Add H100 fused MoE config

[skyloevil]

Purpose

add tuned config for E=128,N=384 on NVIDIA_H100_PCIe with [128,128] block shape so the fused MoE kernel can reuse benchmarked settings vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H100_PCIe,block_shape=[128,128].json

Test Plan

python benchmarks/kernels/benchmark_moe.py --tune --save-dir vllm/model_executor/layers/fused_moe/configs --model Qwen/Qwen3-30B-A3B-FP8 --tp-size 2

[gemini-code-assist]

Code Review

This pull request adds a tuned configuration file for the fused MoE kernel on NVIDIA H100 GPUs. While the configuration values themselves seem reasonable, there is a critical issue with the filename. It appears to be missing the dtype=fp8_w8a8 specifier, which is necessary for vLLM to load this configuration for FP8 models at runtime. Without the correct filename, the system will fall back to default parameters, negating the performance benefits of this tuning.

[gemini-code-assist]

The filename for this configuration is missing the dtype specifier. According to the pull request description, this configuration is for an FP8 model. At runtime, vLLM constructs the configuration filename including a dtype part (e.g., dtype=fp8_w8a8 for FP8 models). Without this in the filename, this configuration file will not be found, and the fused MoE kernel will fall back to default, likely suboptimal, parameters.

To fix this, the file should be renamed to:
E=128,N=384,device_name=NVIDIA_H100_PCIe,dtype=fp8_w8a8,block_shape=[128,128].json

This likely occurred because the benchmark script was run without the --dtype fp8_w8a8 flag, which should be used when tuning for FP8 models.

[samanamp]

Would you share the before and after tuning for the kernels and model e2e execution? AFAIK for H100 the defaults are usually performant.

[skyloevil]

Would you share the before and after tuning for the kernels and model e2e execution? AFAIK for H100 the defaults are usually performant.

ok，we'll share later.