internvl3 support

[Li-Jicheng]

Hi,

I'm trying to load InternVL3 using sglang but encountered issues. I noticed this pull request: #5350 mentions that "InternVL3 includes a flashattention implementation for vision models. However, it doesn't support Tensor Parallelism (TP), which could be a bottleneck".

Does this imply I need to disable TP for a bug-free experience with InternVL3? If yes, are there any known workarounds or fixes planned for future releases?

Thank you!

[LugerW-A]

hi did you solve it?