Add gfx950 mla a8w8 qh32 kernel

[slippedJim]

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 adds support for the gfx950 MLA (Multi-Head Latent Attention) a8w8 (8-bit activation, 8-bit weight) qh32 (32 query heads) kernel. The changes enable a new kernel configuration for fp8 data types with 32 heads in persistent mode with decode sequence length of 4.

Changes:

• Added new kernel entry to the assembly CSV configuration for fp8/fp8 with 32 heads
• Updated metadata generation logic to support 32-head configurations alongside existing 16 and 128-head support
• Modified test harness to disable causal masking and reduce iteration count for testing
• Added debug logging to trace tensor information and kernel arguments

Reviewed changes

Copilot reviewed 8 out of 9 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
hsa/gfx950/mla/mla_asm.csv	Registers new fp8/fp8 kernel with 32-head configuration
csrc/py_itfs_cu/asm_mla.cu	Adds configuration branch for fp8/fp8 32-head decode with qlen=4 and enables debug logging
csrc/kernels/mla/metadata/v1_comm.cuh	Introduces NUM_HEADS dispatcher macro for compile-time head count specialization
csrc/kernels/mla/metadata/v1_2_device.cuh	Extends metadata generation to support 32-head fp8 configurations
aiter/ops/attention.py	Updates metadata calculation to include 32-head case for fp8 data types
aiter/mla.py	Extends native support check and adds tensor debugging utilities
op_tests/test_mla_persistent.py	Disables causal masking for testing new kernel configuration
aiter/test_common.py	Reduces performance test iterations for faster debugging

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.