From 94aeda35513990af3d2854bf3fa57b4006f0b399 Mon Sep 17 00:00:00 2001 From: someone114514 <3407056424@qq.com> Date: Sun, 22 Mar 2026 21:20:23 +0800 Subject: [PATCH] Add non-record EMA and adaptive export explorations --- .../README.md | 76 + .../log.txt | 444 ++++++ .../submission.json | 25 + .../train_gpt.py | 1336 +++++++++++++++++ 4 files changed, 1881 insertions(+) create mode 100644 records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/README.md create mode 100644 records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/log.txt create mode 100644 records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/submission.json create mode 100644 records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/train_gpt.py diff --git a/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/README.md b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/README.md new file mode 100644 index 000000000..925c7bd41 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/README.md @@ -0,0 +1,76 @@ +# Non-Record Submission: Baseline + EMA + Adaptive Export (2×H100) + +## Result + +This run reaches **`1.17251579 val_bpb`** on the final post-quant sliding-window roundtrip evaluation. + +Key numbers: + +- Final post-quant sliding-window: `val_loss=1.97973856`, `val_bpb=1.17251579` +- Pre-export validation at stop: `val_loss=2.0093`, `val_bpb=1.1900` +- Step at stop: `3807` +- Training wallclock: `1200.112s` +- Final eval wallclock: `613.676s` +- Total artifact size: `16,399,881 bytes` +- Budget miss: `449,881 bytes` + +This is therefore **non-record only**: the run produced a strong final score shape, but it does **not** satisfy the 16MB artifact cap yet. + +## What Changed + +This submission starts from the strong `Int6 MLP3x + SmearGate + BigramHash + Muon` baseline and changes only two things: + +1. **Late-stage EMA** + - `EMA_ENABLED=1` + - `EMA_BETA=0.9998` + - `EMA_START_FRAC=0.8` + +2. **Adaptive export-time pruning search** + - `PRUNE_CANDIDATES=0.00,0.01,0.02,0.03,0.04,0.05` + - `TARGET_ARTIFACT_BYTES=15950000` + - select the smallest pruning ratio that meets the target, or the smallest artifact if none do + +The backbone, optimizer family, mixed quantization setup, and final evaluation path were otherwise kept close to the strong merged baseline line. + +## Why This Is Worth Looking At + +The point of this run is not "new backbone, new SOTA." The point is to test a narrower hypothesis: + +> under a hard artifact budget, some of the remaining optimization headroom may come from **weight smoothing before export** and **budget-aware export selection**, not only from changing the training architecture. + +Even though the run is not leaderboard-valid yet, it is still informative for three reasons: + +1. The final score shape is already competitive for a non-record run under limited compute. +2. The failure mode is narrow: the bottleneck is **artifact size**, not a collapse in post-quant quality. +3. The next steps are unusually clear: + - extend pruning search beyond `5%` + - move from global pruning to module-aware budget allocation + - compare against a minimal top-record export-only fork + +That makes this a useful potential-PR result rather than just another failed oversized run. + +## Submission Checklist + +- Training completes under wallclock cap: `yes` +- Final post-quant roundtrip evaluation runs successfully: `yes` +- Sliding-window final eval runs successfully: `yes` +- Code is self-contained in `train_gpt.py`: `yes` +- Artifact is under the 16MB target: `no` +- Multi-seed verification: `no` + +## Compute Limitation + +This run was produced under constrained compute: + +- `2×H100`, not `8×H100` +- single seed only +- no remaining compute budget for additional tuning passes after the first full validation run +- substantial runtime spent in final sliding-window evaluation + +So this should be read as a validated directional result, not as a fully swept or fully scaled submission. + +## Included Files + +- `train_gpt.py` — code snapshot for this non-record run +- `log.txt` — exact training + export + final evaluation log +- `submission.json` — run metadata diff --git a/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/log.txt b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/log.txt new file mode 100644 index 000000000..2912b7f01 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/log.txt @@ -0,0 +1,444 @@ +W0322 12:16:26.305000 23245 torch/distributed/run.py:803] +W0322 12:16:26.305000 23245 torch/distributed/run.py:803] ***************************************** +W0322 12:16:26.305000 23245 torch/distributed/run.py:803] 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. +W0322 12:16:26.305000 23245 torch/distributed/run.py:803] ***************************************** +logs/baseline_ema_export_2gpu.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:22368841 +world_size:2 grad_accum_steps:4 +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.03 matrix_lr:0.02 scalar_lr:0.02 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:1200.000 +ema_enabled:1 ema_beta:0.9998 ema_start_frac:0.8 swa_enabled:0 +target_artifact_bytes:15950000 prune_candidates:0.0000,0.0100,0.0200,0.0300,0.0400,0.0500 +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:1/20000 train_loss:6.9345 train_time:378ms step_avg:378.33ms +step:2/20000 train_loss:8.1631 train_time:684ms step_avg:342.20ms +step:3/20000 train_loss:7.7030 train_time:993ms step_avg:330.92ms +step:4/20000 train_loss:6.9638 train_time:1308ms step_avg:326.91ms +step:5/20000 train_loss:6.7782 train_time:1618ms step_avg:323.68ms +step:6/20000 train_loss:6.5861 train_time:1929ms step_avg:321.54ms +step:7/20000 train_loss:6.5033 train_time:2240ms step_avg:319.97ms +step:8/20000 train_loss:6.4028 train_time:2553ms step_avg:319.07ms +step:9/20000 train_loss:6.3273 train_time:2860ms step_avg:317.81ms +step:10/20000 train_loss:6.1164 train_time:3168ms step_avg:316.77ms +step:50/20000 train_loss:3.6572 train_time:15633ms step_avg:312.67ms +step:100/20000 train_loss:3.2039 train_time:31344ms step_avg:313.44ms +step:150/20000 train_loss:2.9068 train_time:47169ms step_avg:314.46ms +step:200/20000 train_loss:2.6299 train_time:62935ms step_avg:314.68ms +step:250/20000 train_loss:2.7250 train_time:78719ms step_avg:314.87ms +step:300/20000 train_loss:2.4970 train_time:94488ms step_avg:314.96ms +step:350/20000 train_loss:2.5479 train_time:110259ms step_avg:315.02ms +step:400/20000 train_loss:2.4805 train_time:126040ms step_avg:315.10ms +step:450/20000 train_loss:2.3002 train_time:141795ms step_avg:315.10ms +step:500/20000 train_loss:2.3723 train_time:157571ms step_avg:315.14ms +step:550/20000 train_loss:2.4205 train_time:173344ms step_avg:315.17ms +step:600/20000 train_loss:2.3100 train_time:189098ms step_avg:315.16ms +step:650/20000 train_loss:2.3194 train_time:204878ms step_avg:315.20ms +step:700/20000 train_loss:2.3802 train_time:220641ms step_avg:315.20ms +step:750/20000 train_loss:2.3334 train_time:236417ms step_avg:315.22ms +step:800/20000 train_loss:2.3040 train_time:252161ms step_avg:315.20ms +step:850/20000 train_loss:2.2296 train_time:267914ms step_avg:315.19ms +step:900/20000 train_loss:2.1695 train_time:283700ms step_avg:315.22ms +step:950/20000 train_loss:2.4259 train_time:299444ms step_avg:315.20ms +step:1000/20000 train_loss:2.2934 train_time:315200ms step_avg:315.20ms +step:1050/20000 train_loss:2.2071 train_time:331012ms step_avg:315.25ms +step:1100/20000 train_loss:2.2567 train_time:346772ms step_avg:315.25ms +step:1150/20000 train_loss:2.2249 train_time:362511ms step_avg:315.23ms +step:1200/20000 train_loss:2.2613 train_time:378266ms step_avg:315.22ms +step:1250/20000 train_loss:2.2512 train_time:394008ms step_avg:315.21ms +step:1300/20000 train_loss:2.2339 train_time:409767ms step_avg:315.21ms +step:1350/20000 train_loss:2.1877 train_time:425547ms step_avg:315.22ms +step:1400/20000 train_loss:2.2785 train_time:441322ms step_avg:315.23ms +step:1450/20000 train_loss:2.1484 train_time:457077ms step_avg:315.23ms +step:1500/20000 train_loss:2.1739 train_time:472830ms step_avg:315.22ms +step:1550/20000 train_loss:2.1694 train_time:488602ms step_avg:315.23ms +step:1600/20000 train_loss:2.1225 train_time:504347ms step_avg:315.22ms +step:1650/20000 train_loss:2.1518 train_time:520097ms step_avg:315.21ms +step:1700/20000 train_loss:2.1511 train_time:535891ms step_avg:315.23ms +step:1750/20000 train_loss:2.1298 train_time:551627ms step_avg:315.22ms +step:1800/20000 train_loss:2.1738 train_time:567424ms step_avg:315.24ms +step:1850/20000 train_loss:2.1210 train_time:583165ms step_avg:315.22ms +step:1900/20000 train_loss:2.1311 train_time:598931ms step_avg:315.23ms +step:1950/20000 train_loss:2.0931 train_time:614692ms step_avg:315.23ms +step:2000/20000 train_loss:2.0688 train_time:630443ms step_avg:315.22ms +step:2050/20000 train_loss:2.1408 train_time:646183ms step_avg:315.21ms +step:2100/20000 train_loss:2.0210 train_time:661931ms step_avg:315.21ms +step:2150/20000 train_loss:2.1154 train_time:677684ms step_avg:315.20ms +step:2200/20000 train_loss:2.1342 train_time:693454ms step_avg:315.21ms +step:2250/20000 train_loss:2.1240 train_time:709195ms step_avg:315.20ms +step:2300/20000 train_loss:2.0575 train_time:724951ms step_avg:315.20ms +step:2350/20000 train_loss:2.1668 train_time:740704ms step_avg:315.19ms +step:2400/20000 train_loss:2.0960 train_time:756444ms step_avg:315.19ms +step:2450/20000 train_loss:2.1120 train_time:772205ms step_avg:315.19ms +step:2500/20000 train_loss:2.1573 train_time:787957ms step_avg:315.18ms +step:2550/20000 train_loss:2.1415 train_time:803781ms step_avg:315.21ms +step:2600/20000 train_loss:2.1156 train_time:819550ms step_avg:315.21ms +step:2650/20000 train_loss:2.0641 train_time:835302ms step_avg:315.21ms +step:2700/20000 train_loss:2.0541 train_time:851044ms step_avg:315.20ms +step:2750/20000 train_loss:2.0690 train_time:866797ms step_avg:315.20ms +step:2800/20000 train_loss:2.1256 train_time:882559ms step_avg:315.20ms +step:2850/20000 train_loss:1.9869 train_time:898307ms step_avg:315.20ms +step:2900/20000 train_loss:2.0690 train_time:914062ms step_avg:315.19ms +step:2950/20000 train_loss:2.0792 train_time:929842ms step_avg:315.20ms +step:3000/20000 train_loss:2.0628 train_time:945603ms step_avg:315.20ms +step:3050/20000 train_loss:2.0391 train_time:961365ms step_avg:315.20ms +step:3100/20000 train_loss:2.0169 train_time:977127ms step_avg:315.20ms +step:3150/20000 train_loss:1.9379 train_time:992895ms step_avg:315.20ms +step:3200/20000 train_loss:2.0930 train_time:1008659ms step_avg:315.21ms +step:3250/20000 train_loss:2.0195 train_time:1024409ms step_avg:315.20ms +step:3300/20000 train_loss:2.0591 train_time:1040178ms step_avg:315.21ms +step:3350/20000 train_loss:2.0796 train_time:1055969ms step_avg:315.21ms +step:3400/20000 train_loss:2.0558 train_time:1071735ms step_avg:315.22ms +step:3450/20000 train_loss:2.0211 train_time:1087531ms step_avg:315.23ms +step:3500/20000 train_loss:2.0533 train_time:1103290ms step_avg:315.23ms +step:3550/20000 train_loss:1.9659 train_time:1119070ms step_avg:315.23ms +step:3600/20000 train_loss:2.0770 train_time:1134821ms step_avg:315.23ms +step:3650/20000 train_loss:1.8565 train_time:1150605ms step_avg:315.23ms +step:3700/20000 train_loss:2.0143 train_time:1166386ms step_avg:315.24ms +step:3750/20000 train_loss:1.9449 train_time:1182152ms step_avg:315.24ms +step:3800/20000 train_loss:2.0205 train_time:1197899ms step_avg:315.24ms +step:3807/20000 val_loss:2.0093 val_bpb:1.1900 train_time:1200112ms step_avg:315.24ms +stopping_early: wallclock_cap train_time:1200112ms step:3807/20000 +peak memory allocated: 17070 MiB reserved: 17198 MiB +Serialized model: 87413467 bytes +Code size: 57374 bytes +Total submission size: 87470841 bytes +export_search prune:0.0000 model_bytes:16380192 total_bytes:16437566 meets_target:0 +export_search prune:0.0100 model_bytes:16375325 total_bytes:16432699 meets_target:0 +export_search prune:0.0200 model_bytes:16368257 total_bytes:16425631 meets_target:0 +export_search prune:0.0300 model_bytes:16359785 total_bytes:16417159 meets_target:0 +export_search prune:0.0400 model_bytes:16350475 total_bytes:16407849 meets_target:0 +export_search prune:0.0500 model_bytes:16342507 total_bytes:16399881 meets_target:0 +export_search no candidate met target:15950000, using_smallest_total_bytes:16399881 +export_selected prune:0.0500 model_bytes:16342507 total_bytes:16399881 +Serialized model int6+zlib: 16342507 bytes +Total submission size int8+zlib: 16399881 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/484544 windows running_bpb=1.236078 + sliding_eval [ 0.3%] 1632/484544 windows running_bpb=1.168070 + sliding_eval [ 0.7%] 3232/484544 windows running_bpb=1.168556 + sliding_eval [ 1.0%] 4832/484544 windows running_bpb=1.162591 + sliding_eval [ 1.3%] 6432/484544 windows running_bpb=1.174819 + sliding_eval [ 1.7%] 8032/484544 windows running_bpb=1.175749 + sliding_eval [ 2.0%] 9632/484544 windows running_bpb=1.177656 + sliding_eval [ 2.3%] 11232/484544 windows running_bpb=1.173099 + sliding_eval [ 2.6%] 12832/484544 windows running_bpb=1.170567 + sliding_eval [ 3.0%] 14432/484544 windows running_bpb=1.172121 + sliding_eval [ 3.3%] 16032/484544 windows running_bpb=1.180809 + sliding_eval [ 3.6%] 17632/484544 windows running_bpb=1.178949 + sliding_eval [ 4.0%] 19232/484544 windows running_bpb=1.180351 + sliding_eval [ 4.3%] 20832/484544 windows running_bpb=1.178904 + sliding_eval [ 4.6%] 22432/484544 windows running_bpb=1.177503 + sliding_eval [ 5.0%] 24032/484544 windows running_bpb=1.177897 + sliding_eval [ 5.3%] 25632/484544 windows running_bpb=1.179364 + sliding_eval [ 5.6%] 27232/484544 windows running_bpb=1.179831 + sliding_eval [ 6.0%] 28832/484544 windows running_bpb=1.185927 + sliding_eval [ 6.3%] 30432/484544 windows running_bpb=1.183239 + sliding_eval [ 6.6%] 32032/484544 windows running_bpb=1.184274 + sliding_eval [ 6.9%] 33632/484544 windows running_bpb=1.182781 + sliding_eval [ 7.3%] 35232/484544 windows running_bpb=1.182240 + sliding_eval [ 7.6%] 36832/484544 windows running_bpb=1.181703 + sliding_eval [ 7.9%] 38432/484544 windows running_bpb=1.182344 + sliding_eval [ 8.3%] 40032/484544 windows running_bpb=1.179872 + sliding_eval [ 8.6%] 41632/484544 windows running_bpb=1.178948 + sliding_eval [ 8.9%] 43232/484544 windows running_bpb=1.179387 + sliding_eval [ 9.3%] 44832/484544 windows running_bpb=1.178321 + sliding_eval [ 9.6%] 46432/484544 windows running_bpb=1.178161 + sliding_eval [ 9.9%] 48032/484544 windows running_bpb=1.177382 + sliding_eval [ 10.2%] 49632/484544 windows running_bpb=1.178597 + sliding_eval [ 10.6%] 51232/484544 windows running_bpb=1.179711 + sliding_eval [ 10.9%] 52832/484544 windows running_bpb=1.180262 + sliding_eval [ 11.2%] 54432/484544 windows running_bpb=1.179752 + sliding_eval [ 11.6%] 56032/484544 windows running_bpb=1.180132 + sliding_eval [ 11.9%] 57632/484544 windows running_bpb=1.179268 + sliding_eval [ 12.2%] 59232/484544 windows running_bpb=1.175363 + sliding_eval [ 12.6%] 60832/484544 windows running_bpb=1.175477 + sliding_eval [ 12.9%] 62432/484544 windows running_bpb=1.176357 + sliding_eval [ 13.2%] 64032/484544 windows running_bpb=1.176532 + sliding_eval [ 13.5%] 65632/484544 windows running_bpb=1.176365 + sliding_eval [ 13.9%] 67232/484544 windows running_bpb=1.175136 + sliding_eval [ 14.2%] 68832/484544 windows running_bpb=1.174826 + sliding_eval [ 14.5%] 70432/484544 windows running_bpb=1.174091 + sliding_eval [ 14.9%] 72032/484544 windows running_bpb=1.174171 + sliding_eval [ 15.2%] 73632/484544 windows running_bpb=1.174202 + sliding_eval [ 15.5%] 75232/484544 windows running_bpb=1.174362 + sliding_eval [ 15.9%] 76832/484544 windows running_bpb=1.174082 + sliding_eval [ 16.2%] 78432/484544 windows running_bpb=1.174658 + sliding_eval [ 16.5%] 80032/484544 windows running_bpb=1.175009 + sliding_eval [ 16.8%] 81632/484544 windows running_bpb=1.174734 + sliding_eval [ 17.2%] 83232/484544 windows running_bpb=1.175760 + sliding_eval [ 17.5%] 84832/484544 windows running_bpb=1.177722 + sliding_eval [ 17.8%] 86432/484544 windows running_bpb=1.177023 + sliding_eval [ 18.2%] 88032/484544 windows running_bpb=1.177742 + sliding_eval [ 18.5%] 89632/484544 windows running_bpb=1.178107 + sliding_eval [ 18.8%] 91232/484544 windows running_bpb=1.178051 + sliding_eval [ 19.2%] 92832/484544 windows running_bpb=1.177624 + sliding_eval [ 19.5%] 94432/484544 windows running_bpb=1.177899 + sliding_eval [ 19.8%] 96032/484544 windows running_bpb=1.177344 + sliding_eval [ 20.1%] 97632/484544 windows running_bpb=1.180175 + sliding_eval [ 20.5%] 99232/484544 windows running_bpb=1.180168 + sliding_eval [ 20.8%] 100832/484544 windows running_bpb=1.180183 + sliding_eval [ 21.1%] 102432/484544 windows running_bpb=1.179823 + sliding_eval [ 21.5%] 104032/484544 windows running_bpb=1.179350 + sliding_eval [ 21.8%] 105632/484544 windows running_bpb=1.178617 + sliding_eval [ 22.1%] 107232/484544 windows running_bpb=1.178604 + sliding_eval [ 22.5%] 108832/484544 windows running_bpb=1.179247 + sliding_eval [ 22.8%] 110432/484544 windows running_bpb=1.179285 + sliding_eval [ 23.1%] 112032/484544 windows running_bpb=1.179254 + sliding_eval [ 23.5%] 113632/484544 windows running_bpb=1.179734 + sliding_eval [ 23.8%] 115232/484544 windows running_bpb=1.179515 + sliding_eval [ 24.1%] 116832/484544 windows running_bpb=1.179146 + sliding_eval [ 24.4%] 118432/484544 windows running_bpb=1.179493 + sliding_eval [ 24.8%] 120032/484544 windows running_bpb=1.179595 + sliding_eval [ 25.1%] 121632/484544 windows running_bpb=1.179784 + sliding_eval [ 25.4%] 123232/484544 windows running_bpb=1.179821 + sliding_eval [ 25.8%] 124832/484544 windows running_bpb=1.179382 + sliding_eval [ 26.1%] 126432/484544 windows running_bpb=1.179325 + sliding_eval [ 26.4%] 128032/484544 windows running_bpb=1.179286 + sliding_eval [ 26.8%] 129632/484544 windows running_bpb=1.179434 + sliding_eval [ 27.1%] 131232/484544 windows running_bpb=1.179608 + sliding_eval [ 27.4%] 132832/484544 windows running_bpb=1.179195 + sliding_eval [ 27.7%] 134432/484544 windows running_bpb=1.178661 + sliding_eval [ 28.1%] 136032/484544 windows running_bpb=1.177485 + sliding_eval [ 28.4%] 137632/484544 windows running_bpb=1.178006 + sliding_eval [ 28.7%] 139232/484544 windows running_bpb=1.177756 + sliding_eval [ 29.1%] 140832/484544 windows running_bpb=1.178439 + sliding_eval [ 29.4%] 142432/484544 windows running_bpb=1.178835 + sliding_eval [ 29.7%] 144032/484544 windows running_bpb=1.179325 + sliding_eval [ 30.1%] 145632/484544 windows running_bpb=1.179209 + sliding_eval [ 30.4%] 147232/484544 windows running_bpb=1.179098 + sliding_eval [ 30.7%] 148832/484544 windows running_bpb=1.178785 + sliding_eval [ 31.0%] 150432/484544 windows running_bpb=1.178519 + sliding_eval [ 31.4%] 152032/484544 windows running_bpb=1.178270 + sliding_eval [ 31.7%] 153632/484544 windows running_bpb=1.179129 + sliding_eval [ 32.0%] 155232/484544 windows running_bpb=1.179126 + sliding_eval [ 32.4%] 156832/484544 windows running_bpb=1.179650 + sliding_eval [ 32.7%] 158432/484544 windows running_bpb=1.179473 + sliding_eval [ 33.0%] 160032/484544 windows running_bpb=1.179949 + sliding_eval [ 33.4%] 161632/484544 windows running_bpb=1.180099 + sliding_eval [ 33.7%] 163232/484544 windows running_bpb=1.180093 + sliding_eval [ 34.0%] 164832/484544 windows running_bpb=1.180091 + sliding_eval [ 34.3%] 166432/484544 windows running_bpb=1.180168 + sliding_eval [ 34.7%] 168032/484544 windows running_bpb=1.179639 + sliding_eval [ 35.0%] 169632/484544 windows running_bpb=1.179594 + sliding_eval [ 35.3%] 171232/484544 windows running_bpb=1.179289 + sliding_eval [ 35.7%] 172832/484544 windows running_bpb=1.179081 + sliding_eval [ 36.0%] 174432/484544 windows running_bpb=1.179071 + sliding_eval [ 36.3%] 176032/484544 windows running_bpb=1.178952 + sliding_eval [ 36.7%] 177632/484544 windows running_bpb=1.179202 + sliding_eval [ 37.0%] 179232/484544 windows running_bpb=1.179516 + sliding_eval [ 37.3%] 180832/484544 windows running_bpb=1.179948 + sliding_eval [ 37.7%] 182432/484544 windows running_bpb=1.180358 + sliding_eval [ 38.0%] 184032/484544 windows running_bpb=1.180813 + sliding_eval [ 38.3%] 185632/484544 windows running_bpb=1.180474 + sliding_eval [ 38.6%] 187232/484544 windows running_bpb=1.180434 + sliding_eval [ 39.0%] 188832/484544 windows running_bpb=1.180612 + sliding_eval [ 39.3%] 190432/484544 windows running_bpb=1.180608 + sliding_eval [ 39.6%] 192032/484544 windows running_bpb=1.180766 + sliding_eval [ 40.0%] 193632/484544 windows running_bpb=1.180815 + sliding_eval [ 40.3%] 195232/484544 windows running_bpb=1.180360 + sliding_eval [ 40.6%] 196832/484544 windows running_bpb=1.180221 + sliding_eval [ 41.0%] 198432/484544 windows running_bpb=1.180547 + sliding_eval [ 41.3%] 200032/484544 windows running_bpb=1.180766 + sliding_eval [ 41.6%] 201632/484544 windows running_bpb=1.180763 + sliding_eval [ 41.9%] 203232/484544 windows running_bpb=1.180650 + sliding_eval [ 42.3%] 204832/484544 windows running_bpb=1.180475 + sliding_eval [ 42.6%] 206432/484544 windows running_bpb=1.180214 + sliding_eval [ 42.9%] 208032/484544 windows running_bpb=1.179736 + sliding_eval [ 43.3%] 209632/484544 windows running_bpb=1.179510 + sliding_eval [ 43.6%] 211232/484544 windows running_bpb=1.179284 + sliding_eval [ 43.9%] 212832/484544 windows running_bpb=1.179655 + sliding_eval [ 44.3%] 214432/484544 windows running_bpb=1.179442 + sliding_eval [ 44.6%] 216032/484544 windows running_bpb=1.179719 + sliding_eval [ 44.9%] 217632/484544 windows running_bpb=1.179969 + sliding_eval [ 45.2%] 219232/484544 windows running_bpb=1.180068 + sliding_eval [ 45.6%] 220832/484544 windows running_bpb=1.179921 + sliding_eval [ 45.9%] 222432/484544 windows running_bpb=1.179525 + sliding_eval [ 46.2%] 224032/484544 windows running_bpb=1.179653 + sliding_eval [ 46.6%] 225632/484544 windows running_bpb=1.179247 + sliding_eval [ 46.9%] 227232/484544 windows running_bpb=1.179020 + sliding_eval [ 47.2%] 228832/484544 windows running_bpb=1.179526 + sliding_eval [ 47.6%] 230432/484544 windows running_bpb=1.179343 + sliding_eval [ 47.9%] 232032/484544 windows running_bpb=1.179152 + sliding_eval [ 48.2%] 233632/484544 windows running_bpb=1.178897 + sliding_eval [ 48.5%] 235232/484544 windows running_bpb=1.178967 + sliding_eval [ 48.9%] 236832/484544 windows running_bpb=1.179226 + sliding_eval [ 49.2%] 238432/484544 windows running_bpb=1.179295 + sliding_eval [ 49.5%] 240032/484544 windows running_bpb=1.178961 + sliding_eval [ 49.9%] 241632/484544 windows running_bpb=1.178630 + sliding_eval [ 50.2%] 243232/484544 windows running_bpb=1.178472 + sliding_eval [ 50.5%] 244832/484544 windows running_bpb=1.178375 + sliding_eval [ 50.9%] 246432/484544 windows running_bpb=1.178437 + sliding_eval [ 51.2%] 248032/484544 windows running_bpb=1.178028 + sliding_eval [ 51.5%] 249632/484544 windows running_bpb=1.178526 + sliding_eval [ 51.8%] 251232/484544 windows running_bpb=1.178510 + sliding_eval [ 52.2%] 252832/484544 windows running_bpb=1.178719 + sliding_eval [ 52.5%] 254432/484544 windows running_bpb=1.178560 + sliding_eval [ 52.8%] 256032/484544 windows running_bpb=1.178256 + sliding_eval [ 53.2%] 257632/484544 windows running_bpb=1.178134 + sliding_eval [ 53.5%] 259232/484544 windows running_bpb=1.177885 + sliding_eval [ 53.8%] 260832/484544 windows running_bpb=1.177626 + sliding_eval [ 54.2%] 262432/484544 windows running_bpb=1.177602 + sliding_eval [ 54.5%] 264032/484544 windows running_bpb=1.177376 + sliding_eval [ 54.8%] 265632/484544 windows running_bpb=1.177464 + sliding_eval [ 55.2%] 267232/484544 windows running_bpb=1.177112 + sliding_eval [ 55.5%] 268832/484544 windows running_bpb=1.177118 + sliding_eval [ 55.8%] 270432/484544 windows running_bpb=1.177534 + sliding_eval [ 56.1%] 272032/484544 windows running_bpb=1.177911 + sliding_eval [ 56.5%] 273632/484544 windows running_bpb=1.177697 + sliding_eval [ 56.8%] 275232/484544 windows running_bpb=1.177617 + sliding_eval [ 57.1%] 276832/484544 windows running_bpb=1.177931 + sliding_eval [ 57.5%] 278432/484544 windows running_bpb=1.177686 + sliding_eval [ 57.8%] 280032/484544 windows running_bpb=1.177638 + sliding_eval [ 58.1%] 281632/484544 windows running_bpb=1.177363 + sliding_eval [ 58.5%] 283232/484544 windows running_bpb=1.177463 + sliding_eval [ 58.8%] 284832/484544 windows running_bpb=1.177351 + sliding_eval [ 59.1%] 286432/484544 windows running_bpb=1.177219 + sliding_eval [ 59.4%] 288032/484544 windows running_bpb=1.177205 + sliding_eval [ 59.8%] 289632/484544 windows running_bpb=1.177027 + sliding_eval [ 60.1%] 291232/484544 windows running_bpb=1.176753 + sliding_eval [ 60.4%] 292832/484544 windows running_bpb=1.176797 + sliding_eval [ 60.8%] 294432/484544 windows running_bpb=1.176669 + sliding_eval [ 61.1%] 296032/484544 windows running_bpb=1.176761 + sliding_eval [ 61.4%] 297632/484544 windows running_bpb=1.176487 + sliding_eval [ 61.8%] 299232/484544 windows running_bpb=1.176569 + sliding_eval [ 62.1%] 300832/484544 windows running_bpb=1.176263 + sliding_eval [ 62.4%] 302432/484544 windows running_bpb=1.175887 + sliding_eval [ 62.7%] 304032/484544 windows running_bpb=1.176011 + sliding_eval [ 63.1%] 305632/484544 windows running_bpb=1.175996 + sliding_eval [ 63.4%] 307232/484544 windows running_bpb=1.175955 + sliding_eval [ 63.7%] 308832/484544 windows running_bpb=1.175741 + sliding_eval [ 64.1%] 310432/484544 windows running_bpb=1.175712 + sliding_eval [ 64.4%] 312032/484544 windows running_bpb=1.175637 + sliding_eval [ 64.7%] 313632/484544 windows running_bpb=1.175532 + sliding_eval [ 65.1%] 315232/484544 windows running_bpb=1.175563 + sliding_eval [ 65.4%] 316832/484544 windows running_bpb=1.175651 + sliding_eval [ 65.7%] 318432/484544 windows running_bpb=1.175335 + sliding_eval [ 66.0%] 320032/484544 windows running_bpb=1.175270 + sliding_eval [ 66.4%] 321632/484544 windows running_bpb=1.175218 + sliding_eval [ 66.7%] 323232/484544 windows running_bpb=1.174962 + sliding_eval [ 67.0%] 324832/484544 windows running_bpb=1.174670 + sliding_eval [ 67.4%] 326432/484544 windows running_bpb=1.174494 + sliding_eval [ 67.7%] 328032/484544 windows running_bpb=1.174597 + sliding_eval [ 68.0%] 329632/484544 windows running_bpb=1.174716 + sliding_eval [ 68.4%] 331232/484544 windows running_bpb=1.174395 + sliding_eval [ 68.7%] 332832/484544 windows running_bpb=1.174120 + sliding_eval [ 69.0%] 334432/484544 windows running_bpb=1.174008 + sliding_eval [ 69.4%] 336032/484544 windows running_bpb=1.174005 + sliding_eval [ 69.7%] 337632/484544 windows running_bpb=1.173885 + sliding_eval [ 70.0%] 339232/484544 windows running_bpb=1.173988 + sliding_eval [ 70.3%] 340832/484544 windows running_bpb=1.173778 + sliding_eval [ 70.7%] 342432/484544 windows running_bpb=1.173746 + sliding_eval [ 71.0%] 344032/484544 windows running_bpb=1.173383 + sliding_eval [ 71.3%] 345632/484544 windows running_bpb=1.173114 + sliding_eval [ 71.7%] 347232/484544 windows running_bpb=1.173032 + sliding_eval [ 72.0%] 348832/484544 windows running_bpb=1.172852 + sliding_eval [ 72.3%] 350432/484544 windows running_bpb=1.172687 + sliding_eval [ 72.7%] 352032/484544 windows running_bpb=1.172861 + sliding_eval [ 73.0%] 353632/484544 windows running_bpb=1.173080 + sliding_eval [ 73.3%] 355232/484544 windows running_bpb=1.172797 + sliding_eval [ 73.6%] 356832/484544 windows running_bpb=1.172730 + sliding_eval [ 74.0%] 358432/484544 windows running_bpb=1.172425 + sliding_eval [ 74.3%] 360032/484544 windows running_bpb=1.172097 + sliding_eval [ 74.6%] 361632/484544 windows running_bpb=1.171956 + sliding_eval [ 75.0%] 363232/484544 windows running_bpb=1.172263 + sliding_eval [ 75.3%] 364832/484544 windows running_bpb=1.172232 + sliding_eval [ 75.6%] 366432/484544 windows running_bpb=1.172094 + sliding_eval [ 76.0%] 368032/484544 windows running_bpb=1.172021 + sliding_eval [ 76.3%] 369632/484544 windows running_bpb=1.172042 + sliding_eval [ 76.6%] 371232/484544 windows running_bpb=1.172027 + sliding_eval [ 76.9%] 372832/484544 windows running_bpb=1.172095 + sliding_eval [ 77.3%] 374432/484544 windows running_bpb=1.172411 + sliding_eval [ 77.6%] 376032/484544 windows running_bpb=1.172309 + sliding_eval [ 77.9%] 377632/484544 windows running_bpb=1.172438 + sliding_eval [ 78.3%] 379232/484544 windows running_bpb=1.172371 + sliding_eval [ 78.6%] 380832/484544 windows running_bpb=1.172143 + sliding_eval [ 78.9%] 382432/484544 windows running_bpb=1.172136 + sliding_eval [ 79.3%] 384032/484544 windows running_bpb=1.171930 + sliding_eval [ 79.6%] 385632/484544 windows running_bpb=1.172055 + sliding_eval [ 79.9%] 387232/484544 windows running_bpb=1.172059 + sliding_eval [ 80.2%] 388832/484544 windows running_bpb=1.172128 + sliding_eval [ 80.6%] 390432/484544 windows running_bpb=1.172042 + sliding_eval [ 80.9%] 392032/484544 windows running_bpb=1.172036 + sliding_eval [ 81.2%] 393632/484544 windows running_bpb=1.172049 + sliding_eval [ 81.6%] 395232/484544 windows running_bpb=1.171891 + sliding_eval [ 81.9%] 396832/484544 windows running_bpb=1.172118 + sliding_eval [ 82.2%] 398432/484544 windows running_bpb=1.172156 + sliding_eval [ 82.6%] 400032/484544 windows running_bpb=1.172137 + sliding_eval [ 82.9%] 401632/484544 windows running_bpb=1.172112 + sliding_eval [ 83.2%] 403232/484544 windows running_bpb=1.172005 + sliding_eval [ 83.5%] 404832/484544 windows running_bpb=1.172057 + sliding_eval [ 83.9%] 406432/484544 windows running_bpb=1.171870 + sliding_eval [ 84.2%] 408032/484544 windows running_bpb=1.171921 + sliding_eval [ 84.5%] 409632/484544 windows running_bpb=1.171972 + sliding_eval [ 84.9%] 411232/484544 windows running_bpb=1.171830 + sliding_eval [ 85.2%] 412832/484544 windows running_bpb=1.171908 + sliding_eval [ 85.5%] 414432/484544 windows running_bpb=1.171925 + sliding_eval [ 85.9%] 416032/484544 windows running_bpb=1.171919 + sliding_eval [ 86.2%] 417632/484544 windows running_bpb=1.171784 + sliding_eval [ 86.5%] 419232/484544 windows running_bpb=1.171752 + sliding_eval [ 86.9%] 420832/484544 windows running_bpb=1.171934 + sliding_eval [ 87.2%] 422432/484544 windows running_bpb=1.171909 + sliding_eval [ 87.5%] 424032/484544 windows running_bpb=1.171704 + sliding_eval [ 87.8%] 425632/484544 windows running_bpb=1.171639 + sliding_eval [ 88.2%] 427232/484544 windows running_bpb=1.171470 + sliding_eval [ 88.5%] 428832/484544 windows running_bpb=1.171460 + sliding_eval [ 88.8%] 430432/484544 windows running_bpb=1.171413 + sliding_eval [ 89.2%] 432032/484544 windows running_bpb=1.171534 + sliding_eval [ 89.5%] 433632/484544 windows running_bpb=1.171541 + sliding_eval [ 89.8%] 435232/484544 windows running_bpb=1.171416 + sliding_eval [ 90.2%] 436832/484544 windows running_bpb=1.171584 + sliding_eval [ 90.5%] 438432/484544 windows running_bpb=1.171571 + sliding_eval [ 90.8%] 440032/484544 windows running_bpb=1.171540 + sliding_eval [ 91.1%] 441632/484544 windows running_bpb=1.171682 + sliding_eval [ 91.5%] 443232/484544 windows running_bpb=1.171634 + sliding_eval [ 91.8%] 444832/484544 windows running_bpb=1.171706 + sliding_eval [ 92.1%] 446432/484544 windows running_bpb=1.171845 + sliding_eval [ 92.5%] 448032/484544 windows running_bpb=1.171802 + sliding_eval [ 92.8%] 449632/484544 windows running_bpb=1.171803 + sliding_eval [ 93.1%] 451232/484544 windows running_bpb=1.171878 + sliding_eval [ 93.5%] 452832/484544 windows running_bpb=1.171922 + sliding_eval [ 93.8%] 454432/484544 windows running_bpb=1.171702 + sliding_eval [ 94.1%] 456032/484544 windows running_bpb=1.171479 + sliding_eval [ 94.4%] 457632/484544 windows running_bpb=1.171654 + sliding_eval [ 94.8%] 459232/484544 windows running_bpb=1.171587 + sliding_eval [ 95.1%] 460832/484544 windows running_bpb=1.171587 + sliding_eval [ 95.4%] 462432/484544 windows running_bpb=1.171438 + sliding_eval [ 95.8%] 464032/484544 windows running_bpb=1.171330 + sliding_eval [ 96.1%] 465632/484544 windows running_bpb=1.171386 + sliding_eval [ 96.4%] 467232/484544 windows running_bpb=1.171371 + sliding_eval [ 96.8%] 468832/484544 windows running_bpb=1.171463 + sliding_eval [ 97.1%] 470432/484544 windows running_bpb=1.171459 + sliding_eval [ 97.4%] 472032/484544 windows running_bpb=1.171552 + sliding_eval [ 97.7%] 473632/484544 windows running_bpb=1.171419 + sliding_eval [ 98.1%] 475232/484544 windows running_bpb=1.171412 + sliding_eval [ 98.4%] 476832/484544 windows running_bpb=1.171456 + sliding_eval [ 98.7%] 478432/484544 windows running_bpb=1.171349 + sliding_eval [ 99.1%] 480032/484544 windows running_bpb=1.171762 + sliding_eval [ 99.4%] 481632/484544 windows running_bpb=1.171715 + sliding_eval [ 99.7%] 483232/484544 windows running_bpb=1.171813 +final_int8_zlib_roundtrip val_loss:1.9797 val_bpb:1.1725 eval_time:613676ms +final_int8_zlib_roundtrip_exact val_loss:1.97973856 val_bpb:1.17251579 diff --git a/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/submission.json b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/submission.json new file mode 100644 index 000000000..4d7dd4707 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/submission.json @@ -0,0 +1,25 @@ +{ + "author": "someone114514", + "github_id": "someone114514", + "name": "Baseline + EMA + Adaptive Export (Non-Record)", + "blurb": "Non-record run built on the strong Int6 MLP3x + SmearGate + BigramHash + Muon baseline, adding late-stage EMA and adaptive export-time pruning search. Final sliding-window roundtrip reaches 1.1725 val_bpb on 2xH100, but the artifact remains 449,881 bytes over the 16MB target.", + "date": "2026-03-22T00:00:00Z", + "val_loss": 1.97973856, + "val_bpb": 1.17251579, + "pre_quant_val_loss": 2.0093, + "pre_quant_val_bpb": 1.1900, + "step_stop": 3807, + "wallclock_seconds": 1200.112, + "eval_time_seconds": 613.676, + "bytes_total": 16399881, + "bytes_model_int6_zlib": 16342507, + "bytes_code": 57374, + "notes": { + "track": "non-record", + "meets_16mb_budget": false, + "budget_over_bytes": 449881, + "peak_memory_mib": 17070, + "world_size": 2, + "grad_accum_steps": 4 + } +} diff --git a/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/train_gpt.py b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/train_gpt.py new file mode 100644 index 000000000..8088800d5 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_Baseline_EMA_AdaptiveExport/train_gpt.py @@ -0,0 +1,1336 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100)) + + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.01)) + + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 32)) + + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 4096)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "1"))) + ema_beta = float(os.environ.get("EMA_BETA", 0.9998)) + ema_start_frac = float(os.environ.get("EMA_START_FRAC", 0.8)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "0"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.5)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + prune_candidates = os.environ.get("PRUNE_CANDIDATES", "0.00,0.01,0.02,0.03,0.04,0.05") + target_artifact_bytes = int(os.environ.get("TARGET_ARTIFACT_BYTES", 15_950_000)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov, weight_decay=weight_decay), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION +# ----------------------------- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + + +# ----------------------------- +# POST-TRAINING QUANTIZATION (INT8 legacy + INT6 mixed) +# ----------------------------- + +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,smear,bigram.scale", + ).split(",") + if pattern +) +FP16_KEEP_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get("FP16_KEEP_NAME_PATTERNS", "tok_emb,blocks.8.attn.c_k").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 parse_float_list(raw: str) -> list[float]: + values: list[float] = [] + for item in raw.split(","): + item = item.strip() + if item: + values.append(float(item)) + return values + + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1e-12).to(torch.float16) + scale = scale.clamp_min(torch.finfo(torch.float16).tiny) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / 31.0, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8) + return q, scale + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if any(pattern in name for pattern in FP16_KEEP_NAME_PATTERNS): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta[name] + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + + +def build_quantized_blob(state_dict: dict[str, Tensor]) -> tuple[dict[str, Tensor], dict[str, object], bytes]: + quant_result, quant_meta = mixed_quantize_int6(state_dict, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = zlib.compress(quant_raw, 9) + return quant_result, quant_meta, quant_blob + + +def apply_magnitude_pruning(state_dict: dict[str, Tensor], prune_ratio: float) -> dict[str, Tensor]: + if prune_ratio <= 0: + return {name: tensor.clone() for name, tensor in state_dict.items()} + pruned: dict[str, Tensor] = {} + for name, tensor in state_dict.items(): + t = tensor.clone() + if t.is_floating_point() and t.ndim == 2 and t.numel() > 65_536: + threshold = torch.quantile(t.abs().float().flatten(), prune_ratio) + t.masked_fill_(t.abs() < threshold, 0.0) + pruned[name] = t + return pruned + + +def select_export_state( + base_state: dict[str, Tensor], + prune_candidates: list[float], + target_artifact_bytes: int, + code_bytes: int, + log_fn, +) -> tuple[dict[str, Tensor], dict[str, object], bytes, float, int]: + best_state: dict[str, Tensor] | None = None + best_meta: dict[str, object] | None = None + best_blob: bytes | None = None + best_prune_ratio = 0.0 + best_total_bytes: int | None = None + best_meets_target = False + + for prune_ratio in prune_candidates: + export_state = apply_magnitude_pruning(base_state, prune_ratio) + _, quant_meta, quant_blob = build_quantized_blob(export_state) + total_bytes = len(quant_blob) + code_bytes + meets_target = total_bytes <= target_artifact_bytes + log_fn( + f"export_search prune:{prune_ratio:.4f} model_bytes:{len(quant_blob)} " + f"total_bytes:{total_bytes} meets_target:{int(meets_target)}" + ) + should_replace = False + if best_total_bytes is None: + should_replace = True + elif meets_target and not best_meets_target: + should_replace = True + elif meets_target and best_meets_target and prune_ratio < best_prune_ratio: + should_replace = True + elif meets_target == best_meets_target and total_bytes < best_total_bytes: + should_replace = True + if should_replace: + best_state = export_state + best_meta = quant_meta + best_blob = quant_blob + best_prune_ratio = prune_ratio + best_total_bytes = total_bytes + best_meets_target = meets_target + + if best_state is None or best_meta is None or best_blob is None or best_total_bytes is None: + raise RuntimeError("Failed to build export candidates") + if not best_meets_target: + log_fn( + f"export_search no candidate met target:{target_artifact_bytes}, " + f"using_smallest_total_bytes:{best_total_bytes}" + ) + return best_state, best_meta, best_blob, best_prune_ratio, best_total_bytes + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class SmearGate(nn.Module): + """Blend each token's embedding with the previous token's embedding.""" + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + + +class BigramHashEmbedding(nn.Module): + """Hash consecutive token pairs into a learned embedding table.""" + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, rope_base: float, qk_gain_init: float): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> 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: float, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.smear = SmearGate(model_dim) + self.blocks = nn.ModuleList( + [ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init) + for _ in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, +) -> tuple[float, float]: + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if rank == 0 and (bi // batch_seqs) % 50 == 0: + done = min(bi + batch_seqs, len(my_windows)) + pct = done / len(my_windows) * 100 + running_bpb = 0.0 + if token_count.item() > 0: + rl = (loss_sum / token_count).item() + running_bpb = rl / math.log(2.0) * (token_count.item() / byte_count.item()) + print(f" sliding_eval [{pct:5.1f}%] {done}/{len(my_windows)} windows running_bpb={running_bpb:.6f}", flush=True) + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # 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, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.weight_decay, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=0.04, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.weight_decay, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"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"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}" + ) + prune_candidates = parse_float_list(args.prune_candidates) + log0( + f"ema_enabled:{int(args.ema_enabled)} ema_beta:{args.ema_beta} ema_start_frac:{args.ema_start_frac} " + f"swa_enabled:{int(args.swa_enabled)}" + ) + log0( + f"target_artifact_bytes:{args.target_artifact_bytes} " + f"prune_candidates:{','.join(f'{p:.4f}' for p in prune_candidates)}" + ) + log0(f"seed:{args.seed}") + + # DATA LOADER & MODEL WARMUP + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # MAIN TRAINING LOOP + training_time_ms = 0.0 + stop_after_step: int | None = None + ema_state: dict[str, Tensor] | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + 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() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + if args.ema_enabled and step >= max(1, int(args.iterations * args.ema_start_frac)): + current_state = base_model.state_dict() + if ema_state is None: + ema_state = {name: tensor.detach().cpu().clone() for name, tensor in current_state.items()} + log0(f"ema:start step:{step}") + else: + for name, tensor in current_state.items(): + ema_state[name].mul_(args.ema_beta).add_(tensor.detach().cpu(), alpha=1.0 - args.ema_beta) + + # SWA: collect checkpoints during warmdown + if args.swa_enabled and scale < args.swa_start_frac and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + if args.ema_enabled and ema_state is not None: + log0(f"ema:applying beta:{args.ema_beta} start_frac:{args.ema_start_frac}") + current_state = base_model.state_dict() + avg_state = { + name: tensor.to(dtype=current_state[name].dtype) + for name, tensor in ema_state.items() + } + base_model.load_state_dict(avg_state, strict=True) + + # Apply SWA if collected and EMA is disabled + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + current_state = base_model.state_dict() + avg_state = { + name: (tensor / swa_count).to(dtype=current_state[name].dtype) + for name, tensor in swa_state.items() + } + base_model.load_state_dict(avg_state, strict=True) + + # SERIALIZATION + ROUNDTRIP VALIDATION + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + else: + code_bytes = len(code.encode("utf-8")) + + # INT6 mixed quantization + adaptive pruning search + zstd/zlib export + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + export_state, _, quant_blob, selected_prune_ratio, selected_total_bytes = select_export_state( + sd_cpu, prune_candidates, args.target_artifact_bytes, code_bytes, log0 + ) + 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") + log0( + f"export_selected prune:{selected_prune_ratio:.4f} model_bytes:{quant_file_bytes} " + f"total_bytes:{selected_total_bytes}" + ) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + 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() + if _COMPRESSOR == "zstd": + decompressed = zstandard.ZstdDecompressor().decompress(quant_blob_disk) + else: + decompressed = zlib.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu") + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], export_state) + base_model.load_state_dict(deq_state, strict=True) + + # Sliding window eval on int6-roundtripped weights + torch.cuda.synchronize() + t_qeval = time.perf_counter() + if args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + log0(f"final_eval_mode:sliding_window stride:{args.eval_stride} batch_seqs:{args.eval_batch_seqs}") + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, batch_seqs=args.eval_batch_seqs, + ) + else: + log0("final_eval_mode:standard") + q_val_loss, q_val_bpb = eval_val( + args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main()