This pull request includes several changes across multiple files, primarily focusing on refactoring configurations, improving dequantization processes, and updating test cases. The most important changes include adding configuration options for weight transformation, modifying the dequantization process, and updating test configurations and assertions.
This pull request includes several changes across multiple files, primarily focusing on refactoring configurations, improving dequantization processes, and updating test cases. The most important changes include adding configuration options for weight transformation, modifying the dequantization process, and updating test configurations and assertions.
Configuration and Refactoring:
bitblas/ops/general_matmul/tilelang/dense/matmul_tensorcore.py: Addedweight_transform_kindto theMatmulWeightPropagationSchedulerand updated theapply_configmethod to use this new configuration. [1] [2]Dequantization Process:
bitblas/ops/general_matmul/tilelang/dequantize/block_primitive_tensorcore.py: Removedlocal_size_compressedparameter from several methods and updated index calculations accordingly. [1] [2] [3] [4]Testing and Assertions:
testing/python/operators/test_general_matmul_tile_schedule.py: Updated test configurations and assertions for dequantization correctness, including new configurations and additional assertions. [1] [2] [3] [4] [5]testing/python/tilelang/test_tilelang_dequantize_gemm.py: Modified test configurations and added debug prints to verify results. [1] [2] [3]Miscellaneous:
tutorials/.gitignore: Added entries to ignore progress files, debug files, and Jupyter notebook files.tutorials/2.auto_tensorization.ipynb: Added a new tutorial notebook for auto tensorization in BitBLAS.