update vllm to 0.3

[themrzmaster]

0.3 introduce custom fused MOE layers.
vllm-project/vllm#2542
improve performance

[jeffreymeetkai]

Hi, thank you for the notifying us about this improvement!

I've just tested our server inference with vllm migrated to v0.3.0 and it seems like there are some major changes to various parts of vllm such as the introduction of TokenizerGroup, etc. which are breaking our server inference both with and without grammar sampling. More work on the existing server inference code will need to be done before we can safely migrate over to vllm==0.3.0.

I will open a new PR once the modifications to existing code is ready. Thank you once again for this suggestion!