Expand XPU compiler/autograd testing by shangerxin · Pull Request #4894 · intel/torch-xpu-ops

shangerxin · 2026-08-10T07:00:30Z

Description

This PR introduces a SYCL-oriented build system for XPU ops and expands XPU test coverage around compiler infrastructure, activation checkpointing, autograd cache behavior, and PyTorch testing utilities.

Migrate required files in torch repo
test/dynamo/test_activation_checkpointing.py
test/dynamo/test_aot_autograd_cache.py
test/dynamo/test_compiler_bisector.py
test/dynamo/test_debug_utils.py
test/dynamo/test_deviceguard.py
test/dynamo/test_export.py
test/dynamo/test_regional_inductor.py
test/dynamo/test_structured_trace.py
test/dynamo/test_subclasses.py
test/dynamo/test_trace_rules.py
test/test_testing.py

Compare Link

329d891...pull/4894/head

Changes

Added

test/xpu/dynamo/test_activation_checkpointing_xpu.py
New, very large test suite for activation checkpointing, selective checkpointing, rematerialization, region memory budgets, non-strict tracing, nested compile, and many edge cases.

Modified

test/xpu/test_testing_xpu.py
Greatly expands generic testing coverage: assert-close/assert-equal, sparse tensors, quantized tensors, parametrization, opinfo, hardware classification, and import behavior.
test/xpu/skip_list_common.py
Updates skip entries for moved/renamed test files and adds skips for the new XPU dynamo/debug-related tests.

Deleted

test/xpu/dynamo/test_aot_autograd_cache_xpu.py
test/xpu/dynamo/test_compiler_bisector_xpu.py
test/xpu/dynamo/test_deviceguard_xpu.py
test/xpu/dynamo/test_regional_inductor_xpu.py

What this means overall

The PR consolidates older Dynamo/inductor test coverage into a new activation-checkpointing-focused test file.
It also broadens AOT-autograd cache validation and general testing utilities coverage.
Some old test modules are removed because their scenarios were absorbed into the new files.

Bugs

#4943
#4945

shangerxin · 2026-08-11T03:28:15Z

@copilot please fix the lint issue.

Copilot · 2026-08-11T03:31:33Z

@copilot please fix the lint issue.

Fixed in the latest commit. The issues were:

Tab character on line 357 of skip_list_common.py (replaced with spaces)
Unused onlyCUDA import in test_testing_xpu.py (removed)
Missing onlyOn import in test_testing_xpu.py (added to the common_device_type import)

shangerxin · 2026-08-11T06:01:22Z

@copilot please fix the lint job https://github.com/intel/torch-xpu-ops/actions/runs/31455650407/job/93668736816?pr=4894 failures for this pr.

Copilot · 2026-08-11T06:03:36Z

@copilot please fix the lint job https://github.com/intel/torch-xpu-ops/actions/runs/31455650407/job/93668736816?pr=4894 failures for this pr.

Fixed in the latest commit. Two issues were addressed:

Remaining tab character in test/xpu/test_testing_xpu.py line 426 (inside an inline test string)
CLANGTIDY cppcoreguidelines-prefer-member-initializer error in src/ATen/native/xpu/sycl/FFTKernelFunctor.cpp — moved mirror_dim_ initialization to the member initializer list

shangerxin · 2026-08-12T06:02:13Z

@copilot fix the lint issues for the files changed in this PR. The lint job is https://github.com/intel/torch-xpu-ops/actions/runs/31467356782/job/93702938446?pr=4894

Copilot · 2026-08-12T06:03:40Z

@copilot fix the lint issues for the files changed in this PR. The lint job is https://github.com/intel/torch-xpu-ops/actions/runs/31467356782/job/93702938446?pr=4894

Fixed in commit Fix CLANGTIDY mirror_dim_ member initializer lint error. Moved mirror_dim_ initialization to the member initializer list (: mirror_dim_(0)) to satisfy the cppcoreguidelines-prefer-member-initializer rule.

shangerxin · 2026-08-12T06:18:48Z

@copilot fix the lint errors in this https://github.com/intel/torch-xpu-ops/actions/runs/31568622333/job/94025635331?pr=4894 .

Copilot · 2026-08-12T06:20:18Z

@copilot fix the lint errors in this https://github.com/intel/torch-xpu-ops/actions/runs/31568622333/job/94025635331?pr=4894 .

Fixed in the latest commit:

FFTKernelFunctor.cpp: Changed mirror_dim_ to use a default member initializer at declaration (uint32_t mirror_dim_ = 0;) to satisfy modernize-use-default-member-init
test_testing_xpu.py line 426: Replaced remaining tab character with spaces

shangerxin · 2026-08-13T02:18:00Z

@copilot please summarize the top n failures for the files test_* in this pr. the UT jobs are https://github.com/intel/torch-xpu-ops/actions/runs/31571224115/job/94041674226?pr=4894 and https://github.com/intel/torch-xpu-ops/actions/runs/31571224115/job/94073605582?pr=4894

Copilot · 2026-08-13T02:19:18Z

@copilot please summarize the top n failures for the files test_* in this pr. the UT jobs are https://github.com/intel/torch-xpu-ops/actions/runs/31571224115/job/94041674226?pr=4894 and https://github...

Here's a summary of the top failures across both UT jobs:

Job 94073605582 — 35 new failures

1. test_aot_autograd_cache — 10 failures (AOTAutogradCacheTests + AOTAutogradCacheBundledTests)
All are variants of test_cache_hit_across_processes_pre_grad_custom_pass_* (with/without UUID, default/early/late).

2. test_activation_checkpointing_xpu — 2 failures

RematerializeACNodesPassTests.test_attr_compile_submodules_in_checkpoint_wrapper
RematerializeACNodesPassTests.test_ac_rematerialize_with_sdpa_dropout_zero

3. polar op — 13 failures across multiple test files (test_binary_ufuncs_xpu, test_tensor_creation_ops_xpu, test_schema_check, test_decomp_xpu, test_meta_xpu, test_ops_xpu) all for polar_xpu_float16.

4. test_export — 1 failure

ExportTestsSubprocess.test_strict_export_under_pythonoptimize

5. test_aotdispatch_xpu — 1 failure

TestAOTAutogradWithCache.test_input_mutation_noncontiguous

6. test_profiler_xpu — 2 failures

TestProfiler.test_profiler_cuda_sync_events
TestProfiler.test_disable_external_correlation

Job 94041674226 — 1 new failure (op_extended)

7. test_ops_xpu (extended) — 1 new failure

TestCommonXPU.test_compare_cpu_polar_xpu_float16

(4 other failures in test_ops_xpu were pre-existing known issues and skipped.)

Top failure categories by frequency:

polar op on float16 — 14 occurrences across multiple test suites (likely a missing or incorrect XPU kernel implementation)
AOTAutogradCache cross-process cache tests — 10 occurrences
test_activation_checkpointing_xpu — 2 new test failures introduced by this PR
Profiler CUDA-named tests on XPU — 2 failures (likely need to be skipped or adapted for XPU)

shangerxin · 2026-08-13T06:41:14Z

The test_aot_autograd_cache failures are relative to errors and the case of test_export also failed with same error.

Need to check the test infrastructure.

ModuleNotFoundError: No module named 'functorch.compile'

…lready support xpu on upstream Update the skip list with upstream items, removed the one file which already supported in upstream merege stashed changes Fix gpu type is not imported issue Add devices and allow xpu flag Fix the merge issues Fix lint issues: tab indentation and import errors in test files Co-authored-by: shangerxin <3085654+shangerxin@users.noreply.github.com> Fix CLANGTIDY mirror_dim_ member initializer lint error in FFTKernelFunctor.cpp Move mirror_dim_ initialization to member initializer list to fix the cppcoreguidelines-prefer-member-initializer lint error. Test Plan: Lint checked via CI at https://github.com/intel/torch-xpu-ops/actions/runs/31467356782/job/93702938446 This PR was authored with an AI assistant. Co-authored-by: shangerxin <3085654+shangerxin@users.noreply.github.com> Fix remaining lint errors: tabs in test_testing_xpu.py and FFTKernelFunctor.cpp - Replace tab with spaces on line 426 of test/xpu/test_testing_xpu.py - Use default member initializer for mirror_dim_ at declaration (modernize-use-default-member-init) instead of constructor initializer list This PR was authored with an AI assistant. Co-authored-by: shangerxin <3085654+shangerxin@users.noreply.github.com> Fix lint issue and revert the changes fort the test_testing_xpu.py Fix lint issue

shangerxin · 2026-08-14T02:17:40Z

@copilot please summarize the top n failures into a table for the files test_* in this pr. the UT jobs are https://github.com/intel/torch-xpu-ops/actions/runs/31571224115/job/94041674226?pr=4894 and https://github... and https://github.com/intel/torch-xpu-ops/actions/runs/31679057066/job/94447532967?pr=4894

Copilot · 2026-08-14T03:40:41Z

@copilot please summarize the top n failures into a table for the files test_* in this pr. the UT jobs are [https://github.com/intel/torch-xpu-ops/actions/runs/31571224115/job/94041674226?pr=4894 and ...

Here's the failure summary table for the latest UT runs (jobs 94394787611 and 94447532967):

Job 94394787611 (op_extended) — 1 new failure

#	Test File	Test Class	Test Case	Count
1	`test_ops_xpu` (extended)	`TestCommonXPU`	`test_compare_cpu_polar_xpu_float16`	1

Job 94447532967 (op_ut) — 30 new failures

#	Category	Test File	Test Class	Test Cases	Count
1	`polar` op / `float16`	`test_binary_ufuncs_xpu`, `test_tensor_creation_ops_xpu`, `test_schema_check`, `test_decomp_xpu`, `test_meta_xpu`, `test_ops_xpu`	Multiple	`polar_xpu_float16` variants	19
2	AOT Autograd Cache	`test_aot_autograd_cache`	`AOTAutogradCacheTests`, `AOTAutogradCacheBundledTests`	`test_cache_hit_across_processes_pre_grad_custom_pass_*` (default/early/late, with/without uuid)	10
3	Export	`test_export`	`ExportTestsSubprocess`	`test_strict_export_under_pythonoptimize`	1

Observations vs. previous run:

The 2 test_activation_checkpointing_xpu failures and 2 test_profiler_xpu failures from the previous run are now gone (fixed or filtered).
The polar_xpu_float16 failures (19 in op_ut + 1 in op_extended) and the AOTAutogradCache cross-process failures (10) and test_export (1) remain consistent across runs and are not caused by this PR's test files.
The AOTAutogradCache and test_export failures are caused by ModuleNotFoundError: No module named 'functorch.compile' — a test infrastructure issue as noted in comment #5276925378.

shangerxin · 2026-08-14T03:48:16Z

The remain test failures relative to this PR test_aot_autograd_cache.py and test_export.py are due to below error.

The test_aot_autograd_cache failures are relative to errors and the case of test_export also failed with same error.

Need to check the test infrastructure.
ModuleNotFoundError: No module named 'functorch.compile'

shangerxin force-pushed the erxin/migrate_20260810 branch from 7f323a1 to 842bced Compare August 11, 2026 03:01

Copilot started work on behalf of shangerxin August 11, 2026 03:28 View session

Copilot finished work on behalf of shangerxin August 11, 2026 03:31

Copilot started work on behalf of shangerxin August 11, 2026 06:01 View session

Copilot finished work on behalf of shangerxin August 11, 2026 06:03

shangerxin force-pushed the erxin/migrate_20260810 branch from 4df2f05 to 0a2ce0c Compare August 11, 2026 07:02

Copilot started work on behalf of shangerxin August 12, 2026 06:02 View session

Copilot finished work on behalf of shangerxin August 12, 2026 06:03

Copilot started work on behalf of shangerxin August 12, 2026 06:19 View session

Copilot finished work on behalf of shangerxin August 12, 2026 06:20

shangerxin force-pushed the erxin/migrate_20260810 branch from 32655b5 to a08484b Compare August 12, 2026 06:40

Copilot started work on behalf of shangerxin August 13, 2026 02:18 View session

Copilot finished work on behalf of shangerxin August 13, 2026 02:19

shangerxin mentioned this pull request Aug 13, 2026

[Bug Skip]: test_activation_checkpointing_xpu.test_attr_compile_submodules_in_checkpoint_wrapper failed with untimeError( RuntimeError: PyTorch's device state was initialized in the forward pass of a Checkpoint, which is not allowed. #4943

Open

shangerxin added 2 commits August 13, 2026 15:43

Add basic unmigrated files

329d891

shangerxin force-pushed the erxin/migrate_20260810 branch from c2eab35 to f7ad2b0 Compare August 13, 2026 07:44

Copilot started work on behalf of shangerxin August 14, 2026 03:39 View session

Copilot finished work on behalf of shangerxin August 14, 2026 03:40

Uh oh!

Conversation

shangerxin commented Aug 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Compare Link

Changes

Added

Modified

Deleted

What this means overall

Bugs

Uh oh!

shangerxin commented Aug 11, 2026

Uh oh!

Copilot AI commented Aug 11, 2026

Uh oh!

shangerxin commented Aug 11, 2026

Uh oh!

Copilot AI commented Aug 11, 2026

Uh oh!

shangerxin commented Aug 12, 2026

Uh oh!

Copilot AI commented Aug 12, 2026

Uh oh!

shangerxin commented Aug 12, 2026

Uh oh!

Copilot AI commented Aug 12, 2026

Uh oh!

shangerxin commented Aug 13, 2026

Uh oh!

Copilot AI commented Aug 13, 2026

Job 94073605582 — 35 new failures

Job 94041674226 — 1 new failure (op_extended)

Uh oh!

shangerxin commented Aug 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

shangerxin commented Aug 14, 2026

Uh oh!

Copilot AI commented Aug 14, 2026

Job 94394787611 (op_extended) — 1 new failure

Job 94447532967 (op_ut) — 30 new failures

Uh oh!

shangerxin commented Aug 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

shangerxin commented Aug 10, 2026 •

edited

Loading

shangerxin commented Aug 13, 2026 •

edited

Loading