Test Guide

Test Guide#

The MindIE SD test suite supports running with or without Ascend NPU hardware. CPU-compatible tests can be executed even when no NPU hardware is available.

Test environment overview#

Tests fall into two categories:

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

Test entry points#

Option 1: CPU-friendly unit tests (recommended for users without NPU)#

run_UT_test.sh always runs in CPU mode, making it 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

Artifacts are generated under tests/UT/, including:

run_UT.log
final.xml
coverage.xml
htmlcov/

The repository also provides tests/scripts/check_coverage.py for CI coverage gating on newly added Python files.

Option 2: Full tests (three modes)#

run_test.sh accepts a flag to control the test scope. Three modes are available:

1. All tests (default)

Run both CPU-compatible and NPU-dependent tests:

cd tests/
bash run_test.sh --all

When no flag is given, the default is to run all tests:

cd tests/
bash run_test.sh

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

cd tests/
bash run_test.sh --cpu_only

3. NPU-dependent tests only (requires NPU hardware)

cd tests/
bash run_test.sh --npu_only

LA Operator Accuracy Test#

This section describes how to run LA operator accuracy verification in the MindIE SD repository.

If needed, uninstall the currently installed MindIE SD package first:
```
pip uninstall mindiesd
```
Update tests/plugin/la_acc_prof.py, choose Option 1 or Option 2, and load either test_la.csv or enumerated_cases.csv to verify LA accuracy under the required shapes.
- ./tests/plugin/test_la.csv: common input shapes used by SD models
- enumerated_cases.csv: enumerated shape combinations
Run the script:
```
cd tests
python plugin/la_acc_prof.py
```

After the run, result files are generated in the repository root and can be used to inspect similarity between LA and FAScore outputs.

Common Exceptions#

When using MindIE SD for inference, users are responsible for the safety of model files such as weights, configuration files, and model code. Common exceptions include:

If default model configuration values are changed during initialization, interfaces may be affected; excessively large weights or configuration values may trigger out-of-memory errors such as RuntimeError: NPU out of memory. Tried to allocate xxx GiB..
Large tensor shapes during inference may also trigger similar out-of-memory errors.
Invalid input or environment mismatch may raise exceptions that should be handled by upper-layer services.

Exception Type	Description
ZeroDivisionError	Division by zero.
ValueError	Invalid parameter value.