[MI300X]mha fwd v3 gfx942 support hdim192x128

[shay-li77]

Motivation

mha fwd v3 support hdim192x128 for MI300/MI308

Technical Details

Added test coverage for head_dim=192 in smoke tests
Extended kernel dispatcher logic to handle the 192x128 dimension configuration
Updated dimension validation checks to allow hdim_q=192 while maintaining hdim_v=128

Test Plan

MI300/MI308:
python3 test_mha.py
python3 test_mha_varlen.py
bash smoke_test_fwd_v3.sh -a gfx942

Test Result

function test:

test pass

perf test:

ps: CK目前不支持通过API配置rounding mode，需要修改config文件，默认为RTZ。

MI308

causal=True, Rounding mode= RTZ

causal=True, Rounding mode= RTNA

MI300

causal=True, Rounding mode= RTZ

causal=True, Rounding mode= RTNA

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 head dimension 192x128 configuration in the MHA (Multi-Head Attention) forward v3 kernels for gfx942 architecture (MI300/MI308 GPUs).

• Extended kernel dispatcher to handle hdim_q=192 with hdim_v=128
• Added template specializations for all kernel variants (causal/non-causal, group/non-group modes, different rounding modes)
• Updated smoke tests to cover the new dimension configuration

Reviewed changes

Copilot reviewed 3 out of 27 changed files in this pull request and generated no comments.

File	Description
op_tests/cpp/mha/smoke_test_fwd_v3.sh	Added head_dim iteration loop and d_v parameter to test 192x128 configuration
hsa/gfx942/fmha_v3_fwd/codegen.py	Added kernel template specializations and dispatcher logic for hdim 192x128
aiter/ops/mha.py	Removed gfx950-specific restriction, allowing hdim_q=192 for gfx942

---

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