000 train_loss:2.4635 spike_rate:0.0559 train_time:181651ms step_avg:66.05ms +step:2800/20000 train_loss:2.3401 spike_rate:0.0559 train_time:183953ms step_avg:65.70ms +step:2800/20000 val_loss:2.3018 val_bpb:1.3632 train_time:183974ms step_avg:65.71ms +step:2850/20000 train_loss:2.2909 spike_rate:0.0569 train_time:186262ms step_avg:65.36ms +step:2900/20000 train_loss:2.2831 spike_rate:0.0554 train_time:192444ms step_avg:66.36ms +step:2950/20000 train_loss:2.3291 spike_rate:0.0559 train_time:194745ms step_avg:66.02ms +step:3000/20000 train_loss:2.3280 spike_rate:0.0569 train_time:197049ms step_avg:65.68ms +step:3000/20000 val_loss:2.2959 val_bpb:1.3598 train_time:197070ms step_avg:65.69ms +step:3050/20000 train_loss:2.2705 spike_rate:0.0544 train_time:200535ms step_avg:65.75ms +step:3100/20000 train_loss:2.3021 spike_rate:0.0576 train_time:205567ms step_avg:66.31ms +step:3150/20000 train_loss:2.2597 spike_rate:0.0564 train_time:207866ms step_avg:65.99ms +step:3200/20000 train_loss:2.2873 spike_rate:0.0557 train_time:210168ms step_avg:65.68ms +step:3200/20000 val_loss:2.2888 val_bpb:1.3556 train_time:210188ms step_avg:65.68ms +step:3250/20000 train_loss:2.1887 spike_rate:0.0549 train_time:215420ms step_avg:66.28ms +step:3300/20000 train_loss:2.3396 spike_rate:0.0554 train_time:217716ms step_avg:65.97ms +step:3350/20000 train_loss:2.1888 spike_rate:0.0564 train_time:220021ms step_avg:65.68ms +step:3400/20000 train_loss:2.2545 spike_rate:0.0554 train_time:222324ms step_avg:65.39ms +step:3400/20000 val_loss:2.2846 val_bpb:1.3531 train_time:222345ms step_avg:65.40ms +step:3450/20000 train_loss:2.2012 spike_rate:0.0547 train_time:228689ms step_avg:66.29ms +step:3500/20000 train_loss:2.3479 spike_rate:0.0569 train_time:230986ms step_avg:66.00ms +step:3550/20000 train_loss:2.4871 spike_rate:0.0581 train_time:233287ms step_avg:65.71ms +step:3600/20000 train_loss:2.2163 spike_rate:0.0549 train_time:235588ms step_avg:65.44ms +step:3600/20000 val_loss:2.2768 val_bpb:1.3485 train_time:235608ms step_avg:65.45ms +step:3650/20000 train_loss:2.3120 spike_rate:0.0552 train_time:240393ms step_avg:65.86ms +step:3700/20000 train_loss:2.2529 spike_rate:0.0569 train_time:242694ms step_avg:65.59ms +step:3750/20000 train_loss:2.2416 spike_rate:0.0549 train_time:245005ms step_avg:65.33ms +step:3800/20000 train_loss:2.3194 spike_rate:0.0559 train_time:247324ms step_avg:65.09ms +step:3800/20000 val_loss:2.2729 val_bpb:1.3461 train_time:247344ms step_avg:65.09ms +step:3850/20000 train_loss:2.2765 spike_rate:0.0576 train_time:252928ms step_avg:65.70ms +step:3900/20000 train_loss:2.0924 spike_rate:0.0544 train_time:255226ms step_avg:65.44ms +step:3950/20000 train_loss:2.2217 spike_rate:0.0576 train_time:257527ms step_avg:65.20ms +step:4000/20000 train_loss:2.2582 spike_rate:0.0564 train_time:259854ms step_avg:64.96ms +step:4000/20000 val_loss:2.2676 val_bpb:1.3430 train_time:259882ms step_avg:64.97ms +step:4050/20000 train_loss:2.2024 spike_rate:0.0581 train_time:265270ms step_avg:65.50ms +step:4100/20000 train_loss:2.2817 spike_rate:0.0569 train_time:267566ms step_avg:65.26ms +step:4150/20000 train_loss:2.4202 spike_rate:0.0557 train_time:269866ms step_avg:65.03ms +step:4200/20000 train_loss:2.2691 spike_rate:0.0564 train_time:274853ms step_avg:65.44ms +step:4200/20000 val_loss:2.2633 val_bpb:1.3404 train_time:274860ms step_avg:65.44ms +step:4250/20000 train_loss:2.2264 spike_rate:0.0564 train_time:277141ms step_avg:65.21ms +step:4300/20000 train_loss:2.1142 spike_rate:0.0586 train_time:279443ms step_avg:64.99ms +step:4350/20000 train_loss:2.3117 spike_rate:0.0574 train_time:281772ms step_avg:64.78ms +step:4400/20000 train_loss:2.2096 spike_rate:0.0564 train_time:287930ms step_avg:65.44ms +step:4400/20000 val_loss:2.2636 val_bpb:1.3406 train_time:287960ms step_avg:65.45ms +step:4450/20000 train_loss:2.1599 spike_rate:0.0569 train_time:290241ms step_avg:65.22ms +step:4500/20000 train_loss:2.3465 spike_rate:0.0569 train_time:292539ms step_avg:65.01ms +step:4550/20000 train_loss:2.1605 spike_rate:0.0574 train_time:294835ms step_avg:64.80ms +step:4600/20000 train_loss:2.0751 spike_rate:0.0576 train_time:301476ms step_avg:65.54ms +step:4600/20000 val_loss:2.2567 val_bpb:1.3366 train_time:301507ms step_avg:65.55ms +step:4650/20000 train_loss:2.1701 spike_rate:0.0569 train_time:305925ms step_avg:65.79ms +step:4700/20000 train_loss:2.3715 spike_rate:0.0576 train_time:308231ms step_avg:65.58ms +step:4750/20000 train_loss:2.0798 spike_rate:0.0581 train_time:310542ms step_avg:65.38ms +step:4800/20000 train_loss:2.3572 spike_rate:0.0569 train_time:315467ms step_avg:65.72ms +step:4800/20000 val_loss:2.2544 val_bpb:1.3352 train_time:315488ms step_avg:65.73ms +step:4850/20000 train_loss:2.2484 spike_rate:0.0581 train_time:318854ms step_avg:65.74ms +step:4900/20000 train_loss:2.2592 spike_rate:0.0576 train_time:321161ms step_avg:65.54ms +step:4950/20000 train_loss:2.4389 spike_rate:0.0576 train_time:323465ms step_avg:65.35ms +step:5000/20000 train_loss:2.1261 spike_rate:0.0576 train_time:328762ms step_avg:65.75ms +step:5000/20000 val_loss:2.2491 val_bpb:1.3321 train_time:328798ms step_avg:65.76ms +step:5050/20000 train_loss:2.3071 spike_rate:0.0591 train_time:333260ms step_avg:65.99ms +step:5100/20000 train_loss:2.1266 spike_rate:0.0581 train_time:335558ms step_avg:65.80ms +step:5150/20000 train_loss:2.3722 spike_rate:0.0586 train_time:340887ms step_avg:66.19ms +step:5200/20000 train_loss:2.2646 spike_rate:0.0576 train_time:343199ms step_avg:66.00ms +step:5200/20000 val_loss:2.2497 val_bpb:1.3324 train_time:343231ms step_avg:66.01ms +step:5250/20000 train_loss:2.2135 spike_rate:0.0581 train_time:345512ms step_avg:65.81ms +step:5300/20000 train_loss:2.3028 spike_rate:0.0581 train_time:347816ms step_avg:65.63ms +step:5350/20000 train_loss:2.2341 spike_rate:0.0598 train_time:352601ms step_avg:65.91ms +step:5400/20000 train_loss:2.2806 spike_rate:0.0576 train_time:354898ms step_avg:65.72ms +step:5400/20000 val_loss:2.2444 val_bpb:1.3293 train_time:354917ms step_avg:65.73ms +step:5450/20000 train_loss:2.2986 spike_rate:0.0581 train_time:358269ms step_avg:65.74ms +step:5500/20000 train_loss:2.2335 spike_rate:0.0576 train_time:360569ms step_avg:65.56ms +step:5550/20000 train_loss:2.2074 spike_rate:0.0576 train_time:365517ms step_avg:65.86ms +step:5600/20000 train_loss:2.2791 spike_rate:0.0574 train_time:367814ms step_avg:65.68ms +step:5600/20000 val_loss:2.2433 val_bpb:1.3286 train_time:367836ms step_avg:65.69ms +step:5650/20000 train_loss:2.1501 spike_rate:0.0598 train_time:371206ms step_avg:65.70ms +step:5700/20000 train_loss:2.2705 spike_rate:0.0581 train_time:373511ms step_avg:65.53ms +step:5750/20000 train_loss:2.3120 spike_rate:0.0576 train_time:378921ms step_avg:65.90ms +step:5800/20000 train_loss:2.2409 spike_rate:0.0581 train_time:381221ms step_avg:65.73ms +step:5800/20000 val_loss:2.2414 val_bpb:1.3275 train_time:381241ms step_avg:65.73ms +step:5850/20000 train_loss:2.2762 spike_rate:0.0586 train_time:383520ms step_avg:65.56ms +step:5900/20000 train_loss:2.1824 spike_rate:0.0596 train_time:385831ms step_avg:65.40ms +step:5950/20000 train_loss:2.2325 spike_rate:0.0591 train_time:390817ms step_avg:65.68ms +step:6000/20000 train_loss:2.3061 spike_rate:0.0576 train_time:393117ms step_avg:65.52ms +step:6000/20000 val_loss:2.2384 val_bpb:1.3257 train_time:393137ms step_avg:65.52ms +step:6050/20000 train_loss:2.2040 spike_rate:0.0586 train_time:395420ms step_avg:65.36ms +step:6100/20000 train_loss:2.2136 spike_rate:0.0596 train_time:397724ms step_avg:65.20ms +step:6150/20000 train_loss:2.1885 spike_rate:0.0586 train_time:402784ms step_avg:65.49ms +step:6200/20000 train_loss:2.1843 spike_rate:0.0591 train_time:405090ms step_avg:65.34ms +step:6200/20000 val_loss:2.2369 val_bpb:1.3248 train_time:405112ms step_avg:65.34ms +step:6250/20000 train_loss:2.2463 spike_rate:0.0586 train_time:408573ms step_avg:65.37ms +step:6300/20000 train_loss:2.1266 spike_rate:0.0581 train_time:413759ms step_avg:65.68ms +step:6350/20000 train_loss:2.1185 spike_rate:0.0591 train_time:416053ms step_avg:65.52ms +step:6400/20000 train_loss:2.2584 spike_rate:0.0598 train_time:418346ms step_avg:65.37ms +step:6400/20000 val_loss:2.2341 val_bpb:1.3232 train_time:418374ms step_avg:65.37ms +step:6450/20000 train_loss:2.1747 spike_rate:0.0603 train_time:421772ms step_avg:65.39ms +step:6500/20000 train_loss:2.1758 spike_rate:0.0603 train_time:427221ms step_avg:65.73ms +step:6550/20000 train_loss:2.3069 spike_rate:0.0598 train_time:429517ms step_avg:65.58ms +step:6600/20000 train_loss:2.2243 spike_rate:0.0586 train_time:431819ms step_avg:65.43ms +step:6600/20000 val_loss:2.2301 val_bpb:1.3208 train_time:431840ms step_avg:65.43ms +step:6650/20000 train_loss:2.3793 spike_rate:0.0596 train_time:435226ms step_avg:65.45ms +step:6700/20000 train_loss:2.2536 spike_rate:0.0591 train_time:439666ms step_avg:65.62ms +step:6750/20000 train_loss:2.4299 spike_rate:0.0586 train_time:441966ms step_avg:65.48ms +step:6800/20000 train_loss:2.2860 spike_rate:0.0591 train_time:444267ms step_avg:65.33ms +step:6800/20000 val_loss:2.2323 val_bpb:1.3221 train_time:444288ms step_avg:65.34ms +step:6850/20000 train_loss:2.1161 spike_rate:0.0603 train_time:446572ms step_avg:65.19ms +step:6900/20000 train_loss:2.1902 spike_rate:0.0586 train_time:451791ms step_avg:65.48ms +step:6950/20000 train_loss:2.2747 spike_rate:0.0581 train_time:454110ms step_avg:65.34ms +step:7000/20000 train_loss:2.3213 spike_rate:0.0603 train_time:456413ms step_avg:65.20ms +step:7000/20000 val_loss:2.2262 val_bpb:1.3185 train_time:456452ms step_avg:65.21ms +step:7050/20000 train_loss:2.3439 spike_rate:0.0603 train_time:459818ms step_avg:65.22ms +step:7100/20000 train_loss:2.1603 spike_rate:0.0596 train_time:464471ms step_avg:65.42ms +step:7150/20000 train_loss:2.2487 spike_rate:0.0581 train_time:466769ms step_avg:65.28ms +step:7200/20000 train_loss:2.2988 spike_rate:0.0591 train_time:469073ms step_avg:65.15ms +step:7200/20000 val_loss:2.2249 val_bpb:1.3177 train_time:469094ms step_avg:65.15ms +step:7250/20000 train_loss:2.1916 spike_rate:0.0596 train_time:473634ms step_avg:65.33ms +step:7300/20000 train_loss:2.1756 spike_rate:0.0586 train_time:475918ms step_avg:65.19ms +step:7350/20000 train_loss:2.2854 spike_rate:0.0586 train_time:478225ms step_avg:65.06ms +step:7400/20000 train_loss:2.2081 spike_rate:0.0603 train_time:480527ms step_avg:64.94ms +step:7400/20000 val_loss:2.2232 val_bpb:1.3167 train_time:480548ms step_avg:64.94ms +step:7450/20000 train_loss:2.2061 spike_rate:0.0598 train_time:486841ms step_avg:65.35ms +step:7500/20000 train_loss:2.2130 spike_rate:0.0615 train_time:489150ms step_avg:65.22ms +step:7550/20000 train_loss:2.2642 spike_rate:0.0603 train_time:491456ms step_avg:65.09ms +step:7600/20000 train_loss:2.0933 spike_rate:0.0608 train_time:493759ms step_avg:64.97ms +step:7600/20000 val_loss:2.2230 val_bpb:1.3166 train_time:493780ms step_avg:64.97ms +step:7650/20000 train_loss:2.3897 spike_rate:0.0608 train_time:499148ms step_avg:65.25ms +step:7700/20000 train_loss:2.1840 spike_rate:0.0598 train_time:501450ms step_avg:65.12ms +step:7750/20000 train_loss:2.2085 spike_rate:0.0603 train_time:503746ms step_avg:65.00ms +step:7800/20000 train_loss:2.2346 spike_rate:0.0603 train_time:506051ms step_avg:64.88ms +step:7800/20000 val_loss:2.2203 val_bpb:1.3150 train_time:506070ms step_avg:64.88ms +step:7850/20000 train_loss:2.1007 spike_rate:0.0608 train_time:511759ms step_avg:65.19ms +step:7900/20000 train_loss:2.2258 spike_rate:0.0596 train_time:514061ms step_avg:65.07ms +step:7950/20000 train_loss:2.1923 spike_rate:0.0598 train_time:516358ms step_avg:64.95ms +step:8000/20000 train_loss:2.2051 spike_rate:0.0598 train_time:518662ms step_avg:64.83ms +step:8000/20000 val_loss:2.2165 val_bpb:1.3128 train_time:518683ms step_avg:64.84ms +step:8050/20000 train_loss:2.1737 spike_rate:0.0603 train_time:526297ms step_avg:65.38ms +step:8100/20000 train_loss:2.2339 spike_rate:0.0598 train_time:528594ms step_avg:65.26ms +step:8150/20000 train_loss:2.3430 spike_rate:0.0613 train_time:530901ms step_avg:65.14ms +step:8200/20000 train_loss:2.2747 spike_rate:0.0608 train_time:533205ms step_avg:65.02ms +step:8200/20000 val_loss:2.2099 val_bpb:1.3088 train_time:533226ms step_avg:65.03ms +step:8250/20000 train_loss:2.2324 spike_rate:0.0620 train_time:539298ms step_avg:65.37ms +step:8300/20000 train_loss:2.2135 spike_rate:0.0625 train_time:541606ms step_avg:65.25ms +step:8350/20000 train_loss:2.3168 spike_rate:0.0615 train_time:543908ms step_avg:65.14ms +step:8400/20000 train_loss:2.2163 spike_rate:0.0620 train_time:549133ms step_avg:65.37ms +step:8400/20000 val_loss:2.2020 val_bpb:1.3042 train_time:549148ms step_avg:65.37ms +step:8450/20000 train_loss:2.3116 spike_rate:0.0635 train_time:551428ms step_avg:65.26ms +step:8500/20000 train_loss:2.2082 spike_rate:0.0615 train_time:553729ms step_avg:65.14ms +step:8550/20000 train_loss:2.2735 spike_rate:0.0613 train_time:556033ms step_avg:65.03ms +step:8600/20000 train_loss:2.2134 spike_rate:0.0620 train_time:562030ms step_avg:65.35ms +step:8600/20000 val_loss:2.1924 val_bpb:1.2985 train_time:562049ms step_avg:65.35ms +step:8650/20000 train_loss:2.1736 spike_rate:0.0615 train_time:564326ms step_avg:65.24ms +step:8700/20000 train_loss:2.1008 spike_rate:0.0608 train_time:566623ms step_avg:65.13ms +step:8750/20000 train_loss:2.2653 spike_rate:0.0630 train_time:568921ms step_avg:65.02ms +step:8800/20000 train_loss:2.1765 spike_rate:0.0625 train_time:573335ms step_avg:65.15ms +step:8800/20000 val_loss:2.1850 val_bpb:1.2941 train_time:573356ms step_avg:65.15ms +step:8850/20000 train_loss:2.3750 spike_rate:0.0630 train_time:576741ms step_avg:65.17ms +step:8900/20000 train_loss:2.2652 spike_rate:0.0630 train_time:579037ms step_avg:65.06ms +step:8950/20000 train_loss:2.2281 spike_rate:0.0625 train_time:581338ms step_avg:64.95ms +step:9000/20000 train_loss:2.0886 spike_rate:0.0640 train_time:586317ms step_avg:65.15ms +step:9000/20000 val_loss:2.1784 val_bpb:1.2902 train_time:586356ms step_avg:65.15ms +step:9050/20000 train_loss:2.1245 spike_rate:0.0630 train_time:588639ms step_avg:65.04ms +step:9100/20000 train_loss:2.3610 spike_rate:0.0635 train_time:590939ms step_avg:64.94ms +step:9150/20000 train_loss:2.0647 spike_rate:0.0640 train_time:593242ms step_avg:64.84ms +step:9200/20000 train_loss:2.1413 spike_rate:0.0635 train_time:598576ms step_avg:65.06ms +step:9200/20000 val_loss:2.1724 val_bpb:1.2866 train_time:598602ms step_avg:65.07ms +step:9205/20000 val_loss:2.1722 val_bpb:1.2865 train_time:599659ms step_avg:65.14ms +stopping_early: wallclock_cap train_time:599659ms step:9205/20000 +peak memory allocated: 12417 MiB reserved: 12708 MiB +Serialized model: 67233157 bytes +Code size: 52783 bytes +Total submission size: 67285940 bytes +Serialized model int8+zlib: 15723303 bytes (payload:17179020 raw_torch:17231715 payload_ratio:3.91x) +Total submission size int8+zlib: 15776086 bytes +final_int8_zlib_roundtrip val_loss:2.1919 val_bpb:1.2982 eval_time:1435ms +final_int8_zlib_roundtrip_exact val_loss:2.19192126 val_bpb:1.29817924