Quantization and MoE configs for GH200 machines

[arvindsun]

The H200 configs work for the GH200 (it has 92GB gpu memory).

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[robertgshaw2-redhat]

Any end-to-end performance benchmark?

[robertgshaw2-redhat]

Can you post an end-to-end performance benchmark?

These configs are usually tuned on a gpu by gpu basis.

[mgoin]

Is it true that GH200 only comes in single or dual device configurations? So in order to fit deepseek v3 (assuming this since you specify block_shape 128x128) you would need 5 nodes (2xGH200) or 10 nodes (1xGH200) distributed?

[arvindsun]

Is it true that GH200 only comes in single or dual device configurations? So in order to fit deepseek v3 (assuming this since you specify block_shape 128x128) you would need 5 nodes (2xGH200) or 10 nodes (1xGH200) distributed?

We have a configuration with multiple nodes connected over Inifinband. Tested this on a 8 nodex (each with 1xGH200) - can get around 25-27 tokens per second

[arvindsun]

Can you post an end-to-end performance benchmark?

These configs are usually tuned on a gpu by gpu basis.

I am not sure I can do a full e2e perf for all the configs. I am only able to test for 2 of them - the fused_moe and the N=576,K=7168 for the quant config. Should I revert the rest from this PR?