[ozturkosu]

Summary

Extends the amd-smi-first migration (#9731, #10204) from the Tile Engine to the
dispatcher's GPU-architecture detection.

• Adds detect_gpu_arch() to the shared tile_engine/ops/common/smi_utils.py
  wrapper: amd-smi static (TARGET_GRAPHICS_VERSION) first, rocm-smi
  --showproductname (GFX Version) fallback — same amd-smi-first policy as the
  rest of the wrapper.
• dispatcher_common.detect_gpu_arch(): prefers amd-smi via the wrapper, then
  rocminfo, then the caller-supplied default. Also removes a duplicate
  definition of this function.
• gemm_utils._get_arch(): tries amd-smi via the wrapper before rocminfo,
  preserving the raise-on-unresolved (no silent default) behavior.

Motivation

Per the directive behind #9731, device-property access should prefer
amd-smi (rocm-smi is being phased out). The Tile Engine was already migrated
(#9731, #10204); the dispatcher still detected the GPU arch by shelling out to
rocminfo directly. Tracked in JIRA ROCM-28734.

Design note

The dispatcher reaches the sibling tile_engine/ops/common/smi_utils.py through
a guarded import and falls back to rocminfo if the wrapper is unavailable, so
behavior is unchanged on hosts without amd-smi. (Open to moving the shared
wrapper to a neutral location if reviewers prefer avoiding the dispatcher →
tile_engine reach.)

Test plan (MI350X / gfx950)

• New mock-based unit tests for detect_gpu_arch() (amd-smi preference,
  rocm-smi fallback, default) + a live check — 4/4 pass.
• dispatcher_common.detect_gpu_arch() and gemm_utils._get_arch() both
  return gfx950 via amd-smi on the live host.
• test_gemm_utils 21/21 still pass (no arch-detection regression).

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR 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.

🙋 Wish to Override Policy?

[therock-pr-bot]

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

[Copilot]

Pull request overview

This PR extends the existing “amd-smi first, rocm-smi fallback” policy into GPU-architecture detection for the CK Tile Engine and the dispatcher, centralizing arch detection in smi_utils.detect_gpu_arch() and updating dispatcher-side detection to prefer that wrapper before falling back to rocminfo.

Changes:

• Added detect_gpu_arch() to tile_engine/ops/common/smi_utils.py and corresponding unit/live tests.
• Updated dispatcher_common.detect_gpu_arch() to prefer the shared wrapper (and removed a duplicate function definition).
• Updated gemm_utils._get_arch() to try the shared wrapper before rocminfo, while preserving “raise if unresolved” behavior.

Reviewed changes

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

File	Description
projects/composablekernel/tile_engine/ops/common/test_smi_utils.py	Adds unit/live tests for detect_gpu_arch() behavior and preference ordering.
projects/composablekernel/tile_engine/ops/common/smi_utils.py	Introduces detect_gpu_arch() with amd-smi-first / rocm-smi fallback.
projects/composablekernel/dispatcher/python/gemm_utils.py	Tries wrapper-based arch detection before rocminfo in _get_arch().
projects/composablekernel/dispatcher/python/dispatcher_common.py	Prefers wrapper-based arch detection and removes duplicate detect_gpu_arch().

[ozturkosu]

Review self-assessment + follow-up changes (commit 8085f847f3)

I re-reviewed this PR one more than and discovered : it only migrated dispatcher_common.detect_gpu_arch() and gemm_utils._get_arch(), while the busiest entry point — ctypes_utils.detect_gpu_arch(), which every GEMM example calls via default=detect_gpu_arch() — plus grouped_conv_utils were still detecting arch through rocminfo. So the title's claim ("dispatcher GPU-arch detection") wasn't true end-to-end. This commit closes that gap and addresses all Copilot feedback.

What changed

• Completed the migration (the necessity fix). ctypes_utils.detect_gpu_arch() and grouped_conv_utils.detect_gpu_arch() now prefer amd-smi via the shared smi_utils wrapper before falling back to rocminfo, then the default. All Python arch-detection paths in the dispatcher are now amd-smi-first.
• DRY (single source of truth). The amd-smi bridge now lives once in dispatcher_common._detect_gpu_arch_via_amd_smi(); gemm_utils, ctypes_utils, and grouped_conv_utils import it instead of re-implementing the sys.path reach. (Previously the helper was copy-pasted in two files — this PR was removing one duplicate while adding another.)
• Copilot Initial workflows #1 — TestDetectGpuArch is now deterministic: forces CK_SMI_TOOL="" in setUp/tearDown and mocks shutil.which, plus a new tools-absent case.
• Copilot [hipDNN] Enablement #2 — smi_utils.detect_gpu_arch() gains Optional[str] annotations and a shutil.which presence check consistent with the rest of the module.
• Copilot Change shared libs name #3/[PR] [hipSPARSELt] Test monorepo #4 — best-effort broad excepts annotated # noqa: BLE001 with intent comments (or removed via the DRY consolidation).

Verification

• py_compile on all six edited files — OK.
• TestDetectGpuArch — 5/5 pass (mock-based, no GPU needed).
• Live host (gfx90a): all four helpers (dispatcher_common, gemm_utils._get_arch, ctypes_utils, grouped_conv_utils) resolve via amd-smi; gemm_utils reuses the canonical bridge (identity check).

Deliberately left out of scope

• Neutral-location move of smi_utils. The dispatcher still reaches the sibling tile_engine/ops/common/smi_utils via a guarded sys.path import; this is now isolated to the single dispatcher_common helper. Happy to move the wrapper to a shared/neutral module if reviewers prefer avoiding the dispatcher → tile_engine reach — flagging rather than doing it unilaterally.
• fmha_utils ImportError fallback (only used when dispatcher_common isn't importable — a genuine standalone degrade path) and the ck-rocprof bash helper remain on rocminfo; these aren't Python dispatcher detection paths.

Note on motivation: the code being migrated used rocminfo (not rocm-smi), which isn't itself deprecated — so the honest driver here is one consistent amd-smi-first detection policy across TE + dispatcher (per #9731/#10204, JIRA ROCM-28734), plus the duplicate-definition cleanup, rather than rocminfo deprecation.

[ozturkosu]

Wanted to leave a note on why this change is still worth landing on top of #9731 and #10204, since at a glance it can look like more of the same amd-smi migration.

Those two PRs both live entirely in tile_engine/ and both deal with GPU id discovery (detect_gpu_ids, count_gpus, and friends). Neither of them touched the dispatcher, and neither added any way to detect the GPU architecture. On current develop, smi_utils.py has no detect_gpu_arch() at all; the only gfx material in there is the internal parse_amd_gfx / parse_rocm_gfx parsers used by the equivalence tests, which nothing external can call.

Meanwhile every arch-detection entry point in the dispatcher is still shelling out to rocminfo directly:

• dispatcher_common.detect_gpu_arch()
• ctypes_utils.detect_gpu_arch() (the busy one, hit by every GEMM example)
• gemm_utils._get_arch()
• grouped_conv_utils.detect_gpu_arch()

So this PR is doing work the earlier two never covered: it adds the arch wrapper to the shared module and routes all four dispatcher paths through it, so device-property access is amd-smi first everywhere in CK instead of the dispatcher being the one island still on rocminfo. That is the consistency the tech lead asked for under ROCM-28734.

Two other things worth calling out. First, dispatcher_common.py on develop actually defines detect_gpu_arch() twice, and the second definition silently shadows the first. This PR removes the duplicate, so it is a real correctness fix independent of the SMI story. Second, the four hand rolled detectors now share a single _detect_gpu_arch_via_amd_smi bridge rather than each re-implementing the same reach, which is a nice cleanup on its own.

One honest caveat: the "rocm-smi is going away" reasoning from #9731 does not directly apply here, since the dispatcher was on rocminfo, which is a different tool and not on the same deprecation path. rocminfo stays as a fallback, so nothing breaks on hosts without amd-smi. The value here is the end to end consistency plus the duplicate definition fix, not avoiding a future breakage.

Happy to move the shared wrapper to a neutral location if reviewers would rather not have the dispatcher reach into tile_engine/ops/common. That layering is the one open question, and it is easy to address before merge.

[geyyer]

LGTM!