fused_adam.py 920 B

1234567891011121314151617181920212223242526272829303132
  1. """
  2. Copyright 2020 The Microsoft DeepSpeed Team
  3. """
  4. from .builder import CUDAOpBuilder
  5. class FusedAdamBuilder(CUDAOpBuilder):
  6. BUILD_VAR = "DS_BUILD_FUSED_ADAM"
  7. NAME = "fused_adam"
  8. def __init__(self):
  9. super().__init__(name=self.NAME)
  10. def absolute_name(self):
  11. return f'deepspeed.ops.adam.{self.NAME}_op'
  12. def sources(self):
  13. return ['csrc/adam/fused_adam_frontend.cpp', 'csrc/adam/multi_tensor_adam.cu']
  14. def include_paths(self):
  15. return ['csrc/includes', 'csrc/adam']
  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 not self.is_rocm_pytorch():
  22. nvcc_flags.extend(['-lineinfo',
  23. '--use_fast_math'] + self.compute_capability_args())
  24. return nvcc_flags