darkpromise
darkpromise
The authors of LLaVA have tried to create the beta-version of batch inference [https://github.com/haotian-liu/LLaVA/issues/754](url)
> Hi, @darkpromise98 , we will try to include this feature into VLMEvalKit recently. That's great !
> > > > 感谢您的尝试，在inplace前先clone()就可以解决这个问题, 可以同步下huggingface上最新的模型代码，可以参考这个https://huggingface.co/Qwen/Qwen-VL-Chat/commit/e71876524333aa79e65b19c557c639175552e689 > > > > > > > > > 但是我修改后，运行 hugginface会冲掉我的修改，这个怎么解决？感谢 > > > 这个问题已经解决了，要修改原来下载的文件内容而不是在.cache/transformer的文件内容，因为前者会覆盖后者的内容，因此在后面的文件改没有用 > > 您好 请问一下，我只在 .cache目录下 才能找到这个文件，您说的“原来的文件”是手动下载的吗？ “原来的文件” 是你模型文件在本地下载好的路径，例如 /home/your_name/models/Qwen-VL-Chat/modeling_qwen.py",...
+1，looking forward to such a feature.
不微调的话，直接用MLLM output embeddings做相似度度量感觉不太行
When handling multiple images, it's helpful to add **labels (image index)** to the images for better reference. Users can control this behavior with the following settings: `prompt_with_id = processor.apply_chat_template(messages, add_generation_prompt=True,...
We've found that this is a bug caused by our oversight. We'll fix it in future versions.
We found the the parameter `add_vision_id = True` will be ignored by the source code of `transformers/processing_utils.py` (https://github.com/huggingface/transformers/blob/main/src/transformers/processing_utils.py#L1283) To address this bug, please add the parameter definition of `add_vision_id` in...
We recommend to use the transformers with 4.49.0-dev version, which have some differences from version 4.49.0 in the source code. But the reason for this bug is the same.