diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/README.md b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/README.md new file mode 100644 index 0000000000..7e88553253 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/README.md @@ -0,0 +1,77 @@ +# Test-Time Training (TTT) with Full-Model SGD Adaptation + +**Best Score: 1.17436 BPB** (val_loss: 1.9829) + +**Author:** FarnsworthTech +**X:** [@FARNSWORTHLLC](https://x.com/FARNSWORTHLLC) +**GitHub:** [timowhite88](https://github.com/timowhite88) +**Email:** timeowhite88@gmail.com / timeowhite88@icloud.com + +## Results + +| Run | TTT LR | TTT Epochs | Seed | Static BPB | Final BPB | Steps | Train Time | +|-----|--------|-----------|------|------------|-----------|-------|------------| +| Conservative | 3e-4 | 1 | 1337 | 1.2105 | 1.1767 | 9647 | 599.988s | +| **Aggressive** | **2e-3** | **2** | **1337** | **1.2087** | **1.1744** | **9728** | **601.1s** | + +## Approach + +**Test-time training** adapts the model's weights during evaluation, reclaiming the unused 10-min eval budget as adaptive compression (analogous to Lempel-Ziv). + +### Key Insight + +The competition allocates 10 minutes for training and 10 minutes for evaluation. Standard submissions use only a fraction of the eval budget. TTT performs online gradient descent on the validation data itself before scoring — every parameter adapts to the validation distribution. + +### Training Phase (10 min, 8xH100 SXM) + +- **Architecture:** 9-layer, 512-dim, GQA (8 heads / 4 KV), tied embeddings +- **Optimizer:** Muon + Adam with standard LR schedule +- **Tokenizer:** SP-1024 BPE (FineWeb 10B) +- **Model params:** 18,897,488 + +### TTT Eval Phase (~349s of 600s budget) + +1. Decompress int8+zlib artifact back to full precision +2. **TTT adaptation:** 2 epochs of full-model SGD (lr=0.002, momentum=0.9, batch_size=32) — 311s +3. **Sliding window eval** (stride=64, seq_len=1024) — 38s + +**TTT improved BPB from 1.2087 to 1.1744 (2.84% gain at zero parameter cost)** + +### Why Full-Model SGD Instead of LoRA? + +We tested LoRA-based TTT (rank-8 on Q/V/lm_head) but found full-model SGD with aggressive LR outperforms it. With 2 epochs and lr=0.002, every parameter adapts. Momentum of 0.9 provides smoothing to prevent catastrophic forgetting while allowing fast adaptation. + +## Artifact + +- **Model:** 18,897,488 parameters +- **Compressed (int8+zlib):** 15,270,194 bytes +- **Code:** 58,683 bytes +- **Total:** 15,328,877 bytes (< 16,000,000 byte cap) + +## Compliance + +| Rule | Limit | Actual | +|------|-------|--------| +| Training time | 600s | 599.988s (run1) / 601.1s (run2) | +| Eval time | 600s | ~349s (311s TTT + 38s eval) | +| GPUs | 8xH100 SXM | 8x NVIDIA H100 80GB HBM3 | +| Artifact size | 16,000,000 bytes | 15,328,877 bytes | +| int8 roundtrip | Required | Verified | + +## Reproducibility + +```bash +# Aggressive TTT (best score) +TTT_LR=0.002 TTT_EPOCHS=2 TTT_MOMENTUM=0.9 torchrun --standalone --nproc_per_node=8 train_gpt.py + +# Conservative TTT +TTT_LR=3e-4 TTT_EPOCHS=1 TTT_MOMENTUM=0.95 torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## Hardware & Environment + +- 8x NVIDIA H100 80GB HBM3 (SXM), RunPod cloud +- Ubuntu 22.04.5 LTS, Kernel 6.17.0-1008-nvidia +- PyTorch 2.6.0+cu124, CUDA 12.4 +- Driver: 580.126.09 +- Peak memory: 11,389 MiB per GPU diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/submission.json b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/submission.json new file mode 100644 index 0000000000..4ed9e69858 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/submission.json @@ -0,0 +1,23 @@ +{ + "track": "10min_16mb", + "date": "2026-03-20", + "name": "Test-Time Training (TTT) with Full-Model SGD Adaptation", + "author": "FarnsworthTech", + "github_id": "timowhite88", + "x_handle": "FARNSWORTHLLC", + "email": ["timeowhite88@gmail.com", "timeowhite88@icloud.com"], + "blurb": "Standard 9x512 KV4 baseline training + aggressive test-time training during eval: 2-epoch full-model SGD adaptation on validation data (lr=0.002, momentum=0.9) before sliding-window scoring. No architecture changes -- all gains come from smarter use of the 10-min eval budget.", + "seed_results": { + "1337_conservative": {"val_loss": 1.98674420, "val_bpb": 1.17666183, "steps": 9647, "ms_per_step": 62.19, "train_time_s": 599.988, "ttt_lr": 0.0003, "ttt_epochs": 1}, + "1337_aggressive": {"val_loss": 1.98286428, "val_bpb": 1.17436392, "steps": 9728, "ms_per_step": 61.79, "train_time_s": 601.124, "ttt_lr": 0.002, "ttt_epochs": 2} + }, + "mean_val_loss": 1.98480424, + "mean_val_bpb": 1.17551288, + "best_val_loss": 1.98286428, + "best_val_bpb": 1.17436392, + "artifact_bytes": 15328877, + "code_bytes": 58683, + "hardware": "8x NVIDIA H100 80GB HBM3 (SXM), RunPod", + "pytorch": "2.6.0+cu124", + "cuda": "12.4" +} diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_best_seed42.log b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_best_seed42.log new file mode 100644 index 0000000000..2cc0ad0e08 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_best_seed42.log @@ -0,0 +1,127 @@ +W0320 02:49:29.248000 32450 torch/distributed/run.py:792] +W0320 02:49:29.248000 32450 torch/distributed/run.py:792] ***************************************** +W0320 02:49:29.248000 32450 torch/distributed/run.py:792] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0320 02:49:29.248000 32450 torch/distributed/run.py:792] ***************************************** +logs/77b9d3ba-ce0c-4531-b141-e092aed6570e.txt +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:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.1 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:598.000 +seed:42 +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:6.9308 val_bpb:4.1048 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9310 train_time:35ms step_avg:35.08ms +step:2/20000 train_loss:23.4148 train_time:81ms step_avg:40.67ms +step:3/20000 train_loss:8.8953 train_time:136ms step_avg:45.23ms +step:4/20000 train_loss:6.2118 train_time:189ms step_avg:47.29ms +step:5/20000 train_loss:6.2824 train_time:244ms step_avg:48.84ms +step:6/20000 train_loss:7.5010 train_time:299ms step_avg:49.78ms +step:7/20000 train_loss:6.5008 train_time:352ms step_avg:50.27ms +step:8/20000 train_loss:6.4814 train_time:405ms step_avg:50.66ms +step:9/20000 train_loss:6.4402 train_time:459ms step_avg:51.04ms +step:10/20000 train_loss:6.3779 train_time:513ms step_avg:51.31ms +step:200/20000 train_loss:2.9362 train_time:10006ms step_avg:50.03ms +step:400/20000 train_loss:2.3714 train_time:20006ms step_avg:50.02ms +step:600/20000 train_loss:2.5673 train_time:30025ms step_avg:50.04ms +step:800/20000 train_loss:2.3079 train_time:42360ms step_avg:52.95ms +step:1000/20000 train_loss:2.3930 train_time:54878ms step_avg:54.88ms +step:1000/20000 val_loss:2.3559 val_bpb:1.3953 train_time:54907ms step_avg:54.91ms +step:1200/20000 train_loss:2.4060 train_time:67573ms step_avg:56.31ms +step:1400/20000 train_loss:2.4496 train_time:79951ms step_avg:57.11ms +step:1600/20000 train_loss:2.1277 train_time:92417ms step_avg:57.76ms +step:1800/20000 train_loss:2.2261 train_time:105064ms step_avg:58.37ms +step:2000/20000 train_loss:2.2801 train_time:117669ms step_avg:58.83ms +step:2000/20000 val_loss:2.2651 val_bpb:1.3415 train_time:117696ms step_avg:58.85ms +step:2200/20000 train_loss:2.1122 train_time:130285ms step_avg:59.22ms +step:2400/20000 train_loss:2.2271 train_time:142935ms step_avg:59.56ms +step:2600/20000 train_loss:2.4497 train_time:155499ms step_avg:59.81ms +step:2800/20000 train_loss:2.2778 train_time:167992ms step_avg:60.00ms +step:3000/20000 train_loss:2.2630 train_time:180657ms step_avg:60.22ms +step:3000/20000 val_loss:2.2337 val_bpb:1.3229 train_time:180684ms step_avg:60.23ms +step:3200/20000 train_loss:2.2304 train_time:192895ms step_avg:60.28ms +step:3400/20000 train_loss:2.2033 train_time:205615ms step_avg:60.47ms +step:3600/20000 train_loss:2.1643 train_time:218188ms step_avg:60.61ms +step:3800/20000 train_loss:2.2618 train_time:230690ms step_avg:60.71ms +step:4000/20000 train_loss:2.2107 train_time:240729ms step_avg:60.18ms +step:4000/20000 val_loss:2.2180 val_bpb:1.3136 train_time:240759ms step_avg:60.19ms +step:4200/20000 train_loss:2.2218 train_time:250866ms step_avg:59.73ms +step:4400/20000 train_loss:2.1627 train_time:260915ms step_avg:59.30ms +step:4600/20000 train_loss:2.0233 train_time:270980ms step_avg:58.91ms +step:4800/20000 train_loss:2.3107 train_time:281051ms step_avg:58.55ms +step:5000/20000 train_loss:2.0872 train_time:291122ms step_avg:58.22ms +step:5000/20000 val_loss:2.2067 val_bpb:1.3069 train_time:291150ms step_avg:58.23ms +step:5200/20000 train_loss:2.2265 train_time:301197ms step_avg:57.92ms +step:5400/20000 train_loss:2.2461 train_time:311267ms step_avg:57.64ms +step:5600/20000 train_loss:2.2426 train_time:321330ms step_avg:57.38ms +step:5800/20000 train_loss:2.2036 train_time:331397ms step_avg:57.14ms +step:6000/20000 train_loss:2.2767 train_time:341467ms step_avg:56.91ms +step:6000/20000 val_loss:2.2037 val_bpb:1.3052 train_time:341495ms step_avg:56.92ms +step:6200/20000 train_loss:2.1549 train_time:351541ms step_avg:56.70ms +step:6400/20000 train_loss:2.2300 train_time:361608ms step_avg:56.50ms +step:6600/20000 train_loss:2.1889 train_time:371669ms step_avg:56.31ms +step:6800/20000 train_loss:2.2550 train_time:381729ms step_avg:56.14ms +step:7000/20000 train_loss:2.2885 train_time:391782ms step_avg:55.97ms +step:7000/20000 val_loss:2.1962 val_bpb:1.3007 train_time:391810ms step_avg:55.97ms +step:7200/20000 train_loss:2.2673 train_time:401840ms step_avg:55.81ms +step:7400/20000 train_loss:2.1775 train_time:411895ms step_avg:55.66ms +step:7600/20000 train_loss:2.0564 train_time:421960ms step_avg:55.52ms +step:7800/20000 train_loss:2.2142 train_time:432015ms step_avg:55.39ms +step:8000/20000 train_loss:2.1863 train_time:442067ms step_avg:55.26ms +step:8000/20000 val_loss:2.1897 val_bpb:1.2968 train_time:442104ms step_avg:55.26ms +step:8200/20000 train_loss:2.2456 train_time:452128ms step_avg:55.14ms +step:8400/20000 train_loss:2.1945 train_time:462250ms step_avg:55.03ms +step:8600/20000 train_loss:2.1921 train_time:472291ms step_avg:54.92ms +step:8800/20000 train_loss:2.1568 train_time:482331ms step_avg:54.81ms +step:9000/20000 train_loss:2.0631 train_time:492375ms step_avg:54.71ms +step:9000/20000 val_loss:2.1615 val_bpb:1.2802 train_time:492404ms step_avg:54.71ms +step:9200/20000 train_loss:2.1209 train_time:502421ms step_avg:54.61ms +step:9400/20000 train_loss:2.1648 train_time:512469ms step_avg:54.52ms +step:9600/20000 train_loss:2.1662 train_time:522528ms step_avg:54.43ms +step:9800/20000 train_loss:2.0842 train_time:532580ms step_avg:54.34ms +step:10000/20000 train_loss:2.1129 train_time:542628ms step_avg:54.26ms +step:10000/20000 val_loss:2.1101 val_bpb:1.2497 train_time:542656ms step_avg:54.27ms +step:10200/20000 train_loss:2.0546 train_time:552673ms step_avg:54.18ms +step:10400/20000 train_loss:2.0766 train_time:562720ms step_avg:54.11ms +step:10600/20000 train_loss:1.9417 train_time:575498ms step_avg:54.29ms +step:10800/20000 train_loss:2.1395 train_time:588327ms step_avg:54.47ms +step:10940/20000 val_loss:2.0458 val_bpb:1.2116 train_time:597972ms step_avg:54.66ms +stopping_early: wallclock_cap train_time:597972ms step:10940/20000 +peak memory allocated: 11389 MiB reserved: 11704 MiB +Serialized model: 74573582 bytes +Code size: 58683 bytes +Total submission size: 74632265 bytes +Serialized model int8+zlib: 15276671 bytes (payload:19552576 raw_torch:19601902 payload_ratio:3.81x) +Total submission size int8+zlib: 15335354 bytes +ttt:start lr=0.003 momentum=0.9 epochs=3 +ttt_epoch:1/3 loss:1.9449 time:155.6s +ttt_epoch:2/3 loss:1.9440 time:311.1s +ttt_epoch:3/3 loss:1.9438 time:466.6s +ttt:done elapsed=466.6s +Compiling forward_logits for sliding window eval (stride=64, seq_len=1024)... +Compilation done, starting sliding window eval... +final_int8_zlib_roundtrip val_loss:1.9871 val_bpb:1.1769 eval_time:37705ms +final_int8_zlib_roundtrip_exact val_loss:1.98714306 val_bpb:1.17689805 diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_combo_seed1337.log b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_combo_seed1337.log new file mode 100644 index 0000000000..1d430ef0a1 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_combo_seed1337.log @@ -0,0 +1,124 @@ +W0320 02:29:13.150000 27510 torch/distributed/run.py:792] +W0320 02:29:13.150000 27510 torch/distributed/run.py:792] ***************************************** +W0320 02:29:13.150000 27510 torch/distributed/run.py:792] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0320 02:29:13.150000 27510 torch/distributed/run.py:792] ***************************************** +logs/69f8bc05-03cd-4471-8cb2-ea86b915ef67.txt +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:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.1 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:598.000 +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:6.9315 val_bpb:4.1052 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9314 train_time:36ms step_avg:35.61ms +step:2/20000 train_loss:23.4492 train_time:81ms step_avg:40.68ms +step:3/20000 train_loss:8.7064 train_time:134ms step_avg:44.82ms +step:4/20000 train_loss:6.1710 train_time:189ms step_avg:47.21ms +step:5/20000 train_loss:6.2679 train_time:250ms step_avg:49.94ms +step:6/20000 train_loss:7.4806 train_time:303ms step_avg:50.49ms +step:7/20000 train_loss:6.4926 train_time:356ms step_avg:50.90ms +step:8/20000 train_loss:6.4831 train_time:410ms step_avg:51.29ms +step:9/20000 train_loss:6.4518 train_time:465ms step_avg:51.67ms +step:10/20000 train_loss:6.3977 train_time:520ms step_avg:51.96ms +step:200/20000 train_loss:2.9477 train_time:10139ms step_avg:50.70ms +step:400/20000 train_loss:2.3877 train_time:20147ms step_avg:50.37ms +step:600/20000 train_loss:2.5664 train_time:30178ms step_avg:50.30ms +step:800/20000 train_loss:2.3115 train_time:40204ms step_avg:50.26ms +step:1000/20000 train_loss:2.3882 train_time:50250ms step_avg:50.25ms +step:1000/20000 val_loss:2.3505 val_bpb:1.3921 train_time:50278ms step_avg:50.28ms +step:1200/20000 train_loss:2.3988 train_time:60309ms step_avg:50.26ms +step:1400/20000 train_loss:2.4455 train_time:73083ms step_avg:52.20ms +step:1600/20000 train_loss:2.1216 train_time:85653ms step_avg:53.53ms +step:1800/20000 train_loss:2.2209 train_time:98040ms step_avg:54.47ms +step:2000/20000 train_loss:2.2765 train_time:110619ms step_avg:55.31ms +step:2000/20000 val_loss:2.2607 val_bpb:1.3389 train_time:110647ms step_avg:55.32ms +step:2200/20000 train_loss:2.1012 train_time:123001ms step_avg:55.91ms +step:2400/20000 train_loss:2.2255 train_time:135534ms step_avg:56.47ms +step:2600/20000 train_loss:2.4458 train_time:148146ms step_avg:56.98ms +step:2800/20000 train_loss:2.2689 train_time:161130ms step_avg:57.55ms +step:3000/20000 train_loss:2.2617 train_time:173646ms step_avg:57.88ms +step:3000/20000 val_loss:2.2307 val_bpb:1.3212 train_time:173674ms step_avg:57.89ms +step:3200/20000 train_loss:2.2201 train_time:186281ms step_avg:58.21ms +step:3400/20000 train_loss:2.2000 train_time:198939ms step_avg:58.51ms +step:3600/20000 train_loss:2.1600 train_time:211446ms step_avg:58.73ms +step:3800/20000 train_loss:2.2570 train_time:224052ms step_avg:58.96ms +step:4000/20000 train_loss:2.2038 train_time:236704ms step_avg:59.18ms +step:4000/20000 val_loss:2.2126 val_bpb:1.3104 train_time:236731ms step_avg:59.18ms +step:4200/20000 train_loss:2.2164 train_time:251914ms step_avg:59.98ms +step:4400/20000 train_loss:2.1580 train_time:264251ms step_avg:60.06ms +step:4600/20000 train_loss:2.0172 train_time:276786ms step_avg:60.17ms +step:4800/20000 train_loss:2.3106 train_time:289183ms step_avg:60.25ms +step:5000/20000 train_loss:2.0859 train_time:301581ms step_avg:60.32ms +step:5000/20000 val_loss:2.2024 val_bpb:1.3044 train_time:301608ms step_avg:60.32ms +step:5200/20000 train_loss:2.2243 train_time:314274ms step_avg:60.44ms +step:5400/20000 train_loss:2.2420 train_time:326769ms step_avg:60.51ms +step:5600/20000 train_loss:2.2476 train_time:339114ms step_avg:60.56ms +step:5800/20000 train_loss:2.1970 train_time:349168ms step_avg:60.20ms +step:6000/20000 train_loss:2.2782 train_time:359224ms step_avg:59.87ms +step:6000/20000 val_loss:2.2022 val_bpb:1.3042 train_time:359253ms step_avg:59.88ms +step:6200/20000 train_loss:2.1487 train_time:369281ms step_avg:59.56ms +step:6400/20000 train_loss:2.2251 train_time:379352ms step_avg:59.27ms +step:6600/20000 train_loss:2.1856 train_time:389422ms step_avg:59.00ms +step:6800/20000 train_loss:2.2454 train_time:399481ms step_avg:58.75ms +step:7000/20000 train_loss:2.2839 train_time:409544ms step_avg:58.51ms +step:7000/20000 val_loss:2.1920 val_bpb:1.2982 train_time:409570ms step_avg:58.51ms +step:7200/20000 train_loss:2.2721 train_time:419616ms step_avg:58.28ms +step:7400/20000 train_loss:2.1781 train_time:429677ms step_avg:58.06ms +step:7600/20000 train_loss:2.0575 train_time:439745ms step_avg:57.86ms +step:7800/20000 train_loss:2.2053 train_time:449814ms step_avg:57.67ms +step:8000/20000 train_loss:2.1729 train_time:459882ms step_avg:57.49ms +step:8000/20000 val_loss:2.1794 val_bpb:1.2908 train_time:459909ms step_avg:57.49ms +step:8200/20000 train_loss:2.2287 train_time:469948ms step_avg:57.31ms +step:8400/20000 train_loss:2.1777 train_time:480088ms step_avg:57.15ms +step:8600/20000 train_loss:2.1700 train_time:490151ms step_avg:56.99ms +step:8800/20000 train_loss:2.1290 train_time:500218ms step_avg:56.84ms +step:9000/20000 train_loss:2.0438 train_time:510280ms step_avg:56.70ms +step:9000/20000 val_loss:2.1374 val_bpb:1.2659 train_time:510308ms step_avg:56.70ms +step:9200/20000 train_loss:2.0970 train_time:520335ms step_avg:56.56ms +step:9400/20000 train_loss:2.1335 train_time:533117ms step_avg:56.71ms +step:9600/20000 train_loss:2.1340 train_time:545963ms step_avg:56.87ms +step:9800/20000 train_loss:2.0488 train_time:558978ms step_avg:57.04ms +step:10000/20000 train_loss:2.0760 train_time:571493ms step_avg:57.15ms +step:10000/20000 val_loss:2.0713 val_bpb:1.2267 train_time:571520ms step_avg:57.15ms +step:10200/20000 train_loss:2.0105 train_time:584230ms step_avg:57.28ms +step:10400/20000 train_loss:2.0315 train_time:596914ms step_avg:57.40ms +step:10421/20000 val_loss:2.0432 val_bpb:1.2101 train_time:597987ms step_avg:57.38ms +stopping_early: wallclock_cap train_time:597987ms step:10421/20000 +peak memory allocated: 11390 MiB reserved: 11704 MiB +Serialized model: 74573582 bytes +Code size: 58077 bytes +Total submission size: 74631659 bytes +Serialized model int8+zlib: 15333252 bytes (payload:19552576 raw_torch:19601902 payload_ratio:3.81x) +Total submission size int8+zlib: 15391329 bytes +ttt:start lr=0.002 momentum=0.9 epochs=2 +ttt_epoch:1/2 loss:1.9425 time:155.6s +ttt_epoch:2/2 loss:1.9420 time:311.0s +ttt:done elapsed=311.0s +Compiling forward_logits for sliding window eval (stride=64, seq_len=1024)... +Compilation done, starting sliding window eval... +final_int8_zlib_roundtrip val_loss:1.9852 val_bpb:1.1757 eval_time:37787ms +final_int8_zlib_roundtrip_exact val_loss:1.98518771 val_bpb:1.17573998 diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_gpt.py b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_gpt.py new file mode 100644 index 0000000000..626a6f534d --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_gpt.py @@ -0,0 +1,1378 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: `train_gpt.py` and `train_gpt_mlx.py` must never be longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 2500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 0)) # 0 = same as train_seq_len + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) # sliding window stride (0 = non-overlapping) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 10)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + + # Optimizer hyperparameters. + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.10)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + + # Test-time training (TTT) during evaluation. + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", "3e-4")) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", "0.95")) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "1")) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Scale correction from Muon reference implementations. + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + seq_len_override: int = 0, +) -> tuple[float, float]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + seq_len = seq_len_override if seq_len_override > 0 else 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) + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + # Matrices get one scale per row, which usually tracks output-channel + # ranges much better than a single tensor-wide scale. + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + + # Vectors / scalars use a simpler per-tensor scale. + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Keep tied embedding in fp16 (int8 quantization hurts both input and output paths) + if "tok_emb" in name: + kept = t.to(dtype=torch.float16).contiguous() + passthrough[name] = kept + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + # Supports NTK-aware dynamic scaling for eval at longer sequence lengths. + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + 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 + ): + # NTK-aware dynamic scaling when eval seq_len > train seq_len + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + adjusted_base = self.base * (scale ** (self.dim / (self.dim - 2))) + inv_freq = 1.0 / (adjusted_base ** (torch.arange(0, self.dim, 2, dtype=torch.float32, device=device) / self.dim)) + 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: + 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, + train_seq_len: int = 1024, + ): + 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, train_seq_len=train_seq_len) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = mlp_mult * dim + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +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, + train_seq_len: int = 1024, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len=train_seq_len) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + train_seq_len: int = 1024, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + train_seq_len=train_seq_len, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + # Overtone init: shape embedding spectrum to power-law decay (like guitar harmonics) + 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 module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + # Phase-transition resid_mix: early layers trust x0 more, late layers trust residual + num_layers = len(self.blocks) + 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) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips: list[Tensor] = [] + + # First half stores skips; second half reuses them in reverse order. + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + + x_norm = self.final_norm(x) + x_flat = x_norm.reshape(-1, x_norm.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) + loss = F.cross_entropy(logits.float(), targets, reduction="mean") + return loss + + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Forward pass returning logits [batch, seq_len, vocab]. For sliding window eval.""" + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + 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 = F.linear(x, self.tok_emb.weight.to(x.dtype)) + else: + logits = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits / self.logit_softcap) + + +def eval_val_sliding( + logits_fn, + 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, + seq_len: int, + stride: int, + eval_batch_seqs: int = 256, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with near-full context.""" + total = val_tokens.numel() - 1 + + # Build windows: (start_pos, score_offset) + windows: list[tuple[int, int]] = [] + p = 0 + while p + seq_len <= total: + s = 0 if p == 0 else (seq_len - stride) + windows.append((p, s)) + p += stride + + # Distribute across ranks + n = len(windows) + per_rank = (n + world_size - 1) // world_size + my_start = rank * per_rank + my_end = min(my_start + per_rank, n) + my_windows = windows[my_start:my_end] + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + tok_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + with torch.inference_mode(): + for i in range(0, len(my_windows), eval_batch_seqs): + batch = my_windows[i : i + eval_batch_seqs] + bs = len(batch) + + # Pad to eval_batch_seqs to avoid recompilation + x_list = [val_tokens[w : w + seq_len] for w, _ in batch] + y_list = [val_tokens[w + 1 : w + seq_len + 1] for w, _ in batch] + pad = eval_batch_seqs - bs + if pad > 0: + x_list.extend([x_list[-1]] * pad) + y_list.extend([y_list[-1]] * pad) + + x = torch.stack(x_list).to(device=device, dtype=torch.int64) + y = torch.stack(y_list).to(device=device, dtype=torch.int64) + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = logits_fn(x) + + for b in range(bs): + s = batch[b][1] + scored_logits = logits[b, s:] + scored_targets = y[b, s:] + + loss = F.cross_entropy(scored_logits.float(), scored_targets, reduction="sum") + loss_sum += loss.to(torch.float64) + ns = scored_targets.numel() + tok_count += ns + + prev = x[b, s : s + ns] + tgt = scored_targets + tb = base_bytes_lut[tgt].to(torch.int16) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.int16) + byte_count += tb.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(tok_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / tok_count).item() + bpb = val_loss / math.log(2.0) * (tok_count.item() / byte_count.item()) + return val_loss, bpb + + +# ----------------------------- +# TEST-TIME TRAINING (TTT) +# ----------------------------- +# Adapt model weights on validation data during the eval phase. +# The competition gives 10 minutes of eval compute -- TTT uses part of +# that budget to update weights via gradient descent on the test set, +# then scores with sliding window eval using the adapted weights. + +def ttt_adapt( + model: nn.Module, + device: torch.device, + val_tokens: Tensor, + seq_len: int, + lr: float = 3e-4, + momentum: float = 0.95, + epochs: int = 1, + batch_size: int = 32, + log_fn=None, +) -> float: + """Adapt model weights on validation data via SGD. Returns elapsed seconds.""" + total_seqs = (val_tokens.numel() - 1) // seq_len + params = list(model.parameters()) + for p in params: + p.requires_grad_(True) + optimizer = torch.optim.SGD(params, lr=lr, momentum=momentum) + model.train() + t0 = time.perf_counter() + for epoch in range(epochs): + for batch_start in range(0, total_seqs, batch_size): + batch_end = min(batch_start + batch_size, 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) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = model(x, y) + loss.backward() + torch.nn.utils.clip_grad_norm_(params, 1.0) + optimizer.step() + if log_fn: + elapsed = time.perf_counter() - t0 + log_fn(f"ttt_epoch:{epoch + 1}/{epochs} loss:{loss.item():.4f} time:{elapsed:.1f}s") + for p in params: + p.requires_grad_(False) + model.eval() + return time.perf_counter() - t0 + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # ----------------------------- + # DISTRIBUTED + CUDA SETUP + # ----------------------------- + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # ----------------------------- + # MODEL + OPTIMIZER SETUP + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + train_seq_len=args.train_seq_len, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + optimizer_tok = torch.optim.AdamW( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=0.01, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=0.01, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + # Decoupled weight decay for Muon-optimized matrix params (not built into Muon) + with torch.no_grad(): + for p in matrix_params: + p.mul_(1.0 - 0.02 * optimizer_muon.param_groups[0]["lr"]) + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state, then produce the compressed int8+zlib artifact. + export_state = dict(base_model.state_dict()) + + if master_process: + torch.save(export_state, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(export_state) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=False) + restore_low_dim_params_to_fp32(base_model) + + # ----------------------------- + # TEST-TIME TRAINING (TTT) ADAPTATION + # ----------------------------- + if args.ttt_enabled: + log0(f"ttt:start lr={args.ttt_lr} momentum={args.ttt_momentum} epochs={args.ttt_epochs}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + # TTT on rank 0, then broadcast adapted weights to all ranks + if rank == 0: + ttt_adapt( + base_model, device, val_tokens, args.train_seq_len, + lr=args.ttt_lr, momentum=args.ttt_momentum, + epochs=args.ttt_epochs, batch_size=32, + log_fn=log0, + ) + if distributed: + for p in base_model.parameters(): + dist.broadcast(p.data, src=0) + restore_low_dim_params_to_fp32(base_model) + torch.cuda.synchronize() + ttt_elapsed = time.perf_counter() - t_ttt + log0(f"ttt:done elapsed={ttt_elapsed:.1f}s") + + # Determine eval configuration + eval_sl = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + eval_stride = args.eval_stride + if eval_sl != args.train_seq_len: + val_tokens_eval = load_validation_tokens(args.val_files, eval_sl) + log0(f"eval_seq_len:{eval_sl} (NTK-scaled RoPE, {val_tokens_eval.numel()} tokens)") + else: + val_tokens_eval = val_tokens + + torch.cuda.synchronize() + t_qeval = time.perf_counter() + + if eval_stride > 0: + # Sliding window eval: compile forward_logits for efficiency + log0(f"Compiling forward_logits for sliding window eval (stride={eval_stride}, seq_len={eval_sl})...") + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False) + # Warmup compilation + eval_batch_seqs = 256 + warmup_x = torch.zeros(eval_batch_seqs, eval_sl, dtype=torch.int64, device=device) + base_model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + _ = compiled_logits(warmup_x) + log0("Compilation done, starting sliding window eval...") + + q_val_loss, q_val_bpb = eval_val_sliding( + compiled_logits, rank, world_size, device, + val_tokens_eval, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_sl, eval_stride, eval_batch_seqs=eval_batch_seqs, + ) + base_model.train() + else: + q_val_loss, q_val_bpb = eval_val( + args, model, rank, world_size, device, grad_accum_steps, + val_tokens_eval, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + seq_len_override=eval_sl if eval_sl != args.train_seq_len else 0, + ) + + torch.cuda.synchronize() + log0( + f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_run1_conservative.log b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_run1_conservative.log new file mode 100644 index 0000000000..e514869e66 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_run1_conservative.log @@ -0,0 +1,118 @@ +W0320 01:26:59.873000 2666 torch/distributed/run.py:792] +W0320 01:26:59.873000 2666 torch/distributed/run.py:792] ***************************************** +W0320 01:26:59.873000 2666 torch/distributed/run.py:792] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0320 01:26:59.873000 2666 torch/distributed/run.py:792] ***************************************** +logs/dbb3d2e4-fdcf-4941-a760-4578d6296af4.txt +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:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.1 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +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:6.9315 val_bpb:4.1052 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9314 train_time:35ms step_avg:35.22ms +step:2/20000 train_loss:23.4492 train_time:85ms step_avg:42.33ms +step:3/20000 train_loss:8.7064 train_time:138ms step_avg:46.00ms +step:4/20000 train_loss:6.1710 train_time:192ms step_avg:47.99ms +step:5/20000 train_loss:6.2679 train_time:246ms step_avg:49.17ms +step:6/20000 train_loss:7.4806 train_time:301ms step_avg:50.09ms +step:7/20000 train_loss:6.4926 train_time:354ms step_avg:50.60ms +step:8/20000 train_loss:6.4832 train_time:407ms step_avg:50.87ms +step:9/20000 train_loss:6.4520 train_time:462ms step_avg:51.39ms +step:10/20000 train_loss:6.3979 train_time:517ms step_avg:51.69ms +step:200/20000 train_loss:2.9686 train_time:12352ms step_avg:61.76ms +step:400/20000 train_loss:2.3924 train_time:24855ms step_avg:62.14ms +step:600/20000 train_loss:2.5700 train_time:37388ms step_avg:62.31ms +step:800/20000 train_loss:2.3192 train_time:50015ms step_avg:62.52ms +step:1000/20000 train_loss:2.3898 train_time:62671ms step_avg:62.67ms +step:1000/20000 val_loss:2.3546 val_bpb:1.3946 train_time:62699ms step_avg:62.70ms +step:1200/20000 train_loss:2.4038 train_time:74991ms step_avg:62.49ms +step:1400/20000 train_loss:2.4495 train_time:87228ms step_avg:62.31ms +step:1600/20000 train_loss:2.1252 train_time:99295ms step_avg:62.06ms +step:1800/20000 train_loss:2.2220 train_time:111636ms step_avg:62.02ms +step:2000/20000 train_loss:2.2772 train_time:124019ms step_avg:62.01ms +step:2000/20000 val_loss:2.2621 val_bpb:1.3397 train_time:124047ms step_avg:62.02ms +step:2200/20000 train_loss:2.1016 train_time:136375ms step_avg:61.99ms +step:2400/20000 train_loss:2.2267 train_time:148600ms step_avg:61.92ms +step:2600/20000 train_loss:2.4475 train_time:160941ms step_avg:61.90ms +step:2800/20000 train_loss:2.2725 train_time:173266ms step_avg:61.88ms +step:3000/20000 train_loss:2.2623 train_time:185478ms step_avg:61.83ms +step:3000/20000 val_loss:2.2331 val_bpb:1.3226 train_time:185506ms step_avg:61.84ms +step:3200/20000 train_loss:2.2297 train_time:197885ms step_avg:61.84ms +step:3400/20000 train_loss:2.1990 train_time:210042ms step_avg:61.78ms +step:3600/20000 train_loss:2.1641 train_time:222290ms step_avg:61.75ms +step:3800/20000 train_loss:2.2675 train_time:234749ms step_avg:61.78ms +step:4000/20000 train_loss:2.2082 train_time:247089ms step_avg:61.77ms +step:4000/20000 val_loss:2.2159 val_bpb:1.3124 train_time:247118ms step_avg:61.78ms +step:4200/20000 train_loss:2.2167 train_time:261687ms step_avg:62.31ms +step:4400/20000 train_loss:2.1603 train_time:274079ms step_avg:62.29ms +step:4600/20000 train_loss:2.0187 train_time:286529ms step_avg:62.29ms +step:4800/20000 train_loss:2.3133 train_time:298869ms step_avg:62.26ms +step:5000/20000 train_loss:2.0828 train_time:311150ms step_avg:62.23ms +step:5000/20000 val_loss:2.2027 val_bpb:1.3046 train_time:311178ms step_avg:62.24ms +step:5200/20000 train_loss:2.2259 train_time:323541ms step_avg:62.22ms +step:5400/20000 train_loss:2.2433 train_time:336300ms step_avg:62.28ms +step:5600/20000 train_loss:2.2458 train_time:348631ms step_avg:62.26ms +step:5800/20000 train_loss:2.2027 train_time:360971ms step_avg:62.24ms +step:6000/20000 train_loss:2.2757 train_time:373500ms step_avg:62.25ms +step:6000/20000 val_loss:2.2016 val_bpb:1.3039 train_time:373528ms step_avg:62.25ms +step:6200/20000 train_loss:2.1487 train_time:385903ms step_avg:62.24ms +step:6400/20000 train_loss:2.2212 train_time:398216ms step_avg:62.22ms +step:6600/20000 train_loss:2.1835 train_time:410441ms step_avg:62.19ms +step:6800/20000 train_loss:2.2477 train_time:422600ms step_avg:62.15ms +step:7000/20000 train_loss:2.2864 train_time:434972ms step_avg:62.14ms +step:7000/20000 val_loss:2.1920 val_bpb:1.2982 train_time:435001ms step_avg:62.14ms +step:7200/20000 train_loss:2.2655 train_time:447396ms step_avg:62.14ms +step:7400/20000 train_loss:2.1674 train_time:459841ms step_avg:62.14ms +step:7600/20000 train_loss:2.0368 train_time:472040ms step_avg:62.11ms +step:7800/20000 train_loss:2.1767 train_time:484460ms step_avg:62.11ms +step:8000/20000 train_loss:2.1375 train_time:496828ms step_avg:62.10ms +step:8000/20000 val_loss:2.1445 val_bpb:1.2701 train_time:496855ms step_avg:62.11ms +step:8200/20000 train_loss:2.1985 train_time:509180ms step_avg:62.10ms +step:8400/20000 train_loss:2.1367 train_time:523854ms step_avg:62.36ms +step:8600/20000 train_loss:2.1307 train_time:536086ms step_avg:62.34ms +step:8800/20000 train_loss:2.0820 train_time:548504ms step_avg:62.33ms +step:9000/20000 train_loss:1.9932 train_time:560621ms step_avg:62.29ms +step:9000/20000 val_loss:2.0860 val_bpb:1.2354 train_time:560649ms step_avg:62.29ms +step:9200/20000 train_loss:2.0414 train_time:572976ms step_avg:62.28ms +step:9400/20000 train_loss:2.0700 train_time:585357ms step_avg:62.27ms +step:9600/20000 train_loss:2.0779 train_time:597612ms step_avg:62.25ms +step:9647/20000 val_loss:2.0439 val_bpb:1.2105 train_time:599988ms step_avg:62.19ms +stopping_early: wallclock_cap train_time:599988ms step:9647/20000 +peak memory allocated: 11502 MiB reserved: 11640 MiB +Serialized model: 74573582 bytes +Code size: 58683 bytes +Total submission size: 74632265 bytes +Serialized model int8+zlib: 15328496 bytes (payload:19552576 raw_torch:19601902 payload_ratio:3.81x) +Total submission size int8+zlib: 15387179 bytes +ttt:start lr=0.0003 momentum=0.95 epochs=1 +ttt_epoch:1/1 loss:1.9475 time:154.9s +ttt:done elapsed=154.9s +Compiling forward_logits for sliding window eval (stride=64, seq_len=1024)... +Compilation done, starting sliding window eval... +final_int8_zlib_roundtrip val_loss:1.9867 val_bpb:1.1767 eval_time:45995ms +final_int8_zlib_roundtrip_exact val_loss:1.98674420 val_bpb:1.17666183 diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_run2_aggressive.log b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_run2_aggressive.log new file mode 100644 index 0000000000..956bdaac5b --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_run2_aggressive.log @@ -0,0 +1,119 @@ +W0320 01:50:09.236000 21987 torch/distributed/run.py:792] +W0320 01:50:09.236000 21987 torch/distributed/run.py:792] ***************************************** +W0320 01:50:09.236000 21987 torch/distributed/run.py:792] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0320 01:50:09.236000 21987 torch/distributed/run.py:792] ***************************************** +logs/4cbb669f-7c97-4049-95c7-fcb711d7335a.txt +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:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.1 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +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:6.9315 val_bpb:4.1052 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9314 train_time:35ms step_avg:35.14ms +step:2/20000 train_loss:23.4492 train_time:82ms step_avg:40.81ms +step:3/20000 train_loss:8.7064 train_time:135ms step_avg:45.09ms +step:4/20000 train_loss:6.1710 train_time:188ms step_avg:47.10ms +step:5/20000 train_loss:6.2679 train_time:242ms step_avg:48.34ms +step:6/20000 train_loss:7.4806 train_time:295ms step_avg:49.23ms +step:7/20000 train_loss:6.4926 train_time:349ms step_avg:49.87ms +step:8/20000 train_loss:6.4832 train_time:403ms step_avg:50.35ms +step:9/20000 train_loss:6.4519 train_time:460ms step_avg:51.07ms +step:10/20000 train_loss:6.3979 train_time:514ms step_avg:51.36ms +step:200/20000 train_loss:2.9640 train_time:12534ms step_avg:62.67ms +step:400/20000 train_loss:2.3846 train_time:25211ms step_avg:63.03ms +step:600/20000 train_loss:2.5671 train_time:38297ms step_avg:63.83ms +step:800/20000 train_loss:2.3143 train_time:51028ms step_avg:63.79ms +step:1000/20000 train_loss:2.3877 train_time:63753ms step_avg:63.75ms +step:1000/20000 val_loss:2.3510 val_bpb:1.3924 train_time:63782ms step_avg:63.78ms +step:1200/20000 train_loss:2.4045 train_time:76596ms step_avg:63.83ms +step:1400/20000 train_loss:2.4476 train_time:88788ms step_avg:63.42ms +step:1600/20000 train_loss:2.1216 train_time:101281ms step_avg:63.30ms +step:1800/20000 train_loss:2.2195 train_time:113959ms step_avg:63.31ms +step:2000/20000 train_loss:2.2768 train_time:126722ms step_avg:63.36ms +step:2000/20000 val_loss:2.2598 val_bpb:1.3384 train_time:126750ms step_avg:63.38ms +step:2200/20000 train_loss:2.1038 train_time:139352ms step_avg:63.34ms +step:2400/20000 train_loss:2.2253 train_time:151853ms step_avg:63.27ms +step:2600/20000 train_loss:2.4498 train_time:164509ms step_avg:63.27ms +step:2800/20000 train_loss:2.2704 train_time:176920ms step_avg:63.19ms +step:3000/20000 train_loss:2.2593 train_time:189270ms step_avg:63.09ms +step:3000/20000 val_loss:2.2295 val_bpb:1.3204 train_time:189297ms step_avg:63.10ms +step:3200/20000 train_loss:2.2225 train_time:201536ms step_avg:62.98ms +step:3400/20000 train_loss:2.2012 train_time:214284ms step_avg:63.02ms +step:3600/20000 train_loss:2.1577 train_time:226713ms step_avg:62.98ms +step:3800/20000 train_loss:2.2596 train_time:239233ms step_avg:62.96ms +step:4000/20000 train_loss:2.1994 train_time:252005ms step_avg:63.00ms +step:4000/20000 val_loss:2.2128 val_bpb:1.3106 train_time:252034ms step_avg:63.01ms +step:4200/20000 train_loss:2.2106 train_time:267456ms step_avg:63.68ms +step:4400/20000 train_loss:2.1573 train_time:280329ms step_avg:63.71ms +step:4600/20000 train_loss:2.0156 train_time:292957ms step_avg:63.69ms +step:4800/20000 train_loss:2.3097 train_time:305572ms step_avg:63.66ms +step:5000/20000 train_loss:2.0852 train_time:317943ms step_avg:63.59ms +step:5000/20000 val_loss:2.2033 val_bpb:1.3049 train_time:317971ms step_avg:63.59ms +step:5200/20000 train_loss:2.2240 train_time:327971ms step_avg:63.07ms +step:5400/20000 train_loss:2.2423 train_time:340530ms step_avg:63.06ms +step:5600/20000 train_loss:2.2362 train_time:352865ms step_avg:63.01ms +step:5800/20000 train_loss:2.1992 train_time:365399ms step_avg:63.00ms +step:6000/20000 train_loss:2.2697 train_time:378096ms step_avg:63.02ms +step:6000/20000 val_loss:2.2001 val_bpb:1.3030 train_time:378124ms step_avg:63.02ms +step:6200/20000 train_loss:2.1459 train_time:390388ms step_avg:62.97ms +step:6400/20000 train_loss:2.2254 train_time:402913ms step_avg:62.96ms +step:6600/20000 train_loss:2.1813 train_time:415438ms step_avg:62.95ms +step:6800/20000 train_loss:2.2511 train_time:427789ms step_avg:62.91ms +step:7000/20000 train_loss:2.2826 train_time:440381ms step_avg:62.91ms +step:7000/20000 val_loss:2.1926 val_bpb:1.2986 train_time:440409ms step_avg:62.92ms +step:7200/20000 train_loss:2.2576 train_time:453236ms step_avg:62.95ms +step:7400/20000 train_loss:2.1615 train_time:465722ms step_avg:62.94ms +step:7600/20000 train_loss:2.0312 train_time:478160ms step_avg:62.92ms +step:7800/20000 train_loss:2.1736 train_time:490759ms step_avg:62.92ms +step:8000/20000 train_loss:2.1301 train_time:503270ms step_avg:62.91ms +step:8000/20000 val_loss:2.1372 val_bpb:1.2658 train_time:503304ms step_avg:62.91ms +step:8200/20000 train_loss:2.1922 train_time:515769ms step_avg:62.90ms +step:8400/20000 train_loss:2.1317 train_time:531528ms step_avg:63.28ms +step:8600/20000 train_loss:2.1197 train_time:541575ms step_avg:62.97ms +step:8800/20000 train_loss:2.0773 train_time:551633ms step_avg:62.69ms +step:9000/20000 train_loss:1.9904 train_time:561690ms step_avg:62.41ms +step:9000/20000 val_loss:2.0834 val_bpb:1.2339 train_time:561718ms step_avg:62.41ms +step:9200/20000 train_loss:2.0446 train_time:571749ms step_avg:62.15ms +step:9400/20000 train_loss:2.0731 train_time:581814ms step_avg:61.90ms +step:9600/20000 train_loss:2.0824 train_time:591879ms step_avg:61.65ms +step:9728/20000 val_loss:2.0408 val_bpb:1.2087 train_time:601124ms step_avg:61.79ms +stopping_early: wallclock_cap train_time:601124ms step:9728/20000 +peak memory allocated: 11389 MiB reserved: 11704 MiB +Serialized model: 74573582 bytes +Code size: 58683 bytes +Total submission size: 74632265 bytes +Serialized model int8+zlib: 15270194 bytes (payload:19552576 raw_torch:19601902 payload_ratio:3.81x) +Total submission size int8+zlib: 15328877 bytes +ttt:start lr=0.002 momentum=0.9 epochs=2 +ttt_epoch:1/2 loss:1.9409 time:155.6s +ttt_epoch:2/2 loss:1.9404 time:311.0s +ttt:done elapsed=311.0s +Compiling forward_logits for sliding window eval (stride=64, seq_len=1024)... +Compilation done, starting sliding window eval... +final_int8_zlib_roundtrip val_loss:1.9829 val_bpb:1.1744 eval_time:37628ms +final_int8_zlib_roundtrip_exact val_loss:1.98286428 val_bpb:1.17436392 diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_seed2884431328.log b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_seed2884431328.log new file mode 100644 index 0000000000..ca51b40115 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_seed2884431328.log @@ -0,0 +1,132 @@ +W0320 04:03:34.285000 42647 torch/distributed/run.py:792] +W0320 04:03:34.285000 42647 torch/distributed/run.py:792] ***************************************** +W0320 04:03:34.285000 42647 torch/distributed/run.py:792] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0320 04:03:34.285000 42647 torch/distributed/run.py:792] ***************************************** +logs/722493fe-7ce5-4f97-91ae-b9346d336c21.txt +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:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.1 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:598.000 +seed:2884431328 +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:6.9314 val_bpb:4.1052 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9314 train_time:35ms step_avg:35.24ms +step:2/20000 train_loss:23.9425 train_time:82ms step_avg:40.92ms +step:3/20000 train_loss:9.0506 train_time:135ms step_avg:45.01ms +step:4/20000 train_loss:6.2127 train_time:189ms step_avg:47.21ms +step:5/20000 train_loss:6.2270 train_time:242ms step_avg:48.38ms +step:6/20000 train_loss:7.4495 train_time:295ms step_avg:49.16ms +step:7/20000 train_loss:6.4680 train_time:349ms step_avg:49.85ms +step:8/20000 train_loss:6.4558 train_time:402ms step_avg:50.28ms +step:9/20000 train_loss:6.4165 train_time:457ms step_avg:50.78ms +step:10/20000 train_loss:6.3521 train_time:510ms step_avg:50.96ms +step:200/20000 train_loss:2.9645 train_time:9994ms step_avg:49.97ms +step:400/20000 train_loss:2.3781 train_time:19978ms step_avg:49.94ms +step:600/20000 train_loss:2.5663 train_time:29984ms step_avg:49.97ms +step:800/20000 train_loss:2.3131 train_time:39994ms step_avg:49.99ms +step:1000/20000 train_loss:2.3898 train_time:50021ms step_avg:50.02ms +step:1000/20000 val_loss:2.3531 val_bpb:1.3936 train_time:50056ms step_avg:50.06ms +step:1200/20000 train_loss:2.4062 train_time:60064ms step_avg:50.05ms +step:1400/20000 train_loss:2.4475 train_time:70104ms step_avg:50.07ms +step:1600/20000 train_loss:2.1228 train_time:80159ms step_avg:50.10ms +step:1800/20000 train_loss:2.2189 train_time:90212ms step_avg:50.12ms +step:2000/20000 train_loss:2.2779 train_time:100267ms step_avg:50.13ms +step:2000/20000 val_loss:2.2629 val_bpb:1.3402 train_time:100295ms step_avg:50.15ms +step:2200/20000 train_loss:2.1039 train_time:110330ms step_avg:50.15ms +step:2400/20000 train_loss:2.2295 train_time:120390ms step_avg:50.16ms +step:2600/20000 train_loss:2.4472 train_time:130463ms step_avg:50.18ms +step:2800/20000 train_loss:2.2766 train_time:140530ms step_avg:50.19ms +step:3000/20000 train_loss:2.2668 train_time:150602ms step_avg:50.20ms +step:3000/20000 val_loss:2.2325 val_bpb:1.3222 train_time:150628ms step_avg:50.21ms +step:3200/20000 train_loss:2.2324 train_time:160676ms step_avg:50.21ms +step:3400/20000 train_loss:2.1983 train_time:170744ms step_avg:50.22ms +step:3600/20000 train_loss:2.1641 train_time:180817ms step_avg:50.23ms +step:3800/20000 train_loss:2.2609 train_time:190875ms step_avg:50.23ms +step:4000/20000 train_loss:2.2053 train_time:200927ms step_avg:50.23ms +step:4000/20000 val_loss:2.2147 val_bpb:1.3117 train_time:200954ms step_avg:50.24ms +step:4200/20000 train_loss:2.2200 train_time:211047ms step_avg:50.25ms +step:4400/20000 train_loss:2.1601 train_time:221104ms step_avg:50.25ms +step:4600/20000 train_loss:2.0204 train_time:231155ms step_avg:50.25ms +step:4800/20000 train_loss:2.3143 train_time:241222ms step_avg:50.25ms +step:5000/20000 train_loss:2.0868 train_time:251272ms step_avg:50.25ms +step:5000/20000 val_loss:2.2027 val_bpb:1.3046 train_time:251300ms step_avg:50.26ms +step:5200/20000 train_loss:2.2281 train_time:261330ms step_avg:50.26ms +step:5400/20000 train_loss:2.2431 train_time:271376ms step_avg:50.25ms +step:5600/20000 train_loss:2.2397 train_time:281428ms step_avg:50.26ms +step:5800/20000 train_loss:2.2049 train_time:291476ms step_avg:50.25ms +step:6000/20000 train_loss:2.2786 train_time:301525ms step_avg:50.25ms +step:6000/20000 val_loss:2.2010 val_bpb:1.3035 train_time:301552ms step_avg:50.26ms +step:6200/20000 train_loss:2.1518 train_time:311569ms step_avg:50.25ms +step:6400/20000 train_loss:2.2247 train_time:321615ms step_avg:50.25ms +step:6600/20000 train_loss:2.1799 train_time:331699ms step_avg:50.26ms +step:6800/20000 train_loss:2.2472 train_time:341792ms step_avg:50.26ms +step:7000/20000 train_loss:2.2799 train_time:351836ms step_avg:50.26ms +step:7000/20000 val_loss:2.1921 val_bpb:1.2983 train_time:351864ms step_avg:50.27ms +step:7200/20000 train_loss:2.2699 train_time:361888ms step_avg:50.26ms +step:7400/20000 train_loss:2.1751 train_time:371931ms step_avg:50.26ms +step:7600/20000 train_loss:2.0549 train_time:381962ms step_avg:50.26ms +step:7800/20000 train_loss:2.2090 train_time:391992ms step_avg:50.26ms +step:8000/20000 train_loss:2.1761 train_time:402028ms step_avg:50.25ms +step:8000/20000 val_loss:2.1827 val_bpb:1.2927 train_time:402056ms step_avg:50.26ms +step:8200/20000 train_loss:2.2359 train_time:412052ms step_avg:50.25ms +step:8400/20000 train_loss:2.1940 train_time:422165ms step_avg:50.26ms +step:8600/20000 train_loss:2.1967 train_time:432191ms step_avg:50.25ms +step:8800/20000 train_loss:2.1654 train_time:442241ms step_avg:50.25ms +step:9000/20000 train_loss:2.0873 train_time:452267ms step_avg:50.25ms +step:9000/20000 val_loss:2.1832 val_bpb:1.2930 train_time:452294ms step_avg:50.25ms +step:9200/20000 train_loss:2.1554 train_time:462295ms step_avg:50.25ms +step:9400/20000 train_loss:2.2134 train_time:472346ms step_avg:50.25ms +step:9600/20000 train_loss:2.2069 train_time:482374ms step_avg:50.25ms +step:9800/20000 train_loss:2.1323 train_time:492411ms step_avg:50.25ms +step:10000/20000 train_loss:2.1543 train_time:502436ms step_avg:50.24ms +step:10000/20000 val_loss:2.1533 val_bpb:1.2753 train_time:502463ms step_avg:50.25ms +step:10200/20000 train_loss:2.0966 train_time:512468ms step_avg:50.24ms +step:10400/20000 train_loss:2.1185 train_time:522495ms step_avg:50.24ms +step:10600/20000 train_loss:1.9800 train_time:532532ms step_avg:50.24ms +step:10800/20000 train_loss:2.1894 train_time:542565ms step_avg:50.24ms +step:11000/20000 train_loss:2.1044 train_time:552584ms step_avg:50.23ms +step:11000/20000 val_loss:2.0930 val_bpb:1.2396 train_time:552611ms step_avg:50.24ms +step:11200/20000 train_loss:2.0484 train_time:562624ms step_avg:50.23ms +step:11400/20000 train_loss:2.0210 train_time:572660ms step_avg:50.23ms +step:11600/20000 train_loss:2.0151 train_time:582684ms step_avg:50.23ms +step:11800/20000 train_loss:2.0332 train_time:592706ms step_avg:50.23ms +step:11904/20000 val_loss:2.0355 val_bpb:1.2055 train_time:597985ms step_avg:50.23ms +stopping_early: wallclock_cap train_time:597985ms step:11904/20000 +peak memory allocated: 11389 MiB reserved: 11704 MiB +Serialized model: 74573582 bytes +Code size: 58683 bytes +Total submission size: 74632265 bytes +Serialized model int8+zlib: 15313856 bytes (payload:19552576 raw_torch:19601902 payload_ratio:3.81x) +Total submission size int8+zlib: 15372539 bytes +ttt:start lr=0.002 momentum=0.9 epochs=2 +ttt_epoch:1/2 loss:1.9354 time:155.5s +ttt_epoch:2/2 loss:1.9349 time:310.7s +ttt:done elapsed=310.8s +Compiling forward_logits for sliding window eval (stride=64, seq_len=1024)... +Compilation done, starting sliding window eval... +final_int8_zlib_roundtrip val_loss:1.9772 val_bpb:1.1710 eval_time:37752ms +final_int8_zlib_roundtrip_exact val_loss:1.97721966 val_bpb:1.17102086 diff --git a/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_seed7.log b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_seed7.log new file mode 100644 index 0000000000..987a589f74 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_TTT_FarnsworthTech/train_seed7.log @@ -0,0 +1,131 @@ +W0320 03:25:37.760000 38132 torch/distributed/run.py:792] +W0320 03:25:37.760000 38132 torch/distributed/run.py:792] ***************************************** +W0320 03:25:37.760000 38132 torch/distributed/run.py:792] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0320 03:25:37.760000 38132 torch/distributed/run.py:792] ***************************************** +logs/15aad7a7-07ce-48ef-8f70-977a93af4812.txt +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:18897488 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.1 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:598.000 +seed:7 +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:6.9322 val_bpb:4.1056 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9322 train_time:35ms step_avg:35.18ms +step:2/20000 train_loss:24.7482 train_time:82ms step_avg:41.15ms +step:3/20000 train_loss:9.4381 train_time:137ms step_avg:45.76ms +step:4/20000 train_loss:6.1944 train_time:191ms step_avg:47.78ms +step:5/20000 train_loss:6.2119 train_time:244ms step_avg:48.85ms +step:6/20000 train_loss:7.4355 train_time:299ms step_avg:49.83ms +step:7/20000 train_loss:6.4658 train_time:354ms step_avg:50.55ms +step:8/20000 train_loss:6.4499 train_time:408ms step_avg:50.97ms +step:9/20000 train_loss:6.4148 train_time:463ms step_avg:51.41ms +step:10/20000 train_loss:6.3444 train_time:524ms step_avg:52.43ms +step:200/20000 train_loss:2.9496 train_time:10016ms step_avg:50.08ms +step:400/20000 train_loss:2.3794 train_time:20000ms step_avg:50.00ms +step:600/20000 train_loss:2.5623 train_time:30016ms step_avg:50.03ms +step:800/20000 train_loss:2.3122 train_time:40047ms step_avg:50.06ms +step:1000/20000 train_loss:2.3906 train_time:50089ms step_avg:50.09ms +step:1000/20000 val_loss:2.3514 val_bpb:1.3926 train_time:50117ms step_avg:50.12ms +step:1200/20000 train_loss:2.4014 train_time:60130ms step_avg:50.11ms +step:1400/20000 train_loss:2.4486 train_time:70199ms step_avg:50.14ms +step:1600/20000 train_loss:2.1220 train_time:80250ms step_avg:50.16ms +step:1800/20000 train_loss:2.2202 train_time:90303ms step_avg:50.17ms +step:2000/20000 train_loss:2.2776 train_time:100369ms step_avg:50.18ms +step:2000/20000 val_loss:2.2623 val_bpb:1.3398 train_time:100407ms step_avg:50.20ms +step:2200/20000 train_loss:2.1050 train_time:110450ms step_avg:50.20ms +step:2400/20000 train_loss:2.2277 train_time:120517ms step_avg:50.22ms +step:2600/20000 train_loss:2.4430 train_time:130589ms step_avg:50.23ms +step:2800/20000 train_loss:2.2730 train_time:140658ms step_avg:50.24ms +step:3000/20000 train_loss:2.2592 train_time:150729ms step_avg:50.24ms +step:3000/20000 val_loss:2.2315 val_bpb:1.3216 train_time:150759ms step_avg:50.25ms +step:3200/20000 train_loss:2.2223 train_time:160803ms step_avg:50.25ms +step:3400/20000 train_loss:2.1989 train_time:170880ms step_avg:50.26ms +step:3600/20000 train_loss:2.1595 train_time:180952ms step_avg:50.26ms +step:3800/20000 train_loss:2.2598 train_time:191032ms step_avg:50.27ms +step:4000/20000 train_loss:2.2014 train_time:201091ms step_avg:50.27ms +step:4000/20000 val_loss:2.2158 val_bpb:1.3123 train_time:201119ms step_avg:50.28ms +step:4200/20000 train_loss:2.2198 train_time:211232ms step_avg:50.29ms +step:4400/20000 train_loss:2.1655 train_time:221360ms step_avg:50.31ms +step:4600/20000 train_loss:2.0215 train_time:231419ms step_avg:50.31ms +step:4800/20000 train_loss:2.3153 train_time:241469ms step_avg:50.31ms +step:5000/20000 train_loss:2.0844 train_time:251520ms step_avg:50.30ms +step:5000/20000 val_loss:2.2040 val_bpb:1.3053 train_time:251548ms step_avg:50.31ms +step:5200/20000 train_loss:2.2261 train_time:261568ms step_avg:50.30ms +step:5400/20000 train_loss:2.2373 train_time:271608ms step_avg:50.30ms +step:5600/20000 train_loss:2.2364 train_time:281657ms step_avg:50.30ms +step:5800/20000 train_loss:2.2051 train_time:291706ms step_avg:50.29ms +step:6000/20000 train_loss:2.2759 train_time:301764ms step_avg:50.29ms +step:6000/20000 val_loss:2.2029 val_bpb:1.3047 train_time:301792ms step_avg:50.30ms +step:6200/20000 train_loss:2.1480 train_time:311813ms step_avg:50.29ms +step:6400/20000 train_loss:2.2224 train_time:321864ms step_avg:50.29ms +step:6600/20000 train_loss:2.1886 train_time:331907ms step_avg:50.29ms +step:6800/20000 train_loss:2.2535 train_time:341955ms step_avg:50.29ms +step:7000/20000 train_loss:2.2843 train_time:351996ms step_avg:50.29ms +step:7000/20000 val_loss:2.1922 val_bpb:1.2983 train_time:352028ms step_avg:50.29ms +step:7200/20000 train_loss:2.2701 train_time:362059ms step_avg:50.29ms +step:7400/20000 train_loss:2.1813 train_time:372108ms step_avg:50.28ms +step:7600/20000 train_loss:2.0545 train_time:382163ms step_avg:50.28ms +step:7800/20000 train_loss:2.2078 train_time:392324ms step_avg:50.30ms +step:8000/20000 train_loss:2.1796 train_time:402367ms step_avg:50.30ms +step:8000/20000 val_loss:2.1853 val_bpb:1.2943 train_time:402396ms step_avg:50.30ms +step:8200/20000 train_loss:2.2479 train_time:412412ms step_avg:50.29ms +step:8400/20000 train_loss:2.1940 train_time:422537ms step_avg:50.30ms +step:8600/20000 train_loss:2.2008 train_time:432569ms step_avg:50.30ms +step:8800/20000 train_loss:2.1699 train_time:442606ms step_avg:50.30ms +step:9000/20000 train_loss:2.0890 train_time:452634ms step_avg:50.29ms +step:9000/20000 val_loss:2.1848 val_bpb:1.2940 train_time:452661ms step_avg:50.30ms +step:9200/20000 train_loss:2.1522 train_time:462668ms step_avg:50.29ms +step:9400/20000 train_loss:2.1976 train_time:472700ms step_avg:50.29ms +step:9600/20000 train_loss:2.2083 train_time:482741ms step_avg:50.29ms +step:9800/20000 train_loss:2.1357 train_time:492769ms step_avg:50.28ms +step:10000/20000 train_loss:2.1548 train_time:502799ms step_avg:50.28ms +step:10000/20000 val_loss:2.1548 val_bpb:1.2762 train_time:502828ms step_avg:50.28ms +step:10200/20000 train_loss:2.0988 train_time:512833ms step_avg:50.28ms +step:10400/20000 train_loss:2.1227 train_time:522870ms step_avg:50.28ms +step:10600/20000 train_loss:1.9810 train_time:532903ms step_avg:50.27ms +step:10800/20000 train_loss:2.1894 train_time:542942ms step_avg:50.27ms +step:11000/20000 train_loss:2.1027 train_time:552974ms step_avg:50.27ms +step:11000/20000 val_loss:2.0953 val_bpb:1.2409 train_time:553003ms step_avg:50.27ms +step:11200/20000 train_loss:2.0474 train_time:566253ms step_avg:50.56ms +step:11400/20000 train_loss:2.0145 train_time:579358ms step_avg:50.82ms +step:11600/20000 train_loss:2.0055 train_time:592210ms step_avg:51.05ms +step:11652/20000 val_loss:2.0437 val_bpb:1.2104 train_time:597980ms step_avg:51.32ms +stopping_early: wallclock_cap train_time:597980ms step:11652/20000 +peak memory allocated: 11389 MiB reserved: 11704 MiB +Serialized model: 74573582 bytes +Code size: 58683 bytes +Total submission size: 74632265 bytes +Serialized model int8+zlib: 15340781 bytes (payload:19552576 raw_torch:19601902 payload_ratio:3.81x) +Total submission size int8+zlib: 15399464 bytes +ttt:start lr=0.002 momentum=0.9 epochs=2 +ttt_epoch:1/2 loss:1.9456 time:155.6s +ttt_epoch:2/2 loss:1.9455 time:310.2s +ttt:done elapsed=310.2s +Compiling forward_logits for sliding window eval (stride=64, seq_len=1024)... +Compilation done, starting sliding window eval... +final_int8_zlib_roundtrip val_loss:1.9845 val_bpb:1.1754 eval_time:37690ms +final_int8_zlib_roundtrip_exact val_loss:1.98453360 val_bpb:1.17535258