This pull request includes changes to the bitblas library and its associated tests. The most significant changes include enabling debug output in QuickStart.md, modifying the forward method in python/bitblas/module/__init__.py and python/bitblas/ops/general_matmul_splitk.py, and adjusting the testing scripts testing/python/operators/test_general_matmul_fp8.py and testing/python/operators/test_general_matmul_splitk_ops.py.
Debug output:
docs/QuickStart.md: Enabled debug output in three examples using bitblas.set_debug_level("Debug"). [1][2][3]
Codebase modifications:
python/bitblas/module/__init__.py: Modified the forward method to include a stream variable and a stream_handle variable, which is passed to the lib.call method.
python/bitblas/ops/general_matmul_splitk.py: Adjusted the forward method to change the shape of the output tensor, create a new sk_output tensor, and use the torch.sum method to populate the output tensor. [1][2][3]
testing/python/operators/test_general_matmul_splitk_ops.py: Made several changes to the test methods, including adding a SplitK parameter, replacing the get_codegen_result method with a comparison of output_bitblas and output_torch, and adding a map_torch_type method to map input types to torch types. [1][2][3]
This pull request includes changes to the
bitblaslibrary and its associated tests. The most significant changes include enabling debug output inQuickStart.md, modifying theforwardmethod inpython/bitblas/module/__init__.pyandpython/bitblas/ops/general_matmul_splitk.py, and adjusting the testing scriptstesting/python/operators/test_general_matmul_fp8.pyandtesting/python/operators/test_general_matmul_splitk_ops.py.Debug output:
docs/QuickStart.md: Enabled debug output in three examples usingbitblas.set_debug_level("Debug"). [1] [2] [3]Codebase modifications:
python/bitblas/module/__init__.py: Modified theforwardmethod to include astreamvariable and astream_handlevariable, which is passed to thelib.callmethod.python/bitblas/ops/general_matmul_splitk.py: Adjusted theforwardmethod to change the shape of theoutputtensor, create a newsk_outputtensor, and use thetorch.summethod to populate theoutputtensor. [1] [2] [3]Testing script adjustments:
testing/python/operators/test_general_matmul_fp8.py: Commented out the call tobitblas.testing.main()and added a call totest_matmul_torch_forward_weight_dequantize.testing/python/operators/test_general_matmul_splitk_ops.py: Made several changes to the test methods, including adding aSplitKparameter, replacing theget_codegen_resultmethod with a comparison ofoutput_bitblasandoutput_torch, and adding amap_torch_typemethod to map input types to torch types. [1] [2] [3]