Support W8A8 INT8 MoE for compressed-tensors

[mgoin]

Enables INT8 W8A8 channelwise weights and dynamic per-token activations for MoE models quantized with compressed-tensors.

vllm (pretrained=nm-testing/DeepSeek-Coder-V2-Lite-Instruct-quantized.w8a8,max_model_len=2048,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.7513|±  |0.0119|
|     |       |strict-match    |     5|exact_match|↑  |0.7301|±  |0.0122|

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[bnellnm]

LGTM. Is there a test for this?

[ElizaWszola]

Why is this ones? Are the scales not always present?

[mgoin]

Good catch, this is just copied from FP8. I don't think it matters as scales should always be present