Skip to content

[Test] Port roi_align (MPS) to stable ABI. - #9574

Draft
adabeyta wants to merge 2 commits into
adabeyta/roi-align-cuda-stablefrom
adabeyta/roi-align-mps-stable
Draft

[Test] Port roi_align (MPS) to stable ABI. #9574
adabeyta wants to merge 2 commits into
adabeyta/roi-align-cuda-stablefrom
adabeyta/roi-align-mps-stable

Conversation

@adabeyta

@adabeyta adabeyta commented Jul 30, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Ports the roi_align MPS kernels to the stable ABI.

Notes

  • Requires [Stable ABI] Add torch_mps_set_arg_bytes to the MPS shim. pytorch#190932
  • Metal scalar args are version-dispatched. torch ≥ 2.14 binds them with torch_mps_set_arg_bytes, older targets ride them in as 1-element tensors. Today's 2.11 pin builds the tensor path, so no torch floor changes anywhere. Bumping the pin later on switches paths with no code change.
  • Shader source moves to roi_align_metal_shader.h. Once every MPS op migrates, the per-op headers can consolidate to a mps_stable_kernels.h and mps_kernels.h gets dropped.
  • Determinism contract. The MPS backward's alert can't fire from a stable kernel, so the Python backward's gate becomes in ("cuda", "mps").

Stable-ABI audit (torch-abi-audit)

Ran torch-abi-audit (https://github.com/Quansight/torch-abi-audit) on both MPS builds:

      Package: torchvision
        Torch ABI:   UNSTABLE
        CPython ABI: n/a
        Bundled libs: 3
        -- bundled libs --
          [UNSTABLE] [not-abi3] _C.so            (stable_shim=0,  unstable=91)
          [STABLE  ] [not-abi3] _C_stable.so     (stable_shim=77, unstable=0)
          [STABLE  ] [not-abi3] image_stable.so  (stable_shim=70, unstable=0)

@pytorch-bot

pytorch-bot Bot commented Jul 30, 2026 •
edited
Loading

Copy link
Copy Markdown

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9574

Note: Links to docs will display an error until the docs builds have been completed.

❌ 28 New Failures, 1 Unrelated Failure, 1 Unclassified Failure

As of commit b20333b with merge base 0b23e88 (image):

NEW FAILURES - The following jobs have failed:

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

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 meta-cla Bot added the cla signed label Jul 30, 2026
@adabeyta adabeyta changed the title Port roi_align (MPS) to stable ABI. [Test] Port roi_align (MPS) to stable ABI. Jul 30, 2026
NicolasHug
NicolasHug reviewed Jul 31, 2026
View reviewed changes
Comment thread torchvision/csrc/ops/mps/roi_align_kernel.mm Outdated
Comment on lines +19 to +24
#if defined(TORCH_VERSION_2_14_0) && \
TORCH_FEATURE_VERSION >= TORCH_VERSION_2_14_0
#define ROI_ALIGN_MPS_HAS_SET_ARG_BYTES 1
#else
#define ROI_ALIGN_MPS_HAS_SET_ARG_BYTES 0
#endif

@NicolasHug NicolasHug Jul 31, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for going our of your way to support multiple torch versions ! But let's simplify the problem: it's fine to have torch 2.14 as the minimal torch version. We can drop the torch < 2.14 parts

@NicolasHug
NicolasHug force-pushed the adabeyta/roi-align-mps-stable branch from 2bf14a3 to 34a1c35 Compare August 4, 2026 19:05
@adabeyta
adabeyta force-pushed the adabeyta/roi-align-mps-stable branch from 34a1c35 to b20333b Compare August 4, 2026 23:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@NicolasHug NicolasHug NicolasHug left review comments

Assignees

No one assigned

Labels

cla signed

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@adabeyta @NicolasHug