fused_lamb.py 823 B

12345678910111213141516171819202122232425262728293031
  1. """
  2. Copyright 2020 The Microsoft DeepSpeed Team
  3. """
  4. from .builder import CUDAOpBuilder
  5. class FusedLambBuilder(CUDAOpBuilder):
  6. BUILD_VAR = 'DS_BUILD_FUSED_LAMB'
  7. NAME = "fused_lamb"
  8. def __init__(self):
  9. super().__init__(name=self.NAME)
  10. def absolute_name(self):
  11. return f'deepspeed.ops.lamb.{self.NAME}_op'
  12. def sources(self):
  13. return ['csrc/lamb/fused_lamb_cuda.cpp', 'csrc/lamb/fused_lamb_cuda_kernel.cu']
  14. def include_paths(self):
  15. return ['csrc/includes']
  16. def cxx_args(self):
  17. args = super().cxx_args()
  18. return args + self.version_dependent_macros()
  19. def nvcc_args(self):
  20. return ['-lineinfo',
  21. '-O3',
  22. '--use_fast_math'
  23. ] + self.version_dependent_macros() + self.compute_capability_args()