fix(rocke): Remove silent GFX950 fallback in fusion lowering

[ecamartins]

Motivation

In rocKE's fusion_lowering.py, the GPU architecture silently falls back to gfx950, which can mask configuration or environment issues. Hence, this change removes the silent fallback and throws a descriptive error message instead.

Note: This gfx950 fallback exists throughout rocKE. A follow-up ticket will work to remove the remaining fallbacks.

Technical Details

In fusion_lowering.py, there were several cases of the following forms:

 dev = get_device_arch() or "gfx950"
 build_gfx = dev if dev in known_arches() else "gfx950"

In the above case, if get_device_arch() returns None (e.g., due to no GPU) or if the given architecture is not in the set of known architectures, then there is a silent fallback to gfx950. This can mask environment or configuration issues.

Instead of using the silent fallback, we throw a descriptive error message through the addition of the function validate_arch. This will inform the user of how to properly configure their environments.

Test Plan

Addition of the following tests:

• Unit tests in test_rocke_multiarch.py to test validate_arch
• Test cases in test_rocke.py's TestLoweringRegistryBuild to validate cases where validate_arch throws.
  • Despite documentation stating that test_rocke.py does not require a GPU, I discovered that tests that call a lowerer's build method require a GPU (i.e., the build method creates a KernelLauncher object whose constructor calls hipModuleLoad). So, I refactored the TestLoweringRegistryBuild class to mock the calls to the hip functions to remove the GPU dependency.

Test Result

All tests in test_rocke.py and test_rocke_multiarch.py pass.

Submission Checklist

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

JIRA ID : AICK-1541

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[yraparti]

Critical:

• platform/AGENTS.md states under Hard rules: "Default arch is gfx950 … for on-GPU runs, prefer the local device via get_device_arch() and fall back to gfx950." This PR removes that fallback. The change is intentional and a follow-up ticket is noted, but the guidance doc now contradicts the code. Consider updating AGENTS.md in this PR (or the follow-up) so the two don't drift.

Nits:

• validate_arch(arch: str) — the signature says str but None is a first-class, handled input. Type it Optional[str].
• Docstring wording in _build_region_hsaco: says the LLVM lowering "throws an error when … otherwise it throws an error." The error is actually raised by validate_arch before lowering, not by the lowering path. Also, Python convention is "raises" rather than "throws." Small accuracy fix.
• Error message interpolates the raw tuple {known_arches()}; ', '.join(known_arches()) reads a bit cleaner for a user-facing message. Optional.
• elif arch not in known_arches(): follows a raise, so it can just be if — trivial.

[ecamartins]

@yraparti

Critical:

• platform/AGENTS.md states under Hard rules: "Default arch is gfx950 … for on-GPU runs, prefer the local device via get_device_arch() and fall back to gfx950." This PR removes that fallback. The change is intentional and a follow-up ticket is noted, but the guidance doc now contradicts the code. Consider updating AGENTS.md in this PR (or the follow-up) so the two don't drift.

I've updated AGENTS.md to indicate that fusion_lowering.py does not fall back to gfx950.

Nits:

• validate_arch(arch: str) — the signature says str but None is a first-class, handled input. Type it Optional[str].
• Docstring wording in _build_region_hsaco: says the LLVM lowering "throws an error when … otherwise it throws an error." The error is actually raised by validate_arch before lowering, not by the lowering path. Also, Python convention is "raises" rather than "throws." Small accuracy fix.
• Error message interpolates the raw tuple {known_arches()}; ', '.join(known_arches()) reads a bit cleaner for a user-facing message. Optional.
• elif arch not in known_arches(): follows a raise, so it can just be if — trivial.

All required nits have been addressed.

[yraparti]

maybe we can remove this and just add a line saying pass target_architecture for non-gfx950 targets and beware of the silent fallbacks on non-gpus?

[ecamartins]

Sounds good. Done.