Skip to content

Latest commit

 

History

History
67 lines (53 loc) · 2.57 KB

README.md

File metadata and controls

67 lines (53 loc) · 2.57 KB
Raw

Unit test guidelines

The following 2 sections describe existing unit-test cases for the configs supported by the custom ops in the main repository. If a new feature is added (such as num_workers or hw_ctx), the developer is expected to write a pytest compatible unit test case, add it to the existing infrastructure and add the information to the section below.

FAQ

  • Pass --capture=tee-sys --verbose to the pytest command to display more detailed error messages if there aren't any.
  • If you run into this error OMP: Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized, please set KMP_DUPLICATE_LIB_OK=TRUE
  • To run a single test, "-k" option can be used.
  • To add new argument, use conftest.py based on existing examples in the file.
  • Example: pytest --num_workers 2 -k test_QLinear_quantmode1 test_qlinear.py --capture=tee-sys --verbose

Python tests

pytest test_qlinear.py
pytest --w_bit 3 test_qlinear_pergrp.py
pytest --w_bit 4 test_qlinear_pergrp.py

# Flash Attention
pytest --quant_mode w4abf16 test_opt_flash_attention.py
pytest --quant_mode w8a8    test_opt_flash_attention.py

pytest --quant_mode w4abf16 test_llama_flash_attention.py
pytest --quant_mode w8a8    test_llama_flash_attention.py

pytest --quant_mode w4abf16 test_qwen2_flash_attention.py

pytest --quant_mode w4abf16 test_chatglm3_flash_attention.py

pytest --quant_mode w4abf16 test_phi_flash_attention.py

pytest --quant_mode w4abf16 test_mistral_flash_attention.py

# Fast MLP
pytest test_llama_fast_mlp.py --w_bit 3
pytest test_llama_fast_mlp.py --w_bit 4

Additional STX specific tests

pytest --quant_mode w8a16 test_qlinear.py
pytest --quant_mode w8a16 test_opt_flash_attention.py
pytest --quant_mode w8a16 test_llama_flash_attention.py

GEMM DML

To run DirectML GEMM test, please copy DirectML.dll and DirectML.Debug.dll into tests/python and well as into ```root`` folder of conda environment

xcopy ..\..\ops\torch_cpp\external\packages\Microsoft.AI.DirectML.1.15.0\bin\x64-win\*.dll .
xcopy ..\..\ops\torch_cpp\external\packages\Microsoft.AI.DirectML.1.15.0\bin\x64-win\*.dll <user>\AppData\Local\anaconda3\envs\ryzenai-transformers\

pytest test_gemm_dml_torch.py

Tests for CPU and AIE+CPU Experimental ops

Tests for Python tiling, CPU work are here. These do not need to be tested for AIE development but are used for experimental work.

pytest test_tiling.py
pytest test_softmax.py
pytest test_qlinear_cpu.py
pytest test_scalar_mult.py
pytest test_linear_bfloat16.py

# v0 kernel specific
pytest test_qlinear_pytiling.py