Skip to content

[Performance]: supports of fused moe kernel implementation #20176

Open
@oldcpple

Description

@oldcpple

Proposal to improve performance

By reading relative parts of source code and running some test, we find that when launching a MoE model like Qwen3, vLLM seems to use Triton-based fused moe kernel. While other implementations like cutlass or deep gemm is only supported by specific GPU arch like Hopper, or specific quantization method like Compressed Tensor.
Is there a way to specify a type of fused moe kernel to use? For example I might want to compare the performance of Triten-based and Cutlass-based implementation on my A100 GPUs.

Report of performance regression

No response

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

The output of `python collect_env.py`

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.

Metadata

Metadata

Assignees

No one assigned

    Labels

    performancePerformance-related issues

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions