Notice: Upcoming deprecation of CUTLASS Python interface

[mnicely]

Starting with the next release of CUTLASS, the CUTLASS Python interface will be deprecated and no longer maintained. The code will remain in the CUTLASS repository for experimentation, but is not planned to be updated or supported moving forward. The final tag for the nvidia-cutlass PyPI wheel is planned to be 3.8.0.

[manishucsd]

nvidia-cutlass PyPI wheels also have cutlass_library code that we install and use.

>>> import cutlass_library
>>> dir(cutlass_library)
['ArchitectureNames', 'BlasMode', 'BlasModeTag', 'CalculateSmemUsage', 'ComplexMultiplyOp', 'ComplexTransform', 'ComplexTransformTag', 'ComplexTransformTag3x', 'ConvKind', 'ConvKindNames', 'ConvKindTag', 'ConvMode', 'DataType', 'DataTypeNames', 'DataTypeSize', 'DataTypeTag', 'DiagType', 'DiagTypeTag', 'Direct2dConvFixedStrideDilationTileDescription', 'EpilogueFunctor', 'EpilogueFunctor3x', 'EpilogueFunctor3xTag', 'EpilogueFunctorTag', 'EpilogueScheduleSuffixes', 'EpilogueScheduleTag', 'EpilogueScheduleType', 'FillMode', 'FillModeTag', 'GemmKind', 'GemmKindNames', 'GemmUniversalMode', 'GeneratorTarget', 'GeneratorTargetNames', 'GroupMode', 'GroupModeNames', 'GroupModeTag', 'GroupScheduleMode', 'GroupScheduleModeTag', 'IteratorAlgorithm', 'IteratorAlgorithmNames', 'IteratorAlgorithmTag', 'KernelScheduleSuffixes', 'KernelScheduleTag', 'KernelScheduleType', 'LayoutTag', 'LayoutType', 'MathInstruction', 'MathOperation', 'MathOperationTag', 'OpcodeClass', 'OpcodeClassNames', 'OpcodeClassTag', 'OperationKind', 'OperationKindNames', 'RankKKind', 'RankKKindNames', 'RealComplexBijection', 'SharedMemPerCC', 'ShortComplexLayoutNames', 'ShortDataTypeNames', 'ShortDiagTypeNames', 'ShortFillModeNames', 'ShortGroupScheduleModeNames', 'ShortLayoutTypeNames', 'ShortSideModeNames', 'SideMode', 'SideModeTag', 'SplitKMode', 'StrideSupport', 'StrideSupportNames', 'StrideSupportTag', 'SubstituteTemplate', 'SwizzlingFunctor', 'SwizzlingFunctorTag', 'SymmKind', 'SymmKindNames', 'SymmetricTensorDescription', 'Target', 'TensorDescription', 'TileDescription', 'TileSchedulerSuffixes', 'TileSchedulerTag', 'TileSchedulerType', 'TransposedLayout', 'TriangularTensorDescription', 'TrmmKind', 'TrmmKindNames', '__builtins__', '__cached__', '__doc__', '__file__', '__loader__', '__name__', '__package__', '__path__', '__spec__', 'conv2d_operation', 'conv3d_operation', 'conv3x_emitter', 'enum', 'enum_auto', 'gemm_operation', 'generator', 'get_complex_from_real', 'get_real_from_complex', 'install_source_path', 'is_complex', 'library', 'manifest', 'os', 'rank_2k_operation', 'rank_k_operation', 're', 'sm90_shapes', 'sm90_utils', 'source_path', 'symm_operation', 'sys', 'trmm_operation']

All the code in python/cutlass_library, containing many enums, their C++ type tags, and shortnames for generating procedural names, and device-level emitters. One can think of this as a pythonic IR that gives its user access at only the CUTLASS device-level. We would want at least device-level pythonic emitters for Blackwell or even better would be an IR solution not just at the device-level but at the lower-levels. If we have that we can depreciate nvidia-cutlass o.w. it will be good to have it for the functions, enums, classes shared above.

python/cutlass_library should still be maintained because profiler uses and I see Blackwell code in the profiler?