Jim/dev/mi400 mha bwd

[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.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3281 --add-label <label>

[Copilot]

Pull request overview

This PR adds FMHA v3 backward (MHA bwd) ASM-kernel support for the gfx1250 GPU architecture by introducing new gfx1250 kernel/config entries and updating the runtime dispatch logic to recognize and launch them.

Changes:

• Add gfx1250 FMHA v3 bwd config CSVs for ODO (pre), dQ/dK/dV, and dQ convert (post) kernels.
• Extend fmha_v3_bwd dispatch to allow gfx1250, select the appropriate post-processing config, and always enable post-processing on gfx1250.
• Adjust workgroup size (bdx) to 128 on gfx1250 for pre/dqdkdv/post kernel launches.

Reviewed changes

Copilot reviewed 4 out of 8 changed files in this pull request and generated 2 comments.

File	Description
hsa/gfx1250/fmha_v3_bwd/fmha_bwd_odo.csv	Adds gfx1250 ODO (pre-processing) kernel configuration entry.
hsa/gfx1250/fmha_v3_bwd/fmha_bwd_dqdkdv.csv	Adds gfx1250 dQ/dK/dV kernel configuration entries (including causal BR).
hsa/gfx1250/fmha_v3_bwd/fmha_bwd_dq_convert.csv	Adds gfx1250 dQ convert (post-processing) kernel configuration entry.
csrc/cpp_itfs/mha_bwd.cu	Enables gfx1250 in asm v3 bwd dispatch, sets post-processing behavior, and tweaks launch block size.

---

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