Datasets:
File size: 2,674 Bytes
83a780e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | # PanTS Wan2.2 Fine-tune: Cache, Code, and Step-8000 Checkpoint
This package is the reproducibility bundle for the UCSF run `pants_b16_9k_20260519_215532`.
The run targeted 9000 training steps, but the 2-day Slurm allocation ended after step 8016. There was no traceback, OOM, or NaN in the training log. The last complete saved checkpoint pair is step 8000.
## What Is Included
- Final non-EMA checkpoint: `checkpoints/checkpoint-8000/transformer/`
- Distributed resume checkpoint: `checkpoints/checkpoint-8000/distributed_checkpoint/`
- Final EMA checkpoint: `checkpoints/ema_checkpoint-8000/`
- Training cache archive: `cache/wan22_pants_v2_softwin.tar.zst.part-*`
- Two code snapshots: `code/twoframe_fastvideo_code_snapshot.tar.zst`
- Training logs and offline tracker: `logs/`
- Metadata and provenance: `metadata/`
The original raw PanTS train/test data is intentionally packaged separately as:
`/scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411`
That raw package is better uploaded as a Hugging Face dataset repo. This package is better uploaded as a Hugging Face model repo.
## Key Paths From Training
- Raw data root: `/scratch/user/yuhwang/dataset/PanTS`
- Derived cache root: `/scratch/user/yuhwang/dataset/pants-captions-ldm/cache/wan22_pants_v2_softwin`
- Base model path: `/scratch/user/yuhwang/model/Wan2.2-TI2V-5B-Diffusers-merged`
- TwoFrame code: `/scratch/user/yuhwang/code/TwoFrame`
- FastVideo code: `/scratch/user/yuhwang/code/FastVideo`
- Run output: `/scratch/user/yuhwang/artifacts/twoframe/pants_wan22_finetune/pants_b16_9k_20260519_215532`
## Training Summary
See `metadata/train_args.json` and `metadata/train_health.json` for the full captured FastVideo arguments and log health summary.
Important args:
- Model: Wan2.2 TI2V 5B Diffusers merged local directory
- Data cache: `wan22_pants_v2_softwin`
- GPUs: 8
- Batch size: 16
- Precision: bf16
- Optimizer: AdamW
- Learning rate: `1e-6`
- EMA: enabled, decay `0.999`, start step `1`
- Checkpoint interval: 4000 steps
- Last complete checkpoint: 8000
## Restore Cache
```bash
cd cache
cat wan22_pants_v2_softwin.tar.zst.part-* > wan22_pants_v2_softwin.tar.zst
tar --use-compress-program zstd -xf wan22_pants_v2_softwin.tar.zst
```
## Restore Code Snapshot
```bash
mkdir -p /scratch/user/yuhwang/code_restore
cd /scratch/user/yuhwang/code_restore
tar --use-compress-program zstd -xf /path/to/code/twoframe_fastvideo_code_snapshot.tar.zst
```
Then use the FastVideo and TwoFrame trees from that restored directory, or compare against the git status/diff metadata in `metadata/`.
## Check Integrity
```bash
sha256sum -c SHA256SUMS.txt
```
|