modelscope / modelscope/DiffSynth-Studio
strange training speed about full training and lora training in Wan2.1-VACE-14B
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 13.1k
- Forks
- 1.3k
- Avg merge
- 13h 12m
- Merged PRs (30d)
- 45
Description
In Wan2.1-VACE-14B training, I'm running Full Training and LoRA Training separately, but the training speed is not much difference, some example logs:
- full training
0%| | 2/1250 [01:35<15:34:53, 44.95s/it]
0%| | 2/1250 [01:36<15:51:01, 45.72s/it]
0%| | 2/1250 [01:34<15:30:58, 44.76s/it]
0%| | 2/1250 [01:36<15:55:33, 45.94s/it]
0%| | 2/1250 [01:37<15:45:22, 45.45s/it]
- lora training
0%| | 1/1250 [00:46<16:14:52, 46.83s/it]
0%| | 1/1250 [00:46<16:14:52, 46.83s/it]
0%| | 1/1250 [00:46<16:14:54, 46.83s/it]
0%| | 1/1250 [00:46<16:14:54, 46.83s/it]
0%| | 1/1250 [00:46<16:14:59, 46.84s/it]
Is this expected?
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by comparing examples/wanvideo/model_training/full/Wan2.1-VACE-14B.sh with examples/wanvideo/model_training/lora/Wan2.1-VACE-14B.sh and their reported iteration times. Trace the training entry points used by both scripts, then establish whether the similar throughput is expected and document the relevant finding or required change.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100