Fix histc(out=) to reject mismatched out dtype instead of silent cast by N0AHZACH · Pull Request #5380 · intel/torch-xpu-ops

N0AHZACH · 2026-09-15T16:47:04Z

_histc_out_xpu computed the histogram in the input dtype via _histc_xpu
and then did resize_output + result.copy_(ret). copy_ casts silently,
so a float64 histogram written into an int64 out= was truncated with no
error. CPU rejects the same call in histogramdd_prepare_out
(aten/src/ATen/native/Histogram.cpp:132), so this was a CPU-vs-XPU
argument-validation divergence.

Changes:

src/ATen/native/xpu/SummaryOps.cpp: add the CPU-matching TORCH_CHECK
before resize_output so a rejected call leaves the caller tensor
untouched. Message text matches CPU verbatim and matches upstream
Fix histc(out=) dtype validation mismatch between CPU and CUDA pytorch/pytorch#196100 for CUDA.
test/xpu/xpu_test_utils.py: de-scope tuple device_type=("cuda", "xpu")
xfails gated on _cuda_xfail_xpu_pass by dropping only xpu from the
scope. Without this the fixed test reports an unexpected success.
test/xpu/skip_list_common.py: remove test_out_histc_xpu_float32.
test/regressions/test_histc.py: regression coverage for mismatched
out= rejection and matching-dtype success.

This does not depend on pytorch/pytorch#196100 landing and does not
conflict with it.

_histc_out_xpu computed into a fresh tensor of the input dtype and then did result.copy_(ret). copy_ converts, so a float64 histogram written into an int64 out= was silently truncated. CPU rejects the same call in histogramdd_prepare_out, which the XPU out variant does not go through. Add the CPU-matching TORCH_CHECK before resize_output so a rejected call leaves the caller tensor untouched. Remove test_out_histc from the skip list. De-scope the tuple device_type xfail in align_db_decorators so the now-passing test does not report an unexpected success. Fixes intel#5237. Test Plan: ``` python -m py_compile test/regressions/test_histc.py test/xpu/xpu_test_utils.py test/xpu/skip_list_common.py git diff --check ```

N0AHZACH · 2026-09-15T16:58:23Z

@laifenxiawucha could you check this out so that I can get the CI rolling

N0AHZACH · 2026-09-16T15:27:35Z

@laifenxiawucha can we start the CI on this specific commit - I do see a CI workflow has been run but idk where it ran tho.......been searching for a good 15 mins. Could we start the CI on this specific commit. I'm super sorry if its something stupid from my end

laifenxiawucha · 2026-09-17T08:24:50Z

Hi @N0AHZACH .No worries at all — this is likely not an issue on your side. Since the PR is from a fork, some CI jobs may need maintainer approval before they can start. That’s probably why you can see a workflow run but not the full CI for this specific commit.

N0AHZACH · 2026-09-18T03:24:20Z

Thanks @laifenxiawucha. I traced the four failing checks and they are stale, not caused by this change.

The linux-build and windows-build jobs fail before compiling anything (run 35192511198, commit ee3cc7cf):

Linux: python: can't open file '.../pytorch/.ci/manywheel/build_env_setup.py': [Errno 2] No such file or directory
Windows: ModuleNotFoundError: No module named 'build_env_setup'

Both come from PyTorch main relocating the wheel build scripts to .ci/wheel/linux and .ci/wheel/windows (pytorch/pytorch#189275). That breakage was fixed on main by #5417 and #5420, and the current head d355f901 already merges them (main at e9e0bfe4), so .github/scripts/build.sh and build_windows.py now use the new paths.

The change itself is already validated on hardware. In run 35050000160 (272fcabb), where the UT actually ran:

op_regression passed, including both new cases in test/regressions/test_histc.py (test_histc_out_rejects_mismatched_dtype, test_histc_out_matching_dtype_still_works)
test_out_histc_xpu_float32 ran and passed in op_ut (it is no longer skipped, and no longer reports an unexpected success against the upstream ('cuda', 'xpu') xfail)
the remaining op_ut entries in that run are unrelated (dynamo / flex-attention / fx / libtorchbind / env), none of them touch histc

Could you approve/start CI on d355f901? The run for that commit (35302042384) is still waiting on maintainer approval. Thanks!

N0AHZACH · 2026-09-27T05:32:44Z

can we get a CI running on this please

BBBela

This PR will make the test_histc_out_dtype test to fail with Unexpected success - that should be fixed by removing the expectedFailureXPU from the upstream test directly, after this PR is merged - like in this PR pytorch/pytorch#199126

BBBela · 2026-10-01T07:09:42Z

+                    elif (
+                        isinstance(wrapper.device_type, (list, tuple))
+                        and "xpu" in wrapper.device_type
+                        and unittest.expectedFailure in wrapper.decorators
+                        and (op_name, wrapper.test_name) in _cuda_xfail_xpu_pass
+                    ):


This elif: branch is not needed anymore. I suggest to remove it. The upstream ("cuda", "xpu") tuple was removed, so it is now dead code for the failing test case we are talking about.

ahhh okay - gotchu. will push an update shortly.

BBBela · 2026-10-01T07:17:45Z

Please correct me if I am wrong, but it looks like the upstream test test_histc_out_dtype added by you in test_reductions.py file already covers this regression scope - am I right? torch-xpu-ops repo currently imports the upstream test file and launches it, so it also launched this test_histc_out_dtype test. If my understanding above is correct, then I suggest to remove this new regression test from torch-xpu-ops and remove expectedFailureXPU from the upstream test_histc_out_dtype as a follow up PR. Because it now looks like unnecessary code duplication.

okayyyy - I shall fix that then....

BBBela · 2026-10-01T07:25:16Z

It looks like the line:

("histc", "test_out"),

should be removed in this PR also from _cuda_xfail_xpu_pass.

yup will do.

N0AHZACH · 2026-10-02T16:06:14Z

I'm actually still working on this - personal life caught up to me a lil. But I am still working on this !

BBBela · 2026-10-02T16:18:32Z

No problem @N0AHZACH 😉
Please jest let me know when to re-review. I will also launch CI then.

The upstream ("cuda", "xpu") xfail for (histc, test_out) was removed in pytorch/pytorch#196100, so the tuple de-scope branch added to align_db_decorators never matches a wrapper, and ("histc", "test_out") in _cuda_xfail_xpu_pass no longer refers to any DecorateInfo. Remove both, per review. Delete test/regressions/test_histc.py: upstream test_histc_out_dtype in test/test_reductions.py covers the same mismatched out= dtype regression and torch-xpu-ops runs it through test_reductions_xpu.py. The fix itself is unchanged: the CPU-matching TORCH_CHECK in _histc_out_xpu before resize_output, plus dropping test_out_histc_xpu_float32 from the op_ut skip list. Test Plan: python -m py_compile test/xpu/xpu_test_utils.py python -m flake8 test/xpu/xpu_test_utils.py clang-format --dry-run --Werror src/ATen/native/xpu/SummaryOps.cpp git diff --check

N0AHZACH · 2026-10-03T04:59:59Z

@BBBela I pushed some new updates sooo please CI?

N0AHZACH · 2026-10-03T05:02:05Z

No problem @N0AHZACH 😉 Please jest let me know when to re-review. I will also launch CI then.

The job hunts just got real. I got that insane reality check the open-source aint gonna help me keep the lights on😂. 1000+ rejections and a gizzilion more to go.

BBBela · 2026-10-03T07:39:27Z

@BBBela I pushed some new updates sooo please CI?

@N0AHZACH CI launched. Don't bother if there are some failing test cases. They can be sometimes flaky.
On the first look, the PR looks good to me now, but I will double check on Monday morning and then approve if CI is okay 😉

One think to do - please update the PR description as it is outdated after recent changes.

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 Sep 15, 2026

N0AHZACH force-pushed the fix-5237-histc-out-dtype branch from 164c184 to ea4c1d7 Compare September 15, 2026 16:53

N0AHZACH added 2 commits September 16, 2026 08:28

Merge branch 'main' into fix-5237-histc-out-dtype

272fcab

Merge branch 'main' into fix-5237-histc-out-dtype

9bafa70

Merge branch 'main' into fix-5237-histc-out-dtype

ee3cc7c

Merge branch 'main' into fix-5237-histc-out-dtype

d355f90

N0AHZACH added 2 commits September 18, 2026 16:25

Merge branch 'main' into fix-5237-histc-out-dtype

d96cdf8

Merge branch 'main' into fix-5237-histc-out-dtype

d89f76b

Merge branch 'main' into fix-5237-histc-out-dtype

6a7c744

dvrogozh mentioned this pull request Sep 30, 2026

Add proper dtype checks for histc_out. #5554

Open

dvrogozh requested review from AKloniecki, BBBela and guangyey September 30, 2026 19:23

BBBela added the disable_auto Disable auto label workflow for PR label Oct 1, 2026

BBBela mentioned this pull request Oct 1, 2026

[xpu-alignment] histc(out=) on XPU silently casts instead of rejecting a mismatched out dtype #5237

Open

BBBela reviewed Oct 1, 2026

View reviewed changes

Merge branch 'main' into fix-5237-histc-out-dtype

baee108

N0AHZACH added 2 commits October 2, 2026 22:40

Merge branch 'main' into fix-5237-histc-out-dtype

c569fc4

Merge branch 'main' into fix-5237-histc-out-dtype

672fe76

Conversation

N0AHZACH commented Sep 15, 2026

Uh oh!

N0AHZACH commented Sep 15, 2026

Uh oh!

N0AHZACH commented Sep 16, 2026

Uh oh!

laifenxiawucha commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

N0AHZACH commented Sep 18, 2026

Uh oh!

N0AHZACH commented Sep 27, 2026

Uh oh!

BBBela left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

BBBela Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

N0AHZACH Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

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

Uh oh!

Choose a reason for hiding this comment

Uh oh!

N0AHZACH Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

BBBela Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

N0AHZACH Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

N0AHZACH commented Oct 2, 2026

Uh oh!

BBBela commented Oct 2, 2026

Uh oh!

N0AHZACH commented Oct 3, 2026

Uh oh!

N0AHZACH commented Oct 3, 2026

Uh oh!

BBBela commented Oct 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

laifenxiawucha commented Sep 17, 2026 •

edited

Loading

BBBela left a comment •

edited

Loading

BBBela Oct 1, 2026 •

edited

Loading