bugfix: fix fp32 acc threshold for qk using math::inf according to dtype by AIDC-AI

[yongchaoding]

📌 Description

In some models, the acc fp32 of QK will be small than math::inf which leading to bad output

🔍 Related Issues

🚀 Pull Request Checklist

Thank you for contributing to FlashInfer! Before we review your pull request, please make sure the following items are complete.

✅ Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit (or used your preferred method).
• I have installed the hooks with pre-commit install.
• I have run the hooks manually with pre-commit run --all-files and fixed any reported issues.

If you are unsure about how to set up pre-commit, see the pre-commit documentation.

🧪 Tests

• Tests have been added or updated as needed.
• All tests are passing (unittest, etc.).

Reviewer Notes

[gemini-code-assist]

Summary of Changes

Hello @yongchaoding, 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 addresses a subtle but critical bug in the attention mechanism where the representation of negative infinity for FP32 accumulation thresholds was not precise enough. By introducing a type-aware neg_inf utility, the system now correctly initializes and compares against the true negative infinity for floating-point types, resolving potential issues with attention score calculations in certain models, especially when dealing with very small (large negative) values.

Highlights

• Improved Negative Infinity Handling: I've introduced a new templated math::neg_inf<DT>() function. This function now returns std::numeric_limits<DT>::lowest() for float types, ensuring the use of the true negative infinity, and continues to use a predefined large negative constant (-5e4) for other floating-point types like half or bfloat16 where a true lowest() might not be directly applicable or necessary due to their limited range. This change is crucial for accurately representing the lowest possible value in attention score calculations.
• Widespread Application of Corrected Infinity: I've updated numerous attention kernels and related structs across cascade.cuh, decode.cuh, decode_mla_cute_sm80.cuh, hopper/attention_updater.cuh, hopper/epilogue.cuh, hopper/quantization/epilogue.cuh, mla.cuh, mla_hopper.cuh, prefill.cuh, and state.cuh. These modifications ensure that initializations of attention score accumulators (m), mask fill values, and comparisons against negative infinity now consistently use the newly defined type-aware neg_inf or std::numeric_limits::lowest(), preventing potential numerical inaccuracies.
• New Test Case for Validation: A new test case, test_sinqle_prefill_with_paged_kv_cache_neginf, has been added to tests/test_single_prefill.py. This test specifically targets scenarios where query values are large negative numbers, using float16 data types, to verify that the corrected negative infinity handling prevents erroneous outputs in such edge cases, thereby validating the bug fix.

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 configure Gemini 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 replaces the hardcoded -math::inf with math::neg_inf<float>() to address a bug where fp32 accumulation threshold leads to bad output. The changes are applied consistently across the codebase, and a new test case is added to validate the fix.

[gemini-code-assist]

Multiplying a tensor of ones by a scalar is not the most efficient way to create a tensor with a specific value. Consider using torch.full for better performance.

q = torch.full(
        (qo_len, num_qo_heads, head_dim),
        -200,
        device="cuda:0",
        dtype=torch.float16,
    )

[gemini-code-assist]

The neg_inf function provides a type-aware negative infinity. It would be beneficial to add a comment explaining why the inf constant is used for non-float types, and what the implications are for bf16/fp16.

[yongchaoding]

@yzh119 Hi, this is a critical accuracy issue that will trigger related issues in some models after SFT. We want to merge this bugfix into the main branch as soon as possible. If you have any questions, please comment so that I can actively adjust it.