[MPS] Fix dlpack exports/imports for sliced tensors by malfet · Pull Request #169272 · pytorch/pytorch

malfet · 2025-12-01T00:50:17Z

Stack from ghstack (oldest at bottom):

-> [MPS] Fix dlpack exports/imports for sliced tensors #169272

For MPS tensor, one must pass both id<MTL_Buffer> (which is t.storage().data() and t.storage_offset())

Luckily, DLTensor already has byte_offset field, which feels natural to use as product of storage_offset and element_size.

Partially extends #168193, but instead of writing a completely new test, fix both export and import paths of sliced tensor and unskip test_from_dlpack_noncontinguous for MPS

Error out if one is attempting to create tensor with non-zero byte_offsets and no strides, as there are no at::from_blob variant that could be used

Fixes #168177

For MPS tensor, one must pass both `id<MTL_Buffer>` (which is `t.storage().data()` and `t.storage_offset()`) Luckily, DLTensor already has `byte_offset` field, which feels natural to use as product of `storage_offset` and element_size Partially extends #168193, but instead of writing a completely new test, fix both export and import paths of sliced tensor and unskip test_from_dlpack_noncontinguous for MPS Fixes #168177 [ghstack-poisoned]

pytorch-bot · 2025-12-01T00:50:21Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/169272

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki

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

❌ 1 New Failure, 7 Unrelated Failures

As of commit b1bbb85 with merge base a1ab3a0 ():

NEW FAILURE - The following job has failed:

trunk / linux-jammy-cuda12.8-py3.10-gcc11 / test (default, 4, 5, linux.g6.4xlarge.experimental.nvidia.gpu) (gh)
test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained

UNSTABLE - The following jobs are marked as unstable, possibly due to flakiness on trunk:

trunk / linux-jammy-rocm-py3.10 / test (default, 1, 6, linux.rocm.gpu.gfx942.1) (gh) (#169451)
test/inductor/test_aot_inductor.py::AOTInductorTestABICompatibleGpu::test_unbacked_expr_replacements_shift_k_0_use_static_size_False_cuda
trunk / linux-jammy-rocm-py3.10 / test (default, 2, 6, linux.rocm.gpu.gfx942.1) (gh) (#169451)
test/inductor/test_ck_backend.py::TestCKBackend::test_max_autotune_addmm_max_autotune_gemm_backends_CK_x_shape0
trunk / linux-jammy-rocm-py3.10 / test (default, 3, 6, linux.rocm.gpu.gfx942.1) (gh) (#169451)
test/inductor/test_aot_inductor.py::AOTInductorTestABICompatibleGpu::test_non_tensor_input_cuda
trunk / linux-jammy-rocm-py3.10 / test (default, 4, 6, linux.rocm.gpu.gfx942.1) (gh) (#169451)
test/inductor/test_aot_inductor.py::AOTInductorTestABICompatibleGpu::test_unbacked_equals_input_size_runtime_assertion_mark_unbacked_False_cuda
trunk / linux-jammy-rocm-py3.10 / test (default, 5, 6, linux.rocm.gpu.gfx942.1) (gh) (#169451)
test/inductor/test_aot_inductor_package.py::TestAOTInductorPackage_cpu::test_duplicate_calls
trunk / linux-jammy-rocm-py3.10 / test (default, 6, 6, linux.rocm.gpu.gfx942.1) (gh) (#169451)
test/inductor/test_torchinductor_codegen_dynamic_shapes.py::DynamicShapesCodegenGPUTests::test_aoti_eager_override_registration_dynamic_shapes_cuda
trunk / linux-jammy-rocm-py3.10 / test (distributed, 1, 3, linux.rocm.gpu.gfx942.4) (gh) (#169451)
RuntimeError: distributed/test_nvshmem_triton 1/1 failed!

This comment was automatically generated by Dr. CI and updates every 15 minutes.

For MPS tensor, one must pass both `id<MTL_Buffer>` (which is `t.storage().data()` and `t.storage_offset()`) Luckily, DLTensor already has `byte_offset` field, which feels natural to use as product of `storage_offset` and element_size Partially extends #168193, but instead of writing a completely new test, fix both export and import paths of sliced tensor and unskip test_from_dlpack_noncontinguous for MPS Fixes #168177 ghstack-source-id: 61a32f2 Pull Request resolved: #169272

[ghstack-poisoned]

For MPS tensor, one must pass both `id<MTL_Buffer>` (which is `t.storage().data()` and `t.storage_offset()`) Luckily, DLTensor already has `byte_offset` field, which feels natural to use as product of `storage_offset` and element_size Partially extends #168193, but instead of writing a completely new test, fix both export and import paths of sliced tensor and unskip test_from_dlpack_noncontinguous for MPS Fixes #168177 ghstack-source-id: 702dda8 Pull Request resolved: #169272

malfet · 2025-12-02T19:14:01Z

@pytorchbot merge

pytorchmergebot · 2025-12-02T19:15:56Z

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

huydhn · 2025-12-03T00:28:53Z

@pytorchbot revert -m 'I am seeing some ROCm failures in trunk after this lands' -c nosignals

test/inductor/test_aot_inductor.py::AOTInductorTestABICompatibleGpu::test_non_tensor_input_cuda GH job link HUD commit link

test/inductor/test_torchinductor.py::GPUTests::test_aoti_eager_override_registration_cuda GH job link HUD commit link

pytorch-bot · 2025-12-03T00:28:55Z

❌ 🤖 pytorchbot command failed:

@pytorchbot revert: error: argument -c/--classification: invalid choice: 'nosignals' (choose from 'nosignal', 'ignoredsignal', 'landrace', 'weird', 'ghfirst', 'autorevert')

usage: @pytorchbot revert -m MESSAGE -c
                          {nosignal,ignoredsignal,landrace,weird,ghfirst,autorevert}

Try @pytorchbot --help for more info.

huydhn · 2025-12-03T00:34:45Z

@pytorchbot revert -m 'I am seeing some ROCm failures in trunk after this lands' -c nosignal

pytorchmergebot · 2025-12-03T00:36:26Z

@pytorchbot successfully started a revert job. Check the current status here.
Questions? Feedback? Please reach out to the PyTorch DevX Team

This reverts commit 7741edd. Reverted #169272 on behalf of https://github.com/huydhn due to I am seeing some ROCm failures in trunk after this lands ([comment](#169272 (comment)))

pytorchmergebot · 2025-12-03T00:36:34Z

@malfet your PR has been successfully reverted.

huydhn · 2025-12-03T04:51:59Z

@pytorchbot merge -i

Reland this as I don't think the revert helps

pytorchmergebot · 2025-12-03T04:53:53Z

Merge started

Your change will be merged while ignoring the following 8 checks: trunk / linux-jammy-rocm-py3.10 / test (default, 1, 6, linux.rocm.gpu.gfx942.1), trunk / linux-jammy-rocm-py3.10 / test (default, 2, 6, linux.rocm.gpu.gfx942.1), trunk / linux-jammy-rocm-py3.10 / test (default, 6, 6, linux.rocm.gpu.gfx942.1), trunk / linux-jammy-rocm-py3.10 / test (default, 4, 6, linux.rocm.gpu.gfx942.1), trunk / linux-jammy-rocm-py3.10 / test (default, 3, 6, linux.rocm.gpu.gfx942.1), trunk / linux-jammy-rocm-py3.10 / test (default, 5, 6, linux.rocm.gpu.gfx942.1), trunk / linux-jammy-rocm-py3.10 / test (distributed, 1, 3, linux.rocm.gpu.gfx942.4), trunk / linux-jammy-cuda12.8-py3.10-gcc11 / test (default, 4, 5, linux.g6.4xlarge.experimental.nvidia.gpu)

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

For MPS tensor, one must pass both `id<MTL_Buffer>` (which is `t.storage().data()` and `t.storage_offset()`) Luckily, DLTensor already has `byte_offset` field, which feels natural to use as product of `storage_offset` and element_size. Partially extends #168193, but instead of writing a completely new test, fix both export and import paths of sliced tensor and unskip test_from_dlpack_noncontinguous for MPS Error out if one is attempting to create tensor with non-zero `byte_offsets` and no strides, as there are no `at::from_blob` variant that could be used Fixes #168177 Pull Request resolved: #169272 Approved by: https://github.com/ngimel

This reverts commit 7741edd. Reverted #169272 on behalf of https://github.com/huydhn due to I am seeing some ROCm failures in trunk after this lands ([comment](#169272 (comment)))

For MPS tensor, one must pass both `id<MTL_Buffer>` (which is `t.storage().data()` and `t.storage_offset()`) Luckily, DLTensor already has `byte_offset` field, which feels natural to use as product of `storage_offset` and element_size. Partially extends #168193, but instead of writing a completely new test, fix both export and import paths of sliced tensor and unskip test_from_dlpack_noncontinguous for MPS Error out if one is attempting to create tensor with non-zero `byte_offsets` and no strides, as there are no `at::from_blob` variant that could be used Fixes #168177 Pull Request resolved: #169272 Approved by: https://github.com/ngimel

malfet added ciflow/trunk Trigger trunk jobs on your pull request topic: bug fixes topic category release notes: mps Release notes category ciflow/mps Run MPS tests (subset of trunk) labels Dec 1, 2025

malfet requested review from Skylion007, albanD and ezyang December 1, 2025 00:51

Update

b1bbb85

[ghstack-poisoned]

ngimel approved these changes Dec 2, 2025

View reviewed changes

pytorchmergebot added the merging label Dec 2, 2025

pytorchmergebot added the Merged label Dec 2, 2025

pytorchmergebot closed this in 7741edd Dec 2, 2025

pytorchmergebot removed the merging label Dec 2, 2025

pytorchmergebot added Reverted ci-no-td Do not run TD on this PR labels Dec 3, 2025

pytorchmergebot reopened this Dec 3, 2025

pytorchmergebot added the merging label Dec 3, 2025

pytorchmergebot closed this in 5fafc13 Dec 3, 2025

pytorchmergebot removed the merging label Dec 3, 2025

malfet mentioned this pull request Dec 8, 2025

Honor dlpack byte_offset #168193

Closed

github-actions bot deleted the gh/malfet/612/head branch January 3, 2026 02:18

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[MPS] Fix dlpack exports/imports for sliced tensors#169272

[MPS] Fix dlpack exports/imports for sliced tensors#169272
malfet wants to merge 2 commits intogh/malfet/612/basefrom
gh/malfet/612/head

malfet commented Dec 1, 2025 •

edited

Loading

Uh oh!

pytorch-bot bot commented Dec 1, 2025 •

edited

Loading

Uh oh!

malfet commented Dec 2, 2025

Uh oh!

pytorchmergebot commented Dec 2, 2025

Uh oh!

huydhn commented Dec 3, 2025

Uh oh!

pytorch-bot bot commented Dec 3, 2025

Uh oh!

huydhn commented Dec 3, 2025

Uh oh!

pytorchmergebot commented Dec 3, 2025

Uh oh!

pytorchmergebot commented Dec 3, 2025

Uh oh!

huydhn commented Dec 3, 2025

Uh oh!

pytorchmergebot commented Dec 3, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

malfet commented Dec 1, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot bot commented Dec 1, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/169272

❌ 1 New Failure, 7 Unrelated Failures

Uh oh!

malfet commented Dec 2, 2025

Uh oh!

pytorchmergebot commented Dec 2, 2025

Merge started

Uh oh!

huydhn commented Dec 3, 2025

Uh oh!

pytorch-bot bot commented Dec 3, 2025

Uh oh!

huydhn commented Dec 3, 2025

Uh oh!

pytorchmergebot commented Dec 3, 2025

Uh oh!

pytorchmergebot commented Dec 3, 2025

Uh oh!

huydhn commented Dec 3, 2025

Uh oh!

pytorchmergebot commented Dec 3, 2025

Merge started

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

malfet commented Dec 1, 2025 •

edited

Loading

pytorch-bot bot commented Dec 1, 2025 •

edited

Loading