Validate `grad_output` in `grid_sampler_{2d,3d}_backward` by pbielak · Pull Request #5584 · intel/torch-xpu-ops

pbielak · 2026-10-01T15:16:55Z

XPU was the only backend still skipping the grad_output shape check, so a mismatched grad_output made both backward kernels read out of bounds. CPU, CUDA and MPS already use the hardened helpers added upstream by pytorch/pytorch#191915.

2d: replace check_grid_sampler_common + check_grid_sampler_2d with check_grid_sampler_2d_backward
3d: replace the equivalent block with check_grid_sampler_3d_backward and
reject bicubic via a standalone TORCH_CHECK_NOT_IMPLEMENTED, mirroring MPS; the torch check is kept until 5-D bicubic lands (Implement bicubic version of grid_sampler_3d #5511)
port test_grid_sample_backward_error_checking from upstream test_nn.py

XPU was the only backend still skipping the `grad_output` shape check, so a mismatched `grad_output` made both backward kernels read out of bounds. CPU, CUDA and MPS already use the hardened helpers added upstream by pytorch/pytorch#191915. - 2d: switch to `check_grid_sampler_2d_backward` - 3d: add `check_grid_sampler_backward`; the bicubic-rejecting 3-arg `check_grid_sampler_3d` overload is kept until 5-D bicubic lands (#5511) - port `test_grid_sample_backward_error_checking` from upstream `test_nn.py` Requested in review of #5511

dvrogozh

LGTM

dvrogozh

The PR now looks good to me per-se, but I see the CI failure which might be related:

op_extended,third_party.torch-xpu-ops.test.xpu.extended.test_ops_xpu.TestCompositeComplianceXPU,test_forward_ad_nn_functional_grid_sample_xpu_float32

Please, help to investigate.

pbielak · 2026-10-02T14:40:26Z

The PR now looks good to me per-se, but I see the CI failure which might be related:
op_extended,third_party.torch-xpu-ops.test.xpu.extended.test_ops_xpu.TestCompositeComplianceXPU,test_forward_ad_nn_functional_grid_sample_xpu_float32
Please, help to investigate.

Added this test case to the tracking issue: #5452 - all will be fixed with the actual implementation of the Bicubic grid sampler #5511

(This entry also fails in other PR - e.g., https://github.com/intel/torch-xpu-ops/actions/runs/37002662439/job/110852416391?pr=5592 )

dvrogozh · 2026-10-02T14:56:01Z

@pbielak , please, do the similar triage for the recently finished op_ut tests:

op_ut,third_party.torch-xpu-ops.test.xpu.functorch.test_ops_xpu.TestOperatorsXPU,test_vmapjvpall_has_batch_rule_nn_functional_grid_sample_xpu_float32
op_ut,third_party.torch-xpu-ops.test.xpu.functorch.test_ops_xpu.TestOperatorsXPU,test_jvp_nn_functional_grid_sample_xpu_float32
op_ut,third_party.torch-xpu-ops.test.xpu.functorch.test_ops_xpu.TestOperatorsXPU,test_vmapjvpall_nn_functional_grid_sample_xpu_float32
ops.test.xpu.test_ops_fwd_gradients_xpu.TestFwdGradientsXPU,test_forward_mode_AD_nn_functional_grid_sample_xpu_float64
op_ut,third_party.torch-xpu-ops.test.xpu.test_ops_xpu.TestCompositeComplianceXPU,test_forward_ad_nn_functional_grid_sample_xpu_float32

Also, check if other failing tests are a known failures (they don't have grid_sample in their names).

pbielak · 2026-10-02T14:57:44Z

@pbielak , please, do the similar triage for the recently finished op_ut tests:

op_ut,third_party.torch-xpu-ops.test.xpu.functorch.test_ops_xpu.TestOperatorsXPU,test_vmapjvpall_has_batch_rule_nn_functional_grid_sample_xpu_float32
op_ut,third_party.torch-xpu-ops.test.xpu.functorch.test_ops_xpu.TestOperatorsXPU,test_jvp_nn_functional_grid_sample_xpu_float32
op_ut,third_party.torch-xpu-ops.test.xpu.functorch.test_ops_xpu.TestOperatorsXPU,test_vmapjvpall_nn_functional_grid_sample_xpu_float32
ops.test.xpu.test_ops_fwd_gradients_xpu.TestFwdGradientsXPU,test_forward_mode_AD_nn_functional_grid_sample_xpu_float64
op_ut,third_party.torch-xpu-ops.test.xpu.test_ops_xpu.TestCompositeComplianceXPU,test_forward_ad_nn_functional_grid_sample_xpu_float32

Also, check if other failing tests are a known failures (they don't have grid_sample in their names).

I also checked these and I added them to the tracking issue (same problem of missing bicubic implementation)

dvrogozh

LGTM

dvrogozh · 2026-10-02T16:50:29Z

The rest are known failures (same in https://github.com/intel/torch-xpu-ops/actions/runs/36961251529/job/110751670169?pr=5117 for the #5117 which is orthogonal to the PR we review).

dvrogozh · 2026-10-02T16:51:14Z

@torchxpubot merge -f "CI failures are unrelated"

torchxpubot · 2026-10-02T16:51:46Z

Failed to merge PR: Repository rule violations found

A conversation must be resolved before this pull request can be merged.

https://docs.github.com/rest/pulls/pulls#merge-a-pull-request

dvrogozh · 2026-10-02T16:53:42Z

@torchxpubot merge -f "CI failures are unrelated"

torchxpubot · 2026-10-02T16:54:14Z

PR has been successfully merged by @dvrogozh (force merge).

XPU was the only backend still skipping the `grad_output` shape check, so a mismatched `grad_output` made both backward kernels read out of bounds. CPU, CUDA and MPS already use the hardened helpers added upstream by pytorch/pytorch#191915. - 2d: replace `check_grid_sampler_common` + `check_grid_sampler_2d` with `check_grid_sampler_2d_backward` - 3d: replace the equivalent block with `check_grid_sampler_3d_backward` and reject bicubic via a standalone `TORCH_CHECK_NOT_IMPLEMENTED`, mirroring MPS; the torch check is kept until 5-D bicubic lands (#5511) - port `test_grid_sample_backward_error_checking` from upstream `test_nn.py`

XPU was the only backend still skipping the `grad_output` shape check, so a mismatched `grad_output` made both backward kernels read out of bounds. CPU, CUDA and MPS already use the hardened helpers added upstream by pytorch/pytorch#191915. - 2d: replace `check_grid_sampler_common` + `check_grid_sampler_2d` with `check_grid_sampler_2d_backward` - 3d: replace the equivalent block with `check_grid_sampler_3d_backward` and reject bicubic via a standalone `TORCH_CHECK_NOT_IMPLEMENTED`, mirroring MPS; the torch check is kept until 5-D bicubic lands (intel#5511) - port `test_grid_sample_backward_error_checking` from upstream `test_nn.py`

pbielak requested review from dvrogozh and lchen2331 October 1, 2026 15:16

github-actions Bot added disable_e2e Disable all e2e test jobs for the PR disable_distributed Disable distributed UT test jobs for the PR labels Oct 1, 2026

pbielak mentioned this pull request Oct 1, 2026

Implement bicubic version of grid_sampler_3d #5511

Merged

dvrogozh requested changes Oct 1, 2026

View reviewed changes

Comment thread src/ATen/native/xpu/sycl/GridSampler.cpp Resolved

Comment thread test/xpu/test_nn_xpu.py Resolved

pbielak requested a review from dvrogozh October 2, 2026 08:30

tszulist-hbn reviewed Oct 2, 2026

View reviewed changes

Comment thread src/ATen/native/xpu/sycl/GridSampler.cpp Resolved

pbielak force-pushed the harden-grid-sampler-backward-checks branch from 0f0e78a to a8b9dbb Compare October 2, 2026 09:34

Fix dispatch name

7f2d044

pbielak requested a review from tszulist-hbn October 2, 2026 09:37

tszulist-hbn approved these changes Oct 2, 2026

View reviewed changes

dvrogozh approved these changes Oct 2, 2026

View reviewed changes

Comment thread src/ATen/native/xpu/sycl/GridSampler.cpp Resolved

dvrogozh requested changes Oct 2, 2026

View reviewed changes

dvrogozh approved these changes Oct 2, 2026

View reviewed changes

torchxpubot merged commit 9521dfb into main Oct 2, 2026
54 of 58 checks passed

torchxpubot deleted the harden-grid-sampler-backward-checks branch October 2, 2026 16:54

torchxpubot mentioned this pull request Oct 28, 2026

[upstream_ut] XPU backward kernels missing input-validation error checks (replication_pad2d, grid_sample) #5508

Open

Conversation

pbielak commented Oct 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dvrogozh left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

dvrogozh left a comment

Choose a reason for hiding this comment

Uh oh!

pbielak commented Oct 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dvrogozh commented Oct 2, 2026

Uh oh!

pbielak commented Oct 2, 2026

Uh oh!

dvrogozh left a comment

Choose a reason for hiding this comment

Uh oh!

dvrogozh commented Oct 2, 2026

Uh oh!

dvrogozh commented Oct 2, 2026

Uh oh!

torchxpubot commented Oct 2, 2026

Uh oh!

dvrogozh commented Oct 2, 2026

Uh oh!

Uh oh!

torchxpubot commented Oct 2, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

pbielak commented Oct 1, 2026 •

edited

Loading

pbielak commented Oct 2, 2026 •

edited

Loading