Skip to content

[ROCm] Skip unsupported unit tests on MI200#40369

Closed
amd-jianli12 wants to merge 4 commits intoopenxla:mainfrom
ROCm:ci_upstream_ut_fix
Closed

[ROCm] Skip unsupported unit tests on MI200#40369
amd-jianli12 wants to merge 4 commits intoopenxla:mainfrom
ROCm:ci_upstream_ut_fix

Conversation

@amd-jianli12
Copy link
Copy Markdown
Contributor

@amd-jianli12 amd-jianli12 commented Apr 3, 2026

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.
@i-chaochen i-chaochen requested a review from dextero April 3, 2026 21:33
@i-chaochen i-chaochen added the AMD-GPU XLA on AMD GPU label Apr 3, 2026
@github-project-automation github-project-automation bot moved this to Pending Google Review in OpenXLA:GPU - AMD/ROCm Apr 3, 2026
@i-chaochen
Copy link
Copy Markdown
Contributor

I have skipped MOSAIC GPU test from acccb3d as it's NV-only backend.

@i-chaochen
Copy link
Copy Markdown
Contributor

If we can bump up ROCm7.2.1 #40405 CI ROCm JAX will be green as well.

And CI / XLA Linux X86 GPU ONEAPI is irrelevent to this PR.

@i-chaochen
Copy link
Copy Markdown
Contributor

i-chaochen commented Apr 6, 2026

Hi @dimitar-asenov could you review this PR first please? once this is merged we can have a green XLA CI and then we can bazel-cache and CI will be much faster.

Next one is #40405 for JAX UT on XLA


TEST_F(AlgorithmTest, Algorithm6xBF16) {
if (GpuComputeComp().IsRocm()) {
if (GpuComputeComp().rocm_compute_capability()->gfx9_mi200())
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.

nit: the style guide says we should put '{' '}' around if blocks (even if it is just a single line).

Also, why not just make it a && condition together with the previous if?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

Yes, I should. Will update.

copybara-service bot pushed a commit that referenced this pull request Apr 7, 2026
Imported from GitHub PR #40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ce by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f4 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=#40369 from ROCm:ci_upstream_ut_fix b632dc5
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 7, 2026
Imported from GitHub PR openxla/xla#40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ceeca14167b9f024edc9f7ea5e4e1dc62c5 by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f487a59729525713b73369187b5de765554 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5267cacc1ae512727c281acdf1729508b9 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#40369 from ROCm:ci_upstream_ut_fix b632dc5267cacc1ae512727c281acdf1729508b9
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit that referenced this pull request Apr 7, 2026
Imported from GitHub PR #40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ce by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f4 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=#40369 from ROCm:ci_upstream_ut_fix b632dc5
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 7, 2026
Imported from GitHub PR openxla/xla#40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ceeca14167b9f024edc9f7ea5e4e1dc62c5 by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f487a59729525713b73369187b5de765554 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5267cacc1ae512727c281acdf1729508b9 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#40369 from ROCm:ci_upstream_ut_fix b632dc5267cacc1ae512727c281acdf1729508b9
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit that referenced this pull request Apr 7, 2026
Imported from GitHub PR #40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ce by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f4 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=#40369 from ROCm:ci_upstream_ut_fix b632dc5
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 7, 2026
Imported from GitHub PR openxla/xla#40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ceeca14167b9f024edc9f7ea5e4e1dc62c5 by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f487a59729525713b73369187b5de765554 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5267cacc1ae512727c281acdf1729508b9 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#40369 from ROCm:ci_upstream_ut_fix b632dc5267cacc1ae512727c281acdf1729508b9
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit that referenced this pull request Apr 7, 2026
Imported from GitHub PR #40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ce by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f4 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=#40369 from ROCm:ci_upstream_ut_fix b632dc5
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 7, 2026
Imported from GitHub PR openxla/xla#40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ceeca14167b9f024edc9f7ea5e4e1dc62c5 by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f487a59729525713b73369187b5de765554 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5267cacc1ae512727c281acdf1729508b9 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#40369 from ROCm:ci_upstream_ut_fix b632dc5267cacc1ae512727c281acdf1729508b9
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit that referenced this pull request Apr 8, 2026
Imported from GitHub PR #40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ce by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f4 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=#40369 from ROCm:ci_upstream_ut_fix b632dc5
PiperOrigin-RevId: 895851209
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 8, 2026
Imported from GitHub PR openxla/xla#40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ceeca14167b9f024edc9f7ea5e4e1dc62c5 by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f487a59729525713b73369187b5de765554 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5267cacc1ae512727c281acdf1729508b9 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#40369 from ROCm:ci_upstream_ut_fix b632dc5267cacc1ae512727c281acdf1729508b9
PiperOrigin-RevId: 895851209
@copybara-service copybara-service bot closed this in d194a36 Apr 8, 2026
@github-project-automation github-project-automation bot moved this from Pending Google Review to Done in OpenXLA:GPU - AMD/ROCm Apr 8, 2026
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 8, 2026
FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#40369 from ROCm:ci_upstream_ut_fix b632dc5267cacc1ae512727c281acdf1729508b9
PiperOrigin-RevId: 896242389
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 8, 2026
Imported from GitHub PR openxla/xla#40369

Skip multi-pass dot tests for BF16 on MI200 since there's no valid algorithm being returned by hipBlasLt.

📝 Summary of Changes
hipBlasLt is returning empty algorithms for multi-pass bf16_bf16_f32 dot configuration on MI200, so disable these tests.

🎯 Justification
This change is to make the following tests pass on MI200 GPU:
//xla/service/gpu:dot_algorithm_support_test_amdgpu_any
//xla/backends/gpu/codegen/triton:dot_algorithms_test_amdgpu_any

🚀 Kind of Contribution
🐛 Bug Fix

@xla-rotation Would you kindly have a look please?

Copybara import of the project:

--
82352ceeca14167b9f024edc9f7ea5e4e1dc62c5 by Jian Li <Jian.Li@amd.com>:

[ROCm] Skip unsupported unit tests on MI200

Skip multi-pass dot tests for BF16 on MI200 since there's no valid
algorithm being returned by hipBlasLt.

--
e6f64f487a59729525713b73369187b5de765554 by Chao Chen <cchen104@amd.com>:

skipped Mosaic GPU test as it's NV-only

--
b632dc5267cacc1ae512727c281acdf1729508b9 by Jian Li <Jian.Li@amd.com>:

Address review: Enforce style compliance and flatten conditionals

Merging this change closes #40369

PiperOrigin-RevId: 896279515
copybara-service bot pushed a commit to tensorflow/tensorflow that referenced this pull request Apr 8, 2026
FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#40369 from ROCm:ci_upstream_ut_fix b632dc5267cacc1ae512727c281acdf1729508b9
PiperOrigin-RevId: 896242242
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

AMD-GPU XLA on AMD GPU

Projects

Archived in project

Development

Successfully merging this pull request may close these issues.

4 participants