diff --git a/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/README.md b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/README.md new file mode 100644 index 0000000000..a63289795e --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/README.md @@ -0,0 +1,140 @@ +# Late STE QAT + Int6 MLP3x + SmearGate + BigramHash + OrthoInit + Overtone + SWA + SGD TTT + +## Score + +**Measured (seed=1337, single run):** `val_bpb = 1.16292025` · `val_loss = 1.96353693` (after int6+zstd roundtrip + sliding-window eval; see `train.log`). + +Trained on **8×H100 SXM** with a **600s** wallclock cap (`step=5464`). **Total submission size `15,948,643` bytes** (~15.95 MB decimal), **below** the 16,000,000-byte limit — **int6 + zstd-22** artifact plus UTF-8 `train_gpt.py` bytes (`64,426`). + +> *Note:* The template-style multi-seed table below is **not** part of this folder’s logs; only **seed 1337** is recorded here. Re-run with other `SEED` values if you want a proper mean/std. + +## Approach + +Stacked techniques on a **9-layer, 512-dim** GPT-style model, plus **late STE QAT**, **Overtone-style init**, and optional **full-model SGD TTT** (this script defaults to SGD TTT on, LoRA TTT off). + +### 1. Per-row int6 quantization + zstd-22 + +MLP and attention weight matrices are quantized to int6 (roughly `[-32, 31]`) with **per-row scaling**. Tied embeddings stay in a higher-precision path where it matters; the implementation follows the repo’s mixed-quant rules. After `torch.save` of the quantized payload, the blob is compressed with **zstd level 22** (`zstandard`), which is typically a few percent smaller than **zlib-9** on the same bytes — enough here to land **under** the decimal 16MB cap when zlib did not. + +### 2. 3× MLP expansion + +Hidden FFN width **1536** (3×) instead of 2× **1024**, paid for in the budget by int6 + strong compression. + +### 3. SmearGate + +A learned gate blending each token’s embedding with the **previous** token’s embedding for cheap bigram-like signal at the embedding layer (on the order of **~512** extra parameters in the usual setup). + +### 4. BigramHash embedding + +A **4096**-bucket table (e.g. dim **128**, projected to model width) keyed by adjacent token pairs via a small hash of `(prev, curr)`. Adds on the order of **~0.5M** parameters and complements SmearGate with an **additive** bigram path. + +### 5. Orthogonal init (+ muP-style scaling) + +Large matrices initialized orthogonal where applicable; readouts scaled with depth-aware factors consistent with muP-style training in this codebase. + +### 6. Muon + AdamW, weight decay + +**Muon** on matrix blocks with tuned **weight decay** and momentum schedule; scalar/embedding groups use **AdamW** with their own WD. This run uses **`muon_weight_decay=0.038`**, **`matrix_lr=0.025`** (see env overrides in `train_gpt.py`). + +### 7. Stochastic weight averaging (SWA) + +SWA accumulates weights over the **last fraction** of training (default **`swa_start_frac=0.5`**) every **`swa_every`** steps (default **`200`** in this script). The logged run averaged **5** checkpoints before quantization. + +### 8. Late STE QAT (last ~15% of wallclock) + +**Fake-quant (STE)** for int6 is only enabled after **`qat_start_frac≈0.85`** of the wallclock budget, with **`qat_lr_factor=0.5`** on the affected optimizer groups when QAT turns on — so Muon is not fighting quant noise for the whole run. + +### 9. Full-model SGD test-time training (optional) + +A short **SGD** pass on the validation stream (**not** LoRA) to adapt all weights, including gates and bigram paths LoRA often misses. Controlled by **`SGD_TTT_ENABLED`** / **`TTT_LORA_ENABLED`**. + +## Main Hyperparameters + +| Parameter | Value (this script / logged run) | +|-----------|----------------------------------| +| num_layers | 9 | +| model_dim | 512 | +| mlp_mult | 3.0 (hidden=1536) | +| train_seq_len | 2048 | +| train_batch_tokens | 786,432 | +| warmdown_iters | 3000 | +| matrix_lr | 0.025 | +| scalar_lr | 0.02 | +| tied_embed_lr | 0.03 | +| muon_momentum | 0.99 (warmup from 0.92 over 1500 steps) | +| muon_weight_decay | 0.038 | +| weight_decay (AdamW scalars) | 0.01 | +| grad_clip_norm | 0.3 | +| eval_stride | 64 | +| swa_every | 200 | +| swa_start_frac | 0.5 | +| qat_start_frac | 0.85 | +| qat_lr_factor | 0.5 | +| bigram hash buckets | 4096 | +| bigram dim | 128 | +| compressor | **zstd (level 22)** | +| SGD TTT | LR `3e-4`, momentum `0.95` (when enabled) | + +## Key metrics (this snapshot) + +| Item | Value | +|------|--------| +| **val_bpb** | **1.16292025** (`final_int8_zstd_roundtrip_exact`) | +| **val_loss** | **1.96353693** | +| Wallclock cap | 600s | +| Steps completed | 5464 | +| Model params (logged) | ~22.37M | +| **bytes_total** | **15,948,643** (under 16MB cap) | +| **bytes_code** | **64,426** | +| int6+zstd blob (logged) | 15,884,217 bytes | + +## Reproducibility + +**Logged run** (seed **1337**): + +| Seed | val_loss | val_bpb | +|------|----------|---------| +| 1337 | 1.96353693 | 1.16292025 | + +For multiple seeds, re-launch with e.g. `SEED=42`, `SEED=7`, etc. Byte totals and BPB can shift slightly across machines due to GPU non-determinism. + +## Evaluation pipeline (order) + +1. Train until the 600s cap (late QAT only in the tail). +2. Apply SWA checkpoint average. +3. Quantize to int6 + **zstd-22** → `final_model.int8.ptz`. +4. Decompress, dequantize, **sliding-window eval** (`eval_stride=64`). +5. If enabled: **SGD TTT**, then final metrics. + +## How to reproduce + +Install **zstandard** and cache FineWeb (`sp1024`) from the repo root; set **`HF_TOKEN`** if downloads require it. + +```bash +pip install zstandard +export HF_TOKEN="your_token" # if needed +python3 data/cached_challenge_fineweb.py --variant sp1024 +``` + +```bash +cd /path/to/parameter-golf + +RUN_ID=late_qat_sgd_ttt_zstd \ +SEED=1337 \ +DATA_PATH=./data/datasets/fineweb10B_sp1024/ \ +TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \ +VOCAB_SIZE=1024 \ +EVAL_STRIDE=64 \ +SGD_TTT_ENABLED=1 \ +TTT_LORA_ENABLED=0 \ +torchrun --standalone --nproc_per_node=8 \ + old/20/03/26-zstandard/train_gpt.py +``` + +## Files in this folder + +| File | Purpose | +|------|---------| +| `train_gpt.py` | Training + zstd artifact | +| `train.log` | Log for the run above | +| `submission.json` | Summary JSON for the challenge | diff --git a/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/submission.json b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/submission.json new file mode 100644 index 0000000000..cd2f5dac71 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/submission.json @@ -0,0 +1,12 @@ +{ + "track": "10min_16mb", + "date": "2026-03-20", + "name": "Late STE QAT + Int6 MLP3x + SmearGate + BigramHash + OrthoInit + Overtone + SWA + SGD TTT", + "author": "David Puertolas Merenciano", + "github_id": "davidpuertolas", + "blurb": "Late STE QAT (last 15%, per #76) avoids Muon momentum corruption while closing quant gap. Full-model SGD TTT (per #152) replaces LoRA TTT which hurts with SmearGate (#178). WD=0.038 + LR=0.025 from best validated submissions (#179, #194). Artifact: int6+zstd-22, under 16MB cap.", + "val_loss": 1.96353693, + "val_bpb": 1.16292025, + "bytes_total": 15948643, + "bytes_code": 64426 +} diff --git a/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/train.log b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/train.log new file mode 100644 index 0000000000..8d9ad7fed4 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/train.log @@ -0,0 +1,118 @@ +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:22368841 +world_size:8 grad_accum_steps:1 +ste_qat:late_activation at 85% wallclock, lr_factor=0.5 +muon_weight_decay:0.038 matrix_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 max_wallclock:600s +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:6.9311 val_bpb:4.1050 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9332 train_time:137ms step_avg:137.14ms +step:2/20000 train_loss:8.4116 train_time:197ms step_avg:98.25ms +step:3/20000 train_loss:7.4632 train_time:290ms step_avg:96.55ms +step:4/20000 train_loss:7.3430 train_time:383ms step_avg:95.72ms +step:5/20000 train_loss:7.5556 train_time:476ms step_avg:95.25ms +step:6/20000 train_loss:7.6005 train_time:569ms step_avg:94.91ms +step:7/20000 train_loss:7.3236 train_time:664ms step_avg:94.79ms +step:8/20000 train_loss:6.9358 train_time:756ms step_avg:94.55ms +step:9/20000 train_loss:6.5327 train_time:849ms step_avg:94.37ms +step:10/20000 train_loss:6.2320 train_time:942ms step_avg:94.24ms +step:100/20000 train_loss:3.2125 train_time:8846ms step_avg:88.46ms +step:200/20000 train_loss:2.3915 train_time:19276ms step_avg:96.38ms +step:300/20000 train_loss:2.5587 train_time:29727ms step_avg:99.09ms +step:400/20000 train_loss:2.4284 train_time:39989ms step_avg:99.97ms +step:500/20000 train_loss:2.4169 train_time:48791ms step_avg:97.58ms +step:500/20000 val_loss:2.3788 val_bpb:1.4089 train_time:48833ms step_avg:97.67ms +step:600/20000 train_loss:2.3541 train_time:59248ms step_avg:98.75ms +step:700/20000 train_loss:2.3678 train_time:69725ms step_avg:99.61ms +step:800/20000 train_loss:2.2615 train_time:79977ms step_avg:99.97ms +step:900/20000 train_loss:2.1514 train_time:90430ms step_avg:100.48ms +step:1000/20000 train_loss:2.2985 train_time:99231ms step_avg:99.23ms +step:1000/20000 val_loss:2.2502 val_bpb:1.3327 train_time:99274ms step_avg:99.27ms +step:1100/20000 train_loss:2.3433 train_time:109642ms step_avg:99.67ms +step:1200/20000 train_loss:2.3797 train_time:120040ms step_avg:100.03ms +step:1300/20000 train_loss:2.1229 train_time:130463ms step_avg:100.36ms +step:1400/20000 train_loss:2.2070 train_time:140816ms step_avg:100.58ms +step:1500/20000 train_loss:2.2455 train_time:149603ms step_avg:99.74ms +step:1500/20000 val_loss:2.2079 val_bpb:1.3076 train_time:149645ms step_avg:99.76ms +step:1600/20000 train_loss:2.1007 train_time:160074ms step_avg:100.05ms +step:1700/20000 train_loss:2.1664 train_time:170545ms step_avg:100.32ms +step:1800/20000 train_loss:2.1823 train_time:181020ms step_avg:100.57ms +step:1900/20000 train_loss:2.1516 train_time:189803ms step_avg:99.90ms +step:2000/20000 train_loss:2.0915 train_time:200201ms step_avg:100.10ms +step:2000/20000 val_loss:2.1553 val_bpb:1.2765 train_time:200243ms step_avg:100.12ms +step:2100/20000 train_loss:2.0728 train_time:210725ms step_avg:100.35ms +step:2200/20000 train_loss:2.1654 train_time:221232ms step_avg:100.56ms +step:2300/20000 train_loss:2.1299 train_time:231688ms step_avg:100.73ms +step:2400/20000 train_loss:2.0915 train_time:240492ms step_avg:100.20ms +step:2500/20000 train_loss:2.1899 train_time:250967ms step_avg:100.39ms +step:2500/20000 val_loss:2.1292 val_bpb:1.2610 train_time:251009ms step_avg:100.40ms +step:2600/20000 train_loss:2.1329 train_time:261376ms step_avg:100.53ms +step:2700/20000 train_loss:2.1255 train_time:271801ms step_avg:100.67ms +step:2800/20000 train_loss:2.1792 train_time:282179ms step_avg:100.78ms +step:2900/20000 train_loss:2.0496 train_time:290997ms step_avg:100.34ms +step:3000/20000 train_loss:2.1832 train_time:301357ms step_avg:100.45ms +step:3000/20000 val_loss:2.1154 val_bpb:1.2529 train_time:301399ms step_avg:100.47ms +step:3100/20000 train_loss:2.0577 train_time:311765ms step_avg:100.57ms +step:3200/20000 train_loss:2.1901 train_time:322125ms step_avg:100.66ms +step:3300/20000 train_loss:2.0864 train_time:330911ms step_avg:100.28ms +step:3400/20000 train_loss:2.0324 train_time:341219ms step_avg:100.36ms +step:3500/20000 train_loss:2.1902 train_time:351632ms step_avg:100.47ms +step:3500/20000 val_loss:2.0927 val_bpb:1.2394 train_time:351674ms step_avg:100.48ms +step:3600/20000 train_loss:2.1049 train_time:362032ms step_avg:100.56ms +step:3700/20000 train_loss:2.1037 train_time:372242ms step_avg:100.61ms +step:3800/20000 train_loss:2.0835 train_time:381033ms step_avg:100.27ms +step:3900/20000 train_loss:2.0819 train_time:391431ms step_avg:100.37ms +step:4000/20000 train_loss:1.9795 train_time:401794ms step_avg:100.45ms +step:4000/20000 val_loss:2.0721 val_bpb:1.2272 train_time:401836ms step_avg:100.46ms +step:4100/20000 train_loss:2.0220 train_time:412181ms step_avg:100.53ms +step:4200/20000 train_loss:2.1582 train_time:422467ms step_avg:100.59ms +step:4300/20000 train_loss:2.0639 train_time:431253ms step_avg:100.29ms +step:4400/20000 train_loss:2.0361 train_time:441677ms step_avg:100.38ms +step:4500/20000 train_loss:2.1269 train_time:451993ms step_avg:100.44ms +step:4500/20000 val_loss:2.0494 val_bpb:1.2138 train_time:452034ms step_avg:100.45ms +swa:start step:4600 +step:4600/20000 train_loss:1.8460 train_time:462381ms step_avg:100.52ms +step:4700/20000 train_loss:2.2388 train_time:471273ms step_avg:100.27ms +step:4800/20000 train_loss:2.4293 train_time:481774ms step_avg:100.37ms +step:4900/20000 train_loss:2.0529 train_time:492243ms step_avg:100.46ms +step:5000/20000 train_loss:2.1032 train_time:502591ms step_avg:100.52ms +step:5000/20000 val_loss:2.0252 val_bpb:1.1994 train_time:502682ms step_avg:100.54ms +qat:activated step:5084 elapsed:536517ms (89.4% wallclock) +step:5100/20000 train_loss:2.1292 train_time:564634ms step_avg:110.71ms +step:5200/20000 train_loss:2.0344 train_time:573415ms step_avg:110.27ms +step:5300/20000 train_loss:2.0009 train_time:583883ms step_avg:110.17ms +step:5400/20000 train_loss:2.0437 train_time:594307ms step_avg:110.06ms +step:5464/20000 val_loss:1.9995 val_bpb:1.1842 train_time:600015ms step_avg:109.81ms +stopping_early: wallclock_cap train_time:600015ms step:5464/20000 +peak memory allocated: 21470 MiB reserved: 22170 MiB +swa:applying averaged 5 checkpoints +Serialized model: 87413062 bytes +Code size: 64426 bytes +Total submission size: 87477488 bytes +Serialized model int6+zstd: 15884217 bytes +Total submission size: 15948643 bytes +final_eval_mode:sliding_window stride:64 +final_int8_zstd_roundtrip val_loss:1.9635 val_bpb:1.1629 eval_time:235968ms +final_int8_zstd_roundtrip_exact val_loss:1.96353693 val_bpb:1.16292025 diff --git a/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/train_gpt.py b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/train_gpt.py new file mode 100644 index 0000000000..6c71b6b22f --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_STE_QAT_MLP3x_SmearBigram_LoRATTT/train_gpt.py @@ -0,0 +1,1396 @@ +""" +Late STE QAT + Int6 MLP3x + SmearGate + BigramHash + OrthoInit + Overtone + SWA + SGD TTT + +Combines the best proven techniques with evidence-based innovations: +1. Late STE QAT (last 15% of wallclock): avoids corrupting Muon's momentum + subspace while closing the int6 quantization gap. Per #76 (validated 1.1599). +2. Full-model SGD TTT: single-epoch adaptation at eval time. Per #152, this + outperforms LoRA TTT (0.034 BPB gain). LoRA TTT + SmearGate hurt scores (#178). +3. WD=0.038 + LR=0.025 matching best validated submissions (#179, #194). +""" + +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 zstandard + +_COMPRESSOR = "zstd" + +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 +# ----------------------------- + +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", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + 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)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 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", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.01)) + muon_weight_decay = float(os.environ.get("MUON_WEIGHT_DECAY", 0.038)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 32)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 4096)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.5)) + swa_every = int(os.environ.get("SWA_EVERY", 200)) + + qat_start_frac = float(os.environ.get("QAT_START_FRAC", 0.85)) + qat_lr_factor = float(os.environ.get("QAT_LR_FACTOR", 0.5)) + + sgd_ttt_enabled = bool(int(os.environ.get("SGD_TTT_ENABLED", "1"))) + sgd_ttt_lr = float(os.environ.get("SGD_TTT_LR", 3e-4)) + sgd_ttt_momentum = float(os.environ.get("SGD_TTT_MOMENTUM", 0.95)) + sgd_ttt_batch_seqs = int(os.environ.get("SGD_TTT_BATCH_SEQS", 32)) + + ttt_lora_rank = int(os.environ.get("TTT_LORA_RANK", 8)) + ttt_lora_lr = float(os.environ.get("TTT_LORA_LR", 0.01)) + ttt_chunk_size = int(os.environ.get("TTT_CHUNK_SIZE", 256)) + ttt_eval_seq_len = int(os.environ.get("TTT_EVAL_SEQ_LEN", 1024)) + ttt_batch_size = int(os.environ.get("TTT_BATCH_SIZE", 64)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +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 Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + 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) + 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: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + 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 seq_len={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, model: nn.Module, rank: int, world_size: int, + device: torch.device, grad_accum_steps: int, val_tokens: Tensor, + base_bytes_lut: Tensor, has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError("VAL_BATCH_SIZE too small") + 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 (INT6 mixed + zstd) +# ----------------------------- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + p for p 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,smear,bigram.scale", + ).split(",") if p +) +FP16_KEEP_NAME_PATTERNS = tuple( + p for p in os.environ.get("FP16_KEEP_NAME_PATTERNS", "tok_emb,blocks.8.attn.c_k").split(",") if p +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + p for p in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") if p +) +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 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(1e-12).to(torch.float16) + scale = scale.clamp_min(torch.finfo(torch.float16).tiny) + 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(max(amax / 31.0, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8) + 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 mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + 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() <= 65536: + 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 any(p in name for p in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + 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: + t32 = t.float() + clip_abs = torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) if t32.ndim == 2 and t32.numel() else t32.abs().amax(dim=1) if t32.ndim == 2 else torch.tensor(0.0) + if t32.ndim == 2: + clipped = torch.clamp(t32, -clip_abs[:, None], clip_abs[:, None]) + sc = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / sc[:, None]), -127, 127).to(torch.int8) + result[name + ".q"] = q + result[name + ".scale"] = sc.to(torch.float16) + else: + amax = float(t32.abs().max().item()) if t32.numel() else 0.0 + sc = torch.tensor(amax / 127.0 if amax > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -amax, amax) / sc), -127, 127).to(torch.int8) + result[name + ".q"] = q + result[name + ".scale"] = sc + 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] = {} + for name, orig in template_sd.items(): + info = meta[name] + 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 (with STE Int6 QAT) +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +def _ste_fake_quantize_int6(w: Tensor) -> Tensor: + """STE fake-quantize: simulate int6 per-row quantization in the forward pass. + Gradients pass through via straight-through estimator.""" + if w.ndim < 2: + return w + with torch.no_grad(): + row_max = w.abs().amax(dim=1, keepdim=True).clamp_min(1e-12) + scale = row_max / 31.0 + w_q = torch.clamp(torch.round(w / scale), -32, 31) * scale + return w + (w_q - w).detach() + + +class CastedLinear(nn.Linear): + """Weights in fp32, cast to compute dtype at matmul time. + With STE QAT: fake-quantizes weights during training to teach + the model to be resilient to int6 noise.""" + _ste_qat: bool = False + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if self.training and self._ste_qat: + w = _ste_fake_quantize_int6(w) + 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(p in name for p 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): + 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, q_delta=None, v_delta=None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + (q_delta if q_delta is not None else 0) + k = self.c_k(x) + v = self.c_v(x) + (v_delta if v_delta is not None else 0) + q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = k.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + 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] + # Torch in this environment doesn't support enable_gqa kwarg. + # Implement GQA by repeating K/V heads to match Q heads. + if self.num_kv_heads != self.num_heads: + rep = self.num_heads // self.num_kv_heads + k = k.repeat_interleave(rep, dim=1) + v = v.repeat_interleave(rep, dim=1) + y = F.scaled_dot_product_attention(q, k, v, attn_mask=None, is_causal=True) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + 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 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 Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, + rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor, q_delta_fn=None, v_delta_fn=None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x) + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out = self.attn(n, qd, vd) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +class GPT(nn.Module): + def __init__(self, vocab_size: int, num_layers: int, model_dim: int, num_heads: int, + num_kv_heads: int, mlp_mult: float, 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): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + 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.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.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ 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(tied_embed_init_std, num_layers) + + def _init_weights(self, tied_embed_init_std: float, num_layers: int) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=tied_embed_init_std) + with torch.no_grad(): + U, S, V = torch.linalg.svd(self.tok_emb.weight.data, full_matrices=False) + target_S = S[0] * (1.0 / torch.arange(1, S.shape[0] + 1, dtype=S.dtype)) ** 0.5 + self.tok_emb.weight.data = (U * target_S[None, :]) @ V + 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)) + for i, block in enumerate(self.blocks): + with torch.no_grad(): + phase = torch.sigmoid(torch.tensor(3.0 * (i / max(num_layers - 1, 1) - 0.5))) + block.resid_mix.data[0] = phase * torch.ones(block.resid_mix.shape[1]) + block.resid_mix.data[1] = (1 - phase) * torch.ones(block.resid_mix.shape[1]) + + def forward(self, input_ids: Tensor, target_ids: Tensor, lora=None) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + 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): + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x = self.blocks[i](x, x0, qd, vd) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x = self.blocks[bi](x, x0, qd, vd) + 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) + logits = logits_proj + (lora.lm_head_lora(x) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits / self.logit_softcap) + if lora: + bsz, sl, V = logits.shape + return F.cross_entropy( + logits.float().reshape(-1, V), target_ids.reshape(-1), reduction="none").reshape(bsz, sl) + return F.cross_entropy(logits.float().reshape(-1, logits.size(-1)), target_ids.reshape(-1), reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + 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 EVAL +# ----------------------------- + +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, +) -> tuple[float, float]: + seq_len = 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] + my_s = (len(window_starts) * rank) // world_size + my_e = (len(window_starts) * (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() + 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 = base_model.forward_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() + return val_loss, val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + +# ----------------------------- +# TEST-TIME TRAINING (LoRA) +# ----------------------------- + +BOS_ID = 1 + +class BatchedLinearLoRA(nn.Module): + def __init__(self, bsz: int, in_features: int, out_features: int, rank: int): + super().__init__() + self.in_features = in_features + self.A = nn.Parameter(torch.empty(bsz, rank, in_features)) + self.B = nn.Parameter(torch.zeros(bsz, out_features, rank)) + self.reset() + + def forward(self, x: Tensor) -> Tensor: + return (x @ self.A.transpose(1, 2)) @ self.B.transpose(1, 2) + + def reset(self) -> None: + bound = 1.0 / math.sqrt(self.in_features) + with torch.no_grad(): + self.A.uniform_(-bound, bound) + self.B.zero_() + +class BatchedTTTLoRA(nn.Module): + def __init__(self, bsz: int, model: GPT, rank: int): + super().__init__() + dim = model.tok_emb.embedding_dim + vocab = model.tok_emb.num_embeddings + self.lm_head_lora = BatchedLinearLoRA(bsz, dim, vocab, rank) + self.q_loras = nn.ModuleList() + self.v_loras = nn.ModuleList() + for block in model.blocks: + self.q_loras.append(BatchedLinearLoRA(bsz, dim, block.attn.c_q.weight.shape[0], rank)) + self.v_loras.append(BatchedLinearLoRA(bsz, dim, block.attn.c_v.weight.shape[0], rank)) + + def reset(self) -> None: + for m in self.modules(): + if isinstance(m, BatchedLinearLoRA): + m.reset() + +def _reset_ttt_optimizer(opt): + for group in opt.param_groups: + for p in group['params']: + s = opt.state.get(p) + if not s: + continue + s['exp_avg'].zero_() + s['exp_avg_sq'].zero_() + s['step'].fill_(0) + +def _find_docs(all_tokens: Tensor) -> list[tuple[int, int]]: + bos_positions = (all_tokens == BOS_ID).nonzero(as_tuple=True)[0].numpy() + docs = [] + for i in range(len(bos_positions)): + start = int(bos_positions[i]) + end = int(bos_positions[i + 1]) if i + 1 < len(bos_positions) else all_tokens.numel() + if i + 1 < len(bos_positions): + end += 1 + assert end - start >= 2 + docs.append((start, end - start)) + return docs + +def _compute_chunk_window(ci: int, pred_len: int, num_chunks: int, chunk_size: int, eval_seq_len: int): + chunk_start = ci * chunk_size + chunk_end = pred_len if ci == num_chunks - 1 else (ci + 1) * chunk_size + win_start = max(0, chunk_end - eval_seq_len) + win_len = chunk_end - win_start + chunk_offset = chunk_start - win_start + chunk_len = chunk_end - chunk_start + return win_start, win_len, chunk_offset, chunk_len + +def _accumulate_bpb(ptl, x, y, batch_i, chunk_offset, chunk_len, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + loss_sum, byte_sum, token_count): + lbl = ptl[batch_i, chunk_offset:chunk_offset + chunk_len].to(torch.float64) + prev = x[batch_i, chunk_offset:chunk_offset + chunk_len] + tgt = y[batch_i, chunk_offset:chunk_offset + chunk_len] + tok_bytes = base_bytes_lut[tgt].to(torch.float64) + tok_bytes += has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev] + loss_sum += lbl.sum() + byte_sum += tok_bytes.sum() + token_count += chunk_len + +def eval_val_ttt_lora( + args: Hyperparameters, base_model: GPT, rank: int, world_size: int, + device: torch.device, base_bytes_lut: Tensor, has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + files = sorted(glob.glob(args.val_files)) + all_tokens = torch.cat([load_data_shard(Path(f)) for f in files]) + docs = _find_docs(all_tokens) + rank_docs = docs[(len(docs) * rank) // world_size : (len(docs) * (rank + 1)) // world_size] + chunk_size = args.ttt_chunk_size + eval_seq_len = args.ttt_eval_seq_len + batch_size = args.ttt_batch_size + lora_rank = args.ttt_lora_rank + rank_docs.sort(key=lambda d: (d[1] - 2) // chunk_size) + + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(False) + + lora = BatchedTTTLoRA(batch_size, base_model, lora_rank).to(device) + opt = torch.optim.Adam(lora.parameters(), lr=args.ttt_lora_lr, betas=(args.beta1, args.beta2), eps=1e-10) + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + byte_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + + for bi in range(0, len(rank_docs), batch_size): + batch = rank_docs[bi:bi + batch_size] + bsz = len(batch) + if bsz == batch_size: + cur_lora, cur_opt = lora, opt + cur_lora.reset() + _reset_ttt_optimizer(cur_opt) + else: + cur_lora = BatchedTTTLoRA(bsz, base_model, lora_rank).to(device) + cur_opt = torch.optim.Adam(cur_lora.parameters(), lr=args.ttt_lora_lr, betas=(args.beta1, args.beta2), eps=1e-10) + + pred_lens = [doc_len - 1 for _, doc_len in batch] + num_chunks = [(pl + chunk_size - 1) // chunk_size for pl in pred_lens] + max_nc = max(num_chunks) + + for ci in range(max_nc): + chunk_stats = _compute_chunk_window(ci, (ci + 1) * chunk_size, ci + 1, chunk_size, eval_seq_len) + context_size, chunk_offset = chunk_stats[1], chunk_stats[2] + active = [ci < nc for nc in num_chunks] + needs_train = any(ci < nc - 1 for nc in num_chunks) + + x = torch.zeros(bsz, context_size, dtype=torch.int64, device=device) + y = torch.zeros(bsz, context_size, dtype=torch.int64, device=device) + doc_info = [] + for b in range(bsz): + if not active[b]: + doc_info.append((0, 0)) + continue + ds, dl = batch[b] + ws, wl, co, cl = _compute_chunk_window(ci, pred_lens[b], num_chunks[b], chunk_size, eval_seq_len) + chunk = all_tokens[ds + ws: ds + ws + wl + 1] + toks = chunk.to(dtype=torch.int64, device=device) + x[b, :wl] = toks[:-1] + y[b, :wl] = toks[1:] + doc_info.append((co, cl)) + + if needs_train: + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + ptl = base_model(x, y, lora=cur_lora) + else: + with torch.no_grad(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + ptl = base_model(x, y, lora=cur_lora) + + with torch.no_grad(): + for b in range(bsz): + if not active[b]: + continue + co, cl = doc_info[b] + _accumulate_bpb(ptl, x, y, b, co, cl, base_bytes_lut, has_leading_space_lut, + is_boundary_token_lut, loss_sum, byte_sum, token_count) + + if needs_train: + mask = torch.tensor([float(ci < num_chunks[b] - 1) for b in range(bsz)], device=device) + per_doc = ptl[:, chunk_offset:chunk_offset + chunk_size].mean(dim=-1) + cur_opt.zero_grad() + (per_doc * mask).sum().backward() + cur_opt.step() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + + val_loss = float(loss_sum.item() / token_count.item()) + val_bpb = float((loss_sum.item() / math.log(2.0)) / byte_sum.item()) + return val_loss, val_bpb + + +# ----------------------------- +# FULL-MODEL SGD TTT +# ----------------------------- + +def eval_val_sgd_ttt( + args: Hyperparameters, base_model: GPT, 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, + log_fn=None, +) -> tuple[float, float]: + """Full-model SGD test-time training: single epoch over val data, then sliding window eval. + Per PR #152: SGD with LR=3e-4, momentum=0.95 gives ~0.034 BPB improvement.""" + seq_len = args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._ste_qat = False + + base_model.train() + sgd_opt = torch.optim.SGD( + base_model.parameters(), lr=args.sgd_ttt_lr, momentum=args.sgd_ttt_momentum, + ) + + batch_seqs = args.sgd_ttt_batch_seqs + for batch_start in range(0, total_seqs, batch_seqs): + batch_end = min(batch_start + batch_seqs, total_seqs) + raw_start = batch_start * seq_len + raw_end = batch_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + sgd_opt.zero_grad() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + sgd_opt.step() + + if dist.is_available() and dist.is_initialized(): + for p in base_model.parameters(): + dist.all_reduce(p.data, op=dist.ReduceOp.AVG) + + base_model.eval() + return eval_val_sliding( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.eval_batch_seqs, + ) + + +# ----------------------------- +# 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 = "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") + 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) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0(subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, console=False) + log0("=" * 100, console=False) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError(f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}") + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(sp, args.vocab_size, device) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + base_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + ).to(device).bfloat16() + + for name, module in base_model.named_modules(): + if isinstance(module, CastedLinear): + module.float() + cat = _classify_param(name) + module._should_qat = cat in ("mlp", "attn") + module._ste_qat = False + 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 + + 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(pat in name for pat in CONTROL_TENSOR_NAME_PATTERNS)] + scalar_params = [p for name, p in block_named_params if p.ndim < 2 or any(pat in name for pat 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) + + 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) + + optimizer_tok = torch.optim.AdamW(tok_params, betas=(args.beta1, args.beta2), eps=args.adam_eps, weight_decay=args.weight_decay, fused=True) + optimizer_muon = Muon(matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, backend_steps=args.muon_backend_steps, weight_decay=args.muon_weight_decay) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.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.weight_decay, 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(f"ste_qat:late_activation at {args.qat_start_frac*100:.0f}% wallclock, lr_factor={args.qat_lr_factor}") + log0(f"muon_weight_decay:{args.muon_weight_decay} matrix_lr:{args.matrix_lr}") + log0(f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} iterations:{args.iterations} max_wallclock:{args.max_wallclock_seconds:.0f}s") + log0(f"seed:{args.seed}") + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + qat_active = False + qat_base_lrs_saved = False + 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} 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 step:{step}/{args.iterations}") + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + if not qat_active and max_wallclock_ms is not None and elapsed_ms / max_wallclock_ms >= args.qat_start_frac: + qat_active = True + for m in base_model.modules(): + if isinstance(m, CastedLinear) and getattr(m, '_should_qat', False): + m._ste_qat = True + if not qat_base_lrs_saved: + for opt in optimizers: + for group in opt.param_groups: + group["base_lr"] *= args.qat_lr_factor + qat_base_lrs_saved = True + log0(f"qat:activated step:{step} elapsed:{elapsed_ms:.0f}ms ({elapsed_ms/max_wallclock_ms*100:.1f}% wallclock)") + + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + if args.swa_enabled and scale < args.swa_start_frac and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().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] += t.detach().cpu() + 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} 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 reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB") + + if args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + current_state = base_model.state_dict() + avg_state = {name: (tensor / swa_count).to(dtype=current_state[name].dtype) for name, tensor in swa_state.items()} + base_model.load_state_dict(avg_state, strict=True) + + 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") + + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + 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 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")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + decompressed = zstandard.ZstdDecompressor().decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + base_model.load_state_dict(deq_state, strict=True) + + torch.cuda.synchronize() + t_qeval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0(f"final_eval_mode:sliding_window stride:{args.eval_stride}") + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.eval_batch_seqs, + ) + else: + 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} 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 args.sgd_ttt_enabled: + sgd_ttt_sd = {k: v.clone() for k, v in base_model.state_dict().items()} + torch._dynamo.reset() + torch.cuda.synchronize() + t_sgd_ttt = time.perf_counter() + sgd_ttt_loss, sgd_ttt_bpb = eval_val_sgd_ttt( + args, base_model, rank, world_size, device, val_tokens, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + log_fn=log0, + ) + torch.cuda.synchronize() + log0(f"final_sgd_ttt val_loss:{sgd_ttt_loss:.4f} val_bpb:{sgd_ttt_bpb:.4f} eval_time:{1000.0 * (time.perf_counter() - t_sgd_ttt):.0f}ms") + log0(f"final_sgd_ttt_exact val_loss:{sgd_ttt_loss:.8f} val_bpb:{sgd_ttt_bpb:.8f}") + base_model.load_state_dict(sgd_ttt_sd, strict=True) + + ttt_lora_enabled = bool(int(os.environ.get("TTT_LORA_ENABLED", "0"))) + if ttt_lora_enabled: + torch._dynamo.reset() + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt_lora( + args, base_model, rank, world_size, device, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0(f"final_int8_ttt_lora val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main()