modelscope / modelscope/DiffSynth-Studio
开启 deepspeed stage3 对 wan i2v 14b 进行 lora finetune，权重参数和推理结果不正常
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 13.1k
- Forks
- 1.3k
- Avg merge
- 13h 12m
- Merged PRs (30d)
- 45
Description
-
用 zero_to_fp32.py 文件将 lora 权重保存为若干 .safetensors 文件，文件中除了 lora_A 和 lora_B 的权重外，还有 base_layer 的 权重，无论把 base_layer 的权重作为 融合后的权重还是融合前的权重，都对不上，即，base_layer 权重 != 预训练模型中对应层的权重，base_layer - lora_B @ lora_A != 预训练模型中对应层的权重
-
用 1. 中得到的 .safetensors 文件 进行 lora finetune 后的推理 (将 base_layer 去掉则可顺利加载)，推理出的结果很奇怪
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
Reproduce the Wan I2V 14B LoRA fine-tuning flow with DeepSpeed Stage 3 and inspect zero_to_fp32.py output in the generated .safetensors files. Compare base_layer, lora_A, and lora_B against the pretrained model, then test inference with and without base_layer. Done means the reconstructed weights match the pretrained layers and the resulting inference is no longer anomalous.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100