Nikita Vedeneev 46f16b9363 Improve bsr @ strided performance in baddmm for bfloat16/half with Triton kernels. (#88078) ... As per title. Additionally we also introduce support for: - Rectangular block sizes which are powers of 2 and at least 16 (triton's `dot` limitation). - Batch support with broadcasting for either of the arguments. Pull Request resolved: https://github.com/pytorch/pytorch/pull/88078 Approved by: https://github.com/cpuhrsch		2023-01-26 07:58:27 +00:00
..
__init__.py	Revert "Improve bsr @ strided performance in baddmm for bfloat16/half with Triton kernels. (#88078)"	2023-01-19 23:37:59 +00:00
_triton_ops.py	Improve bsr @ strided performance in baddmm for bfloat16/half with Triton kernels. (#88078)	2023-01-26 07:58:27 +00:00