Skip to content

Latest commit

 

History

History
90 lines (62 loc) · 2.46 KB

unit_tests.md

File metadata and controls

90 lines (62 loc) · 2.46 KB

Dependencies

  1. Tests related
  • pytest
  • pytest-xdist : (Optional Plugin) If you wish to run in parallel. A known limitation is that this won't capture the -s flag of pytest, so if you wish to debug and capture the output, it is not recommended to use this.
pip install pytest
# Optional
pip install pytest-xdist

Triton Unit Tests

This test case lies under triton/python/tests/unit/language.

[Note: ] The other tests triton/python/tests/ are not ported support yet.

Important

  1. If you are running tests including math lib, you should set the following flag to the triton lib. It is recommended to set this before a full coverage test.

    export TRITON_LIBDEVICE_PATH={abs-path-to-triton}/python/triton/third_party/xpu/

  2. To run the tests, one needs to manually add the import intel_extension_for_pytorch before using xpu.

  3. If there are cases using triton.compile, one needs to use kernel = triton.compile(name, device_type='xpu') instead of just using triton.compile(name).

Run single test

# Please make sure added `import intel_extension_for_pytorch` to the test_file.py first.
(tests/unit/language)$
pytest -sv -k [your pattern(eg. test_bin_op)] test_file.py --device=xpu

The -s flag will capture all the outputs. If you don't use it, then just run without this flag.

Run all tests

# Please make sure added `import intel_extension_for_pytorch` to the test_file.py first.
(tests/unit/language)$
pytest -v test_file.py --device=xpu

Note that the test_file.py could be ., meaning running all tests on the current folder.

Optional

Run tests in parallel

You need to install pytest-xdist first.

pytest -n 8 ... (The same as before)

Torch Inductor Tests

This test case lies under pytorch/tests/inductor. We are in an active development process, some changes are internally going to be upstreamed to PyTorch shortly. We will update the docs then.

Appendix

Skipping tests

There are two ways to skip:

  1. Skip the whole test function by using a decorator:
@pytest.mark.skip("skip Reason")
# Or use @pytest.mark.skipIf("skip Reason")
def test_my_func():
  1. Skip inside function
def test_my_func():
    pytest.skip("Skip Reason:")
    # OR use skipIf