fused_lamb.py 1.1 KB

1234567891011121314151617181920212223242526272829303132333435363738
  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. nvcc_flags = ['-O3'] + self.version_dependent_macros()
  21. if self.is_rocm_pytorch():
  22. ROCM_MAJOR, ROCM_MINOR = self.installed_rocm_version()
  23. nvcc_flags += [
  24. '-DROCM_VERSION_MAJOR=%s' % ROCM_MAJOR,
  25. '-DROCM_VERSION_MINOR=%s' % ROCM_MINOR
  26. ]
  27. else:
  28. nvcc_flags.extend(['-lineinfo',
  29. '--use_fast_math'] + self.compute_capability_args())
  30. return nvcc_flags