diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/README.md b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/README.md new file mode 100644 index 0000000000..a74590a3e8 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/README.md @@ -0,0 +1,113 @@ +# QAT x SWA Ablation: Antagonistic Interaction in Quantization-Aware Training + +**val_bpb: 1.1402** (mean of 3 seeds, `no_swa_qat` config, 10% magnitude pruning) + +**This is a non-record research submission.** We present a systematic 2x2 factorial ablation of QAT x SWA interaction, revealing that SWA and QAT are antagonistic mechanisms. This finding explains why prior QAT submissions (#117, #139, smeargate_ortho) underperformed non-QAT entries (#180, #162) — they were running both SWA and QAT simultaneously. + +## Run Command + +```bash +# Single experiment (e.g., best config: QAT without SWA) +bash run.sh no_swa_qat 42 + +# Full 2x2 ablation matrix (4 experiments x 2 seeds = 8 runs) +bash run_matrix.sh +``` + +## Key Finding: SWA Sabotages QAT + +### 3-Seed Validation (no_swa_qat vs control) + +| Config | Seed 42 | Seed 1337 | Seed 2024 | Mean | Std | +|--------|---------|-----------|-----------|------|-----| +| **no_swa_qat** | 1.13969 | 1.14010 | 1.14074 | **1.14018** | ±0.00044 | +| control | 1.14335 | 1.14350 | 1.14462 | **1.14382** | ±0.00056 | + +**Delta: -3.64 mBPB** (no_swa_qat beats control, p < 0.01) + +### Full 2x2 Factorial (2-seed means) + +| Config | QAT | SWA | Mean BPB | Delta vs Control | Rank | +|--------|-----|-----|----------|------------------|------| +| **no_swa_qat** | Yes | No | **1.14018** | **-3.64 mBPB** | **1st** | +| control | No | Yes | 1.14382 | baseline | 2nd | +| qat_snap70 | Yes | Yes | 1.14468 | +0.86 mBPB | 3rd | +| no_swa | No | No | 1.14486 | +1.04 mBPB | 4th | + +### Interpretation + +1. **QAT without SWA wins** (-3.64 mBPB vs control). QAT provides genuine benefit when SWA is removed. +2. **SWA + QAT interfere**: When both enabled (`qat_snap70`), the result is worse than either alone. +3. **SWA alone helps modestly**: +1.04 mBPB improvement over no-SWA baseline. +4. **QAT is 3.5x stronger than SWA**: QAT alone saves 3.64 mBPB vs SWA's 1.04 mBPB. +5. **Training val_bpb is misleading for QAT**: QAT shows worse training metrics (1.1623 vs 1.1538) but better post-quantization BPB. The metric that matters is post-quantization. + +### Why SWA and QAT Conflict + +SWA averages checkpoints across the training tail, producing smooth weight distributions that quantize well passively. QAT uses Straight-Through Estimator (STE) fake-quantization during training, actively shaping weights for quantization boundaries. When combined, SWA's averaging dilutes QAT's quantization-aware adjustments — the averaged weights lose the precise boundary alignment that QAT worked to achieve. + +This explains the competition landscape: #180 (no QAT, SWA, 1.1428) beats #139-area (QAT + SWA, 1.1502) not because QAT doesn't work, but because QAT's benefit is cancelled by SWA's averaging. + +## Full Results + +| Experiment | QAT | SWA | Seed | Steps | Training val_bpb | Final BPB | Artifact (bytes) | ms/step | Pruning | +|---|---|---|---|---|---|---|---|---|---| +| control | No | Yes | 42 | 6616 | 1.1538 | 1.14335 | 15,970,722 | 90.70 | 5% | +| control | No | Yes | 1337 | 6616 | 1.1540 | 1.14350 | 16,211,295 | 90.69 | 5% | +| control | No | Yes | 2024 | ~6600 | — | 1.14462 | 15,614,870 | ~90.6 | 5% | +| qat_snap70 | Yes | Yes | 42 | 6501 | 1.1624 | 1.14429 | 16,431,825 | 92.31 | 5% | +| qat_snap70 | Yes | Yes | 1337 | 6497 | 1.1627 | 1.14506 | 15,780,171 | 92.36 | 5% | +| no_swa | No | No | 42 | 6628 | 1.1537 | 1.14475 | 15,814,075 | 90.54 | 5% | +| no_swa | No | No | 1337 | 6622 | 1.1542 | 1.14497 | 15,822,165 | 90.62 | 5% | +| **no_swa_qat** | **Yes** | **No** | **42** | **6502** | **1.1623** | **1.13969** | 16,393,156 | 92.29 | 5% | +| **no_swa_qat** | **Yes** | **No** | **1337** | **6502** | **1.1632** | **1.14010** | 15,853,395 | 92.30 | 5% | +| **no_swa_qat** | **Yes** | **No** | **2024** | **~6400** | — | **1.14074** | **15,787,003** | ~92.3 | **10%** | + +Note: Seeds 42/1337 used 5% pruning (original PG-300 ablation). Seed 2024 used 10% pruning to meet the 16,000,000-byte artifact limit. QAT configs produce less compressible weights, requiring more aggressive pruning. BPB difference from pruning is within seed variance. + +## Architecture + +Based on PR #180 stack (10L/512d/MLP3x): + +``` +Layers: 10, Dim: 512, MLP_MULT: 3 (h=1536) +Heads: 8, KV Heads: 4 (GQA) +Quantization: int5 MLP / int6 attention + zstd-22 +Embedding: FP16 tied +Optimizer: Muon (m=0.99) + AdamW, WD=0.04 +Magnitude pruning: 10% (configurable via PRUNE_PCT) +Wallclock: 600s (10 min) +Eval: Sliding window stride=64 +``` + +### QAT Implementation + +- **Method**: Straight-Through Estimator (STE) fake-quantization +- **Start**: 70% of training (snap at step ~4550) +- **Quantization**: int6 per-row, matching deployment format +- **Gradient**: STE passes gradients through round() operation + +## Hardware + +- **Ablation matrix (PG-300)**: 8xH100 SXM (RunPod), 8 sequential runs, ~1.7 hours +- **3rd seed validation**: 8xH100 SXM (RunPod), 2 runs (control + no_swa_qat) +- **Per-run wallclock**: 600s (enforced cap) + +## Implications for Competition + +Competitors currently using SWA + QAT together should consider removing SWA when QAT is enabled. Based on our ablation, this substitution alone could yield ~3.6 mBPB improvement. + +The top entries (#549 at 1.1194, #374 at 1.1228) use EMA (Exponential Moving Average) instead of SWA. EMA is a different averaging strategy that may interact differently with QAT — this is an open question for future work. + +## Known Limitations + +- **Based on older stack**: Does not include EMA, XSA, Partial RoPE, or other techniques from entries after PR #180. +- **Pruning variance**: QAT configs require 10% pruning to fit under 16MB; non-QAT configs fit at 5%. This is itself an interesting finding — QAT produces less compressible weight distributions. +- **2x2 factorial only**: Did not test QAT start fraction, EMA vs SWA, or other interaction dimensions. + +## Files + +- `train_gpt.py` — Training script with QAT/SWA toggles and configurable pruning +- `run.sh` — Single experiment runner (accepts experiment name + seed) +- `run_matrix.sh` — Full 2x2 ablation matrix runner +- `logs/` — Complete training logs for all runs diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed1337.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed1337.log new file mode 100644 index 0000000000..6733689c57 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed1337.log @@ -0,0 +1,213 @@ +W0321 12:28:06.724000 35866 torch/distributed/run.py:803] +W0321 12:28:06.724000 35866 torch/distributed/run.py:803] ***************************************** +W0321 12:28:06.724000 35866 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. +W0321 12:28:06.724000 35866 torch/distributed/run.py:803] ***************************************** +logs/control_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9301 train_time:132ms step_avg:132.03ms +step:2/20000 train_loss:7.9520 train_time:195ms step_avg:97.26ms +step:3/20000 train_loss:7.5214 train_time:282ms step_avg:94.11ms +step:4/20000 train_loss:6.9070 train_time:370ms step_avg:92.45ms +step:5/20000 train_loss:6.7673 train_time:461ms step_avg:92.27ms +step:6/20000 train_loss:6.7193 train_time:549ms step_avg:91.49ms +step:7/20000 train_loss:6.5967 train_time:638ms step_avg:91.07ms +step:8/20000 train_loss:6.4792 train_time:725ms step_avg:90.68ms +step:9/20000 train_loss:6.2072 train_time:813ms step_avg:90.36ms +step:10/20000 train_loss:5.9989 train_time:902ms step_avg:90.22ms +step:100/20000 train_loss:3.1650 train_time:8982ms step_avg:89.82ms +step:200/20000 train_loss:2.3785 train_time:18006ms step_avg:90.03ms +step:300/20000 train_loss:2.5368 train_time:27037ms step_avg:90.12ms +step:400/20000 train_loss:2.4009 train_time:36080ms step_avg:90.20ms +step:500/20000 train_loss:2.3920 train_time:45088ms step_avg:90.18ms +step:500/20000 val_loss:2.3522 val_bpb:1.3931 train_time:45116ms step_avg:90.23ms +step:600/20000 train_loss:2.3295 train_time:54143ms step_avg:90.24ms +step:700/20000 train_loss:2.3397 train_time:63216ms step_avg:90.31ms +step:800/20000 train_loss:2.2357 train_time:72367ms step_avg:90.46ms +step:900/20000 train_loss:2.1258 train_time:81438ms step_avg:90.49ms +step:1000/20000 train_loss:2.2735 train_time:90455ms step_avg:90.46ms +step:1000/20000 val_loss:2.2262 val_bpb:1.3185 train_time:90481ms step_avg:90.48ms +step:1100/20000 train_loss:2.3276 train_time:99538ms step_avg:90.49ms +step:1200/20000 train_loss:2.3536 train_time:108610ms step_avg:90.51ms +step:1300/20000 train_loss:2.1035 train_time:117692ms step_avg:90.53ms +step:1400/20000 train_loss:2.1875 train_time:126763ms step_avg:90.54ms +step:1500/20000 train_loss:2.2223 train_time:135782ms step_avg:90.52ms +step:1500/20000 val_loss:2.1866 val_bpb:1.2950 train_time:135809ms step_avg:90.54ms +step:1600/20000 train_loss:2.0765 train_time:144853ms step_avg:90.53ms +step:1700/20000 train_loss:2.1454 train_time:153924ms step_avg:90.54ms +step:1800/20000 train_loss:2.1634 train_time:163005ms step_avg:90.56ms +step:1900/20000 train_loss:2.1330 train_time:172029ms step_avg:90.54ms +step:2000/20000 train_loss:2.0711 train_time:181106ms step_avg:90.55ms +step:2000/20000 val_loss:2.1342 val_bpb:1.2640 train_time:181132ms step_avg:90.57ms +step:2100/20000 train_loss:2.0520 train_time:190175ms step_avg:90.56ms +step:2200/20000 train_loss:2.1424 train_time:199249ms step_avg:90.57ms +step:2300/20000 train_loss:2.1146 train_time:208325ms step_avg:90.58ms +step:2400/20000 train_loss:2.0669 train_time:217341ms step_avg:90.56ms +step:2500/20000 train_loss:2.1742 train_time:226412ms step_avg:90.56ms +step:2500/20000 val_loss:2.1095 val_bpb:1.2494 train_time:226440ms step_avg:90.58ms +step:2600/20000 train_loss:2.1096 train_time:235489ms step_avg:90.57ms +step:2700/20000 train_loss:2.1059 train_time:244565ms step_avg:90.58ms +step:2800/20000 train_loss:2.1557 train_time:253644ms step_avg:90.59ms +step:2900/20000 train_loss:2.0285 train_time:262663ms step_avg:90.57ms +step:3000/20000 train_loss:2.1627 train_time:271737ms step_avg:90.58ms +step:3000/20000 val_loss:2.0933 val_bpb:1.2398 train_time:271764ms step_avg:90.59ms +step:3100/20000 train_loss:2.0393 train_time:280810ms step_avg:90.58ms +step:3200/20000 train_loss:2.1754 train_time:289879ms step_avg:90.59ms +step:3300/20000 train_loss:2.0768 train_time:298890ms step_avg:90.57ms +step:3400/20000 train_loss:2.0261 train_time:307961ms step_avg:90.58ms +step:3500/20000 train_loss:2.1880 train_time:317025ms step_avg:90.58ms +step:3500/20000 val_loss:2.0862 val_bpb:1.2356 train_time:317052ms step_avg:90.59ms +step:3600/20000 train_loss:2.0985 train_time:326098ms step_avg:90.58ms +step:3700/20000 train_loss:2.1002 train_time:335170ms step_avg:90.59ms +step:3800/20000 train_loss:2.0764 train_time:344189ms step_avg:90.58ms +step:3900/20000 train_loss:2.0824 train_time:353256ms step_avg:90.58ms +step:4000/20000 train_loss:1.9784 train_time:362333ms step_avg:90.58ms +step:4000/20000 val_loss:2.0707 val_bpb:1.2264 train_time:362361ms step_avg:90.59ms +step:4100/20000 train_loss:2.0183 train_time:371399ms step_avg:90.59ms +step:4200/20000 train_loss:2.1531 train_time:380464ms step_avg:90.59ms +step:4300/20000 train_loss:2.0568 train_time:389475ms step_avg:90.58ms +step:4400/20000 train_loss:2.0354 train_time:398546ms step_avg:90.58ms +step:4500/20000 train_loss:2.1251 train_time:407619ms step_avg:90.58ms +step:4500/20000 val_loss:2.0467 val_bpb:1.2122 train_time:407646ms step_avg:90.59ms +step:4600/20000 train_loss:1.8449 train_time:416686ms step_avg:90.58ms +step:4700/20000 train_loss:2.2358 train_time:425705ms step_avg:90.58ms +step:4800/20000 train_loss:2.4320 train_time:434782ms step_avg:90.58ms +step:4900/20000 train_loss:2.0510 train_time:443862ms step_avg:90.58ms +step:5000/20000 train_loss:2.1055 train_time:452936ms step_avg:90.59ms +step:5000/20000 val_loss:2.0251 val_bpb:1.1994 train_time:452964ms step_avg:90.59ms +step:5100/20000 train_loss:2.1280 train_time:462007ms step_avg:90.59ms +step:5200/20000 train_loss:2.0407 train_time:471027ms step_avg:90.58ms +step:5300/20000 train_loss:2.0074 train_time:480094ms step_avg:90.58ms +step:5400/20000 train_loss:2.0514 train_time:489162ms step_avg:90.59ms +swa:start step:5450 +step:5500/20000 train_loss:2.0172 train_time:498295ms step_avg:90.60ms +step:5500/20000 val_loss:2.0019 val_bpb:1.1856 train_time:498347ms step_avg:90.61ms +step:5600/20000 train_loss:1.9519 train_time:507408ms step_avg:90.61ms +step:5700/20000 train_loss:2.0145 train_time:516483ms step_avg:90.61ms +step:5800/20000 train_loss:1.9969 train_time:525596ms step_avg:90.62ms +step:5900/20000 train_loss:1.9029 train_time:534701ms step_avg:90.63ms +step:6000/20000 train_loss:1.9386 train_time:543844ms step_avg:90.64ms +step:6000/20000 val_loss:1.9775 val_bpb:1.1712 train_time:543895ms step_avg:90.65ms +step:6100/20000 train_loss:1.9148 train_time:552918ms step_avg:90.64ms +step:6200/20000 train_loss:1.9448 train_time:562044ms step_avg:90.65ms +step:6300/20000 train_loss:1.9425 train_time:571160ms step_avg:90.66ms +step:6400/20000 train_loss:1.9975 train_time:580268ms step_avg:90.67ms +step:6500/20000 train_loss:2.0813 train_time:589397ms step_avg:90.68ms +step:6500/20000 val_loss:1.9514 val_bpb:1.1557 train_time:589461ms step_avg:90.69ms +step:6600/20000 train_loss:1.8407 train_time:598474ms step_avg:90.68ms +step:6616/20000 val_loss:1.9485 val_bpb:1.1540 train_time:600033ms step_avg:90.69ms +stopping_early: wallclock_cap train_time:600033ms step:6616/20000 +peak memory allocated: 18866 MiB reserved: 19074 MiB +swa:applying averaged 24 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 16211295 bytes +Total submission size int8+zlib: 16265743 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.209465 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.137885 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.139741 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133056 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.144442 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.145606 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.147273 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.142851 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.140331 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.141986 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.150707 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.148991 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.150429 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.148706 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.147229 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.147571 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.148936 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.149446 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.155566 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.153019 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.154014 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.152689 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.152076 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.151711 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.152401 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.149994 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.149004 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.149338 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.148129 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.147963 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.147219 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.148471 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.149554 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.150032 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.149538 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.149870 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.148986 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.145042 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.145153 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.146115 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.146268 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.146104 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.144895 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.144606 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.143958 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.144052 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.143995 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.144164 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.143894 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.144537 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.144825 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.144501 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.145528 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.147458 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.146786 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.147496 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.147842 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.147821 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.147417 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.147640 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.147031 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.149811 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.149786 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.149802 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.149449 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.148958 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.148230 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.148231 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.148876 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.148897 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.148881 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.149329 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.149070 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.148686 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.148997 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.149083 +final_int8_zlib_roundtrip val_loss:1.9308 val_bpb:1.1435 eval_time:169136ms +final_int8_zlib_roundtrip_exact val_loss:1.93075067 val_bpb:1.14350233 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed2024.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed2024.log new file mode 100644 index 0000000000..47d497fa7e --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed2024.log @@ -0,0 +1,215 @@ +W0327 20:33:18.139000 36639 torch/distributed/run.py:803] +W0327 20:33:18.139000 36639 torch/distributed/run.py:803] ***************************************** +W0327 20:33:18.139000 36639 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. +W0327 20:33:18.139000 36639 torch/distributed/run.py:803] ***************************************** +logs/control_seed2024.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +seed:2024 +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:6.9327 val_bpb:4.1059 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9341 train_time:141ms step_avg:141.19ms +step:2/20000 train_loss:8.1356 train_time:204ms step_avg:102.02ms +step:3/20000 train_loss:7.6976 train_time:292ms step_avg:97.34ms +step:4/20000 train_loss:6.9820 train_time:381ms step_avg:95.19ms +step:5/20000 train_loss:6.7578 train_time:469ms step_avg:93.90ms +step:6/20000 train_loss:6.6205 train_time:558ms step_avg:93.01ms +step:7/20000 train_loss:6.5157 train_time:646ms step_avg:92.29ms +step:8/20000 train_loss:6.5423 train_time:734ms step_avg:91.76ms +step:9/20000 train_loss:6.3051 train_time:822ms step_avg:91.36ms +step:10/20000 train_loss:6.0754 train_time:911ms step_avg:91.13ms +step:100/20000 train_loss:3.1756 train_time:9006ms step_avg:90.06ms +step:200/20000 train_loss:2.3703 train_time:18062ms step_avg:90.31ms +step:300/20000 train_loss:2.5301 train_time:27148ms step_avg:90.49ms +step:400/20000 train_loss:2.4004 train_time:36254ms step_avg:90.63ms +step:500/20000 train_loss:2.3868 train_time:45305ms step_avg:90.61ms +step:500/20000 val_loss:2.3484 val_bpb:1.3909 train_time:45334ms step_avg:90.67ms +step:600/20000 train_loss:2.3280 train_time:54436ms step_avg:90.73ms +step:700/20000 train_loss:2.3419 train_time:63552ms step_avg:90.79ms +step:800/20000 train_loss:2.2328 train_time:72675ms step_avg:90.84ms +step:900/20000 train_loss:2.1254 train_time:81782ms step_avg:90.87ms +step:1000/20000 train_loss:2.2708 train_time:90842ms step_avg:90.84ms +step:1000/20000 val_loss:2.2251 val_bpb:1.3178 train_time:90871ms step_avg:90.87ms +step:1100/20000 train_loss:2.3174 train_time:99962ms step_avg:90.87ms +step:1200/20000 train_loss:2.3510 train_time:109062ms step_avg:90.88ms +step:1300/20000 train_loss:2.1018 train_time:118170ms step_avg:90.90ms +step:1400/20000 train_loss:2.1805 train_time:127263ms step_avg:90.90ms +step:1500/20000 train_loss:2.2219 train_time:136301ms step_avg:90.87ms +step:1500/20000 val_loss:2.1840 val_bpb:1.2935 train_time:136328ms step_avg:90.89ms +step:1600/20000 train_loss:2.0742 train_time:145387ms step_avg:90.87ms +step:1700/20000 train_loss:2.1419 train_time:154480ms step_avg:90.87ms +step:1800/20000 train_loss:2.1600 train_time:163560ms step_avg:90.87ms +step:1900/20000 train_loss:2.1283 train_time:172587ms step_avg:90.84ms +step:2000/20000 train_loss:2.0680 train_time:181658ms step_avg:90.83ms +step:2000/20000 val_loss:2.1315 val_bpb:1.2624 train_time:181685ms step_avg:90.84ms +step:2100/20000 train_loss:2.0449 train_time:190742ms step_avg:90.83ms +step:2200/20000 train_loss:2.1414 train_time:199812ms step_avg:90.82ms +step:2300/20000 train_loss:2.1085 train_time:208886ms step_avg:90.82ms +step:2400/20000 train_loss:2.0629 train_time:217893ms step_avg:90.79ms +step:2500/20000 train_loss:2.1689 train_time:226956ms step_avg:90.78ms +step:2500/20000 val_loss:2.1060 val_bpb:1.2473 train_time:226983ms step_avg:90.79ms +step:2600/20000 train_loss:2.1085 train_time:236020ms step_avg:90.78ms +step:2700/20000 train_loss:2.1037 train_time:245076ms step_avg:90.77ms +step:2800/20000 train_loss:2.1541 train_time:254151ms step_avg:90.77ms +step:2900/20000 train_loss:2.0258 train_time:263146ms step_avg:90.74ms +step:3000/20000 train_loss:2.1604 train_time:272205ms step_avg:90.74ms +step:3000/20000 val_loss:2.0912 val_bpb:1.2385 train_time:272234ms step_avg:90.74ms +step:3100/20000 train_loss:2.0388 train_time:281270ms step_avg:90.73ms +step:3200/20000 train_loss:2.1751 train_time:290323ms step_avg:90.73ms +step:3300/20000 train_loss:2.0723 train_time:299336ms step_avg:90.71ms +step:3400/20000 train_loss:2.0214 train_time:308406ms step_avg:90.71ms +step:3500/20000 train_loss:2.1826 train_time:317460ms step_avg:90.70ms +step:3500/20000 val_loss:2.0840 val_bpb:1.2343 train_time:317486ms step_avg:90.71ms +step:3600/20000 train_loss:2.1012 train_time:326518ms step_avg:90.70ms +step:3700/20000 train_loss:2.0964 train_time:335579ms step_avg:90.70ms +step:3800/20000 train_loss:2.0744 train_time:344621ms step_avg:90.69ms +step:3900/20000 train_loss:2.0758 train_time:353695ms step_avg:90.69ms +step:4000/20000 train_loss:1.9770 train_time:362740ms step_avg:90.68ms +step:4000/20000 val_loss:2.0675 val_bpb:1.2245 train_time:362767ms step_avg:90.69ms +step:4100/20000 train_loss:2.0127 train_time:371805ms step_avg:90.68ms +step:4200/20000 train_loss:2.1528 train_time:380854ms step_avg:90.68ms +step:4300/20000 train_loss:2.0598 train_time:389848ms step_avg:90.66ms +step:4400/20000 train_loss:2.0296 train_time:398902ms step_avg:90.66ms +step:4500/20000 train_loss:2.1209 train_time:407955ms step_avg:90.66ms +step:4500/20000 val_loss:2.0441 val_bpb:1.2106 train_time:407984ms step_avg:90.66ms +step:4600/20000 train_loss:1.8422 train_time:417005ms step_avg:90.65ms +step:4700/20000 train_loss:2.2334 train_time:425998ms step_avg:90.64ms +step:4800/20000 train_loss:2.4242 train_time:435043ms step_avg:90.63ms +step:4900/20000 train_loss:2.0519 train_time:444101ms step_avg:90.63ms +step:5000/20000 train_loss:2.1003 train_time:453156ms step_avg:90.63ms +step:5000/20000 val_loss:2.0228 val_bpb:1.1980 train_time:453182ms step_avg:90.64ms +step:5100/20000 train_loss:2.1210 train_time:462210ms step_avg:90.63ms +step:5200/20000 train_loss:2.0380 train_time:471199ms step_avg:90.62ms +step:5300/20000 train_loss:2.0054 train_time:480250ms step_avg:90.61ms +step:5400/20000 train_loss:2.0481 train_time:489298ms step_avg:90.61ms +swa:start step:5450 +step:5500/20000 train_loss:2.0155 train_time:498418ms step_avg:90.62ms +step:5500/20000 val_loss:1.9994 val_bpb:1.1841 train_time:498472ms step_avg:90.63ms +step:5600/20000 train_loss:1.9554 train_time:507508ms step_avg:90.63ms +step:5700/20000 train_loss:2.0086 train_time:516546ms step_avg:90.62ms +step:5800/20000 train_loss:1.9992 train_time:525641ms step_avg:90.63ms +step:5900/20000 train_loss:1.8970 train_time:534766ms step_avg:90.64ms +step:6000/20000 train_loss:1.9375 train_time:543850ms step_avg:90.64ms +step:6000/20000 val_loss:1.9756 val_bpb:1.1701 train_time:543904ms step_avg:90.65ms +step:6100/20000 train_loss:1.9143 train_time:552891ms step_avg:90.64ms +step:6200/20000 train_loss:1.9462 train_time:562016ms step_avg:90.65ms +step:6300/20000 train_loss:1.9424 train_time:571110ms step_avg:90.65ms +step:6400/20000 train_loss:1.9950 train_time:580197ms step_avg:90.66ms +step:6500/20000 train_loss:2.0758 train_time:589298ms step_avg:90.66ms +step:6500/20000 val_loss:1.9493 val_bpb:1.1545 train_time:589351ms step_avg:90.67ms +step:6600/20000 train_loss:1.8404 train_time:598360ms step_avg:90.66ms +step:6618/20000 val_loss:1.9463 val_bpb:1.1527 train_time:600106ms step_avg:90.68ms +stopping_early: wallclock_cap train_time:600106ms step:6618/20000 +peak memory allocated: 18866 MiB reserved: 19072 MiB +swa:applying averaged 24 checkpoints +Serialized model: 98437419 bytes +Code size: 54942 bytes +Total submission size: 98492361 bytes +magnitude_pruning: 5.0% of weights with >65536 elements +Serialized model int6+zstd: 15614870 bytes +Total submission size int8+zlib: 15669812 bytes +artifact_headroom: 330188 bytes under limit +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.217287 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.138492 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.139814 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133490 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.145611 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.146774 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.148407 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144032 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.141747 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.143430 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.152176 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150458 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.151809 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.150034 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.148539 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.148830 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.150171 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.150697 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.156726 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.154139 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.155155 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.153807 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.153182 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.152773 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.153439 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151006 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150019 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.150372 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.149232 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.149088 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.148333 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.149563 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.150651 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.151151 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.150640 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151029 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.150137 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.146244 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.146353 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.147270 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.147408 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.147232 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.146012 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.145714 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.145021 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.145126 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.145114 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145276 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145025 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.145656 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.145958 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.145623 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.146674 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.148601 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.147888 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.148619 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.148980 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.148974 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.148552 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.148763 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.148156 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.150942 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.150924 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.150966 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.150609 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.150092 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.149346 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149324 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.149947 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.149980 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.149957 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.150397 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.150141 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.149742 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.150042 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.150160 +final_int8_zlib_roundtrip val_loss:1.9326 val_bpb:1.1446 eval_time:171148ms +final_int8_zlib_roundtrip_exact val_loss:1.93264193 val_bpb:1.14462244 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed42.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed42.log new file mode 100644 index 0000000000..07d0e86b78 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/control_seed42.log @@ -0,0 +1,213 @@ +W0321 12:12:47.620000 1963 torch/distributed/run.py:803] +W0321 12:12:47.620000 1963 torch/distributed/run.py:803] ***************************************** +W0321 12:12:47.620000 1963 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. +W0321 12:12:47.620000 1963 torch/distributed/run.py:803] ***************************************** +logs/control_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:135ms step_avg:134.90ms +step:2/20000 train_loss:8.1444 train_time:197ms step_avg:98.56ms +step:3/20000 train_loss:7.6923 train_time:286ms step_avg:95.24ms +step:4/20000 train_loss:6.9899 train_time:374ms step_avg:93.38ms +step:5/20000 train_loss:6.8344 train_time:463ms step_avg:92.61ms +step:6/20000 train_loss:6.6323 train_time:551ms step_avg:91.80ms +step:7/20000 train_loss:6.5366 train_time:639ms step_avg:91.24ms +step:8/20000 train_loss:6.5777 train_time:727ms step_avg:90.85ms +step:9/20000 train_loss:6.3291 train_time:815ms step_avg:90.53ms +step:10/20000 train_loss:6.0567 train_time:903ms step_avg:90.26ms +step:100/20000 train_loss:3.1601 train_time:8958ms step_avg:89.58ms +step:200/20000 train_loss:2.3844 train_time:17970ms step_avg:89.85ms +step:300/20000 train_loss:2.5430 train_time:26996ms step_avg:89.99ms +step:400/20000 train_loss:2.4077 train_time:36037ms step_avg:90.09ms +step:500/20000 train_loss:2.3905 train_time:45043ms step_avg:90.09ms +step:500/20000 val_loss:2.3527 val_bpb:1.3934 train_time:45070ms step_avg:90.14ms +step:600/20000 train_loss:2.3335 train_time:54105ms step_avg:90.18ms +step:700/20000 train_loss:2.3448 train_time:63174ms step_avg:90.25ms +step:800/20000 train_loss:2.2382 train_time:72247ms step_avg:90.31ms +step:900/20000 train_loss:2.1248 train_time:81304ms step_avg:90.34ms +step:1000/20000 train_loss:2.2744 train_time:90328ms step_avg:90.33ms +step:1000/20000 val_loss:2.2265 val_bpb:1.3186 train_time:90355ms step_avg:90.35ms +step:1100/20000 train_loss:2.3193 train_time:99408ms step_avg:90.37ms +step:1200/20000 train_loss:2.3544 train_time:108469ms step_avg:90.39ms +step:1300/20000 train_loss:2.1053 train_time:117546ms step_avg:90.42ms +step:1400/20000 train_loss:2.1873 train_time:126620ms step_avg:90.44ms +step:1500/20000 train_loss:2.2205 train_time:135634ms step_avg:90.42ms +step:1500/20000 val_loss:2.1845 val_bpb:1.2938 train_time:135661ms step_avg:90.44ms +step:1600/20000 train_loss:2.0765 train_time:144716ms step_avg:90.45ms +step:1700/20000 train_loss:2.1429 train_time:153773ms step_avg:90.45ms +step:1800/20000 train_loss:2.1535 train_time:162847ms step_avg:90.47ms +step:1900/20000 train_loss:2.1290 train_time:171871ms step_avg:90.46ms +step:2000/20000 train_loss:2.0701 train_time:180943ms step_avg:90.47ms +step:2000/20000 val_loss:2.1342 val_bpb:1.2640 train_time:180970ms step_avg:90.49ms +step:2100/20000 train_loss:2.0489 train_time:190019ms step_avg:90.49ms +step:2200/20000 train_loss:2.1456 train_time:199095ms step_avg:90.50ms +step:2300/20000 train_loss:2.1103 train_time:208179ms step_avg:90.51ms +step:2400/20000 train_loss:2.0649 train_time:217191ms step_avg:90.50ms +step:2500/20000 train_loss:2.1695 train_time:226254ms step_avg:90.50ms +step:2500/20000 val_loss:2.1067 val_bpb:1.2477 train_time:226280ms step_avg:90.51ms +step:2600/20000 train_loss:2.1101 train_time:235328ms step_avg:90.51ms +step:2700/20000 train_loss:2.1031 train_time:244399ms step_avg:90.52ms +step:2800/20000 train_loss:2.1578 train_time:253479ms step_avg:90.53ms +step:2900/20000 train_loss:2.0272 train_time:262500ms step_avg:90.52ms +step:3000/20000 train_loss:2.1616 train_time:271569ms step_avg:90.52ms +step:3000/20000 val_loss:2.0930 val_bpb:1.2396 train_time:271595ms step_avg:90.53ms +step:3100/20000 train_loss:2.0364 train_time:280640ms step_avg:90.53ms +step:3200/20000 train_loss:2.1766 train_time:289701ms step_avg:90.53ms +step:3300/20000 train_loss:2.0721 train_time:298721ms step_avg:90.52ms +step:3400/20000 train_loss:2.0236 train_time:307797ms step_avg:90.53ms +step:3500/20000 train_loss:2.1874 train_time:316868ms step_avg:90.53ms +step:3500/20000 val_loss:2.0852 val_bpb:1.2350 train_time:316895ms step_avg:90.54ms +step:3600/20000 train_loss:2.1019 train_time:326017ms step_avg:90.56ms +step:3700/20000 train_loss:2.1009 train_time:335083ms step_avg:90.56ms +step:3800/20000 train_loss:2.0758 train_time:344099ms step_avg:90.55ms +step:3900/20000 train_loss:2.0823 train_time:353164ms step_avg:90.55ms +step:4000/20000 train_loss:1.9780 train_time:362234ms step_avg:90.56ms +step:4000/20000 val_loss:2.0692 val_bpb:1.2255 train_time:362260ms step_avg:90.57ms +step:4100/20000 train_loss:2.0168 train_time:371294ms step_avg:90.56ms +step:4200/20000 train_loss:2.1533 train_time:380360ms step_avg:90.56ms +step:4300/20000 train_loss:2.0597 train_time:389383ms step_avg:90.55ms +step:4400/20000 train_loss:2.0389 train_time:398453ms step_avg:90.56ms +step:4500/20000 train_loss:2.1241 train_time:407531ms step_avg:90.56ms +step:4500/20000 val_loss:2.0457 val_bpb:1.2116 train_time:407559ms step_avg:90.57ms +step:4600/20000 train_loss:1.8452 train_time:416595ms step_avg:90.56ms +step:4700/20000 train_loss:2.2369 train_time:425614ms step_avg:90.56ms +step:4800/20000 train_loss:2.4291 train_time:434681ms step_avg:90.56ms +step:4900/20000 train_loss:2.0516 train_time:443749ms step_avg:90.56ms +step:5000/20000 train_loss:2.1045 train_time:452819ms step_avg:90.56ms +step:5000/20000 val_loss:2.0244 val_bpb:1.1990 train_time:452846ms step_avg:90.57ms +step:5100/20000 train_loss:2.1270 train_time:461894ms step_avg:90.57ms +step:5200/20000 train_loss:2.0407 train_time:470913ms step_avg:90.56ms +step:5300/20000 train_loss:2.0090 train_time:479982ms step_avg:90.56ms +step:5400/20000 train_loss:2.0495 train_time:489050ms step_avg:90.56ms +swa:start step:5450 +step:5500/20000 train_loss:2.0156 train_time:498195ms step_avg:90.58ms +step:5500/20000 val_loss:2.0013 val_bpb:1.1853 train_time:498249ms step_avg:90.59ms +step:5600/20000 train_loss:1.9524 train_time:507331ms step_avg:90.59ms +step:5700/20000 train_loss:2.0104 train_time:516411ms step_avg:90.60ms +step:5800/20000 train_loss:1.9942 train_time:525543ms step_avg:90.61ms +step:5900/20000 train_loss:1.9001 train_time:534664ms step_avg:90.62ms +step:6000/20000 train_loss:1.9401 train_time:543796ms step_avg:90.63ms +step:6000/20000 val_loss:1.9773 val_bpb:1.1711 train_time:543850ms step_avg:90.64ms +step:6100/20000 train_loss:1.9157 train_time:552889ms step_avg:90.64ms +step:6200/20000 train_loss:1.9479 train_time:562021ms step_avg:90.65ms +step:6300/20000 train_loss:1.9434 train_time:571156ms step_avg:90.66ms +step:6400/20000 train_loss:1.9968 train_time:580269ms step_avg:90.67ms +step:6500/20000 train_loss:2.0778 train_time:589415ms step_avg:90.68ms +step:6500/20000 val_loss:1.9510 val_bpb:1.1555 train_time:589467ms step_avg:90.69ms +step:6600/20000 train_loss:1.8436 train_time:598496ms step_avg:90.68ms +step:6616/20000 val_loss:1.9482 val_bpb:1.1538 train_time:600051ms step_avg:90.70ms +stopping_early: wallclock_cap train_time:600051ms step:6616/20000 +peak memory allocated: 18870 MiB reserved: 19076 MiB +swa:applying averaged 24 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15970722 bytes +Total submission size int8+zlib: 16025170 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.206476 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.138649 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.139929 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133460 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.145063 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.146251 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.147823 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.143030 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.140661 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.142366 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.150883 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.149249 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.150516 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.148697 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.147189 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.147518 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.148831 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.149330 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.155522 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.152907 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.153822 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.152481 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.151774 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.151406 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.151988 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.149589 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.148664 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.148969 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.147760 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.147616 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.146874 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.148100 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.149195 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.149693 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.149140 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.149477 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.148582 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.144723 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.144835 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.145807 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.145983 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.145848 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.144619 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.144306 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.143621 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.143699 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.143667 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.143869 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.143582 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.144182 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.144503 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.144196 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.145236 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.147147 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.146438 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.147140 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.147471 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.147421 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.146976 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.147200 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.146609 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.149420 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.149414 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.149465 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.149079 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.148589 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.147844 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.147820 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.148442 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.148467 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.148462 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.148902 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.148646 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.148257 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.148581 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.148676 +final_int8_zlib_roundtrip val_loss:1.9305 val_bpb:1.1434 eval_time:170104ms +final_int8_zlib_roundtrip_exact val_loss:1.93049440 val_bpb:1.14335055 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/matrix.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/matrix.log new file mode 100644 index 0000000000..85355d9728 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/matrix.log @@ -0,0 +1,1767 @@ + +============================================ + Running: control (seed=42) +============================================ +========================================== +Experiment: control (seed=42) +ENABLE_QAT=0 SWA_ENABLED=1 +========================================== +GPUs detected: 8 +W0321 12:12:47.620000 1963 torch/distributed/run.py:803] +W0321 12:12:47.620000 1963 torch/distributed/run.py:803] ***************************************** +W0321 12:12:47.620000 1963 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. +W0321 12:12:47.620000 1963 torch/distributed/run.py:803] ***************************************** +logs/control_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:135ms step_avg:134.90ms +step:2/20000 train_loss:8.1444 train_time:197ms step_avg:98.56ms +step:3/20000 train_loss:7.6923 train_time:286ms step_avg:95.24ms +step:4/20000 train_loss:6.9899 train_time:374ms step_avg:93.38ms +step:5/20000 train_loss:6.8344 train_time:463ms step_avg:92.61ms +step:6/20000 train_loss:6.6323 train_time:551ms step_avg:91.80ms +step:7/20000 train_loss:6.5366 train_time:639ms step_avg:91.24ms +step:8/20000 train_loss:6.5777 train_time:727ms step_avg:90.85ms +step:9/20000 train_loss:6.3291 train_time:815ms step_avg:90.53ms +step:10/20000 train_loss:6.0567 train_time:903ms step_avg:90.26ms +step:100/20000 train_loss:3.1601 train_time:8958ms step_avg:89.58ms +step:200/20000 train_loss:2.3844 train_time:17970ms step_avg:89.85ms +step:300/20000 train_loss:2.5430 train_time:26996ms step_avg:89.99ms +step:400/20000 train_loss:2.4077 train_time:36037ms step_avg:90.09ms +step:500/20000 train_loss:2.3905 train_time:45043ms step_avg:90.09ms +step:500/20000 val_loss:2.3527 val_bpb:1.3934 train_time:45070ms step_avg:90.14ms +step:600/20000 train_loss:2.3335 train_time:54105ms step_avg:90.18ms +step:700/20000 train_loss:2.3448 train_time:63174ms step_avg:90.25ms +step:800/20000 train_loss:2.2382 train_time:72247ms step_avg:90.31ms +step:900/20000 train_loss:2.1248 train_time:81304ms step_avg:90.34ms +step:1000/20000 train_loss:2.2744 train_time:90328ms step_avg:90.33ms +step:1000/20000 val_loss:2.2265 val_bpb:1.3186 train_time:90355ms step_avg:90.35ms +step:1100/20000 train_loss:2.3193 train_time:99408ms step_avg:90.37ms +step:1200/20000 train_loss:2.3544 train_time:108469ms step_avg:90.39ms +step:1300/20000 train_loss:2.1053 train_time:117546ms step_avg:90.42ms +step:1400/20000 train_loss:2.1873 train_time:126620ms step_avg:90.44ms +step:1500/20000 train_loss:2.2205 train_time:135634ms step_avg:90.42ms +step:1500/20000 val_loss:2.1845 val_bpb:1.2938 train_time:135661ms step_avg:90.44ms +step:1600/20000 train_loss:2.0765 train_time:144716ms step_avg:90.45ms +step:1700/20000 train_loss:2.1429 train_time:153773ms step_avg:90.45ms +step:1800/20000 train_loss:2.1535 train_time:162847ms step_avg:90.47ms +step:1900/20000 train_loss:2.1290 train_time:171871ms step_avg:90.46ms +step:2000/20000 train_loss:2.0701 train_time:180943ms step_avg:90.47ms +step:2000/20000 val_loss:2.1342 val_bpb:1.2640 train_time:180970ms step_avg:90.49ms +step:2100/20000 train_loss:2.0489 train_time:190019ms step_avg:90.49ms +step:2200/20000 train_loss:2.1456 train_time:199095ms step_avg:90.50ms +step:2300/20000 train_loss:2.1103 train_time:208179ms step_avg:90.51ms +step:2400/20000 train_loss:2.0649 train_time:217191ms step_avg:90.50ms +step:2500/20000 train_loss:2.1695 train_time:226254ms step_avg:90.50ms +step:2500/20000 val_loss:2.1067 val_bpb:1.2477 train_time:226280ms step_avg:90.51ms +step:2600/20000 train_loss:2.1101 train_time:235328ms step_avg:90.51ms +step:2700/20000 train_loss:2.1031 train_time:244399ms step_avg:90.52ms +step:2800/20000 train_loss:2.1578 train_time:253479ms step_avg:90.53ms +step:2900/20000 train_loss:2.0272 train_time:262500ms step_avg:90.52ms +step:3000/20000 train_loss:2.1616 train_time:271569ms step_avg:90.52ms +step:3000/20000 val_loss:2.0930 val_bpb:1.2396 train_time:271595ms step_avg:90.53ms +step:3100/20000 train_loss:2.0364 train_time:280640ms step_avg:90.53ms +step:3200/20000 train_loss:2.1766 train_time:289701ms step_avg:90.53ms +step:3300/20000 train_loss:2.0721 train_time:298721ms step_avg:90.52ms +step:3400/20000 train_loss:2.0236 train_time:307797ms step_avg:90.53ms +step:3500/20000 train_loss:2.1874 train_time:316868ms step_avg:90.53ms +step:3500/20000 val_loss:2.0852 val_bpb:1.2350 train_time:316895ms step_avg:90.54ms +step:3600/20000 train_loss:2.1019 train_time:326017ms step_avg:90.56ms +step:3700/20000 train_loss:2.1009 train_time:335083ms step_avg:90.56ms +step:3800/20000 train_loss:2.0758 train_time:344099ms step_avg:90.55ms +step:3900/20000 train_loss:2.0823 train_time:353164ms step_avg:90.55ms +step:4000/20000 train_loss:1.9780 train_time:362234ms step_avg:90.56ms +step:4000/20000 val_loss:2.0692 val_bpb:1.2255 train_time:362260ms step_avg:90.57ms +step:4100/20000 train_loss:2.0168 train_time:371294ms step_avg:90.56ms +step:4200/20000 train_loss:2.1533 train_time:380360ms step_avg:90.56ms +step:4300/20000 train_loss:2.0597 train_time:389383ms step_avg:90.55ms +step:4400/20000 train_loss:2.0389 train_time:398453ms step_avg:90.56ms +step:4500/20000 train_loss:2.1241 train_time:407531ms step_avg:90.56ms +step:4500/20000 val_loss:2.0457 val_bpb:1.2116 train_time:407559ms step_avg:90.57ms +step:4600/20000 train_loss:1.8452 train_time:416595ms step_avg:90.56ms +step:4700/20000 train_loss:2.2369 train_time:425614ms step_avg:90.56ms +step:4800/20000 train_loss:2.4291 train_time:434681ms step_avg:90.56ms +step:4900/20000 train_loss:2.0516 train_time:443749ms step_avg:90.56ms +step:5000/20000 train_loss:2.1045 train_time:452819ms step_avg:90.56ms +step:5000/20000 val_loss:2.0244 val_bpb:1.1990 train_time:452846ms step_avg:90.57ms +step:5100/20000 train_loss:2.1270 train_time:461894ms step_avg:90.57ms +step:5200/20000 train_loss:2.0407 train_time:470913ms step_avg:90.56ms +step:5300/20000 train_loss:2.0090 train_time:479982ms step_avg:90.56ms +step:5400/20000 train_loss:2.0495 train_time:489050ms step_avg:90.56ms +swa:start step:5450 +step:5500/20000 train_loss:2.0156 train_time:498195ms step_avg:90.58ms +step:5500/20000 val_loss:2.0013 val_bpb:1.1853 train_time:498249ms step_avg:90.59ms +step:5600/20000 train_loss:1.9524 train_time:507331ms step_avg:90.59ms +step:5700/20000 train_loss:2.0104 train_time:516411ms step_avg:90.60ms +step:5800/20000 train_loss:1.9942 train_time:525543ms step_avg:90.61ms +step:5900/20000 train_loss:1.9001 train_time:534664ms step_avg:90.62ms +step:6000/20000 train_loss:1.9401 train_time:543796ms step_avg:90.63ms +step:6000/20000 val_loss:1.9773 val_bpb:1.1711 train_time:543850ms step_avg:90.64ms +step:6100/20000 train_loss:1.9157 train_time:552889ms step_avg:90.64ms +step:6200/20000 train_loss:1.9479 train_time:562021ms step_avg:90.65ms +step:6300/20000 train_loss:1.9434 train_time:571156ms step_avg:90.66ms +step:6400/20000 train_loss:1.9968 train_time:580269ms step_avg:90.67ms +step:6500/20000 train_loss:2.0778 train_time:589415ms step_avg:90.68ms +step:6500/20000 val_loss:1.9510 val_bpb:1.1555 train_time:589467ms step_avg:90.69ms +step:6600/20000 train_loss:1.8436 train_time:598496ms step_avg:90.68ms +step:6616/20000 val_loss:1.9482 val_bpb:1.1538 train_time:600051ms step_avg:90.70ms +stopping_early: wallclock_cap train_time:600051ms step:6616/20000 +peak memory allocated: 18870 MiB reserved: 19076 MiB +swa:applying averaged 24 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15970722 bytes +Total submission size int8+zlib: 16025170 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.206476 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.138649 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.139929 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133460 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.145063 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.146251 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.147823 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.143030 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.140661 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.142366 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.150883 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.149249 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.150516 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.148697 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.147189 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.147518 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.148831 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.149330 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.155522 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.152907 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.153822 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.152481 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.151774 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.151406 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.151988 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.149589 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.148664 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.148969 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.147760 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.147616 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.146874 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.148100 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.149195 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.149693 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.149140 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.149477 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.148582 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.144723 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.144835 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.145807 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.145983 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.145848 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.144619 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.144306 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.143621 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.143699 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.143667 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.143869 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.143582 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.144182 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.144503 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.144196 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.145236 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.147147 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.146438 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.147140 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.147471 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.147421 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.146976 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.147200 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.146609 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.149420 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.149414 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.149465 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.149079 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.148589 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.147844 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.147820 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.148442 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.148467 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.148462 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.148902 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.148646 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.148257 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.148581 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.148676 +final_int8_zlib_roundtrip val_loss:1.9305 val_bpb:1.1434 eval_time:170104ms +final_int8_zlib_roundtrip_exact val_loss:1.93049440 val_bpb:1.14335055 + +============================================ + Running: control (seed=1337) +============================================ +========================================== +Experiment: control (seed=1337) +ENABLE_QAT=0 SWA_ENABLED=1 +========================================== +GPUs detected: 8 +W0321 12:28:06.724000 35866 torch/distributed/run.py:803] +W0321 12:28:06.724000 35866 torch/distributed/run.py:803] ***************************************** +W0321 12:28:06.724000 35866 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. +W0321 12:28:06.724000 35866 torch/distributed/run.py:803] ***************************************** +logs/control_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9301 train_time:132ms step_avg:132.03ms +step:2/20000 train_loss:7.9520 train_time:195ms step_avg:97.26ms +step:3/20000 train_loss:7.5214 train_time:282ms step_avg:94.11ms +step:4/20000 train_loss:6.9070 train_time:370ms step_avg:92.45ms +step:5/20000 train_loss:6.7673 train_time:461ms step_avg:92.27ms +step:6/20000 train_loss:6.7193 train_time:549ms step_avg:91.49ms +step:7/20000 train_loss:6.5967 train_time:638ms step_avg:91.07ms +step:8/20000 train_loss:6.4792 train_time:725ms step_avg:90.68ms +step:9/20000 train_loss:6.2072 train_time:813ms step_avg:90.36ms +step:10/20000 train_loss:5.9989 train_time:902ms step_avg:90.22ms +step:100/20000 train_loss:3.1650 train_time:8982ms step_avg:89.82ms +step:200/20000 train_loss:2.3785 train_time:18006ms step_avg:90.03ms +step:300/20000 train_loss:2.5368 train_time:27037ms step_avg:90.12ms +step:400/20000 train_loss:2.4009 train_time:36080ms step_avg:90.20ms +step:500/20000 train_loss:2.3920 train_time:45088ms step_avg:90.18ms +step:500/20000 val_loss:2.3522 val_bpb:1.3931 train_time:45116ms step_avg:90.23ms +step:600/20000 train_loss:2.3295 train_time:54143ms step_avg:90.24ms +step:700/20000 train_loss:2.3397 train_time:63216ms step_avg:90.31ms +step:800/20000 train_loss:2.2357 train_time:72367ms step_avg:90.46ms +step:900/20000 train_loss:2.1258 train_time:81438ms step_avg:90.49ms +step:1000/20000 train_loss:2.2735 train_time:90455ms step_avg:90.46ms +step:1000/20000 val_loss:2.2262 val_bpb:1.3185 train_time:90481ms step_avg:90.48ms +step:1100/20000 train_loss:2.3276 train_time:99538ms step_avg:90.49ms +step:1200/20000 train_loss:2.3536 train_time:108610ms step_avg:90.51ms +step:1300/20000 train_loss:2.1035 train_time:117692ms step_avg:90.53ms +step:1400/20000 train_loss:2.1875 train_time:126763ms step_avg:90.54ms +step:1500/20000 train_loss:2.2223 train_time:135782ms step_avg:90.52ms +step:1500/20000 val_loss:2.1866 val_bpb:1.2950 train_time:135809ms step_avg:90.54ms +step:1600/20000 train_loss:2.0765 train_time:144853ms step_avg:90.53ms +step:1700/20000 train_loss:2.1454 train_time:153924ms step_avg:90.54ms +step:1800/20000 train_loss:2.1634 train_time:163005ms step_avg:90.56ms +step:1900/20000 train_loss:2.1330 train_time:172029ms step_avg:90.54ms +step:2000/20000 train_loss:2.0711 train_time:181106ms step_avg:90.55ms +step:2000/20000 val_loss:2.1342 val_bpb:1.2640 train_time:181132ms step_avg:90.57ms +step:2100/20000 train_loss:2.0520 train_time:190175ms step_avg:90.56ms +step:2200/20000 train_loss:2.1424 train_time:199249ms step_avg:90.57ms +step:2300/20000 train_loss:2.1146 train_time:208325ms step_avg:90.58ms +step:2400/20000 train_loss:2.0669 train_time:217341ms step_avg:90.56ms +step:2500/20000 train_loss:2.1742 train_time:226412ms step_avg:90.56ms +step:2500/20000 val_loss:2.1095 val_bpb:1.2494 train_time:226440ms step_avg:90.58ms +step:2600/20000 train_loss:2.1096 train_time:235489ms step_avg:90.57ms +step:2700/20000 train_loss:2.1059 train_time:244565ms step_avg:90.58ms +step:2800/20000 train_loss:2.1557 train_time:253644ms step_avg:90.59ms +step:2900/20000 train_loss:2.0285 train_time:262663ms step_avg:90.57ms +step:3000/20000 train_loss:2.1627 train_time:271737ms step_avg:90.58ms +step:3000/20000 val_loss:2.0933 val_bpb:1.2398 train_time:271764ms step_avg:90.59ms +step:3100/20000 train_loss:2.0393 train_time:280810ms step_avg:90.58ms +step:3200/20000 train_loss:2.1754 train_time:289879ms step_avg:90.59ms +step:3300/20000 train_loss:2.0768 train_time:298890ms step_avg:90.57ms +step:3400/20000 train_loss:2.0261 train_time:307961ms step_avg:90.58ms +step:3500/20000 train_loss:2.1880 train_time:317025ms step_avg:90.58ms +step:3500/20000 val_loss:2.0862 val_bpb:1.2356 train_time:317052ms step_avg:90.59ms +step:3600/20000 train_loss:2.0985 train_time:326098ms step_avg:90.58ms +step:3700/20000 train_loss:2.1002 train_time:335170ms step_avg:90.59ms +step:3800/20000 train_loss:2.0764 train_time:344189ms step_avg:90.58ms +step:3900/20000 train_loss:2.0824 train_time:353256ms step_avg:90.58ms +step:4000/20000 train_loss:1.9784 train_time:362333ms step_avg:90.58ms +step:4000/20000 val_loss:2.0707 val_bpb:1.2264 train_time:362361ms step_avg:90.59ms +step:4100/20000 train_loss:2.0183 train_time:371399ms step_avg:90.59ms +step:4200/20000 train_loss:2.1531 train_time:380464ms step_avg:90.59ms +step:4300/20000 train_loss:2.0568 train_time:389475ms step_avg:90.58ms +step:4400/20000 train_loss:2.0354 train_time:398546ms step_avg:90.58ms +step:4500/20000 train_loss:2.1251 train_time:407619ms step_avg:90.58ms +step:4500/20000 val_loss:2.0467 val_bpb:1.2122 train_time:407646ms step_avg:90.59ms +step:4600/20000 train_loss:1.8449 train_time:416686ms step_avg:90.58ms +step:4700/20000 train_loss:2.2358 train_time:425705ms step_avg:90.58ms +step:4800/20000 train_loss:2.4320 train_time:434782ms step_avg:90.58ms +step:4900/20000 train_loss:2.0510 train_time:443862ms step_avg:90.58ms +step:5000/20000 train_loss:2.1055 train_time:452936ms step_avg:90.59ms +step:5000/20000 val_loss:2.0251 val_bpb:1.1994 train_time:452964ms step_avg:90.59ms +step:5100/20000 train_loss:2.1280 train_time:462007ms step_avg:90.59ms +step:5200/20000 train_loss:2.0407 train_time:471027ms step_avg:90.58ms +step:5300/20000 train_loss:2.0074 train_time:480094ms step_avg:90.58ms +step:5400/20000 train_loss:2.0514 train_time:489162ms step_avg:90.59ms +swa:start step:5450 +step:5500/20000 train_loss:2.0172 train_time:498295ms step_avg:90.60ms +step:5500/20000 val_loss:2.0019 val_bpb:1.1856 train_time:498347ms step_avg:90.61ms +step:5600/20000 train_loss:1.9519 train_time:507408ms step_avg:90.61ms +step:5700/20000 train_loss:2.0145 train_time:516483ms step_avg:90.61ms +step:5800/20000 train_loss:1.9969 train_time:525596ms step_avg:90.62ms +step:5900/20000 train_loss:1.9029 train_time:534701ms step_avg:90.63ms +step:6000/20000 train_loss:1.9386 train_time:543844ms step_avg:90.64ms +step:6000/20000 val_loss:1.9775 val_bpb:1.1712 train_time:543895ms step_avg:90.65ms +step:6100/20000 train_loss:1.9148 train_time:552918ms step_avg:90.64ms +step:6200/20000 train_loss:1.9448 train_time:562044ms step_avg:90.65ms +step:6300/20000 train_loss:1.9425 train_time:571160ms step_avg:90.66ms +step:6400/20000 train_loss:1.9975 train_time:580268ms step_avg:90.67ms +step:6500/20000 train_loss:2.0813 train_time:589397ms step_avg:90.68ms +step:6500/20000 val_loss:1.9514 val_bpb:1.1557 train_time:589461ms step_avg:90.69ms +step:6600/20000 train_loss:1.8407 train_time:598474ms step_avg:90.68ms +step:6616/20000 val_loss:1.9485 val_bpb:1.1540 train_time:600033ms step_avg:90.69ms +stopping_early: wallclock_cap train_time:600033ms step:6616/20000 +peak memory allocated: 18866 MiB reserved: 19074 MiB +swa:applying averaged 24 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 16211295 bytes +Total submission size int8+zlib: 16265743 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.209465 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.137885 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.139741 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133056 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.144442 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.145606 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.147273 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.142851 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.140331 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.141986 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.150707 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.148991 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.150429 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.148706 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.147229 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.147571 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.148936 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.149446 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.155566 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.153019 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.154014 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.152689 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.152076 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.151711 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.152401 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.149994 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.149004 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.149338 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.148129 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.147963 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.147219 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.148471 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.149554 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.150032 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.149538 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.149870 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.148986 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.145042 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.145153 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.146115 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.146268 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.146104 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.144895 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.144606 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.143958 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.144052 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.143995 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.144164 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.143894 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.144537 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.144825 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.144501 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.145528 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.147458 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.146786 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.147496 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.147842 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.147821 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.147417 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.147640 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.147031 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.149811 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.149786 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.149802 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.149449 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.148958 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.148230 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.148231 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.148876 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.148897 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.148881 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.149329 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.149070 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.148686 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.148997 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.149083 +final_int8_zlib_roundtrip val_loss:1.9308 val_bpb:1.1435 eval_time:169136ms +final_int8_zlib_roundtrip_exact val_loss:1.93075067 val_bpb:1.14350233 + +============================================ + Running: qat_snap70 (seed=42) +============================================ +========================================== +Experiment: qat_snap70 (seed=42) +ENABLE_QAT=1 SWA_ENABLED=1 +========================================== +GPUs detected: 8 +W0321 12:42:31.273000 36721 torch/distributed/run.py:803] +W0321 12:42:31.273000 36721 torch/distributed/run.py:803] ***************************************** +W0321 12:42:31.273000 36721 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. +W0321 12:42:31.273000 36721 torch/distributed/run.py:803] ***************************************** +logs/qat_snap70_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:133ms step_avg:133.31ms +step:2/20000 train_loss:8.1444 train_time:196ms step_avg:97.88ms +step:3/20000 train_loss:7.6923 train_time:284ms step_avg:94.63ms +step:4/20000 train_loss:6.9899 train_time:372ms step_avg:93.10ms +step:5/20000 train_loss:6.8344 train_time:460ms step_avg:92.00ms +step:6/20000 train_loss:6.6324 train_time:548ms step_avg:91.34ms +step:7/20000 train_loss:6.5370 train_time:636ms step_avg:90.80ms +step:8/20000 train_loss:6.5776 train_time:724ms step_avg:90.49ms +step:9/20000 train_loss:6.3290 train_time:813ms step_avg:90.28ms +step:10/20000 train_loss:6.0567 train_time:901ms step_avg:90.14ms +step:100/20000 train_loss:3.1540 train_time:8977ms step_avg:89.77ms +step:200/20000 train_loss:2.3791 train_time:18000ms step_avg:90.00ms +step:300/20000 train_loss:2.5430 train_time:27040ms step_avg:90.13ms +step:400/20000 train_loss:2.4100 train_time:36074ms step_avg:90.19ms +step:500/20000 train_loss:2.3966 train_time:45083ms step_avg:90.17ms +step:500/20000 val_loss:2.3559 val_bpb:1.3953 train_time:45110ms step_avg:90.22ms +step:600/20000 train_loss:2.3322 train_time:54142ms step_avg:90.24ms +step:700/20000 train_loss:2.3493 train_time:63211ms step_avg:90.30ms +step:800/20000 train_loss:2.2409 train_time:72279ms step_avg:90.35ms +step:900/20000 train_loss:2.1263 train_time:81345ms step_avg:90.38ms +step:1000/20000 train_loss:2.2766 train_time:90371ms step_avg:90.37ms +step:1000/20000 val_loss:2.2277 val_bpb:1.3194 train_time:90399ms step_avg:90.40ms +step:1100/20000 train_loss:2.3252 train_time:99445ms step_avg:90.40ms +step:1200/20000 train_loss:2.3550 train_time:108513ms step_avg:90.43ms +step:1300/20000 train_loss:2.1028 train_time:117585ms step_avg:90.45ms +step:1400/20000 train_loss:2.1857 train_time:126664ms step_avg:90.47ms +step:1500/20000 train_loss:2.2214 train_time:135685ms step_avg:90.46ms +step:1500/20000 val_loss:2.1850 val_bpb:1.2941 train_time:135713ms step_avg:90.48ms +step:1600/20000 train_loss:2.0770 train_time:144761ms step_avg:90.48ms +step:1700/20000 train_loss:2.1429 train_time:153840ms step_avg:90.49ms +step:1800/20000 train_loss:2.1594 train_time:162913ms step_avg:90.51ms +step:1900/20000 train_loss:2.1275 train_time:171951ms step_avg:90.50ms +step:2000/20000 train_loss:2.0708 train_time:181033ms step_avg:90.52ms +step:2000/20000 val_loss:2.1333 val_bpb:1.2634 train_time:181059ms step_avg:90.53ms +step:2100/20000 train_loss:2.0445 train_time:190120ms step_avg:90.53ms +step:2200/20000 train_loss:2.1843 train_time:199192ms step_avg:90.54ms +step:2300/20000 train_loss:2.1116 train_time:208269ms step_avg:90.55ms +step:2400/20000 train_loss:2.0657 train_time:217296ms step_avg:90.54ms +step:2500/20000 train_loss:2.1730 train_time:226363ms step_avg:90.55ms +step:2500/20000 val_loss:2.1068 val_bpb:1.2478 train_time:226390ms step_avg:90.56ms +step:2600/20000 train_loss:2.1099 train_time:235434ms step_avg:90.55ms +step:2700/20000 train_loss:2.1028 train_time:244502ms step_avg:90.56ms +step:2800/20000 train_loss:2.1580 train_time:253573ms step_avg:90.56ms +step:2900/20000 train_loss:2.0257 train_time:262603ms step_avg:90.55ms +step:3000/20000 train_loss:2.1614 train_time:271674ms step_avg:90.56ms +step:3000/20000 val_loss:2.0934 val_bpb:1.2398 train_time:271701ms step_avg:90.57ms +step:3100/20000 train_loss:2.0352 train_time:280747ms step_avg:90.56ms +step:3200/20000 train_loss:2.1750 train_time:289812ms step_avg:90.57ms +step:3300/20000 train_loss:2.0735 train_time:298831ms step_avg:90.55ms +step:3400/20000 train_loss:2.0256 train_time:307903ms step_avg:90.56ms +step:3500/20000 train_loss:2.1846 train_time:316980ms step_avg:90.57ms +step:3500/20000 val_loss:2.0848 val_bpb:1.2347 train_time:317009ms step_avg:90.57ms +step:3600/20000 train_loss:2.0997 train_time:326050ms step_avg:90.57ms +step:3700/20000 train_loss:2.0994 train_time:335115ms step_avg:90.57ms +step:3800/20000 train_loss:2.0756 train_time:344138ms step_avg:90.56ms +step:3900/20000 train_loss:2.0803 train_time:353204ms step_avg:90.57ms +step:4000/20000 train_loss:1.9790 train_time:362276ms step_avg:90.57ms +step:4000/20000 val_loss:2.0689 val_bpb:1.2253 train_time:362303ms step_avg:90.58ms +step:4100/20000 train_loss:2.0183 train_time:371343ms step_avg:90.57ms +step:4200/20000 train_loss:2.1528 train_time:380411ms step_avg:90.57ms +step:4300/20000 train_loss:2.0572 train_time:389433ms step_avg:90.57ms +step:4400/20000 train_loss:2.0344 train_time:398574ms step_avg:90.58ms +step:4500/20000 train_loss:2.1230 train_time:407645ms step_avg:90.59ms +step:4500/20000 val_loss:2.0453 val_bpb:1.2113 train_time:407671ms step_avg:90.59ms +step:4600/20000 train_loss:1.8533 train_time:417138ms step_avg:90.68ms +step:4700/20000 train_loss:2.2426 train_time:426685ms step_avg:90.78ms +step:4800/20000 train_loss:2.4459 train_time:436292ms step_avg:90.89ms +step:4900/20000 train_loss:2.0637 train_time:445889ms step_avg:91.00ms +step:5000/20000 train_loss:2.1141 train_time:455490ms step_avg:91.10ms +step:5000/20000 val_loss:2.0335 val_bpb:1.2044 train_time:455512ms step_avg:91.10ms +step:5100/20000 train_loss:2.1326 train_time:465088ms step_avg:91.19ms +step:5200/20000 train_loss:2.0552 train_time:474644ms step_avg:91.28ms +step:5300/20000 train_loss:2.0146 train_time:484243ms step_avg:91.37ms +swa:start step:5400 +step:5400/20000 train_loss:2.0562 train_time:493849ms step_avg:91.45ms +step:5500/20000 train_loss:2.0247 train_time:503534ms step_avg:91.55ms +step:5500/20000 val_loss:2.0093 val_bpb:1.1900 train_time:503593ms step_avg:91.56ms +step:5600/20000 train_loss:1.9625 train_time:513204ms step_avg:91.64ms +step:5700/20000 train_loss:2.0228 train_time:522811ms step_avg:91.72ms +step:5800/20000 train_loss:2.0070 train_time:532455ms step_avg:91.80ms +step:5900/20000 train_loss:1.9122 train_time:542097ms step_avg:91.88ms +step:6000/20000 train_loss:1.9467 train_time:551743ms step_avg:91.96ms +step:6000/20000 val_loss:1.9843 val_bpb:1.1752 train_time:551788ms step_avg:91.96ms +step:6100/20000 train_loss:1.9258 train_time:561333ms step_avg:92.02ms +step:6200/20000 train_loss:1.9548 train_time:570997ms step_avg:92.10ms +step:6300/20000 train_loss:1.9530 train_time:580667ms step_avg:92.17ms +step:6400/20000 train_loss:2.0045 train_time:590321ms step_avg:92.24ms +step:6500/20000 train_loss:2.0881 train_time:599960ms step_avg:92.30ms +step:6500/20000 val_loss:1.9627 val_bpb:1.1624 train_time:600006ms step_avg:92.31ms +step:6501/20000 val_loss:1.9627 val_bpb:1.1624 train_time:600102ms step_avg:92.31ms +stopping_early: wallclock_cap train_time:600102ms step:6501/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +swa:applying averaged 23 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 16431825 bytes +Total submission size int8+zlib: 16486273 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.209359 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.139071 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.140218 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133889 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.145876 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147242 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.148731 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144021 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.141315 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.142894 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.151671 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150140 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.151429 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.149578 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.148170 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.148474 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.149809 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.150235 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.156352 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.153743 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.154728 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.153359 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.152798 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.152414 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.153049 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.150600 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.149676 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.149971 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.148776 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.148638 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.147929 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.149161 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.150228 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.150731 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.150255 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.150634 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.149746 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.145898 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.146008 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.146925 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.147108 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.146985 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.145750 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.145489 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.144816 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.144888 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.144880 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145066 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.144802 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.145394 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.145694 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.145385 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.146396 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.148301 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.147622 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.148326 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.148677 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.148635 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.148238 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.148430 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.147854 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.150649 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.150677 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.150740 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.150371 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.149881 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.149132 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149093 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.149725 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.149761 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.149757 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.150194 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.149978 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.149574 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.149877 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.149968 +final_int8_zlib_roundtrip val_loss:1.9321 val_bpb:1.1443 eval_time:169979ms +final_int8_zlib_roundtrip_exact val_loss:1.93208752 val_bpb:1.14429409 + +============================================ + Running: qat_snap70 (seed=1337) +============================================ +========================================== +Experiment: qat_snap70 (seed=1337) +ENABLE_QAT=1 SWA_ENABLED=1 +========================================== +GPUs detected: 8 +W0321 12:56:52.719000 37576 torch/distributed/run.py:803] +W0321 12:56:52.719000 37576 torch/distributed/run.py:803] ***************************************** +W0321 12:56:52.719000 37576 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. +W0321 12:56:52.719000 37576 torch/distributed/run.py:803] ***************************************** +logs/qat_snap70_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:133ms step_avg:133.47ms +step:2/20000 train_loss:7.9520 train_time:195ms step_avg:97.71ms +step:3/20000 train_loss:7.5214 train_time:283ms step_avg:94.48ms +step:4/20000 train_loss:6.9071 train_time:372ms step_avg:93.01ms +step:5/20000 train_loss:6.7673 train_time:460ms step_avg:92.07ms +step:6/20000 train_loss:6.7191 train_time:549ms step_avg:91.49ms +step:7/20000 train_loss:6.5969 train_time:636ms step_avg:90.90ms +step:8/20000 train_loss:6.4795 train_time:724ms step_avg:90.51ms +step:9/20000 train_loss:6.2072 train_time:813ms step_avg:90.33ms +step:10/20000 train_loss:5.9981 train_time:901ms step_avg:90.10ms +step:100/20000 train_loss:3.1560 train_time:8969ms step_avg:89.69ms +step:200/20000 train_loss:2.3846 train_time:17989ms step_avg:89.94ms +step:300/20000 train_loss:2.5343 train_time:27029ms step_avg:90.10ms +step:400/20000 train_loss:2.3997 train_time:36075ms step_avg:90.19ms +step:500/20000 train_loss:2.3902 train_time:45082ms step_avg:90.16ms +step:500/20000 val_loss:2.3502 val_bpb:1.3919 train_time:45109ms step_avg:90.22ms +step:600/20000 train_loss:2.3295 train_time:54157ms step_avg:90.26ms +step:700/20000 train_loss:2.3415 train_time:63226ms step_avg:90.32ms +step:800/20000 train_loss:2.2358 train_time:72309ms step_avg:90.39ms +step:900/20000 train_loss:2.1260 train_time:81377ms step_avg:90.42ms +step:1000/20000 train_loss:2.2775 train_time:90412ms step_avg:90.41ms +step:1000/20000 val_loss:2.2275 val_bpb:1.3193 train_time:90439ms step_avg:90.44ms +step:1100/20000 train_loss:2.3213 train_time:99501ms step_avg:90.46ms +step:1200/20000 train_loss:2.3509 train_time:108577ms step_avg:90.48ms +step:1300/20000 train_loss:2.1036 train_time:117655ms step_avg:90.50ms +step:1400/20000 train_loss:2.1834 train_time:126732ms step_avg:90.52ms +step:1500/20000 train_loss:2.2238 train_time:135774ms step_avg:90.52ms +step:1500/20000 val_loss:2.1861 val_bpb:1.2947 train_time:135802ms step_avg:90.53ms +step:1600/20000 train_loss:2.0769 train_time:144908ms step_avg:90.57ms +step:1700/20000 train_loss:2.1443 train_time:153995ms step_avg:90.59ms +step:1800/20000 train_loss:2.1669 train_time:163069ms step_avg:90.59ms +step:1900/20000 train_loss:2.1328 train_time:172098ms step_avg:90.58ms +step:2000/20000 train_loss:2.0709 train_time:181173ms step_avg:90.59ms +step:2000/20000 val_loss:2.1350 val_bpb:1.2645 train_time:181200ms step_avg:90.60ms +step:2100/20000 train_loss:2.0529 train_time:190255ms step_avg:90.60ms +step:2200/20000 train_loss:2.1687 train_time:199329ms step_avg:90.60ms +step:2300/20000 train_loss:2.1136 train_time:208404ms step_avg:90.61ms +step:2400/20000 train_loss:2.0702 train_time:217426ms step_avg:90.59ms +step:2500/20000 train_loss:2.1726 train_time:226503ms step_avg:90.60ms +step:2500/20000 val_loss:2.1103 val_bpb:1.2499 train_time:226530ms step_avg:90.61ms +step:2600/20000 train_loss:2.1121 train_time:235585ms step_avg:90.61ms +step:2700/20000 train_loss:2.1027 train_time:244653ms step_avg:90.61ms +step:2800/20000 train_loss:2.1562 train_time:253720ms step_avg:90.61ms +step:2900/20000 train_loss:2.0286 train_time:262747ms step_avg:90.60ms +step:3000/20000 train_loss:2.1625 train_time:271822ms step_avg:90.61ms +step:3000/20000 val_loss:2.0943 val_bpb:1.2403 train_time:271850ms step_avg:90.62ms +step:3100/20000 train_loss:2.0396 train_time:280903ms step_avg:90.61ms +step:3200/20000 train_loss:2.1752 train_time:289970ms step_avg:90.62ms +step:3300/20000 train_loss:2.0728 train_time:299000ms step_avg:90.61ms +step:3400/20000 train_loss:2.0260 train_time:308087ms step_avg:90.61ms +step:3500/20000 train_loss:2.1890 train_time:317158ms step_avg:90.62ms +step:3500/20000 val_loss:2.0861 val_bpb:1.2355 train_time:317186ms step_avg:90.62ms +step:3600/20000 train_loss:2.1022 train_time:326236ms step_avg:90.62ms +step:3700/20000 train_loss:2.1012 train_time:335305ms step_avg:90.62ms +step:3800/20000 train_loss:2.0781 train_time:344337ms step_avg:90.61ms +step:3900/20000 train_loss:2.0791 train_time:353411ms step_avg:90.62ms +step:4000/20000 train_loss:1.9774 train_time:362475ms step_avg:90.62ms +step:4000/20000 val_loss:2.0700 val_bpb:1.2260 train_time:362501ms step_avg:90.63ms +step:4100/20000 train_loss:2.0160 train_time:371551ms step_avg:90.62ms +step:4200/20000 train_loss:2.1542 train_time:380622ms step_avg:90.62ms +step:4300/20000 train_loss:2.0562 train_time:389650ms step_avg:90.62ms +step:4400/20000 train_loss:2.0337 train_time:398725ms step_avg:90.62ms +step:4500/20000 train_loss:2.1259 train_time:407799ms step_avg:90.62ms +step:4500/20000 val_loss:2.0463 val_bpb:1.2119 train_time:407826ms step_avg:90.63ms +step:4600/20000 train_loss:1.8552 train_time:417315ms step_avg:90.72ms +step:4700/20000 train_loss:2.2464 train_time:426876ms step_avg:90.82ms +step:4800/20000 train_loss:2.4440 train_time:436489ms step_avg:90.94ms +step:4900/20000 train_loss:2.0638 train_time:446100ms step_avg:91.04ms +step:5000/20000 train_loss:2.1133 train_time:455718ms step_avg:91.14ms +step:5000/20000 val_loss:2.0347 val_bpb:1.2051 train_time:455738ms step_avg:91.15ms +step:5100/20000 train_loss:2.1361 train_time:465321ms step_avg:91.24ms +step:5200/20000 train_loss:2.0522 train_time:474880ms step_avg:91.32ms +step:5300/20000 train_loss:2.0168 train_time:484478ms step_avg:91.41ms +swa:start step:5400 +step:5400/20000 train_loss:2.0565 train_time:494093ms step_avg:91.50ms +step:5500/20000 train_loss:2.0252 train_time:503787ms step_avg:91.60ms +step:5500/20000 val_loss:2.0100 val_bpb:1.1904 train_time:503830ms step_avg:91.61ms +step:5600/20000 train_loss:1.9652 train_time:513435ms step_avg:91.68ms +step:5700/20000 train_loss:2.0227 train_time:523044ms step_avg:91.76ms +step:5800/20000 train_loss:2.0123 train_time:532711ms step_avg:91.85ms +step:5900/20000 train_loss:1.9117 train_time:542362ms step_avg:91.93ms +step:6000/20000 train_loss:1.9451 train_time:552010ms step_avg:92.00ms +step:6000/20000 val_loss:1.9853 val_bpb:1.1758 train_time:552052ms step_avg:92.01ms +step:6100/20000 train_loss:1.9225 train_time:561611ms step_avg:92.07ms +step:6200/20000 train_loss:1.9531 train_time:571295ms step_avg:92.14ms +step:6300/20000 train_loss:1.9518 train_time:580981ms step_avg:92.22ms +step:6400/20000 train_loss:2.0050 train_time:590652ms step_avg:92.29ms +step:6497/20000 val_loss:1.9631 val_bpb:1.1627 train_time:600038ms step_avg:92.36ms +stopping_early: wallclock_cap train_time:600038ms step:6497/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +swa:applying averaged 22 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15780171 bytes +Total submission size int8+zlib: 15834619 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.210436 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.140179 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.140921 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.134716 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.146665 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147428 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.148850 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144470 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.141933 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.143594 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.152524 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150837 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.152232 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.150451 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.149066 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.149485 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.150837 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.151285 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.157308 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.154666 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.155724 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.154386 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.153808 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.153359 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.153999 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151579 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150586 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.150912 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.149717 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.149572 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.148825 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.150031 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.151123 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.151644 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.151132 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151516 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.150598 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.146696 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.146815 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.147742 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.147904 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.147748 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.146530 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.146251 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.145570 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.145642 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.145627 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145806 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145526 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.146127 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.146406 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.146103 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.147148 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.149092 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.148402 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.149102 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.149467 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.149442 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.149066 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.149313 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.148726 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.151526 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.151536 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.151542 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.151159 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.150671 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.149938 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149941 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.150560 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.150584 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.150564 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.150973 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.150721 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.150323 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.150632 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.150727 +final_int8_zlib_roundtrip val_loss:1.9334 val_bpb:1.1451 eval_time:169723ms +final_int8_zlib_roundtrip_exact val_loss:1.93338725 val_bpb:1.14506387 + +============================================ + Running: no_swa (seed=42) +============================================ +========================================== +Experiment: no_swa (seed=42) +ENABLE_QAT=0 SWA_ENABLED=0 +========================================== +GPUs detected: 8 +W0321 13:11:11.416000 38463 torch/distributed/run.py:803] +W0321 13:11:11.416000 38463 torch/distributed/run.py:803] ***************************************** +W0321 13:11:11.416000 38463 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. +W0321 13:11:11.416000 38463 torch/distributed/run.py:803] ***************************************** +logs/no_swa_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:132ms step_avg:132.17ms +step:2/20000 train_loss:8.1444 train_time:196ms step_avg:97.96ms +step:3/20000 train_loss:7.6923 train_time:284ms step_avg:94.64ms +step:4/20000 train_loss:6.9899 train_time:372ms step_avg:93.10ms +step:5/20000 train_loss:6.8345 train_time:460ms step_avg:92.01ms +step:6/20000 train_loss:6.6323 train_time:548ms step_avg:91.27ms +step:7/20000 train_loss:6.5364 train_time:637ms step_avg:90.94ms +step:8/20000 train_loss:6.5772 train_time:724ms step_avg:90.49ms +step:9/20000 train_loss:6.3297 train_time:813ms step_avg:90.30ms +step:10/20000 train_loss:6.0565 train_time:901ms step_avg:90.13ms +step:100/20000 train_loss:3.1658 train_time:8958ms step_avg:89.58ms +step:200/20000 train_loss:2.3797 train_time:17972ms step_avg:89.86ms +step:300/20000 train_loss:2.5399 train_time:27010ms step_avg:90.03ms +step:400/20000 train_loss:2.4091 train_time:36042ms step_avg:90.11ms +step:500/20000 train_loss:2.3978 train_time:45051ms step_avg:90.10ms +step:500/20000 val_loss:2.3547 val_bpb:1.3946 train_time:45079ms step_avg:90.16ms +step:600/20000 train_loss:2.3326 train_time:54116ms step_avg:90.19ms +step:700/20000 train_loss:2.3399 train_time:63177ms step_avg:90.25ms +step:800/20000 train_loss:2.2367 train_time:72241ms step_avg:90.30ms +step:900/20000 train_loss:2.1285 train_time:81297ms step_avg:90.33ms +step:1000/20000 train_loss:2.2713 train_time:90327ms step_avg:90.33ms +step:1000/20000 val_loss:2.2263 val_bpb:1.3185 train_time:90354ms step_avg:90.35ms +step:1100/20000 train_loss:2.3167 train_time:99409ms step_avg:90.37ms +step:1200/20000 train_loss:2.3542 train_time:108467ms step_avg:90.39ms +step:1300/20000 train_loss:2.1079 train_time:117534ms step_avg:90.41ms +step:1400/20000 train_loss:2.1851 train_time:126605ms step_avg:90.43ms +step:1500/20000 train_loss:2.2227 train_time:135625ms step_avg:90.42ms +step:1500/20000 val_loss:2.1850 val_bpb:1.2941 train_time:135651ms step_avg:90.43ms +step:1600/20000 train_loss:2.0774 train_time:144691ms step_avg:90.43ms +step:1700/20000 train_loss:2.1434 train_time:153766ms step_avg:90.45ms +step:1800/20000 train_loss:2.1582 train_time:162832ms step_avg:90.46ms +step:1900/20000 train_loss:2.1298 train_time:171854ms step_avg:90.45ms +step:2000/20000 train_loss:2.0703 train_time:180919ms step_avg:90.46ms +step:2000/20000 val_loss:2.1333 val_bpb:1.2635 train_time:180947ms step_avg:90.47ms +step:2100/20000 train_loss:2.0475 train_time:189972ms step_avg:90.46ms +step:2200/20000 train_loss:2.1359 train_time:199033ms step_avg:90.47ms +step:2300/20000 train_loss:2.1103 train_time:208102ms step_avg:90.48ms +step:2400/20000 train_loss:2.0634 train_time:217118ms step_avg:90.47ms +step:2500/20000 train_loss:2.1689 train_time:226169ms step_avg:90.47ms +step:2500/20000 val_loss:2.1070 val_bpb:1.2479 train_time:226198ms step_avg:90.48ms +step:2600/20000 train_loss:2.1090 train_time:235230ms step_avg:90.47ms +step:2700/20000 train_loss:2.1030 train_time:244291ms step_avg:90.48ms +step:2800/20000 train_loss:2.1550 train_time:253361ms step_avg:90.49ms +step:2900/20000 train_loss:2.0264 train_time:262372ms step_avg:90.47ms +step:3000/20000 train_loss:2.1590 train_time:271437ms step_avg:90.48ms +step:3000/20000 val_loss:2.0930 val_bpb:1.2396 train_time:271463ms step_avg:90.49ms +step:3100/20000 train_loss:2.0377 train_time:280504ms step_avg:90.49ms +step:3200/20000 train_loss:2.1728 train_time:289563ms step_avg:90.49ms +step:3300/20000 train_loss:2.0774 train_time:298579ms step_avg:90.48ms +step:3400/20000 train_loss:2.0220 train_time:307648ms step_avg:90.48ms +step:3500/20000 train_loss:2.1821 train_time:316712ms step_avg:90.49ms +step:3500/20000 val_loss:2.0850 val_bpb:1.2348 train_time:316741ms step_avg:90.50ms +step:3600/20000 train_loss:2.1017 train_time:325784ms step_avg:90.50ms +step:3700/20000 train_loss:2.0988 train_time:334847ms step_avg:90.50ms +step:3800/20000 train_loss:2.0763 train_time:343855ms step_avg:90.49ms +step:3900/20000 train_loss:2.0801 train_time:352923ms step_avg:90.49ms +step:4000/20000 train_loss:1.9802 train_time:361985ms step_avg:90.50ms +step:4000/20000 val_loss:2.0702 val_bpb:1.2261 train_time:362012ms step_avg:90.50ms +step:4100/20000 train_loss:2.0188 train_time:371048ms step_avg:90.50ms +step:4200/20000 train_loss:2.1545 train_time:380114ms step_avg:90.50ms +step:4300/20000 train_loss:2.0620 train_time:389136ms step_avg:90.50ms +step:4400/20000 train_loss:2.0341 train_time:398208ms step_avg:90.50ms +step:4500/20000 train_loss:2.1217 train_time:407264ms step_avg:90.50ms +step:4500/20000 val_loss:2.0464 val_bpb:1.2120 train_time:407290ms step_avg:90.51ms +step:4600/20000 train_loss:1.8442 train_time:416324ms step_avg:90.51ms +step:4700/20000 train_loss:2.2348 train_time:425344ms step_avg:90.50ms +step:4800/20000 train_loss:2.4339 train_time:434402ms step_avg:90.50ms +step:4900/20000 train_loss:2.0514 train_time:443468ms step_avg:90.50ms +step:5000/20000 train_loss:2.1046 train_time:452534ms step_avg:90.51ms +step:5000/20000 val_loss:2.0245 val_bpb:1.1990 train_time:452561ms step_avg:90.51ms +step:5100/20000 train_loss:2.1233 train_time:461599ms step_avg:90.51ms +step:5200/20000 train_loss:2.0415 train_time:470676ms step_avg:90.51ms +step:5300/20000 train_loss:2.0101 train_time:479746ms step_avg:90.52ms +step:5400/20000 train_loss:2.0477 train_time:488820ms step_avg:90.52ms +step:5500/20000 train_loss:2.0177 train_time:497887ms step_avg:90.52ms +step:5500/20000 val_loss:2.0015 val_bpb:1.1854 train_time:497914ms step_avg:90.53ms +step:5600/20000 train_loss:1.9552 train_time:506948ms step_avg:90.53ms +step:5700/20000 train_loss:2.0121 train_time:515965ms step_avg:90.52ms +step:5800/20000 train_loss:1.9973 train_time:525029ms step_avg:90.52ms +step:5900/20000 train_loss:1.9013 train_time:534104ms step_avg:90.53ms +step:6000/20000 train_loss:1.9422 train_time:543166ms step_avg:90.53ms +step:6000/20000 val_loss:1.9777 val_bpb:1.1713 train_time:543193ms step_avg:90.53ms +step:6100/20000 train_loss:1.9139 train_time:552186ms step_avg:90.52ms +step:6200/20000 train_loss:1.9460 train_time:561253ms step_avg:90.52ms +step:6300/20000 train_loss:1.9427 train_time:570318ms step_avg:90.53ms +step:6400/20000 train_loss:1.9974 train_time:579382ms step_avg:90.53ms +step:6500/20000 train_loss:2.0819 train_time:588446ms step_avg:90.53ms +step:6500/20000 val_loss:1.9513 val_bpb:1.1556 train_time:588473ms step_avg:90.53ms +step:6600/20000 train_loss:1.8436 train_time:597464ms step_avg:90.52ms +step:6628/20000 val_loss:1.9480 val_bpb:1.1537 train_time:600067ms step_avg:90.54ms +stopping_early: wallclock_cap train_time:600067ms step:6628/20000 +peak memory allocated: 18866 MiB reserved: 19074 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15814075 bytes +Total submission size int8+zlib: 15868523 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.210598 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.139989 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.141509 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.134515 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.146334 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147660 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.149130 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144544 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.142216 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.143973 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.152774 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150963 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.152287 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.150562 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.149028 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.149300 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.150625 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.151187 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.157323 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.154733 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.155720 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.154314 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.153615 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.153161 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.153855 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151514 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150511 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.150859 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.149576 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.149450 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.148716 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.149910 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.150948 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.151476 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.150953 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151313 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.150423 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.146568 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.146692 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.147611 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.147732 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.147587 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.146373 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.146057 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.145363 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.145418 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.145379 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145535 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145254 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.145859 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.146146 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.145872 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.146919 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.148871 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.148161 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.148916 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.149245 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.149202 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.148790 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.149007 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.148403 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.151212 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.151209 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.151250 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.150882 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.150389 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.149653 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149614 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.150256 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.150286 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.150297 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.150735 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.150494 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.150083 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.150406 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.150485 +final_int8_zlib_roundtrip val_loss:1.9329 val_bpb:1.1448 eval_time:170013ms +final_int8_zlib_roundtrip_exact val_loss:1.93286440 val_bpb:1.14475420 + +============================================ + Running: no_swa (seed=1337) +============================================ +========================================== +Experiment: no_swa (seed=1337) +ENABLE_QAT=0 SWA_ENABLED=0 +========================================== +GPUs detected: 8 +W0321 13:25:32.981000 39318 torch/distributed/run.py:803] +W0321 13:25:32.981000 39318 torch/distributed/run.py:803] ***************************************** +W0321 13:25:32.981000 39318 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. +W0321 13:25:32.981000 39318 torch/distributed/run.py:803] ***************************************** +logs/no_swa_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:133ms step_avg:133.49ms +step:2/20000 train_loss:7.9520 train_time:196ms step_avg:97.76ms +step:3/20000 train_loss:7.5214 train_time:285ms step_avg:94.95ms +step:4/20000 train_loss:6.9071 train_time:372ms step_avg:93.09ms +step:5/20000 train_loss:6.7673 train_time:460ms step_avg:92.06ms +step:6/20000 train_loss:6.7191 train_time:550ms step_avg:91.66ms +step:7/20000 train_loss:6.5968 train_time:638ms step_avg:91.12ms +step:8/20000 train_loss:6.4794 train_time:726ms step_avg:90.71ms +step:9/20000 train_loss:6.2073 train_time:814ms step_avg:90.44ms +step:10/20000 train_loss:5.9979 train_time:903ms step_avg:90.29ms +step:100/20000 train_loss:3.1588 train_time:8968ms step_avg:89.68ms +step:200/20000 train_loss:2.3729 train_time:17999ms step_avg:89.99ms +step:300/20000 train_loss:2.5396 train_time:27044ms step_avg:90.15ms +step:400/20000 train_loss:2.4056 train_time:36085ms step_avg:90.21ms +step:500/20000 train_loss:2.3899 train_time:45097ms step_avg:90.19ms +step:500/20000 val_loss:2.3511 val_bpb:1.3924 train_time:45124ms step_avg:90.25ms +step:600/20000 train_loss:2.3331 train_time:54163ms step_avg:90.27ms +step:700/20000 train_loss:2.3443 train_time:63242ms step_avg:90.35ms +step:800/20000 train_loss:2.2374 train_time:72315ms step_avg:90.39ms +step:900/20000 train_loss:2.1257 train_time:81386ms step_avg:90.43ms +step:1000/20000 train_loss:2.2763 train_time:90422ms step_avg:90.42ms +step:1000/20000 val_loss:2.2287 val_bpb:1.3200 train_time:90449ms step_avg:90.45ms +step:1100/20000 train_loss:2.3257 train_time:99497ms step_avg:90.45ms +step:1200/20000 train_loss:2.3534 train_time:108569ms step_avg:90.47ms +step:1300/20000 train_loss:2.1039 train_time:117652ms step_avg:90.50ms +step:1400/20000 train_loss:2.1843 train_time:126725ms step_avg:90.52ms +step:1500/20000 train_loss:2.2245 train_time:135752ms step_avg:90.50ms +step:1500/20000 val_loss:2.1867 val_bpb:1.2951 train_time:135780ms step_avg:90.52ms +step:1600/20000 train_loss:2.0760 train_time:144831ms step_avg:90.52ms +step:1700/20000 train_loss:2.1430 train_time:153904ms step_avg:90.53ms +step:1800/20000 train_loss:2.1647 train_time:162983ms step_avg:90.55ms +step:1900/20000 train_loss:2.1351 train_time:172015ms step_avg:90.53ms +step:2000/20000 train_loss:2.0711 train_time:181081ms step_avg:90.54ms +step:2000/20000 val_loss:2.1351 val_bpb:1.2645 train_time:181107ms step_avg:90.55ms +step:2100/20000 train_loss:2.0523 train_time:190153ms step_avg:90.55ms +step:2200/20000 train_loss:2.1625 train_time:199232ms step_avg:90.56ms +step:2300/20000 train_loss:2.1125 train_time:208300ms step_avg:90.57ms +step:2400/20000 train_loss:2.0678 train_time:217327ms step_avg:90.55ms +step:2500/20000 train_loss:2.1711 train_time:226449ms step_avg:90.58ms +step:2500/20000 val_loss:2.1090 val_bpb:1.2491 train_time:226477ms step_avg:90.59ms +step:2600/20000 train_loss:2.1122 train_time:235525ms step_avg:90.59ms +step:2700/20000 train_loss:2.1076 train_time:244599ms step_avg:90.59ms +step:2800/20000 train_loss:2.1582 train_time:253677ms step_avg:90.60ms +step:2900/20000 train_loss:2.0270 train_time:262707ms step_avg:90.59ms +step:3000/20000 train_loss:2.1652 train_time:271779ms step_avg:90.59ms +step:3000/20000 val_loss:2.0947 val_bpb:1.2406 train_time:271807ms step_avg:90.60ms +step:3100/20000 train_loss:2.0422 train_time:280861ms step_avg:90.60ms +step:3200/20000 train_loss:2.1762 train_time:289938ms step_avg:90.61ms +step:3300/20000 train_loss:2.0756 train_time:298966ms step_avg:90.60ms +step:3400/20000 train_loss:2.0245 train_time:308032ms step_avg:90.60ms +step:3500/20000 train_loss:2.1869 train_time:317109ms step_avg:90.60ms +step:3500/20000 val_loss:2.0866 val_bpb:1.2358 train_time:317137ms step_avg:90.61ms +step:3600/20000 train_loss:2.1030 train_time:326185ms step_avg:90.61ms +step:3700/20000 train_loss:2.0983 train_time:335255ms step_avg:90.61ms +step:3800/20000 train_loss:2.0795 train_time:344279ms step_avg:90.60ms +step:3900/20000 train_loss:2.0787 train_time:353351ms step_avg:90.60ms +step:4000/20000 train_loss:1.9777 train_time:362411ms step_avg:90.60ms +step:4000/20000 val_loss:2.0705 val_bpb:1.2263 train_time:362438ms step_avg:90.61ms +step:4100/20000 train_loss:2.0156 train_time:371480ms step_avg:90.60ms +step:4200/20000 train_loss:2.1560 train_time:380547ms step_avg:90.61ms +step:4300/20000 train_loss:2.0616 train_time:389574ms step_avg:90.60ms +step:4400/20000 train_loss:2.0360 train_time:398656ms step_avg:90.60ms +step:4500/20000 train_loss:2.1284 train_time:407727ms step_avg:90.61ms +step:4500/20000 val_loss:2.0471 val_bpb:1.2124 train_time:407754ms step_avg:90.61ms +step:4600/20000 train_loss:1.8495 train_time:416800ms step_avg:90.61ms +step:4700/20000 train_loss:2.2364 train_time:425823ms step_avg:90.60ms +step:4800/20000 train_loss:2.4339 train_time:434895ms step_avg:90.60ms +step:4900/20000 train_loss:2.0486 train_time:443970ms step_avg:90.61ms +step:5000/20000 train_loss:2.1054 train_time:453048ms step_avg:90.61ms +step:5000/20000 val_loss:2.0254 val_bpb:1.1996 train_time:453077ms step_avg:90.62ms +step:5100/20000 train_loss:2.1278 train_time:462125ms step_avg:90.61ms +step:5200/20000 train_loss:2.0443 train_time:471145ms step_avg:90.60ms +step:5300/20000 train_loss:2.0105 train_time:480223ms step_avg:90.61ms +step:5400/20000 train_loss:2.0501 train_time:489300ms step_avg:90.61ms +step:5500/20000 train_loss:2.0159 train_time:498370ms step_avg:90.61ms +step:5500/20000 val_loss:2.0021 val_bpb:1.1858 train_time:498398ms step_avg:90.62ms +step:5600/20000 train_loss:1.9537 train_time:507445ms step_avg:90.62ms +step:5700/20000 train_loss:2.0109 train_time:516483ms step_avg:90.61ms +step:5800/20000 train_loss:1.9985 train_time:525550ms step_avg:90.61ms +step:5900/20000 train_loss:1.9004 train_time:534620ms step_avg:90.61ms +step:6000/20000 train_loss:1.9401 train_time:543696ms step_avg:90.62ms +step:6000/20000 val_loss:1.9781 val_bpb:1.1715 train_time:543723ms step_avg:90.62ms +step:6100/20000 train_loss:1.9140 train_time:552726ms step_avg:90.61ms +step:6200/20000 train_loss:1.9456 train_time:561796ms step_avg:90.61ms +step:6300/20000 train_loss:1.9436 train_time:570870ms step_avg:90.61ms +step:6400/20000 train_loss:1.9969 train_time:579955ms step_avg:90.62ms +step:6500/20000 train_loss:2.0813 train_time:589032ms step_avg:90.62ms +step:6500/20000 val_loss:1.9519 val_bpb:1.1560 train_time:589058ms step_avg:90.62ms +step:6600/20000 train_loss:1.8445 train_time:598048ms step_avg:90.61ms +step:6622/20000 val_loss:1.9488 val_bpb:1.1542 train_time:600114ms step_avg:90.62ms +stopping_early: wallclock_cap train_time:600114ms step:6622/20000 +peak memory allocated: 18866 MiB reserved: 19074 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15822165 bytes +Total submission size int8+zlib: 15876613 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.203304 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.138641 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.140397 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133944 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.146099 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147628 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.149099 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144731 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.142173 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.143833 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.152537 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150620 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.152017 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.150250 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.148829 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.149001 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.150497 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.151096 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.157300 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.154768 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.155795 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.154491 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.153844 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.153414 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.154179 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151857 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150873 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.151216 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.150040 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.149853 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.149120 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.150347 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.151387 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.151883 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.151357 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151739 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.150859 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.146918 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.147017 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.147949 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.148103 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.147957 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.146732 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.146452 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.145744 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.145784 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.145756 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145924 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145653 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.146300 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.146592 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.146292 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.147334 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.149253 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.148533 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.149257 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.149590 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.149565 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.149153 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.149354 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.148746 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.151527 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.151531 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.151574 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.151217 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.150730 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.150000 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149989 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.150615 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.150666 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.150642 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.151069 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.150820 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.150432 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.150715 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.150817 +final_int8_zlib_roundtrip val_loss:1.9332 val_bpb:1.1450 eval_time:170113ms +final_int8_zlib_roundtrip_exact val_loss:1.93323306 val_bpb:1.14497254 + +============================================ + Running: no_swa_qat (seed=42) +============================================ +========================================== +Experiment: no_swa_qat (seed=42) +ENABLE_QAT=1 SWA_ENABLED=0 +========================================== +GPUs detected: 8 +W0321 13:39:54.884000 40210 torch/distributed/run.py:803] +W0321 13:39:54.884000 40210 torch/distributed/run.py:803] ***************************************** +W0321 13:39:54.884000 40210 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. +W0321 13:39:54.884000 40210 torch/distributed/run.py:803] ***************************************** +logs/no_swa_qat_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:133ms step_avg:133.01ms +step:2/20000 train_loss:8.1444 train_time:195ms step_avg:97.69ms +step:3/20000 train_loss:7.6923 train_time:283ms step_avg:94.38ms +step:4/20000 train_loss:6.9899 train_time:371ms step_avg:92.85ms +step:5/20000 train_loss:6.8345 train_time:460ms step_avg:92.09ms +step:6/20000 train_loss:6.6323 train_time:548ms step_avg:91.40ms +step:7/20000 train_loss:6.5368 train_time:638ms step_avg:91.11ms +step:8/20000 train_loss:6.5773 train_time:726ms step_avg:90.75ms +step:9/20000 train_loss:6.3278 train_time:814ms step_avg:90.42ms +step:10/20000 train_loss:6.0561 train_time:901ms step_avg:90.13ms +step:100/20000 train_loss:3.1709 train_time:8969ms step_avg:89.69ms +step:200/20000 train_loss:2.3853 train_time:17989ms step_avg:89.95ms +step:300/20000 train_loss:2.5382 train_time:27025ms step_avg:90.08ms +step:400/20000 train_loss:2.4047 train_time:36061ms step_avg:90.15ms +step:500/20000 train_loss:2.3883 train_time:45080ms step_avg:90.16ms +step:500/20000 val_loss:2.3517 val_bpb:1.3928 train_time:45107ms step_avg:90.21ms +step:600/20000 train_loss:2.3279 train_time:54156ms step_avg:90.26ms +step:700/20000 train_loss:2.3403 train_time:63225ms step_avg:90.32ms +step:800/20000 train_loss:2.2361 train_time:72294ms step_avg:90.37ms +step:900/20000 train_loss:2.1274 train_time:81356ms step_avg:90.40ms +step:1000/20000 train_loss:2.2733 train_time:90387ms step_avg:90.39ms +step:1000/20000 val_loss:2.2252 val_bpb:1.3179 train_time:90414ms step_avg:90.41ms +step:1100/20000 train_loss:2.3200 train_time:99467ms step_avg:90.42ms +step:1200/20000 train_loss:2.3560 train_time:108541ms step_avg:90.45ms +step:1300/20000 train_loss:2.1075 train_time:117609ms step_avg:90.47ms +step:1400/20000 train_loss:2.1833 train_time:126680ms step_avg:90.49ms +step:1500/20000 train_loss:2.2230 train_time:135703ms step_avg:90.47ms +step:1500/20000 val_loss:2.1854 val_bpb:1.2943 train_time:135731ms step_avg:90.49ms +step:1600/20000 train_loss:2.0758 train_time:144775ms step_avg:90.48ms +step:1700/20000 train_loss:2.1434 train_time:153843ms step_avg:90.50ms +step:1800/20000 train_loss:2.1632 train_time:162918ms step_avg:90.51ms +step:1900/20000 train_loss:2.1286 train_time:171949ms step_avg:90.50ms +step:2000/20000 train_loss:2.0680 train_time:181015ms step_avg:90.51ms +step:2000/20000 val_loss:2.1321 val_bpb:1.2627 train_time:181042ms step_avg:90.52ms +step:2100/20000 train_loss:2.0516 train_time:190097ms step_avg:90.52ms +step:2200/20000 train_loss:2.1474 train_time:199172ms step_avg:90.53ms +step:2300/20000 train_loss:2.1132 train_time:208236ms step_avg:90.54ms +step:2400/20000 train_loss:2.0646 train_time:217255ms step_avg:90.52ms +step:2500/20000 train_loss:2.1703 train_time:226322ms step_avg:90.53ms +step:2500/20000 val_loss:2.1072 val_bpb:1.2480 train_time:226350ms step_avg:90.54ms +step:2600/20000 train_loss:2.1081 train_time:235396ms step_avg:90.54ms +step:2700/20000 train_loss:2.1040 train_time:244471ms step_avg:90.54ms +step:2800/20000 train_loss:2.1558 train_time:253538ms step_avg:90.55ms +step:2900/20000 train_loss:2.0261 train_time:262563ms step_avg:90.54ms +step:3000/20000 train_loss:2.1609 train_time:271634ms step_avg:90.54ms +step:3000/20000 val_loss:2.0934 val_bpb:1.2398 train_time:271661ms step_avg:90.55ms +step:3100/20000 train_loss:2.0350 train_time:280716ms step_avg:90.55ms +step:3200/20000 train_loss:2.1724 train_time:289780ms step_avg:90.56ms +step:3300/20000 train_loss:2.0731 train_time:298804ms step_avg:90.55ms +step:3400/20000 train_loss:2.0256 train_time:307864ms step_avg:90.55ms +step:3500/20000 train_loss:2.1852 train_time:316933ms step_avg:90.55ms +step:3500/20000 val_loss:2.0848 val_bpb:1.2347 train_time:316960ms step_avg:90.56ms +step:3600/20000 train_loss:2.1010 train_time:325999ms step_avg:90.56ms +step:3700/20000 train_loss:2.0990 train_time:335064ms step_avg:90.56ms +step:3800/20000 train_loss:2.0791 train_time:344090ms step_avg:90.55ms +step:3900/20000 train_loss:2.0846 train_time:353157ms step_avg:90.55ms +step:4000/20000 train_loss:1.9802 train_time:362223ms step_avg:90.56ms +step:4000/20000 val_loss:2.0691 val_bpb:1.2254 train_time:362250ms step_avg:90.56ms +step:4100/20000 train_loss:2.0179 train_time:371295ms step_avg:90.56ms +step:4200/20000 train_loss:2.1516 train_time:380367ms step_avg:90.56ms +step:4300/20000 train_loss:2.0571 train_time:389393ms step_avg:90.56ms +step:4400/20000 train_loss:2.0354 train_time:398454ms step_avg:90.56ms +step:4500/20000 train_loss:2.1229 train_time:407523ms step_avg:90.56ms +step:4500/20000 val_loss:2.0462 val_bpb:1.2119 train_time:407549ms step_avg:90.57ms +step:4600/20000 train_loss:1.8540 train_time:417026ms step_avg:90.66ms +step:4700/20000 train_loss:2.2482 train_time:426605ms step_avg:90.77ms +step:4800/20000 train_loss:2.4444 train_time:436231ms step_avg:90.88ms +step:4900/20000 train_loss:2.0630 train_time:445855ms step_avg:90.99ms +step:5000/20000 train_loss:2.1143 train_time:455478ms step_avg:91.10ms +step:5000/20000 val_loss:2.0339 val_bpb:1.2046 train_time:455499ms step_avg:91.10ms +step:5100/20000 train_loss:2.1348 train_time:465116ms step_avg:91.20ms +step:5200/20000 train_loss:2.0521 train_time:474699ms step_avg:91.29ms +step:5300/20000 train_loss:2.0180 train_time:484327ms step_avg:91.38ms +step:5400/20000 train_loss:2.0598 train_time:493963ms step_avg:91.47ms +step:5500/20000 train_loss:2.0274 train_time:503603ms step_avg:91.56ms +step:5500/20000 val_loss:2.0090 val_bpb:1.1899 train_time:503624ms step_avg:91.57ms +step:5600/20000 train_loss:1.9595 train_time:513242ms step_avg:91.65ms +step:5700/20000 train_loss:2.0196 train_time:522817ms step_avg:91.72ms +step:5800/20000 train_loss:2.0053 train_time:532442ms step_avg:91.80ms +step:5900/20000 train_loss:1.9092 train_time:542097ms step_avg:91.88ms +step:6000/20000 train_loss:1.9488 train_time:551790ms step_avg:91.97ms +step:6000/20000 val_loss:1.9846 val_bpb:1.1754 train_time:551810ms step_avg:91.97ms +step:6100/20000 train_loss:1.9229 train_time:561370ms step_avg:92.03ms +step:6200/20000 train_loss:1.9552 train_time:571007ms step_avg:92.10ms +step:6300/20000 train_loss:1.9502 train_time:580630ms step_avg:92.16ms +step:6400/20000 train_loss:2.0035 train_time:590260ms step_avg:92.23ms +step:6500/20000 train_loss:2.0889 train_time:599888ms step_avg:92.29ms +step:6500/20000 val_loss:1.9625 val_bpb:1.1623 train_time:599908ms step_avg:92.29ms +step:6502/20000 val_loss:1.9625 val_bpb:1.1623 train_time:600100ms step_avg:92.29ms +stopping_early: wallclock_cap train_time:600100ms step:6502/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 16393156 bytes +Total submission size int8+zlib: 16447604 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.206283 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.135907 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.136591 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.129658 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.141151 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.142378 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.144109 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.139585 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.136884 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.138391 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.147117 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.145611 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.146881 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.145199 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.143626 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.144003 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.145315 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.145847 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.152028 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.149442 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.150336 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.148930 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.148287 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.147887 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.148527 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.146133 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.145167 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.145547 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.144320 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.144190 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.143482 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.144732 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.145789 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.146258 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.145700 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.146105 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.145208 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.141312 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.141387 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.142336 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.142506 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.142356 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.141132 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.140850 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.140186 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.140266 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.140238 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.140382 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.140119 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.140727 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.141034 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.140765 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.141793 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.143709 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.142966 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.143698 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.144053 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.144032 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.143610 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.143826 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.143238 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.146037 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.146030 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.146065 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.145701 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.145216 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.144467 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.144439 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.145064 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.145094 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.145079 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.145541 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.145286 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.144900 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.145218 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.145311 +final_int8_zlib_roundtrip val_loss:1.9243 val_bpb:1.1397 eval_time:170272ms +final_int8_zlib_roundtrip_exact val_loss:1.92431130 val_bpb:1.13968856 + +============================================ + Running: no_swa_qat (seed=1337) +============================================ +========================================== +Experiment: no_swa_qat (seed=1337) +ENABLE_QAT=1 SWA_ENABLED=0 +========================================== +GPUs detected: 8 +W0321 13:54:19.915000 41065 torch/distributed/run.py:803] +W0321 13:54:19.915000 41065 torch/distributed/run.py:803] ***************************************** +W0321 13:54:19.915000 41065 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. +W0321 13:54:19.915000 41065 torch/distributed/run.py:803] ***************************************** +logs/no_swa_qat_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:134ms step_avg:133.72ms +step:2/20000 train_loss:7.9520 train_time:196ms step_avg:98.09ms +step:3/20000 train_loss:7.5214 train_time:284ms step_avg:94.62ms +step:4/20000 train_loss:6.9070 train_time:372ms step_avg:92.91ms +step:5/20000 train_loss:6.7672 train_time:460ms step_avg:92.04ms +step:6/20000 train_loss:6.7192 train_time:548ms step_avg:91.39ms +step:7/20000 train_loss:6.5968 train_time:637ms step_avg:90.94ms +step:8/20000 train_loss:6.4791 train_time:725ms step_avg:90.68ms +step:9/20000 train_loss:6.2073 train_time:813ms step_avg:90.35ms +step:10/20000 train_loss:5.9982 train_time:901ms step_avg:90.09ms +step:100/20000 train_loss:3.1553 train_time:8969ms step_avg:89.69ms +step:200/20000 train_loss:2.3704 train_time:17997ms step_avg:89.98ms +step:300/20000 train_loss:2.5352 train_time:27044ms step_avg:90.15ms +step:400/20000 train_loss:2.4038 train_time:36093ms step_avg:90.23ms +step:500/20000 train_loss:2.3895 train_time:45107ms step_avg:90.21ms +step:500/20000 val_loss:2.3499 val_bpb:1.3917 train_time:45134ms step_avg:90.27ms +step:600/20000 train_loss:2.3338 train_time:54178ms step_avg:90.30ms +step:700/20000 train_loss:2.3440 train_time:63256ms step_avg:90.37ms +step:800/20000 train_loss:2.2364 train_time:72325ms step_avg:90.41ms +step:900/20000 train_loss:2.1279 train_time:81395ms step_avg:90.44ms +step:1000/20000 train_loss:2.2741 train_time:90427ms step_avg:90.43ms +step:1000/20000 val_loss:2.2255 val_bpb:1.3180 train_time:90455ms step_avg:90.45ms +step:1100/20000 train_loss:2.3269 train_time:99507ms step_avg:90.46ms +step:1200/20000 train_loss:2.3551 train_time:108580ms step_avg:90.48ms +step:1300/20000 train_loss:2.1018 train_time:117667ms step_avg:90.51ms +step:1400/20000 train_loss:2.1864 train_time:126741ms step_avg:90.53ms +step:1500/20000 train_loss:2.2216 train_time:135776ms step_avg:90.52ms +step:1500/20000 val_loss:2.1864 val_bpb:1.2949 train_time:135804ms step_avg:90.54ms +step:1600/20000 train_loss:2.0797 train_time:144856ms step_avg:90.53ms +step:1700/20000 train_loss:2.1454 train_time:153927ms step_avg:90.55ms +step:1800/20000 train_loss:2.1668 train_time:163003ms step_avg:90.56ms +step:1900/20000 train_loss:2.1333 train_time:172026ms step_avg:90.54ms +step:2000/20000 train_loss:2.0729 train_time:181104ms step_avg:90.55ms +step:2000/20000 val_loss:2.1347 val_bpb:1.2643 train_time:181131ms step_avg:90.57ms +step:2100/20000 train_loss:2.0504 train_time:190187ms step_avg:90.57ms +step:2200/20000 train_loss:2.1622 train_time:199259ms step_avg:90.57ms +step:2300/20000 train_loss:2.1125 train_time:208345ms step_avg:90.58ms +step:2400/20000 train_loss:2.0682 train_time:217372ms step_avg:90.57ms +step:2500/20000 train_loss:2.1739 train_time:226442ms step_avg:90.58ms +step:2500/20000 val_loss:2.1093 val_bpb:1.2492 train_time:226471ms step_avg:90.59ms +step:2600/20000 train_loss:2.1129 train_time:235517ms step_avg:90.58ms +step:2700/20000 train_loss:2.1023 train_time:244583ms step_avg:90.59ms +step:2800/20000 train_loss:2.1587 train_time:253661ms step_avg:90.59ms +step:2900/20000 train_loss:2.0285 train_time:262692ms step_avg:90.58ms +step:3000/20000 train_loss:2.1631 train_time:271766ms step_avg:90.59ms +step:3000/20000 val_loss:2.0946 val_bpb:1.2406 train_time:271794ms step_avg:90.60ms +step:3100/20000 train_loss:2.0382 train_time:280850ms step_avg:90.60ms +step:3200/20000 train_loss:2.1755 train_time:289946ms step_avg:90.61ms +step:3300/20000 train_loss:2.0764 train_time:299020ms step_avg:90.61ms +step:3400/20000 train_loss:2.0281 train_time:308102ms step_avg:90.62ms +step:3500/20000 train_loss:2.1858 train_time:317185ms step_avg:90.62ms +step:3500/20000 val_loss:2.0861 val_bpb:1.2355 train_time:317213ms step_avg:90.63ms +step:3600/20000 train_loss:2.1034 train_time:326264ms step_avg:90.63ms +step:3700/20000 train_loss:2.1020 train_time:335341ms step_avg:90.63ms +step:3800/20000 train_loss:2.0768 train_time:344360ms step_avg:90.62ms +step:3900/20000 train_loss:2.0792 train_time:353434ms step_avg:90.62ms +step:4000/20000 train_loss:1.9752 train_time:362509ms step_avg:90.63ms +step:4000/20000 val_loss:2.0699 val_bpb:1.2259 train_time:362535ms step_avg:90.63ms +step:4100/20000 train_loss:2.0198 train_time:371581ms step_avg:90.63ms +step:4200/20000 train_loss:2.1518 train_time:380657ms step_avg:90.63ms +step:4300/20000 train_loss:2.0618 train_time:389687ms step_avg:90.62ms +step:4400/20000 train_loss:2.0344 train_time:398776ms step_avg:90.63ms +step:4500/20000 train_loss:2.1240 train_time:407851ms step_avg:90.63ms +step:4500/20000 val_loss:2.0472 val_bpb:1.2124 train_time:407878ms step_avg:90.64ms +step:4600/20000 train_loss:1.8566 train_time:417451ms step_avg:90.75ms +step:4700/20000 train_loss:2.2488 train_time:427042ms step_avg:90.86ms +step:4800/20000 train_loss:2.4450 train_time:436685ms step_avg:90.98ms +step:4900/20000 train_loss:2.0646 train_time:446315ms step_avg:91.08ms +step:5000/20000 train_loss:2.1151 train_time:455933ms step_avg:91.19ms +step:5000/20000 val_loss:2.0346 val_bpb:1.2050 train_time:455953ms step_avg:91.19ms +step:5100/20000 train_loss:2.1385 train_time:465542ms step_avg:91.28ms +step:5200/20000 train_loss:2.0534 train_time:475104ms step_avg:91.37ms +step:5300/20000 train_loss:2.0168 train_time:484715ms step_avg:91.46ms +step:5400/20000 train_loss:2.0596 train_time:494316ms step_avg:91.54ms +step:5500/20000 train_loss:2.0259 train_time:503934ms step_avg:91.62ms +step:5500/20000 val_loss:2.0106 val_bpb:1.1908 train_time:503955ms step_avg:91.63ms +step:5600/20000 train_loss:1.9655 train_time:513543ms step_avg:91.70ms +step:5700/20000 train_loss:2.0231 train_time:523101ms step_avg:91.77ms +step:5800/20000 train_loss:2.0066 train_time:532700ms step_avg:91.84ms +step:5900/20000 train_loss:1.9116 train_time:542308ms step_avg:91.92ms +step:6000/20000 train_loss:1.9476 train_time:551909ms step_avg:91.98ms +step:6000/20000 val_loss:1.9860 val_bpb:1.1762 train_time:551930ms step_avg:91.99ms +step:6100/20000 train_loss:1.9256 train_time:561465ms step_avg:92.04ms +step:6200/20000 train_loss:1.9565 train_time:571071ms step_avg:92.11ms +step:6300/20000 train_loss:1.9526 train_time:580675ms step_avg:92.17ms +step:6400/20000 train_loss:2.0045 train_time:590285ms step_avg:92.23ms +step:6500/20000 train_loss:2.0877 train_time:599900ms step_avg:92.29ms +step:6500/20000 val_loss:1.9639 val_bpb:1.1632 train_time:599922ms step_avg:92.30ms +step:6502/20000 val_loss:1.9639 val_bpb:1.1632 train_time:600111ms step_avg:92.30ms +stopping_early: wallclock_cap train_time:600111ms step:6502/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15853395 bytes +Total submission size int8+zlib: 15907843 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.205687 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.135544 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.136189 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.129847 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.141728 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.143025 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.144538 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.140012 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.137477 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.139089 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.147794 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.145777 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.147077 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.145298 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.143821 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.144103 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.145428 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.145891 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.152036 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.149468 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.150505 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.149183 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.148568 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.148195 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.148855 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.146443 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.145513 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.145858 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.144618 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.144500 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.143746 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.144948 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.146024 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.146580 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.146054 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.146445 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.145529 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.141630 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.141714 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.142689 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.142847 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.142679 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.141474 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.141202 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.140519 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.140605 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.140559 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.140750 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.140464 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.141103 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.141405 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.141119 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.142138 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.144041 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.143356 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.144084 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.144443 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.144435 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.144012 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.144218 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.143637 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.146437 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.146447 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.146484 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.146123 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.145610 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.144888 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.144865 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.145488 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.145526 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.145492 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.145933 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.145689 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.145306 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.145592 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.145673 +final_int8_zlib_roundtrip val_loss:1.9250 val_bpb:1.1401 eval_time:170052ms +final_int8_zlib_roundtrip_exact val_loss:1.92500262 val_bpb:1.14009800 + +============================================ + ALL EXPERIMENTS COMPLETE +============================================ +Logs in: /workspace/parameter-golf/experiments/alex_qat/logs/ diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed1337.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed1337.log new file mode 100644 index 0000000000..ff26ff62dc --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed1337.log @@ -0,0 +1,210 @@ +W0321 13:54:19.915000 41065 torch/distributed/run.py:803] +W0321 13:54:19.915000 41065 torch/distributed/run.py:803] ***************************************** +W0321 13:54:19.915000 41065 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. +W0321 13:54:19.915000 41065 torch/distributed/run.py:803] ***************************************** +logs/no_swa_qat_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:134ms step_avg:133.72ms +step:2/20000 train_loss:7.9520 train_time:196ms step_avg:98.09ms +step:3/20000 train_loss:7.5214 train_time:284ms step_avg:94.62ms +step:4/20000 train_loss:6.9070 train_time:372ms step_avg:92.91ms +step:5/20000 train_loss:6.7672 train_time:460ms step_avg:92.04ms +step:6/20000 train_loss:6.7192 train_time:548ms step_avg:91.39ms +step:7/20000 train_loss:6.5968 train_time:637ms step_avg:90.94ms +step:8/20000 train_loss:6.4791 train_time:725ms step_avg:90.68ms +step:9/20000 train_loss:6.2073 train_time:813ms step_avg:90.35ms +step:10/20000 train_loss:5.9982 train_time:901ms step_avg:90.09ms +step:100/20000 train_loss:3.1553 train_time:8969ms step_avg:89.69ms +step:200/20000 train_loss:2.3704 train_time:17997ms step_avg:89.98ms +step:300/20000 train_loss:2.5352 train_time:27044ms step_avg:90.15ms +step:400/20000 train_loss:2.4038 train_time:36093ms step_avg:90.23ms +step:500/20000 train_loss:2.3895 train_time:45107ms step_avg:90.21ms +step:500/20000 val_loss:2.3499 val_bpb:1.3917 train_time:45134ms step_avg:90.27ms +step:600/20000 train_loss:2.3338 train_time:54178ms step_avg:90.30ms +step:700/20000 train_loss:2.3440 train_time:63256ms step_avg:90.37ms +step:800/20000 train_loss:2.2364 train_time:72325ms step_avg:90.41ms +step:900/20000 train_loss:2.1279 train_time:81395ms step_avg:90.44ms +step:1000/20000 train_loss:2.2741 train_time:90427ms step_avg:90.43ms +step:1000/20000 val_loss:2.2255 val_bpb:1.3180 train_time:90455ms step_avg:90.45ms +step:1100/20000 train_loss:2.3269 train_time:99507ms step_avg:90.46ms +step:1200/20000 train_loss:2.3551 train_time:108580ms step_avg:90.48ms +step:1300/20000 train_loss:2.1018 train_time:117667ms step_avg:90.51ms +step:1400/20000 train_loss:2.1864 train_time:126741ms step_avg:90.53ms +step:1500/20000 train_loss:2.2216 train_time:135776ms step_avg:90.52ms +step:1500/20000 val_loss:2.1864 val_bpb:1.2949 train_time:135804ms step_avg:90.54ms +step:1600/20000 train_loss:2.0797 train_time:144856ms step_avg:90.53ms +step:1700/20000 train_loss:2.1454 train_time:153927ms step_avg:90.55ms +step:1800/20000 train_loss:2.1668 train_time:163003ms step_avg:90.56ms +step:1900/20000 train_loss:2.1333 train_time:172026ms step_avg:90.54ms +step:2000/20000 train_loss:2.0729 train_time:181104ms step_avg:90.55ms +step:2000/20000 val_loss:2.1347 val_bpb:1.2643 train_time:181131ms step_avg:90.57ms +step:2100/20000 train_loss:2.0504 train_time:190187ms step_avg:90.57ms +step:2200/20000 train_loss:2.1622 train_time:199259ms step_avg:90.57ms +step:2300/20000 train_loss:2.1125 train_time:208345ms step_avg:90.58ms +step:2400/20000 train_loss:2.0682 train_time:217372ms step_avg:90.57ms +step:2500/20000 train_loss:2.1739 train_time:226442ms step_avg:90.58ms +step:2500/20000 val_loss:2.1093 val_bpb:1.2492 train_time:226471ms step_avg:90.59ms +step:2600/20000 train_loss:2.1129 train_time:235517ms step_avg:90.58ms +step:2700/20000 train_loss:2.1023 train_time:244583ms step_avg:90.59ms +step:2800/20000 train_loss:2.1587 train_time:253661ms step_avg:90.59ms +step:2900/20000 train_loss:2.0285 train_time:262692ms step_avg:90.58ms +step:3000/20000 train_loss:2.1631 train_time:271766ms step_avg:90.59ms +step:3000/20000 val_loss:2.0946 val_bpb:1.2406 train_time:271794ms step_avg:90.60ms +step:3100/20000 train_loss:2.0382 train_time:280850ms step_avg:90.60ms +step:3200/20000 train_loss:2.1755 train_time:289946ms step_avg:90.61ms +step:3300/20000 train_loss:2.0764 train_time:299020ms step_avg:90.61ms +step:3400/20000 train_loss:2.0281 train_time:308102ms step_avg:90.62ms +step:3500/20000 train_loss:2.1858 train_time:317185ms step_avg:90.62ms +step:3500/20000 val_loss:2.0861 val_bpb:1.2355 train_time:317213ms step_avg:90.63ms +step:3600/20000 train_loss:2.1034 train_time:326264ms step_avg:90.63ms +step:3700/20000 train_loss:2.1020 train_time:335341ms step_avg:90.63ms +step:3800/20000 train_loss:2.0768 train_time:344360ms step_avg:90.62ms +step:3900/20000 train_loss:2.0792 train_time:353434ms step_avg:90.62ms +step:4000/20000 train_loss:1.9752 train_time:362509ms step_avg:90.63ms +step:4000/20000 val_loss:2.0699 val_bpb:1.2259 train_time:362535ms step_avg:90.63ms +step:4100/20000 train_loss:2.0198 train_time:371581ms step_avg:90.63ms +step:4200/20000 train_loss:2.1518 train_time:380657ms step_avg:90.63ms +step:4300/20000 train_loss:2.0618 train_time:389687ms step_avg:90.62ms +step:4400/20000 train_loss:2.0344 train_time:398776ms step_avg:90.63ms +step:4500/20000 train_loss:2.1240 train_time:407851ms step_avg:90.63ms +step:4500/20000 val_loss:2.0472 val_bpb:1.2124 train_time:407878ms step_avg:90.64ms +step:4600/20000 train_loss:1.8566 train_time:417451ms step_avg:90.75ms +step:4700/20000 train_loss:2.2488 train_time:427042ms step_avg:90.86ms +step:4800/20000 train_loss:2.4450 train_time:436685ms step_avg:90.98ms +step:4900/20000 train_loss:2.0646 train_time:446315ms step_avg:91.08ms +step:5000/20000 train_loss:2.1151 train_time:455933ms step_avg:91.19ms +step:5000/20000 val_loss:2.0346 val_bpb:1.2050 train_time:455953ms step_avg:91.19ms +step:5100/20000 train_loss:2.1385 train_time:465542ms step_avg:91.28ms +step:5200/20000 train_loss:2.0534 train_time:475104ms step_avg:91.37ms +step:5300/20000 train_loss:2.0168 train_time:484715ms step_avg:91.46ms +step:5400/20000 train_loss:2.0596 train_time:494316ms step_avg:91.54ms +step:5500/20000 train_loss:2.0259 train_time:503934ms step_avg:91.62ms +step:5500/20000 val_loss:2.0106 val_bpb:1.1908 train_time:503955ms step_avg:91.63ms +step:5600/20000 train_loss:1.9655 train_time:513543ms step_avg:91.70ms +step:5700/20000 train_loss:2.0231 train_time:523101ms step_avg:91.77ms +step:5800/20000 train_loss:2.0066 train_time:532700ms step_avg:91.84ms +step:5900/20000 train_loss:1.9116 train_time:542308ms step_avg:91.92ms +step:6000/20000 train_loss:1.9476 train_time:551909ms step_avg:91.98ms +step:6000/20000 val_loss:1.9860 val_bpb:1.1762 train_time:551930ms step_avg:91.99ms +step:6100/20000 train_loss:1.9256 train_time:561465ms step_avg:92.04ms +step:6200/20000 train_loss:1.9565 train_time:571071ms step_avg:92.11ms +step:6300/20000 train_loss:1.9526 train_time:580675ms step_avg:92.17ms +step:6400/20000 train_loss:2.0045 train_time:590285ms step_avg:92.23ms +step:6500/20000 train_loss:2.0877 train_time:599900ms step_avg:92.29ms +step:6500/20000 val_loss:1.9639 val_bpb:1.1632 train_time:599922ms step_avg:92.30ms +step:6502/20000 val_loss:1.9639 val_bpb:1.1632 train_time:600111ms step_avg:92.30ms +stopping_early: wallclock_cap train_time:600111ms step:6502/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15853395 bytes +Total submission size int8+zlib: 15907843 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.205687 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.135544 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.136189 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.129847 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.141728 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.143025 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.144538 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.140012 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.137477 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.139089 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.147794 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.145777 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.147077 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.145298 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.143821 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.144103 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.145428 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.145891 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.152036 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.149468 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.150505 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.149183 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.148568 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.148195 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.148855 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.146443 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.145513 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.145858 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.144618 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.144500 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.143746 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.144948 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.146024 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.146580 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.146054 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.146445 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.145529 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.141630 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.141714 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.142689 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.142847 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.142679 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.141474 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.141202 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.140519 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.140605 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.140559 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.140750 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.140464 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.141103 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.141405 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.141119 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.142138 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.144041 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.143356 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.144084 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.144443 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.144435 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.144012 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.144218 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.143637 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.146437 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.146447 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.146484 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.146123 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.145610 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.144888 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.144865 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.145488 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.145526 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.145492 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.145933 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.145689 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.145306 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.145592 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.145673 +final_int8_zlib_roundtrip val_loss:1.9250 val_bpb:1.1401 eval_time:170052ms +final_int8_zlib_roundtrip_exact val_loss:1.92500262 val_bpb:1.14009800 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed2024.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed2024.log new file mode 100644 index 0000000000..82acd8573e --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed2024.log @@ -0,0 +1,210 @@ +W0327 20:47:48.973000 37571 torch/distributed/run.py:803] +W0327 20:47:48.973000 37571 torch/distributed/run.py:803] ***************************************** +W0327 20:47:48.973000 37571 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. +W0327 20:47:48.973000 37571 torch/distributed/run.py:803] ***************************************** +logs/no_swa_qat_seed2024.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +seed:2024 +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:6.9327 val_bpb:4.1059 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9341 train_time:140ms step_avg:139.78ms +step:2/20000 train_loss:8.1356 train_time:203ms step_avg:101.68ms +step:3/20000 train_loss:7.6976 train_time:292ms step_avg:97.21ms +step:4/20000 train_loss:6.9820 train_time:380ms step_avg:95.09ms +step:5/20000 train_loss:6.7577 train_time:469ms step_avg:93.73ms +step:6/20000 train_loss:6.6206 train_time:557ms step_avg:92.87ms +step:7/20000 train_loss:6.5157 train_time:645ms step_avg:92.14ms +step:8/20000 train_loss:6.5426 train_time:734ms step_avg:91.80ms +step:9/20000 train_loss:6.3052 train_time:823ms step_avg:91.49ms +step:10/20000 train_loss:6.0757 train_time:911ms step_avg:91.14ms +step:100/20000 train_loss:3.1568 train_time:9005ms step_avg:90.05ms +step:200/20000 train_loss:2.3789 train_time:18065ms step_avg:90.33ms +step:300/20000 train_loss:2.5414 train_time:27165ms step_avg:90.55ms +step:400/20000 train_loss:2.4059 train_time:36265ms step_avg:90.66ms +step:500/20000 train_loss:2.3921 train_time:45325ms step_avg:90.65ms +step:500/20000 val_loss:2.3519 val_bpb:1.3930 train_time:45353ms step_avg:90.71ms +step:600/20000 train_loss:2.3265 train_time:54449ms step_avg:90.75ms +step:700/20000 train_loss:2.3483 train_time:63569ms step_avg:90.81ms +step:800/20000 train_loss:2.2306 train_time:72679ms step_avg:90.85ms +step:900/20000 train_loss:2.1226 train_time:81769ms step_avg:90.85ms +step:1000/20000 train_loss:2.2760 train_time:90893ms step_avg:90.89ms +step:1000/20000 val_loss:2.2237 val_bpb:1.3170 train_time:90920ms step_avg:90.92ms +step:1100/20000 train_loss:2.3259 train_time:99993ms step_avg:90.90ms +step:1200/20000 train_loss:2.3524 train_time:109087ms step_avg:90.91ms +step:1300/20000 train_loss:2.0951 train_time:118189ms step_avg:90.91ms +step:1400/20000 train_loss:2.1785 train_time:127299ms step_avg:90.93ms +step:1500/20000 train_loss:2.2207 train_time:136324ms step_avg:90.88ms +step:1500/20000 val_loss:2.1820 val_bpb:1.2923 train_time:136351ms step_avg:90.90ms +step:1600/20000 train_loss:2.0771 train_time:145393ms step_avg:90.87ms +step:1700/20000 train_loss:2.1407 train_time:154472ms step_avg:90.87ms +step:1800/20000 train_loss:2.1593 train_time:163544ms step_avg:90.86ms +step:1900/20000 train_loss:2.1279 train_time:172551ms step_avg:90.82ms +step:2000/20000 train_loss:2.0670 train_time:181597ms step_avg:90.80ms +step:2000/20000 val_loss:2.1341 val_bpb:1.2639 train_time:181623ms step_avg:90.81ms +step:2100/20000 train_loss:2.0441 train_time:190667ms step_avg:90.79ms +step:2200/20000 train_loss:2.1448 train_time:199724ms step_avg:90.78ms +step:2300/20000 train_loss:2.1093 train_time:208782ms step_avg:90.77ms +step:2400/20000 train_loss:2.0658 train_time:217785ms step_avg:90.74ms +step:2500/20000 train_loss:2.1692 train_time:226848ms step_avg:90.74ms +step:2500/20000 val_loss:2.1067 val_bpb:1.2477 train_time:226875ms step_avg:90.75ms +step:2600/20000 train_loss:2.1088 train_time:235897ms step_avg:90.73ms +step:2700/20000 train_loss:2.1021 train_time:244944ms step_avg:90.72ms +step:2800/20000 train_loss:2.1538 train_time:254008ms step_avg:90.72ms +step:2900/20000 train_loss:2.0265 train_time:263003ms step_avg:90.69ms +step:3000/20000 train_loss:2.1583 train_time:272055ms step_avg:90.68ms +step:3000/20000 val_loss:2.0910 val_bpb:1.2384 train_time:272082ms step_avg:90.69ms +step:3100/20000 train_loss:2.0363 train_time:281105ms step_avg:90.68ms +step:3200/20000 train_loss:2.1734 train_time:290145ms step_avg:90.67ms +step:3300/20000 train_loss:2.0709 train_time:299138ms step_avg:90.65ms +step:3400/20000 train_loss:2.0205 train_time:308191ms step_avg:90.64ms +step:3500/20000 train_loss:2.1840 train_time:317240ms step_avg:90.64ms +step:3500/20000 val_loss:2.0833 val_bpb:1.2338 train_time:317269ms step_avg:90.65ms +step:3600/20000 train_loss:2.0994 train_time:326280ms step_avg:90.63ms +step:3700/20000 train_loss:2.0956 train_time:335334ms step_avg:90.63ms +step:3800/20000 train_loss:2.0775 train_time:344331ms step_avg:90.61ms +step:3900/20000 train_loss:2.0770 train_time:353392ms step_avg:90.61ms +step:4000/20000 train_loss:1.9768 train_time:362446ms step_avg:90.61ms +step:4000/20000 val_loss:2.0676 val_bpb:1.2245 train_time:362473ms step_avg:90.62ms +step:4100/20000 train_loss:2.0181 train_time:371489ms step_avg:90.61ms +step:4200/20000 train_loss:2.1524 train_time:380536ms step_avg:90.60ms +step:4300/20000 train_loss:2.0571 train_time:389528ms step_avg:90.59ms +step:4400/20000 train_loss:2.0336 train_time:398576ms step_avg:90.59ms +step:4500/20000 train_loss:2.1207 train_time:407619ms step_avg:90.58ms +step:4500/20000 val_loss:2.0445 val_bpb:1.2109 train_time:407646ms step_avg:90.59ms +step:4600/20000 train_loss:1.8551 train_time:417225ms step_avg:90.70ms +step:4700/20000 train_loss:2.2499 train_time:426891ms step_avg:90.83ms +step:4800/20000 train_loss:2.4371 train_time:436632ms step_avg:90.96ms +step:4900/20000 train_loss:2.0623 train_time:446376ms step_avg:91.10ms +step:5000/20000 train_loss:2.1136 train_time:456111ms step_avg:91.22ms +step:5000/20000 val_loss:2.0325 val_bpb:1.2037 train_time:456133ms step_avg:91.23ms +step:5100/20000 train_loss:2.1377 train_time:465861ms step_avg:91.35ms +step:5200/20000 train_loss:2.0502 train_time:475543ms step_avg:91.45ms +step:5300/20000 train_loss:2.0127 train_time:485274ms step_avg:91.56ms +step:5400/20000 train_loss:2.0558 train_time:495032ms step_avg:91.67ms +step:5500/20000 train_loss:2.0245 train_time:504789ms step_avg:91.78ms +step:5500/20000 val_loss:2.0074 val_bpb:1.1889 train_time:504810ms step_avg:91.78ms +step:5600/20000 train_loss:1.9653 train_time:514532ms step_avg:91.88ms +step:5700/20000 train_loss:2.0166 train_time:524209ms step_avg:91.97ms +step:5800/20000 train_loss:2.0060 train_time:533977ms step_avg:92.07ms +step:5900/20000 train_loss:1.9096 train_time:543721ms step_avg:92.16ms +step:6000/20000 train_loss:1.9437 train_time:553479ms step_avg:92.25ms +step:6000/20000 val_loss:1.9830 val_bpb:1.1744 train_time:553499ms step_avg:92.25ms +step:6100/20000 train_loss:1.9181 train_time:563174ms step_avg:92.32ms +step:6200/20000 train_loss:1.9532 train_time:572938ms step_avg:92.41ms +step:6300/20000 train_loss:1.9483 train_time:582682ms step_avg:92.49ms +step:6400/20000 train_loss:2.0016 train_time:592449ms step_avg:92.57ms +step:6478/20000 val_loss:1.9616 val_bpb:1.1618 train_time:600030ms step_avg:92.63ms +stopping_early: wallclock_cap train_time:600030ms step:6478/20000 +peak memory allocated: 18958 MiB reserved: 19104 MiB +Serialized model: 98437419 bytes +Code size: 54942 bytes +Total submission size: 98492361 bytes +magnitude_pruning: 5.0% of weights with >65536 elements +Serialized model int6+zstd: 15983180 bytes +Total submission size int8+zlib: 16038122 bytes +WARNING: artifact 16038122 exceeds 16,000,000 byte limit by 38122 bytes! +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.201744 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.133229 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.134908 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.128424 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.140327 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.141421 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.143029 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.138434 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.135849 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.137481 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.146279 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.144553 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.145808 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.143937 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.142338 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.142696 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.144089 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.144618 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.150736 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.148247 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.149204 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.147953 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.147319 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.147017 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.147694 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.145291 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.144259 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.144629 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.143405 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.143312 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.142583 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.143729 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.144760 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.145268 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.144799 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.145192 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.144332 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.140481 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.140560 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.141525 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.141716 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.141550 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.140263 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.139979 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.139314 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.139406 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.139368 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.139529 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.139284 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.139869 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.140154 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.139871 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.140921 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.142828 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.142116 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.142831 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.143198 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.143177 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.142755 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.142961 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.142365 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.145155 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.145160 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.145204 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.144850 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.144364 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.143616 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.143608 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.144253 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.144284 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.144264 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.144712 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.144453 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.144071 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.144391 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.144474 +final_int8_zlib_roundtrip val_loss:1.9230 val_bpb:1.1389 eval_time:170559ms +final_int8_zlib_roundtrip_exact val_loss:1.92304451 val_bpb:1.13893829 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed2024_v3.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed2024_v3.log new file mode 100644 index 0000000000..313fb2a8ed --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed2024_v3.log @@ -0,0 +1,215 @@ +========================================== +Experiment: no_swa_qat (seed=2024) +ENABLE_QAT=1 SWA_ENABLED=0 +========================================== +GPUs detected: 8 +W0327 21:30:19.810000 34157 torch/distributed/run.py:803] +W0327 21:30:19.810000 34157 torch/distributed/run.py:803] ***************************************** +W0327 21:30:19.810000 34157 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. +W0327 21:30:19.810000 34157 torch/distributed/run.py:803] ***************************************** +logs/no_swa_qat_seed2024.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +seed:2024 +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:6.9327 val_bpb:4.1059 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9341 train_time:145ms step_avg:144.56ms +step:2/20000 train_loss:8.1356 train_time:207ms step_avg:103.44ms +step:3/20000 train_loss:7.6976 train_time:295ms step_avg:98.37ms +step:4/20000 train_loss:6.9820 train_time:385ms step_avg:96.26ms +step:5/20000 train_loss:6.7578 train_time:474ms step_avg:94.75ms +step:6/20000 train_loss:6.6204 train_time:562ms step_avg:93.65ms +step:7/20000 train_loss:6.5159 train_time:650ms step_avg:92.92ms +step:8/20000 train_loss:6.5426 train_time:739ms step_avg:92.32ms +step:9/20000 train_loss:6.3042 train_time:827ms step_avg:91.92ms +step:10/20000 train_loss:6.0756 train_time:915ms step_avg:91.52ms +step:100/20000 train_loss:3.1598 train_time:8994ms step_avg:89.94ms +step:200/20000 train_loss:2.3875 train_time:18047ms step_avg:90.24ms +step:300/20000 train_loss:2.5358 train_time:27120ms step_avg:90.40ms +step:400/20000 train_loss:2.4049 train_time:36208ms step_avg:90.52ms +step:500/20000 train_loss:2.3886 train_time:45249ms step_avg:90.50ms +step:500/20000 val_loss:2.3488 val_bpb:1.3911 train_time:45276ms step_avg:90.55ms +step:600/20000 train_loss:2.3246 train_time:54364ms step_avg:90.61ms +step:700/20000 train_loss:2.3401 train_time:63479ms step_avg:90.68ms +step:800/20000 train_loss:2.2268 train_time:72595ms step_avg:90.74ms +step:900/20000 train_loss:2.1247 train_time:81700ms step_avg:90.78ms +step:1000/20000 train_loss:2.2714 train_time:90754ms step_avg:90.75ms +step:1000/20000 val_loss:2.2239 val_bpb:1.3171 train_time:90781ms step_avg:90.78ms +step:1100/20000 train_loss:2.3227 train_time:99864ms step_avg:90.79ms +step:1200/20000 train_loss:2.3487 train_time:108967ms step_avg:90.81ms +step:1300/20000 train_loss:2.1000 train_time:118072ms step_avg:90.82ms +step:1400/20000 train_loss:2.1827 train_time:127173ms step_avg:90.84ms +step:1500/20000 train_loss:2.2219 train_time:136206ms step_avg:90.80ms +step:1500/20000 val_loss:2.1834 val_bpb:1.2931 train_time:136234ms step_avg:90.82ms +step:1600/20000 train_loss:2.0764 train_time:145310ms step_avg:90.82ms +step:1700/20000 train_loss:2.1433 train_time:154389ms step_avg:90.82ms +step:1800/20000 train_loss:2.1600 train_time:163465ms step_avg:90.81ms +step:1900/20000 train_loss:2.1278 train_time:172482ms step_avg:90.78ms +step:2000/20000 train_loss:2.0673 train_time:181552ms step_avg:90.78ms +step:2000/20000 val_loss:2.1315 val_bpb:1.2624 train_time:181578ms step_avg:90.79ms +step:2100/20000 train_loss:2.0474 train_time:190611ms step_avg:90.77ms +step:2200/20000 train_loss:2.1525 train_time:199664ms step_avg:90.76ms +step:2300/20000 train_loss:2.1101 train_time:208731ms step_avg:90.75ms +step:2400/20000 train_loss:2.0654 train_time:217733ms step_avg:90.72ms +step:2500/20000 train_loss:2.1719 train_time:226780ms step_avg:90.71ms +step:2500/20000 val_loss:2.1061 val_bpb:1.2473 train_time:226806ms step_avg:90.72ms +step:2600/20000 train_loss:2.1073 train_time:235854ms step_avg:90.71ms +step:2700/20000 train_loss:2.0981 train_time:244890ms step_avg:90.70ms +step:2800/20000 train_loss:2.1533 train_time:253953ms step_avg:90.70ms +step:2900/20000 train_loss:2.0227 train_time:262941ms step_avg:90.67ms +step:3000/20000 train_loss:2.1591 train_time:271993ms step_avg:90.66ms +step:3000/20000 val_loss:2.0909 val_bpb:1.2384 train_time:272020ms step_avg:90.67ms +step:3100/20000 train_loss:2.0348 train_time:281039ms step_avg:90.66ms +step:3200/20000 train_loss:2.1720 train_time:290083ms step_avg:90.65ms +step:3300/20000 train_loss:2.0712 train_time:299086ms step_avg:90.63ms +step:3400/20000 train_loss:2.0238 train_time:308136ms step_avg:90.63ms +step:3500/20000 train_loss:2.1849 train_time:317180ms step_avg:90.62ms +step:3500/20000 val_loss:2.0839 val_bpb:1.2342 train_time:317206ms step_avg:90.63ms +step:3600/20000 train_loss:2.1000 train_time:326225ms step_avg:90.62ms +step:3700/20000 train_loss:2.0973 train_time:335272ms step_avg:90.61ms +step:3800/20000 train_loss:2.0722 train_time:344265ms step_avg:90.60ms +step:3900/20000 train_loss:2.0762 train_time:353317ms step_avg:90.59ms +step:4000/20000 train_loss:1.9804 train_time:362357ms step_avg:90.59ms +step:4000/20000 val_loss:2.0681 val_bpb:1.2249 train_time:362383ms step_avg:90.60ms +step:4100/20000 train_loss:2.0191 train_time:371392ms step_avg:90.58ms +step:4200/20000 train_loss:2.1536 train_time:380438ms step_avg:90.58ms +step:4300/20000 train_loss:2.0599 train_time:389435ms step_avg:90.57ms +step:4400/20000 train_loss:2.0358 train_time:398473ms step_avg:90.56ms +step:4500/20000 train_loss:2.1223 train_time:407515ms step_avg:90.56ms +step:4500/20000 val_loss:2.0449 val_bpb:1.2111 train_time:407543ms step_avg:90.57ms +step:4600/20000 train_loss:1.8566 train_time:417099ms step_avg:90.67ms +step:4700/20000 train_loss:2.2462 train_time:426762ms step_avg:90.80ms +step:4800/20000 train_loss:2.4413 train_time:436488ms step_avg:90.94ms +step:4900/20000 train_loss:2.0589 train_time:446209ms step_avg:91.06ms +step:5000/20000 train_loss:2.1134 train_time:455937ms step_avg:91.19ms +step:5000/20000 val_loss:2.0325 val_bpb:1.2038 train_time:455957ms step_avg:91.19ms +step:5100/20000 train_loss:2.1335 train_time:465690ms step_avg:91.31ms +step:5200/20000 train_loss:2.0489 train_time:475369ms step_avg:91.42ms +step:5300/20000 train_loss:2.0184 train_time:485090ms step_avg:91.53ms +step:5400/20000 train_loss:2.0552 train_time:494815ms step_avg:91.63ms +step:5500/20000 train_loss:2.0265 train_time:504548ms step_avg:91.74ms +step:5500/20000 val_loss:2.0088 val_bpb:1.1897 train_time:504569ms step_avg:91.74ms +step:5600/20000 train_loss:1.9629 train_time:514380ms step_avg:91.85ms +step:5700/20000 train_loss:2.0215 train_time:524052ms step_avg:91.94ms +step:5800/20000 train_loss:2.0031 train_time:533796ms step_avg:92.03ms +step:5900/20000 train_loss:1.9095 train_time:543531ms step_avg:92.12ms +step:6000/20000 train_loss:1.9460 train_time:553260ms step_avg:92.21ms +step:6000/20000 val_loss:1.9829 val_bpb:1.1744 train_time:553280ms step_avg:92.21ms +step:6100/20000 train_loss:1.9196 train_time:562939ms step_avg:92.29ms +step:6200/20000 train_loss:1.9524 train_time:572674ms step_avg:92.37ms +step:6300/20000 train_loss:1.9477 train_time:582407ms step_avg:92.45ms +step:6400/20000 train_loss:2.0032 train_time:592152ms step_avg:92.52ms +step:6482/20000 val_loss:1.9620 val_bpb:1.1620 train_time:600104ms step_avg:92.58ms +stopping_early: wallclock_cap train_time:600104ms step:6482/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +Serialized model: 98437419 bytes +Code size: 54950 bytes +Total submission size: 98492369 bytes +magnitude_pruning: 10.0% of weights with >65536 elements +Serialized model int6+zstd: 15787003 bytes +Total submission size int8+zlib: 15841953 bytes +artifact_headroom: 158047 bytes under limit +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.205114 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.135388 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.137099 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.130797 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.142976 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.144216 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.145329 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.140887 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.138390 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.139832 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.148519 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.146583 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.147937 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.146253 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.144729 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.145050 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.146400 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.146969 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.153114 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.150541 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.151476 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.150108 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.149448 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.149029 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.149754 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.147351 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.146319 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.146675 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.145450 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.145353 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.144567 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.145716 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.146750 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.147247 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.146737 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.147160 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.146328 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.142452 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.142558 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.143516 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.143643 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.143496 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.142231 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.141950 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.141261 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.141315 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.141285 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.141458 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.141179 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.141789 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.142081 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.141775 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.142823 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.144755 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.144065 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.144804 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.145159 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.145151 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.144751 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.144986 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.144407 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.147180 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.147165 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.147203 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.146834 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.146341 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.145599 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.145591 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.146236 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.146274 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.146258 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.146691 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.146440 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.146065 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.146396 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.146465 +final_int8_zlib_roundtrip val_loss:1.9261 val_bpb:1.1407 eval_time:171330ms +final_int8_zlib_roundtrip_exact val_loss:1.92608597 val_bpb:1.14073962 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed42.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed42.log new file mode 100644 index 0000000000..d6011be80d --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_qat_seed42.log @@ -0,0 +1,210 @@ +W0321 13:39:54.884000 40210 torch/distributed/run.py:803] +W0321 13:39:54.884000 40210 torch/distributed/run.py:803] ***************************************** +W0321 13:39:54.884000 40210 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. +W0321 13:39:54.884000 40210 torch/distributed/run.py:803] ***************************************** +logs/no_swa_qat_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:133ms step_avg:133.01ms +step:2/20000 train_loss:8.1444 train_time:195ms step_avg:97.69ms +step:3/20000 train_loss:7.6923 train_time:283ms step_avg:94.38ms +step:4/20000 train_loss:6.9899 train_time:371ms step_avg:92.85ms +step:5/20000 train_loss:6.8345 train_time:460ms step_avg:92.09ms +step:6/20000 train_loss:6.6323 train_time:548ms step_avg:91.40ms +step:7/20000 train_loss:6.5368 train_time:638ms step_avg:91.11ms +step:8/20000 train_loss:6.5773 train_time:726ms step_avg:90.75ms +step:9/20000 train_loss:6.3278 train_time:814ms step_avg:90.42ms +step:10/20000 train_loss:6.0561 train_time:901ms step_avg:90.13ms +step:100/20000 train_loss:3.1709 train_time:8969ms step_avg:89.69ms +step:200/20000 train_loss:2.3853 train_time:17989ms step_avg:89.95ms +step:300/20000 train_loss:2.5382 train_time:27025ms step_avg:90.08ms +step:400/20000 train_loss:2.4047 train_time:36061ms step_avg:90.15ms +step:500/20000 train_loss:2.3883 train_time:45080ms step_avg:90.16ms +step:500/20000 val_loss:2.3517 val_bpb:1.3928 train_time:45107ms step_avg:90.21ms +step:600/20000 train_loss:2.3279 train_time:54156ms step_avg:90.26ms +step:700/20000 train_loss:2.3403 train_time:63225ms step_avg:90.32ms +step:800/20000 train_loss:2.2361 train_time:72294ms step_avg:90.37ms +step:900/20000 train_loss:2.1274 train_time:81356ms step_avg:90.40ms +step:1000/20000 train_loss:2.2733 train_time:90387ms step_avg:90.39ms +step:1000/20000 val_loss:2.2252 val_bpb:1.3179 train_time:90414ms step_avg:90.41ms +step:1100/20000 train_loss:2.3200 train_time:99467ms step_avg:90.42ms +step:1200/20000 train_loss:2.3560 train_time:108541ms step_avg:90.45ms +step:1300/20000 train_loss:2.1075 train_time:117609ms step_avg:90.47ms +step:1400/20000 train_loss:2.1833 train_time:126680ms step_avg:90.49ms +step:1500/20000 train_loss:2.2230 train_time:135703ms step_avg:90.47ms +step:1500/20000 val_loss:2.1854 val_bpb:1.2943 train_time:135731ms step_avg:90.49ms +step:1600/20000 train_loss:2.0758 train_time:144775ms step_avg:90.48ms +step:1700/20000 train_loss:2.1434 train_time:153843ms step_avg:90.50ms +step:1800/20000 train_loss:2.1632 train_time:162918ms step_avg:90.51ms +step:1900/20000 train_loss:2.1286 train_time:171949ms step_avg:90.50ms +step:2000/20000 train_loss:2.0680 train_time:181015ms step_avg:90.51ms +step:2000/20000 val_loss:2.1321 val_bpb:1.2627 train_time:181042ms step_avg:90.52ms +step:2100/20000 train_loss:2.0516 train_time:190097ms step_avg:90.52ms +step:2200/20000 train_loss:2.1474 train_time:199172ms step_avg:90.53ms +step:2300/20000 train_loss:2.1132 train_time:208236ms step_avg:90.54ms +step:2400/20000 train_loss:2.0646 train_time:217255ms step_avg:90.52ms +step:2500/20000 train_loss:2.1703 train_time:226322ms step_avg:90.53ms +step:2500/20000 val_loss:2.1072 val_bpb:1.2480 train_time:226350ms step_avg:90.54ms +step:2600/20000 train_loss:2.1081 train_time:235396ms step_avg:90.54ms +step:2700/20000 train_loss:2.1040 train_time:244471ms step_avg:90.54ms +step:2800/20000 train_loss:2.1558 train_time:253538ms step_avg:90.55ms +step:2900/20000 train_loss:2.0261 train_time:262563ms step_avg:90.54ms +step:3000/20000 train_loss:2.1609 train_time:271634ms step_avg:90.54ms +step:3000/20000 val_loss:2.0934 val_bpb:1.2398 train_time:271661ms step_avg:90.55ms +step:3100/20000 train_loss:2.0350 train_time:280716ms step_avg:90.55ms +step:3200/20000 train_loss:2.1724 train_time:289780ms step_avg:90.56ms +step:3300/20000 train_loss:2.0731 train_time:298804ms step_avg:90.55ms +step:3400/20000 train_loss:2.0256 train_time:307864ms step_avg:90.55ms +step:3500/20000 train_loss:2.1852 train_time:316933ms step_avg:90.55ms +step:3500/20000 val_loss:2.0848 val_bpb:1.2347 train_time:316960ms step_avg:90.56ms +step:3600/20000 train_loss:2.1010 train_time:325999ms step_avg:90.56ms +step:3700/20000 train_loss:2.0990 train_time:335064ms step_avg:90.56ms +step:3800/20000 train_loss:2.0791 train_time:344090ms step_avg:90.55ms +step:3900/20000 train_loss:2.0846 train_time:353157ms step_avg:90.55ms +step:4000/20000 train_loss:1.9802 train_time:362223ms step_avg:90.56ms +step:4000/20000 val_loss:2.0691 val_bpb:1.2254 train_time:362250ms step_avg:90.56ms +step:4100/20000 train_loss:2.0179 train_time:371295ms step_avg:90.56ms +step:4200/20000 train_loss:2.1516 train_time:380367ms step_avg:90.56ms +step:4300/20000 train_loss:2.0571 train_time:389393ms step_avg:90.56ms +step:4400/20000 train_loss:2.0354 train_time:398454ms step_avg:90.56ms +step:4500/20000 train_loss:2.1229 train_time:407523ms step_avg:90.56ms +step:4500/20000 val_loss:2.0462 val_bpb:1.2119 train_time:407549ms step_avg:90.57ms +step:4600/20000 train_loss:1.8540 train_time:417026ms step_avg:90.66ms +step:4700/20000 train_loss:2.2482 train_time:426605ms step_avg:90.77ms +step:4800/20000 train_loss:2.4444 train_time:436231ms step_avg:90.88ms +step:4900/20000 train_loss:2.0630 train_time:445855ms step_avg:90.99ms +step:5000/20000 train_loss:2.1143 train_time:455478ms step_avg:91.10ms +step:5000/20000 val_loss:2.0339 val_bpb:1.2046 train_time:455499ms step_avg:91.10ms +step:5100/20000 train_loss:2.1348 train_time:465116ms step_avg:91.20ms +step:5200/20000 train_loss:2.0521 train_time:474699ms step_avg:91.29ms +step:5300/20000 train_loss:2.0180 train_time:484327ms step_avg:91.38ms +step:5400/20000 train_loss:2.0598 train_time:493963ms step_avg:91.47ms +step:5500/20000 train_loss:2.0274 train_time:503603ms step_avg:91.56ms +step:5500/20000 val_loss:2.0090 val_bpb:1.1899 train_time:503624ms step_avg:91.57ms +step:5600/20000 train_loss:1.9595 train_time:513242ms step_avg:91.65ms +step:5700/20000 train_loss:2.0196 train_time:522817ms step_avg:91.72ms +step:5800/20000 train_loss:2.0053 train_time:532442ms step_avg:91.80ms +step:5900/20000 train_loss:1.9092 train_time:542097ms step_avg:91.88ms +step:6000/20000 train_loss:1.9488 train_time:551790ms step_avg:91.97ms +step:6000/20000 val_loss:1.9846 val_bpb:1.1754 train_time:551810ms step_avg:91.97ms +step:6100/20000 train_loss:1.9229 train_time:561370ms step_avg:92.03ms +step:6200/20000 train_loss:1.9552 train_time:571007ms step_avg:92.10ms +step:6300/20000 train_loss:1.9502 train_time:580630ms step_avg:92.16ms +step:6400/20000 train_loss:2.0035 train_time:590260ms step_avg:92.23ms +step:6500/20000 train_loss:2.0889 train_time:599888ms step_avg:92.29ms +step:6500/20000 val_loss:1.9625 val_bpb:1.1623 train_time:599908ms step_avg:92.29ms +step:6502/20000 val_loss:1.9625 val_bpb:1.1623 train_time:600100ms step_avg:92.29ms +stopping_early: wallclock_cap train_time:600100ms step:6502/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 16393156 bytes +Total submission size int8+zlib: 16447604 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.206283 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.135907 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.136591 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.129658 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.141151 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.142378 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.144109 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.139585 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.136884 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.138391 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.147117 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.145611 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.146881 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.145199 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.143626 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.144003 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.145315 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.145847 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.152028 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.149442 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.150336 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.148930 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.148287 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.147887 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.148527 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.146133 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.145167 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.145547 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.144320 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.144190 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.143482 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.144732 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.145789 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.146258 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.145700 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.146105 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.145208 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.141312 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.141387 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.142336 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.142506 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.142356 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.141132 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.140850 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.140186 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.140266 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.140238 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.140382 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.140119 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.140727 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.141034 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.140765 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.141793 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.143709 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.142966 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.143698 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.144053 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.144032 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.143610 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.143826 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.143238 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.146037 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.146030 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.146065 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.145701 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.145216 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.144467 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.144439 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.145064 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.145094 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.145079 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.145541 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.145286 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.144900 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.145218 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.145311 +final_int8_zlib_roundtrip val_loss:1.9243 val_bpb:1.1397 eval_time:170272ms +final_int8_zlib_roundtrip_exact val_loss:1.92431130 val_bpb:1.13968856 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_seed1337.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_seed1337.log new file mode 100644 index 0000000000..bf4e1c980f --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_seed1337.log @@ -0,0 +1,211 @@ +W0321 13:25:32.981000 39318 torch/distributed/run.py:803] +W0321 13:25:32.981000 39318 torch/distributed/run.py:803] ***************************************** +W0321 13:25:32.981000 39318 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. +W0321 13:25:32.981000 39318 torch/distributed/run.py:803] ***************************************** +logs/no_swa_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:133ms step_avg:133.49ms +step:2/20000 train_loss:7.9520 train_time:196ms step_avg:97.76ms +step:3/20000 train_loss:7.5214 train_time:285ms step_avg:94.95ms +step:4/20000 train_loss:6.9071 train_time:372ms step_avg:93.09ms +step:5/20000 train_loss:6.7673 train_time:460ms step_avg:92.06ms +step:6/20000 train_loss:6.7191 train_time:550ms step_avg:91.66ms +step:7/20000 train_loss:6.5968 train_time:638ms step_avg:91.12ms +step:8/20000 train_loss:6.4794 train_time:726ms step_avg:90.71ms +step:9/20000 train_loss:6.2073 train_time:814ms step_avg:90.44ms +step:10/20000 train_loss:5.9979 train_time:903ms step_avg:90.29ms +step:100/20000 train_loss:3.1588 train_time:8968ms step_avg:89.68ms +step:200/20000 train_loss:2.3729 train_time:17999ms step_avg:89.99ms +step:300/20000 train_loss:2.5396 train_time:27044ms step_avg:90.15ms +step:400/20000 train_loss:2.4056 train_time:36085ms step_avg:90.21ms +step:500/20000 train_loss:2.3899 train_time:45097ms step_avg:90.19ms +step:500/20000 val_loss:2.3511 val_bpb:1.3924 train_time:45124ms step_avg:90.25ms +step:600/20000 train_loss:2.3331 train_time:54163ms step_avg:90.27ms +step:700/20000 train_loss:2.3443 train_time:63242ms step_avg:90.35ms +step:800/20000 train_loss:2.2374 train_time:72315ms step_avg:90.39ms +step:900/20000 train_loss:2.1257 train_time:81386ms step_avg:90.43ms +step:1000/20000 train_loss:2.2763 train_time:90422ms step_avg:90.42ms +step:1000/20000 val_loss:2.2287 val_bpb:1.3200 train_time:90449ms step_avg:90.45ms +step:1100/20000 train_loss:2.3257 train_time:99497ms step_avg:90.45ms +step:1200/20000 train_loss:2.3534 train_time:108569ms step_avg:90.47ms +step:1300/20000 train_loss:2.1039 train_time:117652ms step_avg:90.50ms +step:1400/20000 train_loss:2.1843 train_time:126725ms step_avg:90.52ms +step:1500/20000 train_loss:2.2245 train_time:135752ms step_avg:90.50ms +step:1500/20000 val_loss:2.1867 val_bpb:1.2951 train_time:135780ms step_avg:90.52ms +step:1600/20000 train_loss:2.0760 train_time:144831ms step_avg:90.52ms +step:1700/20000 train_loss:2.1430 train_time:153904ms step_avg:90.53ms +step:1800/20000 train_loss:2.1647 train_time:162983ms step_avg:90.55ms +step:1900/20000 train_loss:2.1351 train_time:172015ms step_avg:90.53ms +step:2000/20000 train_loss:2.0711 train_time:181081ms step_avg:90.54ms +step:2000/20000 val_loss:2.1351 val_bpb:1.2645 train_time:181107ms step_avg:90.55ms +step:2100/20000 train_loss:2.0523 train_time:190153ms step_avg:90.55ms +step:2200/20000 train_loss:2.1625 train_time:199232ms step_avg:90.56ms +step:2300/20000 train_loss:2.1125 train_time:208300ms step_avg:90.57ms +step:2400/20000 train_loss:2.0678 train_time:217327ms step_avg:90.55ms +step:2500/20000 train_loss:2.1711 train_time:226449ms step_avg:90.58ms +step:2500/20000 val_loss:2.1090 val_bpb:1.2491 train_time:226477ms step_avg:90.59ms +step:2600/20000 train_loss:2.1122 train_time:235525ms step_avg:90.59ms +step:2700/20000 train_loss:2.1076 train_time:244599ms step_avg:90.59ms +step:2800/20000 train_loss:2.1582 train_time:253677ms step_avg:90.60ms +step:2900/20000 train_loss:2.0270 train_time:262707ms step_avg:90.59ms +step:3000/20000 train_loss:2.1652 train_time:271779ms step_avg:90.59ms +step:3000/20000 val_loss:2.0947 val_bpb:1.2406 train_time:271807ms step_avg:90.60ms +step:3100/20000 train_loss:2.0422 train_time:280861ms step_avg:90.60ms +step:3200/20000 train_loss:2.1762 train_time:289938ms step_avg:90.61ms +step:3300/20000 train_loss:2.0756 train_time:298966ms step_avg:90.60ms +step:3400/20000 train_loss:2.0245 train_time:308032ms step_avg:90.60ms +step:3500/20000 train_loss:2.1869 train_time:317109ms step_avg:90.60ms +step:3500/20000 val_loss:2.0866 val_bpb:1.2358 train_time:317137ms step_avg:90.61ms +step:3600/20000 train_loss:2.1030 train_time:326185ms step_avg:90.61ms +step:3700/20000 train_loss:2.0983 train_time:335255ms step_avg:90.61ms +step:3800/20000 train_loss:2.0795 train_time:344279ms step_avg:90.60ms +step:3900/20000 train_loss:2.0787 train_time:353351ms step_avg:90.60ms +step:4000/20000 train_loss:1.9777 train_time:362411ms step_avg:90.60ms +step:4000/20000 val_loss:2.0705 val_bpb:1.2263 train_time:362438ms step_avg:90.61ms +step:4100/20000 train_loss:2.0156 train_time:371480ms step_avg:90.60ms +step:4200/20000 train_loss:2.1560 train_time:380547ms step_avg:90.61ms +step:4300/20000 train_loss:2.0616 train_time:389574ms step_avg:90.60ms +step:4400/20000 train_loss:2.0360 train_time:398656ms step_avg:90.60ms +step:4500/20000 train_loss:2.1284 train_time:407727ms step_avg:90.61ms +step:4500/20000 val_loss:2.0471 val_bpb:1.2124 train_time:407754ms step_avg:90.61ms +step:4600/20000 train_loss:1.8495 train_time:416800ms step_avg:90.61ms +step:4700/20000 train_loss:2.2364 train_time:425823ms step_avg:90.60ms +step:4800/20000 train_loss:2.4339 train_time:434895ms step_avg:90.60ms +step:4900/20000 train_loss:2.0486 train_time:443970ms step_avg:90.61ms +step:5000/20000 train_loss:2.1054 train_time:453048ms step_avg:90.61ms +step:5000/20000 val_loss:2.0254 val_bpb:1.1996 train_time:453077ms step_avg:90.62ms +step:5100/20000 train_loss:2.1278 train_time:462125ms step_avg:90.61ms +step:5200/20000 train_loss:2.0443 train_time:471145ms step_avg:90.60ms +step:5300/20000 train_loss:2.0105 train_time:480223ms step_avg:90.61ms +step:5400/20000 train_loss:2.0501 train_time:489300ms step_avg:90.61ms +step:5500/20000 train_loss:2.0159 train_time:498370ms step_avg:90.61ms +step:5500/20000 val_loss:2.0021 val_bpb:1.1858 train_time:498398ms step_avg:90.62ms +step:5600/20000 train_loss:1.9537 train_time:507445ms step_avg:90.62ms +step:5700/20000 train_loss:2.0109 train_time:516483ms step_avg:90.61ms +step:5800/20000 train_loss:1.9985 train_time:525550ms step_avg:90.61ms +step:5900/20000 train_loss:1.9004 train_time:534620ms step_avg:90.61ms +step:6000/20000 train_loss:1.9401 train_time:543696ms step_avg:90.62ms +step:6000/20000 val_loss:1.9781 val_bpb:1.1715 train_time:543723ms step_avg:90.62ms +step:6100/20000 train_loss:1.9140 train_time:552726ms step_avg:90.61ms +step:6200/20000 train_loss:1.9456 train_time:561796ms step_avg:90.61ms +step:6300/20000 train_loss:1.9436 train_time:570870ms step_avg:90.61ms +step:6400/20000 train_loss:1.9969 train_time:579955ms step_avg:90.62ms +step:6500/20000 train_loss:2.0813 train_time:589032ms step_avg:90.62ms +step:6500/20000 val_loss:1.9519 val_bpb:1.1560 train_time:589058ms step_avg:90.62ms +step:6600/20000 train_loss:1.8445 train_time:598048ms step_avg:90.61ms +step:6622/20000 val_loss:1.9488 val_bpb:1.1542 train_time:600114ms step_avg:90.62ms +stopping_early: wallclock_cap train_time:600114ms step:6622/20000 +peak memory allocated: 18866 MiB reserved: 19074 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15822165 bytes +Total submission size int8+zlib: 15876613 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.203304 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.138641 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.140397 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133944 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.146099 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147628 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.149099 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144731 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.142173 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.143833 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.152537 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150620 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.152017 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.150250 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.148829 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.149001 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.150497 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.151096 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.157300 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.154768 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.155795 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.154491 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.153844 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.153414 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.154179 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151857 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150873 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.151216 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.150040 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.149853 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.149120 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.150347 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.151387 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.151883 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.151357 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151739 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.150859 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.146918 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.147017 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.147949 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.148103 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.147957 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.146732 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.146452 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.145744 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.145784 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.145756 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145924 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145653 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.146300 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.146592 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.146292 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.147334 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.149253 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.148533 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.149257 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.149590 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.149565 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.149153 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.149354 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.148746 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.151527 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.151531 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.151574 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.151217 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.150730 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.150000 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149989 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.150615 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.150666 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.150642 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.151069 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.150820 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.150432 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.150715 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.150817 +final_int8_zlib_roundtrip val_loss:1.9332 val_bpb:1.1450 eval_time:170113ms +final_int8_zlib_roundtrip_exact val_loss:1.93323306 val_bpb:1.14497254 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_seed42.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_seed42.log new file mode 100644 index 0000000000..a277d49bf4 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/no_swa_seed42.log @@ -0,0 +1,211 @@ +W0321 13:11:11.416000 38463 torch/distributed/run.py:803] +W0321 13:11:11.416000 38463 torch/distributed/run.py:803] ***************************************** +W0321 13:11:11.416000 38463 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. +W0321 13:11:11.416000 38463 torch/distributed/run.py:803] ***************************************** +logs/no_swa_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:132ms step_avg:132.17ms +step:2/20000 train_loss:8.1444 train_time:196ms step_avg:97.96ms +step:3/20000 train_loss:7.6923 train_time:284ms step_avg:94.64ms +step:4/20000 train_loss:6.9899 train_time:372ms step_avg:93.10ms +step:5/20000 train_loss:6.8345 train_time:460ms step_avg:92.01ms +step:6/20000 train_loss:6.6323 train_time:548ms step_avg:91.27ms +step:7/20000 train_loss:6.5364 train_time:637ms step_avg:90.94ms +step:8/20000 train_loss:6.5772 train_time:724ms step_avg:90.49ms +step:9/20000 train_loss:6.3297 train_time:813ms step_avg:90.30ms +step:10/20000 train_loss:6.0565 train_time:901ms step_avg:90.13ms +step:100/20000 train_loss:3.1658 train_time:8958ms step_avg:89.58ms +step:200/20000 train_loss:2.3797 train_time:17972ms step_avg:89.86ms +step:300/20000 train_loss:2.5399 train_time:27010ms step_avg:90.03ms +step:400/20000 train_loss:2.4091 train_time:36042ms step_avg:90.11ms +step:500/20000 train_loss:2.3978 train_time:45051ms step_avg:90.10ms +step:500/20000 val_loss:2.3547 val_bpb:1.3946 train_time:45079ms step_avg:90.16ms +step:600/20000 train_loss:2.3326 train_time:54116ms step_avg:90.19ms +step:700/20000 train_loss:2.3399 train_time:63177ms step_avg:90.25ms +step:800/20000 train_loss:2.2367 train_time:72241ms step_avg:90.30ms +step:900/20000 train_loss:2.1285 train_time:81297ms step_avg:90.33ms +step:1000/20000 train_loss:2.2713 train_time:90327ms step_avg:90.33ms +step:1000/20000 val_loss:2.2263 val_bpb:1.3185 train_time:90354ms step_avg:90.35ms +step:1100/20000 train_loss:2.3167 train_time:99409ms step_avg:90.37ms +step:1200/20000 train_loss:2.3542 train_time:108467ms step_avg:90.39ms +step:1300/20000 train_loss:2.1079 train_time:117534ms step_avg:90.41ms +step:1400/20000 train_loss:2.1851 train_time:126605ms step_avg:90.43ms +step:1500/20000 train_loss:2.2227 train_time:135625ms step_avg:90.42ms +step:1500/20000 val_loss:2.1850 val_bpb:1.2941 train_time:135651ms step_avg:90.43ms +step:1600/20000 train_loss:2.0774 train_time:144691ms step_avg:90.43ms +step:1700/20000 train_loss:2.1434 train_time:153766ms step_avg:90.45ms +step:1800/20000 train_loss:2.1582 train_time:162832ms step_avg:90.46ms +step:1900/20000 train_loss:2.1298 train_time:171854ms step_avg:90.45ms +step:2000/20000 train_loss:2.0703 train_time:180919ms step_avg:90.46ms +step:2000/20000 val_loss:2.1333 val_bpb:1.2635 train_time:180947ms step_avg:90.47ms +step:2100/20000 train_loss:2.0475 train_time:189972ms step_avg:90.46ms +step:2200/20000 train_loss:2.1359 train_time:199033ms step_avg:90.47ms +step:2300/20000 train_loss:2.1103 train_time:208102ms step_avg:90.48ms +step:2400/20000 train_loss:2.0634 train_time:217118ms step_avg:90.47ms +step:2500/20000 train_loss:2.1689 train_time:226169ms step_avg:90.47ms +step:2500/20000 val_loss:2.1070 val_bpb:1.2479 train_time:226198ms step_avg:90.48ms +step:2600/20000 train_loss:2.1090 train_time:235230ms step_avg:90.47ms +step:2700/20000 train_loss:2.1030 train_time:244291ms step_avg:90.48ms +step:2800/20000 train_loss:2.1550 train_time:253361ms step_avg:90.49ms +step:2900/20000 train_loss:2.0264 train_time:262372ms step_avg:90.47ms +step:3000/20000 train_loss:2.1590 train_time:271437ms step_avg:90.48ms +step:3000/20000 val_loss:2.0930 val_bpb:1.2396 train_time:271463ms step_avg:90.49ms +step:3100/20000 train_loss:2.0377 train_time:280504ms step_avg:90.49ms +step:3200/20000 train_loss:2.1728 train_time:289563ms step_avg:90.49ms +step:3300/20000 train_loss:2.0774 train_time:298579ms step_avg:90.48ms +step:3400/20000 train_loss:2.0220 train_time:307648ms step_avg:90.48ms +step:3500/20000 train_loss:2.1821 train_time:316712ms step_avg:90.49ms +step:3500/20000 val_loss:2.0850 val_bpb:1.2348 train_time:316741ms step_avg:90.50ms +step:3600/20000 train_loss:2.1017 train_time:325784ms step_avg:90.50ms +step:3700/20000 train_loss:2.0988 train_time:334847ms step_avg:90.50ms +step:3800/20000 train_loss:2.0763 train_time:343855ms step_avg:90.49ms +step:3900/20000 train_loss:2.0801 train_time:352923ms step_avg:90.49ms +step:4000/20000 train_loss:1.9802 train_time:361985ms step_avg:90.50ms +step:4000/20000 val_loss:2.0702 val_bpb:1.2261 train_time:362012ms step_avg:90.50ms +step:4100/20000 train_loss:2.0188 train_time:371048ms step_avg:90.50ms +step:4200/20000 train_loss:2.1545 train_time:380114ms step_avg:90.50ms +step:4300/20000 train_loss:2.0620 train_time:389136ms step_avg:90.50ms +step:4400/20000 train_loss:2.0341 train_time:398208ms step_avg:90.50ms +step:4500/20000 train_loss:2.1217 train_time:407264ms step_avg:90.50ms +step:4500/20000 val_loss:2.0464 val_bpb:1.2120 train_time:407290ms step_avg:90.51ms +step:4600/20000 train_loss:1.8442 train_time:416324ms step_avg:90.51ms +step:4700/20000 train_loss:2.2348 train_time:425344ms step_avg:90.50ms +step:4800/20000 train_loss:2.4339 train_time:434402ms step_avg:90.50ms +step:4900/20000 train_loss:2.0514 train_time:443468ms step_avg:90.50ms +step:5000/20000 train_loss:2.1046 train_time:452534ms step_avg:90.51ms +step:5000/20000 val_loss:2.0245 val_bpb:1.1990 train_time:452561ms step_avg:90.51ms +step:5100/20000 train_loss:2.1233 train_time:461599ms step_avg:90.51ms +step:5200/20000 train_loss:2.0415 train_time:470676ms step_avg:90.51ms +step:5300/20000 train_loss:2.0101 train_time:479746ms step_avg:90.52ms +step:5400/20000 train_loss:2.0477 train_time:488820ms step_avg:90.52ms +step:5500/20000 train_loss:2.0177 train_time:497887ms step_avg:90.52ms +step:5500/20000 val_loss:2.0015 val_bpb:1.1854 train_time:497914ms step_avg:90.53ms +step:5600/20000 train_loss:1.9552 train_time:506948ms step_avg:90.53ms +step:5700/20000 train_loss:2.0121 train_time:515965ms step_avg:90.52ms +step:5800/20000 train_loss:1.9973 train_time:525029ms step_avg:90.52ms +step:5900/20000 train_loss:1.9013 train_time:534104ms step_avg:90.53ms +step:6000/20000 train_loss:1.9422 train_time:543166ms step_avg:90.53ms +step:6000/20000 val_loss:1.9777 val_bpb:1.1713 train_time:543193ms step_avg:90.53ms +step:6100/20000 train_loss:1.9139 train_time:552186ms step_avg:90.52ms +step:6200/20000 train_loss:1.9460 train_time:561253ms step_avg:90.52ms +step:6300/20000 train_loss:1.9427 train_time:570318ms step_avg:90.53ms +step:6400/20000 train_loss:1.9974 train_time:579382ms step_avg:90.53ms +step:6500/20000 train_loss:2.0819 train_time:588446ms step_avg:90.53ms +step:6500/20000 val_loss:1.9513 val_bpb:1.1556 train_time:588473ms step_avg:90.53ms +step:6600/20000 train_loss:1.8436 train_time:597464ms step_avg:90.52ms +step:6628/20000 val_loss:1.9480 val_bpb:1.1537 train_time:600067ms step_avg:90.54ms +stopping_early: wallclock_cap train_time:600067ms step:6628/20000 +peak memory allocated: 18866 MiB reserved: 19074 MiB +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15814075 bytes +Total submission size int8+zlib: 15868523 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.210598 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.139989 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.141509 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.134515 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.146334 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147660 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.149130 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144544 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.142216 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.143973 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.152774 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150963 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.152287 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.150562 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.149028 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.149300 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.150625 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.151187 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.157323 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.154733 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.155720 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.154314 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.153615 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.153161 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.153855 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151514 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150511 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.150859 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.149576 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.149450 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.148716 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.149910 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.150948 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.151476 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.150953 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151313 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.150423 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.146568 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.146692 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.147611 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.147732 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.147587 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.146373 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.146057 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.145363 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.145418 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.145379 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145535 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145254 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.145859 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.146146 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.145872 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.146919 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.148871 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.148161 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.148916 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.149245 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.149202 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.148790 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.149007 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.148403 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.151212 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.151209 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.151250 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.150882 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.150389 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.149653 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149614 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.150256 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.150286 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.150297 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.150735 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.150494 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.150083 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.150406 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.150485 +final_int8_zlib_roundtrip val_loss:1.9329 val_bpb:1.1448 eval_time:170013ms +final_int8_zlib_roundtrip_exact val_loss:1.93286440 val_bpb:1.14475420 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/qat_snap70_seed1337.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/qat_snap70_seed1337.log new file mode 100644 index 0000000000..6c2a1c486b --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/qat_snap70_seed1337.log @@ -0,0 +1,210 @@ +W0321 12:56:52.719000 37576 torch/distributed/run.py:803] +W0321 12:56:52.719000 37576 torch/distributed/run.py:803] ***************************************** +W0321 12:56:52.719000 37576 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. +W0321 12:56:52.719000 37576 torch/distributed/run.py:803] ***************************************** +logs/qat_snap70_seed1337.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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: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:6.9294 val_bpb:4.1040 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9301 train_time:133ms step_avg:133.47ms +step:2/20000 train_loss:7.9520 train_time:195ms step_avg:97.71ms +step:3/20000 train_loss:7.5214 train_time:283ms step_avg:94.48ms +step:4/20000 train_loss:6.9071 train_time:372ms step_avg:93.01ms +step:5/20000 train_loss:6.7673 train_time:460ms step_avg:92.07ms +step:6/20000 train_loss:6.7191 train_time:549ms step_avg:91.49ms +step:7/20000 train_loss:6.5969 train_time:636ms step_avg:90.90ms +step:8/20000 train_loss:6.4795 train_time:724ms step_avg:90.51ms +step:9/20000 train_loss:6.2072 train_time:813ms step_avg:90.33ms +step:10/20000 train_loss:5.9981 train_time:901ms step_avg:90.10ms +step:100/20000 train_loss:3.1560 train_time:8969ms step_avg:89.69ms +step:200/20000 train_loss:2.3846 train_time:17989ms step_avg:89.94ms +step:300/20000 train_loss:2.5343 train_time:27029ms step_avg:90.10ms +step:400/20000 train_loss:2.3997 train_time:36075ms step_avg:90.19ms +step:500/20000 train_loss:2.3902 train_time:45082ms step_avg:90.16ms +step:500/20000 val_loss:2.3502 val_bpb:1.3919 train_time:45109ms step_avg:90.22ms +step:600/20000 train_loss:2.3295 train_time:54157ms step_avg:90.26ms +step:700/20000 train_loss:2.3415 train_time:63226ms step_avg:90.32ms +step:800/20000 train_loss:2.2358 train_time:72309ms step_avg:90.39ms +step:900/20000 train_loss:2.1260 train_time:81377ms step_avg:90.42ms +step:1000/20000 train_loss:2.2775 train_time:90412ms step_avg:90.41ms +step:1000/20000 val_loss:2.2275 val_bpb:1.3193 train_time:90439ms step_avg:90.44ms +step:1100/20000 train_loss:2.3213 train_time:99501ms step_avg:90.46ms +step:1200/20000 train_loss:2.3509 train_time:108577ms step_avg:90.48ms +step:1300/20000 train_loss:2.1036 train_time:117655ms step_avg:90.50ms +step:1400/20000 train_loss:2.1834 train_time:126732ms step_avg:90.52ms +step:1500/20000 train_loss:2.2238 train_time:135774ms step_avg:90.52ms +step:1500/20000 val_loss:2.1861 val_bpb:1.2947 train_time:135802ms step_avg:90.53ms +step:1600/20000 train_loss:2.0769 train_time:144908ms step_avg:90.57ms +step:1700/20000 train_loss:2.1443 train_time:153995ms step_avg:90.59ms +step:1800/20000 train_loss:2.1669 train_time:163069ms step_avg:90.59ms +step:1900/20000 train_loss:2.1328 train_time:172098ms step_avg:90.58ms +step:2000/20000 train_loss:2.0709 train_time:181173ms step_avg:90.59ms +step:2000/20000 val_loss:2.1350 val_bpb:1.2645 train_time:181200ms step_avg:90.60ms +step:2100/20000 train_loss:2.0529 train_time:190255ms step_avg:90.60ms +step:2200/20000 train_loss:2.1687 train_time:199329ms step_avg:90.60ms +step:2300/20000 train_loss:2.1136 train_time:208404ms step_avg:90.61ms +step:2400/20000 train_loss:2.0702 train_time:217426ms step_avg:90.59ms +step:2500/20000 train_loss:2.1726 train_time:226503ms step_avg:90.60ms +step:2500/20000 val_loss:2.1103 val_bpb:1.2499 train_time:226530ms step_avg:90.61ms +step:2600/20000 train_loss:2.1121 train_time:235585ms step_avg:90.61ms +step:2700/20000 train_loss:2.1027 train_time:244653ms step_avg:90.61ms +step:2800/20000 train_loss:2.1562 train_time:253720ms step_avg:90.61ms +step:2900/20000 train_loss:2.0286 train_time:262747ms step_avg:90.60ms +step:3000/20000 train_loss:2.1625 train_time:271822ms step_avg:90.61ms +step:3000/20000 val_loss:2.0943 val_bpb:1.2403 train_time:271850ms step_avg:90.62ms +step:3100/20000 train_loss:2.0396 train_time:280903ms step_avg:90.61ms +step:3200/20000 train_loss:2.1752 train_time:289970ms step_avg:90.62ms +step:3300/20000 train_loss:2.0728 train_time:299000ms step_avg:90.61ms +step:3400/20000 train_loss:2.0260 train_time:308087ms step_avg:90.61ms +step:3500/20000 train_loss:2.1890 train_time:317158ms step_avg:90.62ms +step:3500/20000 val_loss:2.0861 val_bpb:1.2355 train_time:317186ms step_avg:90.62ms +step:3600/20000 train_loss:2.1022 train_time:326236ms step_avg:90.62ms +step:3700/20000 train_loss:2.1012 train_time:335305ms step_avg:90.62ms +step:3800/20000 train_loss:2.0781 train_time:344337ms step_avg:90.61ms +step:3900/20000 train_loss:2.0791 train_time:353411ms step_avg:90.62ms +step:4000/20000 train_loss:1.9774 train_time:362475ms step_avg:90.62ms +step:4000/20000 val_loss:2.0700 val_bpb:1.2260 train_time:362501ms step_avg:90.63ms +step:4100/20000 train_loss:2.0160 train_time:371551ms step_avg:90.62ms +step:4200/20000 train_loss:2.1542 train_time:380622ms step_avg:90.62ms +step:4300/20000 train_loss:2.0562 train_time:389650ms step_avg:90.62ms +step:4400/20000 train_loss:2.0337 train_time:398725ms step_avg:90.62ms +step:4500/20000 train_loss:2.1259 train_time:407799ms step_avg:90.62ms +step:4500/20000 val_loss:2.0463 val_bpb:1.2119 train_time:407826ms step_avg:90.63ms +step:4600/20000 train_loss:1.8552 train_time:417315ms step_avg:90.72ms +step:4700/20000 train_loss:2.2464 train_time:426876ms step_avg:90.82ms +step:4800/20000 train_loss:2.4440 train_time:436489ms step_avg:90.94ms +step:4900/20000 train_loss:2.0638 train_time:446100ms step_avg:91.04ms +step:5000/20000 train_loss:2.1133 train_time:455718ms step_avg:91.14ms +step:5000/20000 val_loss:2.0347 val_bpb:1.2051 train_time:455738ms step_avg:91.15ms +step:5100/20000 train_loss:2.1361 train_time:465321ms step_avg:91.24ms +step:5200/20000 train_loss:2.0522 train_time:474880ms step_avg:91.32ms +step:5300/20000 train_loss:2.0168 train_time:484478ms step_avg:91.41ms +swa:start step:5400 +step:5400/20000 train_loss:2.0565 train_time:494093ms step_avg:91.50ms +step:5500/20000 train_loss:2.0252 train_time:503787ms step_avg:91.60ms +step:5500/20000 val_loss:2.0100 val_bpb:1.1904 train_time:503830ms step_avg:91.61ms +step:5600/20000 train_loss:1.9652 train_time:513435ms step_avg:91.68ms +step:5700/20000 train_loss:2.0227 train_time:523044ms step_avg:91.76ms +step:5800/20000 train_loss:2.0123 train_time:532711ms step_avg:91.85ms +step:5900/20000 train_loss:1.9117 train_time:542362ms step_avg:91.93ms +step:6000/20000 train_loss:1.9451 train_time:552010ms step_avg:92.00ms +step:6000/20000 val_loss:1.9853 val_bpb:1.1758 train_time:552052ms step_avg:92.01ms +step:6100/20000 train_loss:1.9225 train_time:561611ms step_avg:92.07ms +step:6200/20000 train_loss:1.9531 train_time:571295ms step_avg:92.14ms +step:6300/20000 train_loss:1.9518 train_time:580981ms step_avg:92.22ms +step:6400/20000 train_loss:2.0050 train_time:590652ms step_avg:92.29ms +step:6497/20000 val_loss:1.9631 val_bpb:1.1627 train_time:600038ms step_avg:92.36ms +stopping_early: wallclock_cap train_time:600038ms step:6497/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +swa:applying averaged 22 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 15780171 bytes +Total submission size int8+zlib: 15834619 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.210436 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.140179 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.140921 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.134716 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.146665 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147428 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.148850 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144470 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.141933 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.143594 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.152524 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150837 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.152232 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.150451 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.149066 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.149485 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.150837 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.151285 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.157308 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.154666 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.155724 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.154386 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.153808 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.153359 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.153999 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.151579 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.150586 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.150912 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.149717 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.149572 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.148825 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.150031 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.151123 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.151644 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.151132 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.151516 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.150598 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.146696 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.146815 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.147742 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.147904 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.147748 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.146530 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.146251 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.145570 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.145642 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.145627 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145806 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.145526 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.146127 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.146406 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.146103 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.147148 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.149092 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.148402 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.149102 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.149467 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.149442 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.149066 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.149313 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.148726 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.151526 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.151536 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.151542 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.151159 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.150671 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.149938 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149941 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.150560 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.150584 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.150564 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.150973 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.150721 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.150323 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.150632 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.150727 +final_int8_zlib_roundtrip val_loss:1.9334 val_bpb:1.1451 eval_time:169723ms +final_int8_zlib_roundtrip_exact val_loss:1.93338725 val_bpb:1.14506387 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/qat_snap70_seed42.log b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/qat_snap70_seed42.log new file mode 100644 index 0000000000..4f3719ab31 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/logs/qat_snap70_seed42.log @@ -0,0 +1,212 @@ +W0321 12:42:31.273000 36721 torch/distributed/run.py:803] +W0321 12:42:31.273000 36721 torch/distributed/run.py:803] ***************************************** +W0321 12:42:31.273000 36721 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. +W0321 12:42:31.273000 36721 torch/distributed/run.py:803] ***************************************** +logs/qat_snap70_seed42.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=/workspace/parameter-golf/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:25517137 +world_size:8 grad_accum_steps:1 +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:600.000 +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:6.9323 val_bpb:4.1057 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9334 train_time:133ms step_avg:133.31ms +step:2/20000 train_loss:8.1444 train_time:196ms step_avg:97.88ms +step:3/20000 train_loss:7.6923 train_time:284ms step_avg:94.63ms +step:4/20000 train_loss:6.9899 train_time:372ms step_avg:93.10ms +step:5/20000 train_loss:6.8344 train_time:460ms step_avg:92.00ms +step:6/20000 train_loss:6.6324 train_time:548ms step_avg:91.34ms +step:7/20000 train_loss:6.5370 train_time:636ms step_avg:90.80ms +step:8/20000 train_loss:6.5776 train_time:724ms step_avg:90.49ms +step:9/20000 train_loss:6.3290 train_time:813ms step_avg:90.28ms +step:10/20000 train_loss:6.0567 train_time:901ms step_avg:90.14ms +step:100/20000 train_loss:3.1540 train_time:8977ms step_avg:89.77ms +step:200/20000 train_loss:2.3791 train_time:18000ms step_avg:90.00ms +step:300/20000 train_loss:2.5430 train_time:27040ms step_avg:90.13ms +step:400/20000 train_loss:2.4100 train_time:36074ms step_avg:90.19ms +step:500/20000 train_loss:2.3966 train_time:45083ms step_avg:90.17ms +step:500/20000 val_loss:2.3559 val_bpb:1.3953 train_time:45110ms step_avg:90.22ms +step:600/20000 train_loss:2.3322 train_time:54142ms step_avg:90.24ms +step:700/20000 train_loss:2.3493 train_time:63211ms step_avg:90.30ms +step:800/20000 train_loss:2.2409 train_time:72279ms step_avg:90.35ms +step:900/20000 train_loss:2.1263 train_time:81345ms step_avg:90.38ms +step:1000/20000 train_loss:2.2766 train_time:90371ms step_avg:90.37ms +step:1000/20000 val_loss:2.2277 val_bpb:1.3194 train_time:90399ms step_avg:90.40ms +step:1100/20000 train_loss:2.3252 train_time:99445ms step_avg:90.40ms +step:1200/20000 train_loss:2.3550 train_time:108513ms step_avg:90.43ms +step:1300/20000 train_loss:2.1028 train_time:117585ms step_avg:90.45ms +step:1400/20000 train_loss:2.1857 train_time:126664ms step_avg:90.47ms +step:1500/20000 train_loss:2.2214 train_time:135685ms step_avg:90.46ms +step:1500/20000 val_loss:2.1850 val_bpb:1.2941 train_time:135713ms step_avg:90.48ms +step:1600/20000 train_loss:2.0770 train_time:144761ms step_avg:90.48ms +step:1700/20000 train_loss:2.1429 train_time:153840ms step_avg:90.49ms +step:1800/20000 train_loss:2.1594 train_time:162913ms step_avg:90.51ms +step:1900/20000 train_loss:2.1275 train_time:171951ms step_avg:90.50ms +step:2000/20000 train_loss:2.0708 train_time:181033ms step_avg:90.52ms +step:2000/20000 val_loss:2.1333 val_bpb:1.2634 train_time:181059ms step_avg:90.53ms +step:2100/20000 train_loss:2.0445 train_time:190120ms step_avg:90.53ms +step:2200/20000 train_loss:2.1843 train_time:199192ms step_avg:90.54ms +step:2300/20000 train_loss:2.1116 train_time:208269ms step_avg:90.55ms +step:2400/20000 train_loss:2.0657 train_time:217296ms step_avg:90.54ms +step:2500/20000 train_loss:2.1730 train_time:226363ms step_avg:90.55ms +step:2500/20000 val_loss:2.1068 val_bpb:1.2478 train_time:226390ms step_avg:90.56ms +step:2600/20000 train_loss:2.1099 train_time:235434ms step_avg:90.55ms +step:2700/20000 train_loss:2.1028 train_time:244502ms step_avg:90.56ms +step:2800/20000 train_loss:2.1580 train_time:253573ms step_avg:90.56ms +step:2900/20000 train_loss:2.0257 train_time:262603ms step_avg:90.55ms +step:3000/20000 train_loss:2.1614 train_time:271674ms step_avg:90.56ms +step:3000/20000 val_loss:2.0934 val_bpb:1.2398 train_time:271701ms step_avg:90.57ms +step:3100/20000 train_loss:2.0352 train_time:280747ms step_avg:90.56ms +step:3200/20000 train_loss:2.1750 train_time:289812ms step_avg:90.57ms +step:3300/20000 train_loss:2.0735 train_time:298831ms step_avg:90.55ms +step:3400/20000 train_loss:2.0256 train_time:307903ms step_avg:90.56ms +step:3500/20000 train_loss:2.1846 train_time:316980ms step_avg:90.57ms +step:3500/20000 val_loss:2.0848 val_bpb:1.2347 train_time:317009ms step_avg:90.57ms +step:3600/20000 train_loss:2.0997 train_time:326050ms step_avg:90.57ms +step:3700/20000 train_loss:2.0994 train_time:335115ms step_avg:90.57ms +step:3800/20000 train_loss:2.0756 train_time:344138ms step_avg:90.56ms +step:3900/20000 train_loss:2.0803 train_time:353204ms step_avg:90.57ms +step:4000/20000 train_loss:1.9790 train_time:362276ms step_avg:90.57ms +step:4000/20000 val_loss:2.0689 val_bpb:1.2253 train_time:362303ms step_avg:90.58ms +step:4100/20000 train_loss:2.0183 train_time:371343ms step_avg:90.57ms +step:4200/20000 train_loss:2.1528 train_time:380411ms step_avg:90.57ms +step:4300/20000 train_loss:2.0572 train_time:389433ms step_avg:90.57ms +step:4400/20000 train_loss:2.0344 train_time:398574ms step_avg:90.58ms +step:4500/20000 train_loss:2.1230 train_time:407645ms step_avg:90.59ms +step:4500/20000 val_loss:2.0453 val_bpb:1.2113 train_time:407671ms step_avg:90.59ms +step:4600/20000 train_loss:1.8533 train_time:417138ms step_avg:90.68ms +step:4700/20000 train_loss:2.2426 train_time:426685ms step_avg:90.78ms +step:4800/20000 train_loss:2.4459 train_time:436292ms step_avg:90.89ms +step:4900/20000 train_loss:2.0637 train_time:445889ms step_avg:91.00ms +step:5000/20000 train_loss:2.1141 train_time:455490ms step_avg:91.10ms +step:5000/20000 val_loss:2.0335 val_bpb:1.2044 train_time:455512ms step_avg:91.10ms +step:5100/20000 train_loss:2.1326 train_time:465088ms step_avg:91.19ms +step:5200/20000 train_loss:2.0552 train_time:474644ms step_avg:91.28ms +step:5300/20000 train_loss:2.0146 train_time:484243ms step_avg:91.37ms +swa:start step:5400 +step:5400/20000 train_loss:2.0562 train_time:493849ms step_avg:91.45ms +step:5500/20000 train_loss:2.0247 train_time:503534ms step_avg:91.55ms +step:5500/20000 val_loss:2.0093 val_bpb:1.1900 train_time:503593ms step_avg:91.56ms +step:5600/20000 train_loss:1.9625 train_time:513204ms step_avg:91.64ms +step:5700/20000 train_loss:2.0228 train_time:522811ms step_avg:91.72ms +step:5800/20000 train_loss:2.0070 train_time:532455ms step_avg:91.80ms +step:5900/20000 train_loss:1.9122 train_time:542097ms step_avg:91.88ms +step:6000/20000 train_loss:1.9467 train_time:551743ms step_avg:91.96ms +step:6000/20000 val_loss:1.9843 val_bpb:1.1752 train_time:551788ms step_avg:91.96ms +step:6100/20000 train_loss:1.9258 train_time:561333ms step_avg:92.02ms +step:6200/20000 train_loss:1.9548 train_time:570997ms step_avg:92.10ms +step:6300/20000 train_loss:1.9530 train_time:580667ms step_avg:92.17ms +step:6400/20000 train_loss:2.0045 train_time:590321ms step_avg:92.24ms +step:6500/20000 train_loss:2.0881 train_time:599960ms step_avg:92.30ms +step:6500/20000 val_loss:1.9627 val_bpb:1.1624 train_time:600006ms step_avg:92.31ms +step:6501/20000 val_loss:1.9627 val_bpb:1.1624 train_time:600102ms step_avg:92.31ms +stopping_early: wallclock_cap train_time:600102ms step:6501/20000 +peak memory allocated: 18959 MiB reserved: 19106 MiB +swa:applying averaged 23 checkpoints +Serialized model: 98437419 bytes +Code size: 54448 bytes +Total submission size: 98491867 bytes +Serialized model int6+zstd: 16431825 bytes +Total submission size int8+zlib: 16486273 bytes +final_eval_mode:sliding_window stride:64 batch_seqs:32 + sliding_eval [ 0.0%] 32/121136 windows running_bpb=1.209359 + sliding_eval [ 1.3%] 1632/121136 windows running_bpb=1.139071 + sliding_eval [ 2.7%] 3232/121136 windows running_bpb=1.140218 + sliding_eval [ 4.0%] 4832/121136 windows running_bpb=1.133889 + sliding_eval [ 5.3%] 6432/121136 windows running_bpb=1.145876 + sliding_eval [ 6.6%] 8032/121136 windows running_bpb=1.147242 + sliding_eval [ 8.0%] 9632/121136 windows running_bpb=1.148731 + sliding_eval [ 9.3%] 11232/121136 windows running_bpb=1.144021 + sliding_eval [ 10.6%] 12832/121136 windows running_bpb=1.141315 + sliding_eval [ 11.9%] 14432/121136 windows running_bpb=1.142894 + sliding_eval [ 13.2%] 16032/121136 windows running_bpb=1.151671 + sliding_eval [ 14.6%] 17632/121136 windows running_bpb=1.150140 + sliding_eval [ 15.9%] 19232/121136 windows running_bpb=1.151429 + sliding_eval [ 17.2%] 20832/121136 windows running_bpb=1.149578 + sliding_eval [ 18.5%] 22432/121136 windows running_bpb=1.148170 + sliding_eval [ 19.8%] 24032/121136 windows running_bpb=1.148474 + sliding_eval [ 21.2%] 25632/121136 windows running_bpb=1.149809 + sliding_eval [ 22.5%] 27232/121136 windows running_bpb=1.150235 + sliding_eval [ 23.8%] 28832/121136 windows running_bpb=1.156352 + sliding_eval [ 25.1%] 30432/121136 windows running_bpb=1.153743 + sliding_eval [ 26.4%] 32032/121136 windows running_bpb=1.154728 + sliding_eval [ 27.8%] 33632/121136 windows running_bpb=1.153359 + sliding_eval [ 29.1%] 35232/121136 windows running_bpb=1.152798 + sliding_eval [ 30.4%] 36832/121136 windows running_bpb=1.152414 + sliding_eval [ 31.7%] 38432/121136 windows running_bpb=1.153049 + sliding_eval [ 33.0%] 40032/121136 windows running_bpb=1.150600 + sliding_eval [ 34.4%] 41632/121136 windows running_bpb=1.149676 + sliding_eval [ 35.7%] 43232/121136 windows running_bpb=1.149971 + sliding_eval [ 37.0%] 44832/121136 windows running_bpb=1.148776 + sliding_eval [ 38.3%] 46432/121136 windows running_bpb=1.148638 + sliding_eval [ 39.7%] 48032/121136 windows running_bpb=1.147929 + sliding_eval [ 41.0%] 49632/121136 windows running_bpb=1.149161 + sliding_eval [ 42.3%] 51232/121136 windows running_bpb=1.150228 + sliding_eval [ 43.6%] 52832/121136 windows running_bpb=1.150731 + sliding_eval [ 44.9%] 54432/121136 windows running_bpb=1.150255 + sliding_eval [ 46.3%] 56032/121136 windows running_bpb=1.150634 + sliding_eval [ 47.6%] 57632/121136 windows running_bpb=1.149746 + sliding_eval [ 48.9%] 59232/121136 windows running_bpb=1.145898 + sliding_eval [ 50.2%] 60832/121136 windows running_bpb=1.146008 + sliding_eval [ 51.5%] 62432/121136 windows running_bpb=1.146925 + sliding_eval [ 52.9%] 64032/121136 windows running_bpb=1.147108 + sliding_eval [ 54.2%] 65632/121136 windows running_bpb=1.146985 + sliding_eval [ 55.5%] 67232/121136 windows running_bpb=1.145750 + sliding_eval [ 56.8%] 68832/121136 windows running_bpb=1.145489 + sliding_eval [ 58.1%] 70432/121136 windows running_bpb=1.144816 + sliding_eval [ 59.5%] 72032/121136 windows running_bpb=1.144888 + sliding_eval [ 60.8%] 73632/121136 windows running_bpb=1.144880 + sliding_eval [ 62.1%] 75232/121136 windows running_bpb=1.145066 + sliding_eval [ 63.4%] 76832/121136 windows running_bpb=1.144802 + sliding_eval [ 64.7%] 78432/121136 windows running_bpb=1.145394 + sliding_eval [ 66.1%] 80032/121136 windows running_bpb=1.145694 + sliding_eval [ 67.4%] 81632/121136 windows running_bpb=1.145385 + sliding_eval [ 68.7%] 83232/121136 windows running_bpb=1.146396 + sliding_eval [ 70.0%] 84832/121136 windows running_bpb=1.148301 + sliding_eval [ 71.4%] 86432/121136 windows running_bpb=1.147622 + sliding_eval [ 72.7%] 88032/121136 windows running_bpb=1.148326 + sliding_eval [ 74.0%] 89632/121136 windows running_bpb=1.148677 + sliding_eval [ 75.3%] 91232/121136 windows running_bpb=1.148635 + sliding_eval [ 76.6%] 92832/121136 windows running_bpb=1.148238 + sliding_eval [ 78.0%] 94432/121136 windows running_bpb=1.148430 + sliding_eval [ 79.3%] 96032/121136 windows running_bpb=1.147854 + sliding_eval [ 80.6%] 97632/121136 windows running_bpb=1.150649 + sliding_eval [ 81.9%] 99232/121136 windows running_bpb=1.150677 + sliding_eval [ 83.2%] 100832/121136 windows running_bpb=1.150740 + sliding_eval [ 84.6%] 102432/121136 windows running_bpb=1.150371 + sliding_eval [ 85.9%] 104032/121136 windows running_bpb=1.149881 + sliding_eval [ 87.2%] 105632/121136 windows running_bpb=1.149132 + sliding_eval [ 88.5%] 107232/121136 windows running_bpb=1.149093 + sliding_eval [ 89.8%] 108832/121136 windows running_bpb=1.149725 + sliding_eval [ 91.2%] 110432/121136 windows running_bpb=1.149761 + sliding_eval [ 92.5%] 112032/121136 windows running_bpb=1.149757 + sliding_eval [ 93.8%] 113632/121136 windows running_bpb=1.150194 + sliding_eval [ 95.1%] 115232/121136 windows running_bpb=1.149978 + sliding_eval [ 96.4%] 116832/121136 windows running_bpb=1.149574 + sliding_eval [ 97.8%] 118432/121136 windows running_bpb=1.149877 + sliding_eval [ 99.1%] 120032/121136 windows running_bpb=1.149968 +final_int8_zlib_roundtrip val_loss:1.9321 val_bpb:1.1443 eval_time:169979ms +final_int8_zlib_roundtrip_exact val_loss:1.93208752 val_bpb:1.14429409 diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/run.sh b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/run.sh new file mode 100644 index 0000000000..578ac22c93 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/run.sh @@ -0,0 +1,83 @@ +#!/usr/bin/env bash +set -euo pipefail + +# Usage: bash run.sh [seed] +# Experiments: control, qat_snap70, no_swa, no_swa_qat +# Default seed: 42 + +EXPERIMENT="${1:?Usage: bash run.sh [seed]}" +SEED="${2:-42}" + +SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" +REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)" + +# Common settings (identical across all experiments) +export DATA_PATH="$REPO_ROOT/data/datasets/fineweb10B_sp1024" +export TOKENIZER_PATH="$REPO_ROOT/data/tokenizers/fineweb_1024_bpe.model" +export VOCAB_SIZE=1024 +export SEED="$SEED" +export RUN_ID="${EXPERIMENT}_seed${SEED}" +export VAL_LOSS_EVERY=500 +export TRAIN_LOG_EVERY=100 + +# Defaults from #180 (unchanged) +export NUM_LAYERS=10 +export MODEL_DIM=512 +export MLP_MULT=3.0 +export NUM_HEADS=8 +export NUM_KV_HEADS=4 +export WEIGHT_DECAY=0.04 +export MUON_MOMENTUM=0.99 +export GRAD_CLIP_NORM=0.3 +export EVAL_STRIDE=64 +export BIGRAM_VOCAB_SIZE=10240 +export BIGRAM_DIM=128 +export TRAIN_SEQ_LEN=2048 +export TRAIN_BATCH_TOKENS=786432 + +# 2x2 Matrix: QAT (on/off) x SWA (on/off) +case "$EXPERIMENT" in + control) + # Cell [QAT=off, SWA=on] — this is the #180 baseline + export ENABLE_QAT=0 + export SWA_ENABLED=1 + ;; + qat_snap70) + # Cell [QAT=on, SWA=on] — QAT activates in last 70% of warmdown + export ENABLE_QAT=1 + export QAT_START_FRAC=0.7 + export SWA_ENABLED=1 + ;; + no_swa) + # Cell [QAT=off, SWA=off] + export ENABLE_QAT=0 + export SWA_ENABLED=0 + ;; + no_swa_qat) + # Cell [QAT=on, SWA=off] — QAT without SWA + export ENABLE_QAT=1 + export QAT_START_FRAC=0.7 + export SWA_ENABLED=0 + ;; + *) + echo "Unknown experiment: $EXPERIMENT" + echo "Valid: control, qat_snap70, no_swa, no_swa_qat" + exit 1 + ;; +esac + +echo "==========================================" +echo "Experiment: $EXPERIMENT (seed=$SEED)" +echo "ENABLE_QAT=${ENABLE_QAT} SWA_ENABLED=${SWA_ENABLED}" +echo "==========================================" + +# Detect GPU count +NGPU=$(nvidia-smi -L | wc -l) +echo "GPUs detected: $NGPU" + +mkdir -p "$SCRIPT_DIR/logs" + +cd "$REPO_ROOT" +torchrun --standalone --nproc_per_node="$NGPU" \ + experiments/alex_qat/train_gpt.py \ + 2>&1 | tee "experiments/alex_qat/logs/${EXPERIMENT}_seed${SEED}.log" diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/run_matrix.sh b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/run_matrix.sh new file mode 100644 index 0000000000..9bd77b5bd1 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/run_matrix.sh @@ -0,0 +1,27 @@ +#!/usr/bin/env bash +set -euo pipefail + +# Run the full 2x2 matrix with 2 seeds each (8 total runs) +# Usage: bash run_matrix.sh + +SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" +mkdir -p "$SCRIPT_DIR/logs" + +EXPERIMENTS=(control qat_snap70 no_swa no_swa_qat) +SEEDS=(42 1337) + +for exp in "${EXPERIMENTS[@]}"; do + for seed in "${SEEDS[@]}"; do + echo "" + echo "============================================" + echo " Running: $exp (seed=$seed)" + echo "============================================" + bash "$SCRIPT_DIR/run.sh" "$exp" "$seed" + done +done + +echo "" +echo "============================================" +echo " ALL EXPERIMENTS COMPLETE" +echo "============================================" +echo "Logs in: $SCRIPT_DIR/logs/" diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/submission.json b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/submission.json new file mode 100644 index 0000000000..69b876a034 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/submission.json @@ -0,0 +1,20 @@ +{ + "name": "QAT x SWA Ablation: Antagonistic Interaction in Quantization-Aware Training", + "author": "Alexander Perry", + "github_id": "alexanderaperry-arch", + "date": "2026-03-28", + "track": "10min-16mb", + "val_bpb": 1.14018, + "val_loss": 1.9250, + "bytes_total": 15787003, + "blurb": "2x2 factorial ablation of QAT x SWA interaction on PR #180 stack. Key finding: SWA and QAT are antagonistic — QAT alone (1.14018, 3-seed mean) beats SWA alone (1.14382) by 3.64 mBPB, but combining them (1.14468) is worse than either. Explains why prior QAT entries underperformed non-QAT submissions.", + "submission_type": "non-record-research", + "seeds": [42, 1337, 2024], + "seed_results": { + "no_swa_qat": {"seed_42": 1.13969, "seed_1337": 1.14010, "seed_2024": 1.14074, "mean": 1.14018, "std": 0.00044}, + "control": {"seed_42": 1.14335, "seed_1337": 1.14350, "seed_2024": 1.14462, "mean": 1.14382, "std": 0.00056} + }, + "hardware": "8xH100 SXM (RunPod)", + "wallclock_seconds": 600, + "based_on": "PR #180 (10L Int5-MLP + BigramHash + SWA + WD=0.04)" +} diff --git a/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/train_gpt.py b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/train_gpt.py new file mode 100644 index 0000000000..2ba46bb000 --- /dev/null +++ b/records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/train_gpt.py @@ -0,0 +1,1267 @@ +""" +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", 42)) + + 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", 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 = 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", 0.04)) + + 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", 10240)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + enable_qat = bool(int(os.environ.get("ENABLE_QAT", "0"))) + qat_start_frac = float(os.environ.get("QAT_START_FRAC", "0.7")) + prune_pct = float(os.environ.get("PRUNE_PCT", 0.10)) # magnitude pruning % (10% for <16MB decimal limit) + +# ----------------------------- +# 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 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 "bigram" in name: + return "bigram" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip_range).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]), -(clip_range+1), clip_range).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range+1), clip_range).to(torch.int8) + return q, scale + +def fake_quantize_per_row(weight: Tensor, clip_range: int) -> Tensor: + """Simulate per-row intN quantization: quantize then dequantize (STE).""" + if weight.ndim != 2: + return weight + q, scale = quantize_intN_per_row(weight, clip_range=clip_range) + return (q.float() * scale.float()[:, None]).to(weight.dtype) + +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() <= 8192: + 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: + clip = 15 if cat == "mlp" else 31 # int5 for MLP, int6 for attention + q, s = quantize_intN_per_row(t, clip_range=clip) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{5 if cat == 'mlp' else 6}"} + 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 + + +# ----------------------------- +# 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 >= stride or ws == 0] + 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}" + ) + 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 + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + _qat_active = False + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + 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) + if args.enable_qat: + _qat_active = args.qat_start_frac <= 0.0 or scale < args.qat_start_frac + else: + _qat_active = False + 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) + _saved_weights: dict[str, Tensor] = {} + if _qat_active: + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + cat = _classify_param(name) + if cat in ("mlp", "attn", "bigram"): + _saved_weights[name] = param.data.clone() + clip_range = 15 if cat == "mlp" else 31 + param.data.copy_(fake_quantize_per_row(param.data, clip_range)) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if _saved_weights: + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if name in _saved_weights: + param.data.copy_(_saved_weights[name]) + 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) + + # 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" + ) + + # Apply SWA if collected + if args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + current_state = base_model.state_dict() + avg_state = { + name: (tensor / swa_count).to(dtype=current_state[name].dtype) + for name, tensor in swa_state.items() + } + base_model.load_state_dict(avg_state, strict=True) + + # 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") + + # Magnitude pruning: zero out smallest weights to improve compression + log0(f"magnitude_pruning: {args.prune_pct*100:.1f}% of weights with >65536 elements") + with torch.no_grad(): + for name, param in base_model.named_parameters(): + if param.ndim == 2 and param.numel() > 65536: + threshold = torch.quantile(param.abs().float().flatten(), args.prune_pct) + mask = param.abs() < threshold + param.masked_fill_(mask, 0.0) + + # INT6 mixed quantization + zstd/zlib export + sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn", "bigram"}) + 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) + 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")) + total_bytes = quant_file_bytes + code_bytes + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int8+zlib: {total_bytes} bytes") + if total_bytes > 16_000_000: + log0(f"WARNING: artifact {total_bytes} exceeds 16,000,000 byte limit by {total_bytes - 16_000_000} bytes!") + else: + log0(f"artifact_headroom: {16_000_000 - total_bytes} bytes under limit") + + 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"], sd_cpu) + 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() +# fixes applied +# tuned