Repository navigation
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/25017
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 Cancelled Job, 1 Unrelated FailureAs of commit c73d218 with merge base f53710a ( CANCELLED JOB - The following job was cancelled. Please retry:
BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
|
@Gasoonjia has exported this pull request. If you are a Meta employee, you can view the originating Diff in D131591208. |
This PR needs a
|
Summary: Inductor offers its grouped GEMM Triton template only on sm_90 and above, and this backend autotunes with Triton only, so on an A100 (sm_80) a model using `aten._grouped_mm` fails to export with `NoValidChoicesError: No choices to select` (pytorch#24902). The template uses TMA only on sm_90 and above; its remaining path works on sm_80. During AOTInductor compilation, a `_CompilePatch` now reports `has_grouped_mm_triton_support()` as true when the current device is sm_80 or newer, following the existing small-GPU override for Inductor's 68-SM floor. The override is scoped to the threads inside the compile, is skipped on ROCm, and leaves devices below sm_80 refused. Reviewed By: shoumikhin Differential Revision: D131591208
meta-codesync
Bot
changed the title
5a12cb7 to
c73d218
Compare
Summary:
Inductor offers its grouped GEMM Triton template only on sm_90 and
above, and this backend autotunes with Triton only, so on an A100
(sm_80) a model using
aten._grouped_mmfails to export withNoValidChoicesError: No choices to select(#24902).The template uses TMA only on sm_90 and above; its remaining path
works on sm_80. During AOTInductor compilation, a
_CompilePatchnow reports
has_grouped_mm_triton_support()as true when thecurrent device is sm_80 or newer, following the existing small-GPU
override for Inductor's 68-SM floor. The override is scoped to the
threads inside the compile, is skipped on ROCm, and leaves devices
below sm_80 refused.
Reviewed By: shoumikhin
Differential Revision: D131591208