Wrap cudnn backend to unified interface

[cyx-6]

📌 Description

Add cudnn backend to BatchPrefillWithPagedKVCacheWrapper and BatchDecodeWithPagedKVCacheWrapper.

🔍 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 @cyx-6, 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 integrates the cuDNN backend into FlashInfer's unified attention wrappers for both prefill and decode operations. This expands the available high-performance backend options, allowing users to leverage cuDNN for optimized attention computations within the existing API framework.

Highlights

• Unified Interface Extension: Extended BatchPrefillWithPagedKVCacheWrapper and BatchDecodeWithPagedKVCacheWrapper to support the cuDNN backend, providing a unified interface for different attention implementations.
• cuDNN Backend Integration: Introduced new CudnnPrefillModule and CudnnDecodeModule classes to encapsulate cuDNN-specific logic for prefill and decode operations, respectively.
• Backend Abstraction Refinement: Refactored the module retrieval mechanism (get_trtllm_gen_decode_module renamed to get_decode_module) to dynamically load the appropriate backend module based on configuration.
• Test Coverage Expansion: Updated existing test suites (test_cudnn_decode.py, test_cudnn_prefill.py) to validate the functionality and correctness of the cuDNN backend through the newly integrated unified wrappers.

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 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 or fill out our survey 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 introduces a cudnn backend to the BatchPrefillWithPagedKVCacheWrapper and BatchDecodeWithPagedKVCacheWrapper, unifying the interface to support multiple backends. The changes are extensive, touching core logic, JIT compilation files, and tests.

Overall, the changes are well-structured and move towards a more extensible design. However, I've identified a critical bug in flashinfer/sparse.py where a function is called with incorrect arguments, which will lead to a runtime error. I've also found a potential UnboundLocalError in flashinfer/decode.py due to the use of consecutive if statements instead of if/elif for backend dispatching. Additionally, there are some maintainability concerns, such as hardcoded values in the new CudnnPrefillModule, which could limit its applicability.

Please address the critical and high-severity issues. The medium-severity issues are suggestions for improving code quality and maintainability.

[gemini-code-assist]

The function get_batch_decode_module is being called with self._backend as the first argument. However, its definition in flashinfer/decode.py does not accept a backend string; it expects dtype_q as the first argument. This will cause a runtime error. Please remove self._backend from this call.

            self._cached_module = get_batch_decode_module(
                q_data_type,
                kv_data_type,
                self._o_dtype,

[gemini-code-assist]

Using consecutive if statements here for backend dispatching could lead to an UnboundLocalError if a backend other than "trtllm-gen" or "cudnn" is passed, as module would not be initialized. Using elif would make the logic more robust and prevent this potential issue.

Suggested change

	if backend == "cudnn":
	module = CudnnDecodeModule()
	elif backend == "cudnn":
	module = CudnnDecodeModule()

[gemini-code-assist]

The logic for block_tables is hardcoded for specific head dimensions (d_qk of 192 and 128). This limits the general applicability of the cudnn backend. Consider making this logic more generic or adding a comment to document this limitation.