Isotr0py

> I think the config is too complicated, and it will also request hf multiple times. I think we just need to redirect to local location when initializing `ModelConfig`, so...

> Otherwise, log output and server name will be very weird. IMO, if we redirect the model_repo to local directory manually, we should also make sure the model's name updated...

> 1. Redirection to local location when initializing ModelConfig. Log output and server names will be very weird. It may even trigger strange bugs @noooop I prefer 1 TBH. Redirection...

The entrypoint test failure should be unrelated, I can confirm it's passed locally. The V1 test is flaky currently. 😅

> same err here > The error can be resolved by installing the specific version vllm==0.10.1.1. Sorry for breaking this. https://github.com/vllm-project/vllm/pull/25146 should fix this.

@Journey7331 I tried but unable to reproduce the tensor schema issue on vLLM's current main branch. The processed pixvel_values from `OpenGVLab/InternVL3_5-30B-A3B-Instruct` indeed has 448x448 size per patch on my side....

Hmmm, I remember that pytorch 2.8 has deprecated cu118 support...

> BTH, pixel_values error only exists with ckpt downloaded from OpenGVLab/InternVL3_5-30B-A3B-HF HF version Oh, I see. HF format models use a different model implementation compared to GitHub format, will take...

Oh, I totally forgot this... 😅 Let me port the ViT for the florence models to finish this.

This should has been supported, you can try serving the AWQ model with this command: ```shell vllm serve OpenGVLab/InternVL2_5-78B-AWQ --quantization awq --dtype half --max-model-len 4096 ```