diff --git a/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/README.md b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/README.md new file mode 100644 index 000000000..1471e9cc4 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/README.md @@ -0,0 +1,126 @@ +# Non-Record Submission: PrismLM v3 — DiffTransformer V2 + NorMuon + TrigramHash + +## Score: val_bpb = 1.1715 (post-quant int6+zstd, no sliding window) + +Trained on 8×H100 SXM in 600 seconds. 15.59MB artifact (int6+zstd-22). Single seed run. + +This is a non-record submission exploring **three novel techniques** not yet attempted in any merged or open PR, built on top of the proven technique stack from PR #315. + +## Novel Contributions + +### 1. DiffTransformer V2 Attention (Last 2 Layers) + +Based on [Differential Transformer](https://arxiv.org/abs/2410.05258) (Microsoft, ICLR 2025 Oral). Computes two separate softmax attention maps and subtracts them, cancelling noise in the attention pattern: + +``` +attn = softmax(Q1 @ K1^T) - λ · softmax(Q2 @ K2^T) +``` + +Applied only to the last 2 layers where attention refinement matters most. The scalar `λ` is learned per-head via `lambda_init` reparameterization. Remaining layers use standard GQA + XSA. + +### 2. NorMuon Optimizer + +Replaces standard Muon with NorMuon ([Keller Jordan, Oct 2025](https://kellerjordan.github.io/posts/muon/)), which adds **per-neuron row normalization** after the Newton-Schulz orthogonalization step. This normalizes gradient updates by the second moment of each row, giving ~11% better compute efficiency. Uses `beta2=0.95` for the second moment EMA. + +### 3. TrigramHash + Context-Aware N-gram Gating + +Extends BigramHash with a TrigramHash table (2048 buckets, dim 64) that captures three-token patterns via `(t0 * 961 + t1 * 31 + t2) % (vocab_size - 1) + 1`. Both n-gram signals are modulated by a **context-aware gate** (inspired by [DeepSeek Engram](https://github.com/deepseek-ai/Engram)) that learns when to rely on n-gram vs. neural predictions: + +``` +gate = sigmoid(linear(hidden_state)) +output = hidden + gate * (bigram_signal + trigram_signal) +``` + +## Full Architecture + +| Component | Value | +|-----------|-------| +| Layers | 11 | +| Model dim | 512 | +| Heads / KV heads | 8 / 4 (GQA) | +| MLP expansion | 3× (hidden=1536), ReLU² | +| XSA layers | Last 6 | +| DiffAttn layers | Last 2 | +| Partial RoPE | 16/64 dims (25%) | +| LN depth scaling | 1/√(layer+1) | +| SmearGate | Yes | +| BigramHash | 2048 buckets, dim 128 | +| TrigramHash | 2048 buckets, dim 64 | +| N-gram gating | Context-aware sigmoid gate | +| U-Net skips | Yes | +| Logit softcap | 30.0 | +| Tied embeddings | Yes (FP16) | + +## Training Configuration + +| Parameter | Value | +|-----------|-------| +| Optimizer (matrices) | NorMuon (lr=0.04, momentum=0.95, WD=0.02, beta2=0.95) | +| Optimizer (embeddings/scalars) | AdamW (lr=0.04, WD=0.01) | +| Tied embed LR | 0.05 | +| Batch size | 786,432 tokens | +| Sequence length | 2048 | +| Warmdown iters | 1200 | +| Grad clip | 0.3 | +| SWA | Enabled (every 200 steps) | +| Late QAT | Enabled (when lr_scale < 0.1) | +| Warmup | 20 steps | + +## Quantization & Compression + +- **Int6** per-row quantization on MLP and attention weight matrices +- **FP16** for tied embeddings +- **3% magnitude pruning** before quantization (adaptive up to 15% if over budget) +- **zstd level 22** compression +- Flash Attention 3 fallback to `F.scaled_dot_product_attention` (FA3 not available in our environment) + +## Key Metrics + +- **val_bpb (post-quant): 1.1715** (standard eval, no sliding window) +- Pre-quant val_bpb: 1.1607 +- Quantization penalty: ~0.011 bpb +- Steps completed: 4,600 / 20,000 (wallclock-capped at 600s) +- Step average: 130.43 ms/step +- Model params: 27,518,587 +- Artifact size: 15,586,651 bytes (15.59MB) + - Model int6+zstd: 15,521,912 bytes + - Code: 64,739 bytes +- Peak memory: 25,921 MiB allocated, 26,460 MiB reserved +- GPU: 8×H100 SXM (Modal) + +## Training Progression + +| Step | val_loss | val_bpb | +|------|----------|---------| +| 0 | 6.9300 | 4.1043 | +| 1000 | 2.2005 | 1.3033 | +| 2000 | 2.1133 | 1.2516 | +| 3000 | 2.0876 | 1.2364 | +| 4000 | 2.0209 | 1.1969 | +| 4600 (final) | 1.9598 | 1.1607 | +| **Post-quant** | **1.9780** | **1.1715** | + +## Gap Analysis vs. SOTA + +Our score of 1.1715 is ~0.029 bpb behind the merged SOTA (1.1428) and ~0.047 bpb behind the unmerged frontier (1.1248). Key factors: + +1. **No sliding window eval** — was disabled to save eval time. Sliding window typically gives ~0.03 bpb improvement; re-enabled in the submitted code for future runs. +2. **Small n-gram tables** — BigramHash(2048) vs. the SOTA's BigramHash(10240). Larger tables are worth ~0.005 bpb. +3. **NorMuon hyperparameters** — momentum=0.95 vs. proven momentum=0.99. The lower momentum may have hurt convergence in the warmdown phase. +4. **DiffAttn parameter overhead** — 1.5× attention parameters on 2 layers reduces capacity available for other components. The noise-cancellation benefit at this scale is unclear. +5. **SDPA fallback** — Flash Attention 3 was unavailable; SDPA is functionally equivalent but ~10% slower, meaning fewer training steps. + +## What We'd Change With More Compute + +1. Increase BigramHash to 10240 buckets (~0.005 bpb) +2. Re-enable sliding window eval (~0.03 bpb) +3. Tune NorMuon momentum to 0.99 +4. Try EMA instead of SWA (works better with XSA per community data) +5. Ablate DiffAttn vs. standard attention to quantify its contribution +6. Increase TrigramHash to 8192 buckets + +## Included Files + +- `train_gpt.py` — Self-contained training + evaluation script (bug-fixed: correct 16MB decimal limit, sliding window eval re-enabled) +- `train.log` — Training log from the 8×H100 run (seed 1337) +- `submission.json` — Leaderboard metadata diff --git a/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/submission.json b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/submission.json new file mode 100644 index 000000000..d9dc5a5a4 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/submission.json @@ -0,0 +1,18 @@ +{ + "author": "Yash Verma", + "github_id": "yasboop", + "name": "PrismLM v3 — DiffTransformer V2 + NorMuon + TrigramHash", + "blurb": "Non-record 8xH100 submission exploring 3 novel techniques on top of PR #315's technique stack: (1) DiffTransformer V2 attention on last 2 layers for noise-cancelled attention maps, (2) NorMuon optimizer with per-neuron row normalization for 11% better convergence, (3) TrigramHash + context-aware n-gram gating. 11 layers, 512 dim, XSA on 6 layers, Partial RoPE (16 dims), LN depth scaling, SmearGate, BigramHash(2048), int6+zstd-22 quantization, SWA, late QAT. Post-quant val_bpb 1.1715 (without sliding window eval).", + "date": "2026-03-22T00:00:00Z", + "track": "non-record-16mb", + "val_loss": 1.97797457, + "val_bpb": 1.17146796, + "pre_quant_val_loss": 1.9598, + "pre_quant_val_bpb": 1.1607, + "step_stop": 4600, + "wallclock_seconds": 599.974, + "bytes_total": 15586651, + "bytes_model_int6_zstd": 15521912, + "bytes_code": 64739, + "gpu": "8xH100 SXM (Modal)" +} diff --git a/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/train.log b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/train.log new file mode 100644 index 000000000..0c8d65ee3 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/train.log @@ -0,0 +1,85 @@ +logs/prism_seed1337_gpu8.txt +flash_attn_3:False +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:100 +val_loader:shards pattern=/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +torch.compile(fullgraph=True): registered (compiles on first real forward) +model_params:27518587 +diff_attn_last_n:2 xsa_last_n:6 +bigram_vocab:2048 trigram_vocab:2048 +optimizer:NorMuon matrix_lr:0.04 muon_momentum:0.95 +world_size:8 grad_accum_steps:1 +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9300 val_bpb:4.1043 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9319 train_time:185ms step_avg:185.24ms +step:2/20000 train_loss:10.2254 train_time:311ms step_avg:155.55ms +step:3/20000 train_loss:8.5880 train_time:439ms step_avg:146.27ms +step:4/20000 train_loss:7.7303 train_time:573ms step_avg:143.23ms +step:5/20000 train_loss:7.1825 train_time:706ms step_avg:141.23ms +step:6/20000 train_loss:6.8526 train_time:836ms step_avg:139.38ms +step:7/20000 train_loss:6.7896 train_time:964ms step_avg:137.72ms +step:8/20000 train_loss:6.7501 train_time:1096ms step_avg:137.04ms +step:9/20000 train_loss:6.4141 train_time:1227ms step_avg:136.36ms +step:10/20000 train_loss:5.9556 train_time:1354ms step_avg:135.37ms +step:200/20000 train_loss:2.3882 train_time:26134ms step_avg:130.67ms +step:400/20000 train_loss:2.4336 train_time:52379ms step_avg:130.95ms +step:600/20000 train_loss:2.3296 train_time:78406ms step_avg:130.68ms +step:800/20000 train_loss:2.2175 train_time:104554ms step_avg:130.69ms +step:1000/20000 train_loss:2.2512 train_time:130559ms step_avg:130.56ms +step:1000/20000 val_loss:2.2005 val_bpb:1.3033 train_time:130567ms step_avg:130.57ms +step:1200/20000 train_loss:2.3233 train_time:156718ms step_avg:130.60ms +step:1400/20000 train_loss:2.1494 train_time:182849ms step_avg:130.61ms +step:1600/20000 train_loss:2.0411 train_time:208894ms step_avg:130.56ms +step:1800/20000 train_loss:2.1322 train_time:235041ms step_avg:130.58ms +step:2000/20000 train_loss:2.0504 train_time:261028ms step_avg:130.51ms +step:2000/20000 val_loss:2.1133 val_bpb:1.2516 train_time:261033ms step_avg:130.52ms +step:2200/20000 train_loss:2.1147 train_time:287151ms step_avg:130.52ms +step:2400/20000 train_loss:2.0524 train_time:313147ms step_avg:130.48ms +step:2600/20000 train_loss:2.1013 train_time:339231ms step_avg:130.47ms +step:2800/20000 train_loss:2.1496 train_time:365361ms step_avg:130.49ms +step:3000/20000 train_loss:2.1562 train_time:391369ms step_avg:130.46ms +step:3000/20000 val_loss:2.0876 val_bpb:1.2364 train_time:391375ms step_avg:130.46ms +step:3200/20000 train_loss:2.1661 train_time:417498ms step_avg:130.47ms +step:3400/20000 train_loss:2.0158 train_time:443423ms step_avg:130.42ms +step:3600/20000 train_loss:2.0787 train_time:469562ms step_avg:130.43ms +step:3800/20000 train_loss:2.0383 train_time:495533ms step_avg:130.40ms +step:4000/20000 train_loss:1.9308 train_time:521677ms step_avg:130.42ms +step:4000/20000 val_loss:2.0209 val_bpb:1.1969 train_time:521687ms step_avg:130.42ms +swa:start step:4200 +step:4200/20000 train_loss:2.0923 train_time:547809ms step_avg:130.43ms +step:4400/20000 train_loss:1.9583 train_time:573818ms step_avg:130.41ms +late_qat:enabled step:4481 scale:0.0994 +step:4600/20000 train_loss:1.7641 train_time:599968ms step_avg:130.43ms +step:4600/20000 val_loss:1.9598 val_bpb:1.1607 train_time:599974ms step_avg:130.43ms +stopping_early: wallclock_cap train_time:599974ms step:4600/20000 +peak memory allocated: 25921 MiB reserved: 26460 MiB +swa:applying averaged 3 checkpoints +Serialized model: 108278477 bytes +Code size: 64739 bytes +Serialized model int6+zstd: 15521912 bytes +Total submission size int6+zstd: 15586651 bytes +Under 16MB: True +final_int6_roundtrip val_loss:1.9780 val_bpb:1.1715 eval_time:22821ms +final_int6_roundtrip_exact val_loss:1.97797457 val_bpb:1.17146796 diff --git a/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/train_gpt.py b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/train_gpt.py new file mode 100644 index 000000000..7d86e3575 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_PrismLM_v3_DiffAttn_NorMuon_TrigramHash/train_gpt.py @@ -0,0 +1,1539 @@ +""" +PrismLM v3 — 3 novel additions to PR #315's proven 1.1248 bpb baseline: + 1. DiffTransformer V2 attention on last 2 layers (Microsoft, Jan 2026) + 2. NorMuon optimizer — 11% more efficient than Muon (Oct 2025) + 3. TrigramHash + Context-Aware N-gram Gating +11 layers × 512d. Flash Attention 3, torch.compile(fullgraph=True), PyTorch 2.9.1. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _HAS_FLASH3 = True +except ImportError: + _HAS_FLASH3 = False + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + 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)) + + 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", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + 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", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + muon_wd = float(os.environ.get("MUON_WD", 0.02)) + adam_wd = float(os.environ.get("ADAM_WD", 0.01)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 200)) + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + late_qat = bool(int(os.environ.get("LATE_QAT", "1"))) + + # PrismLM v3 innovations + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + trigram_vocab_size = int(os.environ.get("TRIGRAM_VOCAB_SIZE", 2048)) + trigram_dim = int(os.environ.get("TRIGRAM_DIM", 64)) + diff_attn_last_n = int(os.environ.get("DIFF_ATTN_LAST_N", 2)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 6)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + + +# ----------------------------- +# NORMUON OPTIMIZER (replaces Muon — 11% more efficient) +# ----------------------------- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class NorMuon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0, beta2: float = 0.95): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay, beta2=beta2), + ) + + @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"] + beta2 = group["beta2"] + + 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) + state["second_momentum"] = torch.zeros(g.size(0), 1, device=g.device, dtype=torch.float32) + 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) + vnorm = g.norm(dim=(-2, -1), keepdim=True) + v_mean = torch.mean(g * g, dim=-1, keepdim=True).float() + state["second_momentum"].lerp_(v_mean, 1 - beta2) + step_size = 1.0 / state["second_momentum"].sqrt().add_(1e-10) + g = g * step_size.to(dtype=g.dtype) + vnorm_new = g.norm(dim=(-2, -1), keepdim=True) + g = g * (vnorm / vnorm_new.add_(1e-10)) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + 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 +# ----------------------------- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < 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}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // 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 * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, 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) + + +# ----------------------------- +# QUANTIZATION (Int6 mixed from PR #315) +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,mlp_scale,resid_mix,q_gain,skip_weight,skip_weights,smear,lambda_init,ngram_gate", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + + +def quantize_int6_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(amax / 31.0 if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8) + return q, scale + + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], prune_frac: float = 0.0): + if prune_frac > 0: + for name, t in state_dict.items(): + if t.is_floating_point() and t.ndim == 2 and t.numel() > 65536: + threshold = torch.quantile(t.abs().float().flatten(), prune_frac) + state_dict[name] = t.masked_fill(t.abs() < threshold, 0.0) + + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + passthrough_orig_dtypes = result.get("__passthrough_orig_dtypes__", {}) + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.rope_dims = rope_dims if rope_dims > 0 else dim + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + rd = self.rope_dims + inv_freq = 1.0 / (base ** (torch.arange(0, rd, 2, dtype=torch.float32) / rd)) + 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 + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + 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: + rd = cos.size(-1) * 2 + if rd < x.size(-1): + x_rope, x_pass = x[..., :rd], x[..., rd:] + half = rd // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rot = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rot, x_pass), dim=-1) + 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, + rope_dims: int = 0, + use_diff_attn: bool = False, + ): + 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") + self.use_diff_attn = use_diff_attn + if use_diff_attn: + self.num_q_heads = 2 * num_heads + else: + self.num_q_heads = num_heads + q_dim = self.num_q_heads * self.head_dim + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, q_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((self.num_q_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = rope_dims + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.use_xsa = False + if use_diff_attn: + self.lambda_init = nn.Parameter(torch.zeros(num_heads, dtype=torch.float32)) + + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_q_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if _HAS_FLASH3: + y = flash_attn_3_func(q, k, v, causal=True) + else: + Hq, Hkv = q.size(2), k.size(2) + qt = q.transpose(1, 2) + kt = k.transpose(1, 2) + vt = v.transpose(1, 2) + if Hq != Hkv: + kt = kt.repeat_interleave(Hq // Hkv, dim=1) + vt = vt.repeat_interleave(Hq // Hkv, dim=1) + y = F.scaled_dot_product_attention(qt, kt, vt, is_causal=True).transpose(1, 2) + if self.use_diff_attn: + y1 = y[:, :, 0::2, :] + y2 = y[:, :, 1::2, :] + lam = torch.sigmoid(self.lambda_init.to(dtype=y1.dtype))[None, None, :, None] + y = y1 - lam * y2 + elif self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) + + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class TrigramHashEmbedding(nn.Module): + def __init__(self, trigram_vocab_size: int, trigram_dim: int, model_dim: int): + super().__init__() + self.trigram_vocab_size = trigram_vocab_size + self.embed = nn.Embedding(trigram_vocab_size, trigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(trigram_dim, model_dim, bias=False) if trigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.03, dtype=torch.float32)) + + def trigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.trigram_vocab_size - 1 + out = torch.full_like(t, mod) + out[..., 2:] = ( + torch.bitwise_xor( + torch.bitwise_xor(48271 * t[..., 2:], 31547 * t[..., 1:-1]), + 17389 * t[..., :-2], + ) % mod + ) + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.trigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class NgramContextGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate_proj = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + self.bias = nn.Parameter(torch.tensor(1.0, dtype=torch.float32)) + + def forward(self, hidden: Tensor, ngram_signal: Tensor) -> Tensor: + gate = torch.sigmoid( + (hidden * self.gate_proj.to(dtype=hidden.dtype)[None, None, :]).sum(dim=-1, keepdim=True) + + self.bias.to(dtype=hidden.dtype) + ) + return gate * ngram_signal + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + rope_dims: int = 0, + layer_idx: int = 0, + ln_scale: bool = False, + use_diff_attn: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention( + dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + rope_dims=rope_dims, use_diff_attn=use_diff_attn, + ) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + s = self.ln_scale_factor + attn_out = self.attn(self.attn_norm(x) * s) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x) * s) + return x + + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + trigram_vocab_size: int = 0, + trigram_dim: int = 64, + xsa_last_n: int = 0, + diff_attn_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + ): + 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.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.trigram = TrigramHashEmbedding(trigram_vocab_size, trigram_dim, model_dim) if trigram_vocab_size > 0 else None + self.ngram_gate = NgramContextGate(model_dim) if (bigram_vocab_size > 0 or trigram_vocab_size > 0) else None + self.smear = SmearGate(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)) + diff_start = max(0, num_layers - diff_attn_last_n) if diff_attn_last_n > 0 else num_layers + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + rope_dims=rope_dims, + layer_idx=i, + ln_scale=ln_scale, + use_diff_attn=(i >= diff_start), + ) + 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 + if xsa_last_n > 0: + xsa_start = max(0, num_layers - xsa_last_n) + for i in range(xsa_start, num_layers): + if i < diff_start: + self.blocks[i].attn.use_xsa = 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) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + ngram_signal = torch.zeros_like(x) + if self.bigram is not None: + ngram_signal = ngram_signal + self.bigram(input_ids) + if self.trigram is not None: + ngram_signal = ngram_signal + self.trigram(input_ids) + if self.ngram_gate is not None: + ngram_signal = self.ngram_gate(x, ngram_signal) + x = x + ngram_signal + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + ngram_signal = torch.zeros_like(x) + if self.bigram is not None: + ngram_signal = ngram_signal + self.bigram(input_ids) + if self.trigram is not None: + ngram_signal = ngram_signal + self.trigram(input_ids) + if self.ngram_gate is not None: + ngram_signal = self.ngram_gate(x, ngram_signal) + x = x + ngram_signal + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + +# ----------------------------- +# SLIDING WINDOW EVALUATION +# ----------------------------- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + try: + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + except Exception: + pass + + # 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 + + 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, flush=True) + 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(f"flash_attn_3:{_HAS_FLASH3}") + 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"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_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 + CastedLinear._qat_enabled = False + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + trigram_vocab_size=args.trigram_vocab_size, + trigram_dim=args.trigram_dim, + xsa_last_n=args.xsa_last_n, + diff_attn_last_n=args.diff_attn_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + try: + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + log0("torch.compile(fullgraph=True): registered (compiles on first real forward)") + except Exception as e: + log0(f"torch.compile FAILED: {e}") + log0("FALLING BACK to eager mode (slower but functional)") + compiled_model = base_model + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(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) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + if base_model.trigram is not None: + scalar_params.append(base_model.trigram.scale) + if base_model.ngram_gate is not None: + scalar_params.append(base_model.ngram_gate.gate_proj) + scalar_params.append(base_model.ngram_gate.bias) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.trigram is not None: + tok_params.append({"params": [base_model.trigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.trigram.proj is not None: + matrix_params.append(base_model.trigram.proj.weight) + + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = NorMuon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + beta2=args.beta2, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + 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"diff_attn_last_n:{args.diff_attn_last_n} xsa_last_n:{args.xsa_last_n}") + log0(f"bigram_vocab:{args.bigram_vocab_size} trigram_vocab:{args.trigram_vocab_size}") + log0(f"optimizer:NorMuon matrix_lr:{args.matrix_lr} muon_momentum:{args.muon_momentum}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # DATA LOADER & MODEL WARMUP + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # MAIN TRAINING LOOP + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state: dict[str, Tensor] | None = None + if args.ema_enabled: + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + + 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) + qat_threshold = float(os.environ.get("QAT_THRESHOLD", "0.1")) + if args.late_qat and scale < qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + if ema_state is not None: + d = args.ema_decay + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(d).add_(t.detach().float(), alpha=1.0 - d) + + if args.swa_enabled and not args.ema_enabled and scale < 0.5 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name].add_(t.detach().float()) + swa_count += 1 + + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + if ema_state is not None: + log0("ema:applying EMA weights") + avg_state = {name: t.to(dtype=base_model.state_dict()[name].dtype) + for name, t in ema_state.items()} + del ema_state + base_model.load_state_dict(avg_state, strict=True) + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + avg_state = {name: (t / swa_count).to(dtype=base_model.state_dict()[name].dtype) + for name, t in swa_state.items()} + del swa_state + base_model.load_state_dict(avg_state, strict=True) + + # SERIALIZATION + ROUNDTRIP VALIDATION + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items()} + + if master_process: + torch.save(export_sd, "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") + + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + code_bytes = len(code.encode("utf-8")) + budget_bytes = 16_000_000 - code_bytes + + prune_frac = 0.03 + quant_result, quant_meta = mixed_quantize_int6(dict(sd_cpu), {"mlp", "attn"}, prune_frac=prune_frac) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + + for extra_prune in [0.05, 0.08, 0.10, 0.15]: + if len(quant_blob) <= budget_bytes: + break + log0(f"prune:over_budget at {prune_frac*100:.0f}%, retrying with {extra_prune*100:.0f}%") + prune_frac = extra_prune + sd_cpu_fresh = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu_fresh, {"mlp", "attn"}, prune_frac=prune_frac) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + log0(f"Under 16MB: {(quant_file_bytes + code_bytes) < 16_000_000}") + sys.stdout.flush() + + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + trigram_vocab_size=args.trigram_vocab_size, trigram_dim=args.trigram_dim, + xsa_last_n=args.xsa_last_n, diff_attn_last_n=args.diff_attn_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, eval_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_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_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if args.eval_stride > 0 and args.eval_stride < args.eval_seq_len: + log0(f"final_eval_mode:sliding_window stride:{args.eval_stride}") + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=32, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main()