feat: update cubins for dynamic scale factor interface and one-copy kvcache on trtllm-gen mla

[yyihuang]

📌 Description

Update the static scaling factor interface to take:
q_scale, k_scale, v_scale, sm_scale, o_scale.

🔍 Related Issues

TODOs in #1222, last comment.

🚀 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 @yyihuang, 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 more explicit and flexible interface for handling scaling factors within the trtllm-gen mla kernels. By replacing a single scale parameter with a set of five specific scales (for query, key, value, softmax, and output), the changes enable finer-grained control over quantization, particularly in preparation for future FP8 quantization support. The modifications span both the Python API and the underlying C++ CUDA kernels, ensuring consistent behavior and improved extensibility for advanced quantization schemes.

Highlights

• API Interface Update: The trtllm_batch_decode_with_kv_cache_mla function in flashinfer/decode.py and its corresponding C++ kernel launcher (trtllm_paged_attention_mla_launcher in csrc/trtllm_mla_kernel_launcher.cu) have been refactored to replace a single generic scale parameter with five distinct, more granular scaling factors: q_scale, k_scale, v_scale, sm_scale, and o_scale.
• Dynamic Scale Factor Derivation: The bmm1_scale and bmm2_scale values, which are consumed by the C++ kernel, are now dynamically computed within the Python wrapper (flashinfer/decode.py) based on the newly introduced granular input scales. Specifically, bmm1_scale is derived from q_scale, k_scale, and sm_scale (and the head dimension), while bmm2_scale is derived from v_scale and o_scale.
• Internal Kernel Logic Adjustment: In the C++ kernel (csrc/trtllm_mla_kernel_launcher.cu), the runner_params.mScaleQ (previously controlled by the generic scale parameter) is now hardcoded to 1.0. The runner_params.scaleSoftmaxLog2 calculation has been updated to bmm1_scale * std::sqrt((float)(runner_params.mHeadDimQk)), aligning with the new scale derivation logic from the Python frontend.
• Test Suite Adaptation: The test cases in tests/test_trtllm_gen_decode.py have been updated to reflect the new function signature and parameter passing for trtllm_batch_decode_with_kv_cache_mla, ensuring the correctness and compatibility of the changes.

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 refactors the scaling factor interface for the TRT-LLM MLA kernels, moving from a single scale parameter to a more detailed set of quantization and model scales (q_scale, k_scale, v_scale, sm_scale, o_scale). This improves clarity and flexibility. The changes are mostly in the C++ kernel launcher and the Python interface.

I've found two high-severity issues that should be addressed:

1. In csrc/trtllm_mla_kernel_launcher.cu, there's a bug in the calculation of the softmax scaling factor, which uses an incorrect head dimension.
2. In flashinfer/decode.py, there's an incorrect shape and dtype check for the pre-allocated output tensor, which would cause runtime errors.

[yyihuang]

Close to skip rebasing on last refactor. Moved to #1342.