Testing

The MindIE SD test suite supports running with or without Ascend NPU hardware. CPU-compatible tests can be executed even without NPU hardware.

Test Environment Description

Tests are divided into two categories:

Type	Description	Requires NPU
CPU-compatible tests	Configuration parsing, utility functions, quantization parameter validation, compilation logic, etc.	No
NPU-dependent tests	Custom operator accuracy, Flash Attention, model tensor operations, etc.	Yes

Test Entry Points

Entry Point 1: CPU-Friendly Unit Tests (Recommended for users without NPU)

run_UT_test.sh always runs in CPU mode, suitable for development environments without NPU hardware.

python -m pip install -r requirements.txt
python -m pip install -r requirements-test.txt
bash tests/run_UT_test.sh

Default artifacts are generated in the tests/UT/ directory, including:

• run_UT.log
• final.xml
• coverage.xml
• htmlcov/

The tests/scripts/check_coverage.py script in the repository is used in CI to verify the coverage gate for newly added Python files.

Entry Point 2: Full Tests (Supports three modes)

run_test.sh controls the test scope via parameters and supports the following three modes:

1. Full tests (default)

Run all CPU-compatible tests and NPU-dependent tests:

cd tests/
bash run_test.sh --all

When no parameters are passed, full tests are executed by default:

cd tests/
bash run_test.sh

2. Run CPU-compatible tests only (no NPU hardware required)

cd tests/
bash run_test.sh --cpu_only

3. Run NPU-dependent tests only (NPU hardware required)

cd tests/
bash run_test.sh --npu_only

LA Operator Accuracy Testing

This section describes the method for self-testing the accuracy of LA operators in the MindIE SD repository.

1. To switch the installed version, first uninstall the current MindIE SD:

   python -m pip uninstall mindiesd
   

2. Modify the tests/plugin/la_acc_prof.py file, select Option 1 or Option 2, and load the test_la.csv or enumerated_cases.csv file to test the accuracy of LA operators under the specified shapes.

   • ./tests/plugin/test_la.csv: Contains input shapes for commonly used SD models.
   • enumerated_cases.csv: Contains various enumerated shapes.

3. After making the modifications, run the following command:

   cd tests
   python plugin/la_acc_prof.py
   

Upon successful execution, a result file is generated in the repository directory, recording the similarity between LA and FAScore. This can be used to inspect the operator's accuracy performance under the target shapes.

Test Writing Conventions

Test Organization Principles

Test directories under tests/ correspond to the mindiesd/ module hierarchy:

Source Module	Corresponding Test Directory
mindiesd/layers/	tests/layers/
mindiesd/cache/	tests/cache/
mindiesd/compilation/	tests/compilation/
mindiesd/quantization/	tests/quantization/

When adding new code, check whether the corresponding test directory already has test files, and append to existing files or create new ones to cover the new functionality.

Test Performance Conventions

• If a single test file takes more than 500 seconds to run, split it into multiple files
• Prefer reusing existing device/model loading to reduce repeated initialization overhead
• CPU-friendly tests should preferably be verified using bash tests/run_UT_test.sh