microsoft / BitBLAS

BitBLAS is a library to support mixed-precision matrix multiplications, especially for quantized LLM deployment.
MIT License
190 stars 21 forks source link

[BugFix] Fix a bug in Static shape build #53

Closed LeiWang1999 closed 3 weeks ago

LeiWang1999 commented 3 weeks ago

This pull request includes changes to versioning and improvements to the handling of different cases in several functions across the bitblas Python package. The version number has been updated in the VERSION file and in the __version__ variable in python/bitblas/__init__.py. The apply_config function in python/bitblas/gpu/gemv.py and the get_vectorize_factor function in python/bitblas/gpu/gemv_dequantize.py have been updated to handle cases where the length of block_info.iters and sch.get_loops(block_b) is 4. In python/bitblas/wrapper/general.py, the legalize_c function has been updated to handle cases where dynamic_symbolic_set is not empty.

Here are the most important changes:

Versioning:

Handling of different cases: