[Unit Test] Add unit test for deep gemm #20090
Conversation
Signed-off-by: yewentao256 <zhyanwentao@126.com>
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run 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 either: Add 🚀 |
There was a problem hiding this comment.
Summary of Changes
Hello @yewentao256, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request introduces a new unit test to validate the deep_gemm FP8 kernels within the vLLM framework. The test ensures numerical correctness by comparing deep_gemm's output against an existing Triton-based fallback, covering various configurations to enhance the reliability of the Mixture of Experts (MoE) implementation.
Highlights
- New Unit Test: Added a new unit test file,
tests/kernels/moe/test_deepgemm.py, specifically for thedeep_gemmFP8 kernels. - Numerical Validation: The test validates the numerical accuracy of the
deep_gemmimplementation by comparing its output against the existing Triton fallback withinvLLM'sfused_expertsfunction. - Comprehensive Coverage: The test suite is parameterized to cover various matrix dimensions (M, N, K),
topkvalues, and number of experts, ensuring robust validation across different configurations.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request adds a unit test for the deepgemm kernel, comparing its output against a Triton fallback. The changes include creating FP8 block-quantized weights, running the computation with both DeepGEMM and Triton, and asserting that the results are within tolerance. The review suggests adding more descriptive messages to the assertions and ensuring that the input tensors are on the same device.
Update through gemini's suggestion Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Signed-off-by: yewentao256 <zhyanwentao@126.com>
Update through gemini's suggestion Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Signed-off-by: yewentao256 <zhyanwentao@126.com>
ce2260c to
d2f4905
Compare
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
| w2_scale=w2_s, | ||
| a1_scale=a1_scale, | ||
| block_shape=block_size, | ||
| allow_deep_gemm=True, |
There was a problem hiding this comment.
We don't have a way to prove that DeepGEMM was used, right? For instance I see this case where we won't be using DG for the test cases here where N=512
vllm/vllm/model_executor/layers/fused_moe/fused_moe.py
Lines 1163 to 1167 in c6c9830
There was a problem hiding this comment.
Sounds good, use monkeypatch to make sure deepgemm is really called now.
tests/kernels/moe/test_deepgemm.py
Outdated
| # ----- Compare ----- | ||
| rel_diff = (torch.mean( | ||
| torch.abs( | ||
| out_deepgemm.to(torch.float32) - out_triton.to(torch.float32))) / | ||
| torch.mean(torch.abs(out_triton.to(torch.float32)))) | ||
|
|
||
| assert rel_diff < 0.005, \ | ||
| f'Relative error: {rel_diff:.5f} (m={m}, k={k}, n={n})' | ||
|
|
||
| diff = calc_diff(out_deepgemm, out_triton) | ||
| assert diff < 0.001, f'Dice error: {diff:.5f} (m={m}, k={k}, n={n})' |
There was a problem hiding this comment.
Is there a reason not to use torch.testing.assert_close as we generally do?
There was a problem hiding this comment.
atol originally may cause some bad case, so I didn't use it first.
> torch.testing.assert_close(
out_deepgemm.to(torch.float32),
out_triton.to(torch.float32),
rtol=0.06,
atol=0.25,
)
E AssertionError: Tensor-likes are not close!
E
E Mismatched elements: 15 / 1048576 (0.0%)
E Greatest absolute difference: 0.5 at index (129, 241) (up to 0.25 allowed)
E Greatest relative difference: inf at index (175, 1594) (up to 0.06 allowed)But I think you are right, since torch.testing.assert_close is more welcome, so I use the dynamic atol to pass
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Signed-off-by: avigny <47987522+avigny@users.noreply.github.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Purpose
We have unit test for deepep, deepep+deepgemm, but we are lacking the unit test for deepgemm only, this pr adds the unit test for that.
Test