[Model] Use the same fused_moe configs for all H200 devices

[bufferoverflow]

Purpose

The main difference of H200 and H200 NVL is the interface (SXM5, vs PCIe), so the configs can be the same. This PR adds a copy of the H200 configs for H200_NVL. see https://www.nvidia.com/en-us/data-center/h200/ and https://resources.nvidia.com/en-us-data-center-overview-mc/en-us-data-center-overview/grace-hopper-superchip-datasheet-partner (GH200)

Test Plan

run qwen3-coder-30b-a3b-instruct on H200 NVL, no warning such as

WARNING 2025-08-25 01:54:03,624 [fused_moe.py:727] Using default MoE config. Performance might be sub-optimal! Config file not found at ['/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H200_NVL.json'] 

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

⚒️ with ❤️ by https://github.com/siemens

[gemini-code-assist]

Code Review

This pull request adds tuned configurations for the H200_NVL GPU by copying the existing configurations for the H200. While this approach enables support for the new hardware, it introduces a significant amount of file duplication, which can become a maintenance burden. I've left a comment on one of the new files with a suggestion for a more scalable approach by implementing a fallback in the configuration loading logic. This would avoid the need to duplicate configuration files for hardware with similar performance characteristics.

[bufferoverflow]

@simon-mo could you have a look at this?

[tomschelsen]

In the same vein, what about the GH200 ? Running Qwen/Qwen3-Coder-30B-A3B-Instruct on an Nvidia GH200 NVL2 system, I get the following warning :

WARNING 10-10 07:30:42 [fused_moe.py:798] Using default MoE config. Performance might be sub-optimal! Config file not found at [‘/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_GH200_144G_HBM3e.json’] 

So here device_name = NVIDIA_GH200_144G_HBM3e

There exist a config for the same E and N for the H200 : vllm/vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H200.json at main · vllm-project/vllm · GitHub

So maybe this PR could be extended to all device names containing H200, unless there is a reason to differentiate between H200 and GH200 ?

[bufferoverflow]

@tomschelsen I changed to use the same config for all H200 devices

[Isotr0py]

Look reasonable to me!

[simon-mo]

Sorry about the late reply, there's GH200 as well which is an H100 chip under the hood.

[robertgshaw2-redhat]

This PR completely breaks the H200 triton kernel config lookup

It needs to be device_name=NVIDIA_H200

[robertgshaw2-redhat]

fixed: #28349

[tomschelsen]

And so @simon-mo see my message above how could I get the right config for the device that I have then ? Thanks