[Feature]: Add moe_wna16 kernel as a backend for CompressedTensorsWNA16MoEMethod

[mgoin]

🚀 The feature, motivation and pitch

A triton implementation to support MoE layers quantized with GPTQ or AWQ was implemented in #12185

It is more performant than the current Marlin MoE kernel in the case where there are many, small experts - which is why I ported it to be the default in the case of num_experts > 32 for AWQ and GPTQMarlin configs #13236

We should also propagate the usage of this kernel to compressed-tensors that have mixed precision.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[jennifurhe]

Hi! I'm exploring potential issues to contribute to, and I noticed this one. It looks like PR #16038 might already address this. Just wanted to check if this issue is still open or needs further work. Thanks!

[mgoin]

Thanks for checking @panyuhe, yes this issue has basically been addressed. I'll try to make some more issues for folks to pick up! Feel free to ping me (@mgoin) in the vLLM Slack if you have specific ideas for discussion!