[Question] Does TransformerLens support LVLM like Qwen2-VL?

[Tizzzzy]

Question

Hi, I was wondering if TransformerLens support loading LVLM like Qwen2.5-VL?

Thank you

[bryce13950]

TransformerLens just started officially supporting vision models in the last month. There is a way to basically make it happen without needing to wait for a specific model to be added, but there is a little bit of work that would need to be done on the users part. It's not too bad, since we already support QWen2, so if you are interested in using TransformerLens with Qwen2-VL ASAP, let me know, and we can potentially look at what needs to be done to put it together.

[fzyzcjy]

+1 it would be great if vision models like qwen2 vl can be supported!

[Tizzzzy]

TransformerLens just started officially supporting vision models in the last month. There is a way to basically make it happen without needing to wait for a specific model to be added, but there is a little bit of work that would need to be done on the users part. It's not too bad, since we already support QWen2, so if you are interested in using TransformerLens with Qwen2-VL ASAP, let me know, and we can potentially look at what needs to be done to put it together.

Hi, Thank you for your reply. Is it possible to support Qwen2.5-VL? Let me know if I need to code anything on my end.

[ExplorerFreda]

@Tizzzzy @fzyzcjy shameless self-advertisement, but I think you should try out our tool (https://github.com/compling-wat/vlm-lens) if there's still a need!