多图推理OOM

[whatever-wlb]

当进行多图推理时，会爆显存（单卡，32G显存） 提升使用的推理卡数量（2卡，4卡均有尝试），依然爆显存，请问为什么会出现该问题

[jklj077]

Hi, we need more information to understand your situation:

• which framework were you using? transformers or vllm? was flash_attention used?
• how did you run the inference? what resolutions were the images? did you try lowering max_pixels ($$x$$ pixels = $$x/784$$ tokens) as in the README?

[whatever-wlb]

model: Qwen2-VL-7B-Instruct framework: transformers no flash_attention how did you run the inference? : use demo code in readme resolutions: [1436, 717] or higher did you try lowering max_pixels (x pixels = x/784 tokens) as in the README? no, I'm worried that I'll lose image details this way

[sharonsalabiglossai]

@jklj077 where did you see that (x pixels = x/784 tokens)? can we control it during finetuning will llama factory?