[Bugfix][1/n] Fix the speculative decoding test by setting the target dtype

[houseroad]

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.

Purpose

The tests are broken on trunk, and #18751 was the culprit.
So setting the precision back to float32 mitigate the issue.

Test Plan

pytest tests/spec_decode/e2e/test_multistep_correctness.py
CI

Test Result

============================================ 64 passed in 1955.55s (0:32:35) =============================================

(Optional) Documentation Update

N/A

[github-actions]

👋 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 fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

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 ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @houseroad, 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 aims to fix broken speculative decoding end-to-end tests (test_multistep_correctness.py) that were failing on trunk. The issue was traced back to a previous PR (18751). The fix involves ensuring the tests use the original float32 data type for the model, which resolves numerical stability problems that were causing the failures.

Highlights

• Test Fix: Explicitly set the model dtype to float32 in several speculative decoding test configurations within test_multistep_correctness.py. This is intended to restore numerical stability and fix recent test failures.

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

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configureGemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request addresses a bug in speculative decoding tests by explicitly setting the dtype to float32 for numerical stability. The change is applied consistently across various test parameterizations.

My primary feedback focuses on improving maintainability by reducing the repetition of this dtype setting. I've suggested defining a shared configuration dictionary that can be reused across the different test cases. This will make future modifications to this common parameter easier and less error-prone.

Please also remember to update the 'Test Result' section in the PR description, as it's currently marked as 'TODO'.

[gemini-code-assist]

This dtype setting and its accompanying comment are repeated in several places throughout this file. To enhance maintainability and adhere to the DRY (Don't Repeat Yourself) principle, consider defining these common keyword arguments in a shared dictionary at the module level. This dictionary can then be spread into each specific common_llm_kwargs dictionary where this setting is needed.

For instance, you could define a constant like this at the module scope:

_COMMON_FLOAT32_KWARGS = {
    # The original model is float32, keep it for numerical stability.
    "dtype": "float32",
}

And then utilize it within the parametrize decorator as shown in the suggestion below. This approach centralizes the configuration, making future updates simpler and more consistent.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

Similar to the previous comment, this dtype configuration is repeated. Applying the shared dictionary pattern (e.g., using **_COMMON_FLOAT32_KWARGS) here would improve maintainability by centralizing this common setting.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

This is another instance of the repeated dtype configuration. Using a shared dictionary (e.g., **_COMMON_FLOAT32_KWARGS) would be beneficial for maintainability here as well.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

The dtype configuration is repeated here. Consider using the suggested shared dictionary pattern (e.g., **_COMMON_FLOAT32_KWARGS) to centralize this setting for better maintainability.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

This dtype setting is repeated. Applying the shared dictionary pattern (e.g., **_COMMON_FLOAT32_KWARGS) would improve code maintainability.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

Again, the dtype configuration is duplicated. Using a shared dictionary (e.g., **_COMMON_FLOAT32_KWARGS) would help in centralizing this common parameter.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

This dtype setting is repeated. Employing the shared dictionary pattern (e.g., **_COMMON_FLOAT32_KWARGS) would enhance maintainability.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

The dtype configuration appears again. Centralizing this with a shared dictionary (e.g., **_COMMON_FLOAT32_KWARGS) is recommended for better maintainability.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[gemini-code-assist]

This is the final instance of the repeated dtype configuration in this PR. Using a shared dictionary (e.g., **_COMMON_FLOAT32_KWARGS) would be beneficial for maintainability.

Suggested change

	# The original model is float32, keep it for numerical stability.
	"dtype": "float32",
	**_COMMON_FLOAT32_KWARGS,

[ywang96]

Thanks for the fix!

[DarkLight1337]

Thanks for taking the time to investigate this!

[zixi-qi]

LGTM!