windreamer

> The poor performance is relative though, take the RTX 6000 - if you have a 110GB model that you want to use with the 96GB VRAM, offloading that final...

``` File "/home/li_mingze/.local/lib/python3.12/site-packages/transformers/models/auto/processing_auto.py", line 28, in from ...processing_utils import ProcessorMixin File "/home/li_mingze/.local/lib/python3.12/site-packages/transformers/processing_utils.py", line 34, in from .audio_utils import load_audio File "/home/li_mingze/.local/lib/python3.12/site-packages/transformers/audio_utils.py", line 42, in import soundfile as sf File "/home/li_mingze/.local/lib/python3.12/site-packages/soundfile.py",...

@HsinHui-Tseng could you please try #4106 to test if this can fix the issue? Thanks.

It seems we didn't find corresponding lora parameters in your lora model. Could you please share your lora model or list your lora parameters so that we can reproduce this...

May I know how do you train your lora model? It seems the parameter name is not what LMDeploy expected. For lora models of internvl, the parameters should be named...

Actrually I am not quite falimlar with this topic. You can check if the following link helps: https://internvl.readthedocs.io/en/latest/tutorials/coco_caption_finetune.html

You may try [Xtuner]() for LoRA finetuning of InternVL3: Here is an example script for internlm2-chat-7b: https://github.com/InternLM/xtuner/blob/main/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_7b_qlora_custom_sft_e1.py And you can also read this Chinese doc for more details: https://github.com/InternLM/xtuner/blob/main/docs/zh_cn/legacy/training/custom_sft_dataset.rst Hope...

This may have been fixed in #4029 in release v0.10.2 . You can check if this release fix your issue.