Mla splitkv enhance split alg inte

[valarLip]

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull Request Overview

This PR enhances the MLA (Multi-head Latent Attention) split key-value algorithm with significant new functionality including persistent thread group support, sparse attention capabilities, and fp8 quantization support. The changes introduce metadata generation for optimized work distribution and a reduce kernel for merging partial results.

• Adds persistent thread group implementation for variable query/output lengths
• Implements sparse attention with top-k token selection
• Integrates fp8 quantization support for both Q and KV tensors

Reviewed Changes

Copilot reviewed 52 out of 85 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
op_tests/test_mla_sparse.py	New test file for sparse MLA attention with top-k token selection
op_tests/test_mla_persistent*.py	New test files for persistent thread group MLA implementation
csrc/py_itfs_cu/asm_mla.cu	Updated to support persistent mode, fp8 datatypes, and new metadata parameters
csrc/kernels/mla/reduce.cu	New reduce kernel for merging partial attention outputs
csrc/kernels/mla/metadata*.cuh	New metadata generation kernels for work distribution
aiter/mla.py	Updated decode forward pass to use new metadata and reduce operations
csrc/include/mla.h	New header defining MLA data structures and function signatures
Various copyright headers	Updated copyright format from (c) to (C)

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}