[menglcai]

Motivation

Split out from #4306, with none of the CK backward/PREBUILD changes from that PR.

cmdGenFunc_mha_bwd and cmdGenFunc_mha_varlen_bwd return "flags_extra_cc": ["'-DONLY_FAV3=0'"]. That value reaches the compiler through extra_cflags, which is passed as an argv list, so the quotes are part of the flag rather than shell syntax and the compiler reads ''-DONLY_FAV3=0'' as a filename.

Linux never sees this, because ninja runs each command through /bin/sh -c and the shell strips the quotes. Windows uses CreateProcess with no shell, so they survive and the backward module fails to compile:

clang: error: no such file or directory: ''-DONLY_FAV3=0''

Technical Details

The flag is written into build.ninja the same way on both platforms; only the execution differs, which is why the bug is invisible on Linux.

The change drops the quotes at the two sites that carry this flag:

-        "flags_extra_cc": ["'-DONLY_FAV3=0'"],
+        "flags_extra_cc": ["-DONLY_FAV3=0"],

ONLY_FAV3 only appears on the two backward generators, so nothing else is affected. On Linux the change is a no-op, since the shell was already removing the quotes.

Test Plan

Four machines, covering both operating systems and two GPU architectures:

host	OS	GPU	ROCm / torch
AMD Ryzen™ AI Max+ 395	Ubuntu 24.04	gfx1151	torch 2.13.0+ROCm 7.15
AMD Ryzen™ AI Max+ 395	Windows 11 25H2	gfx1151	torch 2.14.0+ROCm 10.2.0
AMD Radeon™ AI PRO R9700	Windows 11 25H2	gfx1201	torch 2.14.0+ROCm 10.2.0
AMD Radeon™ AI PRO R9700	Ubuntu 24.04	gfx1201	torch 2.14.0+ROCm 10.1

Three checks:

1. Compiler behaviour, no shell involved. Calls the HIP compiler directly so the quoting can be judged on its own, independently of platform:

import subprocess, shutil, tempfile, os
cc = shutil.which("hipcc") or shutil.which("clang++")
d = tempfile.mkdtemp(); src = os.path.join(d, "t.cpp")
open(src, "w").write("int main(){return ONLY_FAV3;}\n")
for flag in ["-DONLY_FAV3=0", "'-DONLY_FAV3=0'"]:
    r = subprocess.run([cc, flag, "-c", src, "-o", d + "/t.o"], capture_output=True, text=True)
    print(repr(flag), "rc =", r.returncode, r.stderr.strip()[:80])

2. mha_bwd JIT build, before and after, on Windows and Linux. The op is called directly rather than through test_mha.py. On a non-gfx9 host that test stops at the forward step with invalid argument for fmha_fwd, because ck-tile's generate.py defaults to --targets gfx9,gfx950 and emits no kernel the runtime arch can dispatch, so it never reaches the backward path this change affects.

3. Linux regression check. Full test_mha.py run with and without the change, to confirm it is a no-op there.

Test Result

1. Compiler behaviour — identical on all four hosts, regardless of OS, GPU or ROCm version:

'-DONLY_FAV3=0'    rc = 0
"'-DONLY_FAV3=0'"  rc = 1  clang++: error: no such file or directory: ''-DONLY_FAV3=0''

2. mha_bwd JIT build:

host	before	after
AMD Ryzen™ AI Max+ 395 (Windows, gfx1151)	fails, all 166 translation units	succeeds
AMD Ryzen™ AI Max+ 395 (Ubuntu, gfx1151)	succeeds	succeeds
AMD Radeon™ AI PRO R9700 (Windows, gfx1201)	fails, 25 translation units report the error	succeeds
AMD Radeon™ AI PRO R9700 (Ubuntu, gfx1201)	succeeds	succeeds

The split follows the operating system, not the GPU: both Windows hosts fail and both Linux hosts build either way, which is what the shell difference predicts.

3. Linux regression check: test_mha.py gives identical numerics either way (Output and dQ max diff 0.0009765625 in both).

Scope

This is a build-correctness fix and does not enable anything on its own: backward on a non-gfx9 host still fails to dispatch afterwards, for the same CK target reason noted above. That is out of scope here per the decision on #4306.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.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:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs (added automatically when gfx942 configs change); main branch always runs both MI35X and MI300X
ci:triton-355	Run the full Triton test suite on MI35X, not only the tests the change affects
ci:windows	Windows JIT op smoke test on the gfx1201 ROCm runner
ci:multi-gpu	Aiter multi-GPU tests on the 8-GPU runner
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: DeepSeek-V4-Pro, MiniMax-M3-MXFP4
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 6616 --add-label <label>

One backend per PR:
A PR changes one kernel backend: [Triton/Gluon] (Triton and Gluon count as one), [HIP], [ASM], [CK], [OPUS] or [FlyDSL]. If the title ends up with two backend tags, split the PR -- as stacked pull requests when one part cannot merge without the other.

PR title tags & labels:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title and as PR labels automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf], op tags like [MLA], and human labels (ci:*) are left untouched. Add the no-auto-title label to stop the title rewrites; labels stay in sync either way.

[zufayu]

The quoting fix is correctly isolated from the Windows CK support work, and the full CI is green. LGTM.