[Model] Add Support for GPTQ Fused MOE

[izhuhaoran]

This PR adds support for GPTQ Fused Mixture-of-Experts (MoE) ，and test passed for Qwen2-57B-A14B-Instruct-GPTQ-Int4 and Qwen1.5-MoE-A2.7B-Chat-GPTQ-Int4 on A100/A800 GPUs

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[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 trigger fastcheck CI to run, which consists only a small and essential subset of tests to quickly catch errors with the flexibility to run extra individual tests on top (you can do this by unblocking test steps in the Buildkite run).

Full CI run is still required to merge this PR so once the PR is ready to go, please make sure to run it. If you need all test signals in between PR commits, you can trigger full CI as well.

To run full CI, you can do one of these:

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

🚀

[izhuhaoran]

@robertgshaw2-neuralmagic Sorry to bother you, do you have any suggestions/comments for this pr which will guide me to modify it further!

[qibaoyuan]

Any problem on a800-80g, please refer to this PR.
BTW, you should upgrade your Triton to 3.0.0 to support interleave op.

[superawind]

@izhuhaoran Hi, When I reason about the qwen-moe-gptq-int4 model, it always prompts triton.runtime.errors.OutOfResources: out of resource: shared memory, Error, how to solve it

[sighingnow]

@izhuhaoran Hi, When I reason about the qwen-moe-gptq-int4 model, it always prompts triton.runtime.errors.OutOfResources: out of resource: shared memory, Error, how to solve it

We calibrate the triton kernel configs for A100 and H800 and for qwen2_moe model using https://github.com/vllm-project/vllm/blob/main/benchmarks/kernels/benchmark_moe.py.

To avoid out of shared memory error, you could try calibrate your own config.

[superawind]

@izhuhaoran Hi, When I reason about the qwen-moe-gptq-int4 model, it always prompts triton.runtime.errors.OutOfResources: out of resource: shared memory, Error, how to solve it

We calibrate the triton kernel configs for A100 and H800 and for qwen2_moe model using https://github.com/vllm-project/vllm/blob/main/benchmarks/kernels/benchmark_moe.py.

To avoid out of shared memory error, you could try calibrate your own config.

Thank you. I'll try

[jeejeelee]

Is this PR currently in a state of moving forward? @robertgshaw2-neuralmagic @izhuhaoran

[izhuhaoran]

Is this PR currently in a state of moving forward? @robertgshaw2-neuralmagic @izhuhaoran

Yes, still progressing. BTW, I just updated the kernel implementation, looking forward to any comments from reviewers!

[jeejeelee]

Is this PR currently in a state of moving forward? @robertgshaw2-neuralmagic @izhuhaoran

Yes, still progressing. BTW, I just updated the kernel implementation, looking forward to any comments from reviewers!

Thank you very much, I'm glad to see this PR is still moving forward.

BTW, @mgoin @robertgshaw2-neuralmagic I'd like to know your considerations on this PR. There have indeed been some similar efforts recently.

[donpromax]

@izhuhaoran Hi, When I reason about the qwen-moe-gptq-int4 model, it always prompts triton.runtime.errors.OutOfResources: out of resource: shared memory, Error, how to solve it

We calibrate the triton kernel configs for A100 and H800 and for qwen2_moe model using https://github.com/vllm-project/vllm/blob/main/benchmarks/kernels/benchmark_moe.py.

To avoid out of shared memory error, you could try calibrate your own config.

I noticed that the script located at /benchmarks/kernels/benchmark_moe.py doesn't currently support qwen-moe-gptq-int4 model. I would greatly appreciate it if you could share how you use this script to generate configuration files, such as E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB,dtype=a16w4.json. Thank you for your assistance!

[donpromax]

@izhuhaoran

Bug Report

I would like to bring to your attention a potential bug in the code. It seems that the line const_kwargs.update(config) should be placed before const_kwargs.update({"BLOCK_SIZE_K": k_per_scale}).

I observed that my inference results were incorrect, and upon further investigation, I found that when the config file is not located in fused_moe/configs and BLOCK_SIZE_K is set to 128 (group size), it may lead to incorrect results in the kernel computation. This issue arises because BLOCK_SIZE_K is being overridden by the config file or the default value.

[mgoin]

@izhuhaoran @jeejeelee Sorry for the lack of review! There is a PR porting the Marlin kernel to grouped gemm (#7527) so we should have GPTQ Fused MoE support for most models with this. I think we could consider this kernel separately (especially if it supports more quantization configs), but I would like to see some performance comparison

[jeejeelee]

@izhuhaoran @mgoin We have conducted some performance comparisons. Among them:

• moe_triton denotes this PR, and the model used is Qwen2-57B-A14B-Instruct-GPTQ-Int4

• moe_marlin denotes [Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel #7766 , and this quantized model is created by ourselves using llm_compressor

[sighingnow]

Thanks a lot for your feedback @jeejeelee !

[hmellor]

Since the existing Marlin kernels are consistently faster, I'm going to close this.