[torch.compile] register allreduce operations as custom ops (#8526)
This commit is contained in:
@@ -870,12 +870,6 @@ def init_custom_ar(meta: torch.Tensor, rank_data: torch.Tensor,
|
||||
offsets, rank, full_nvlink)
|
||||
|
||||
|
||||
def should_custom_ar(inp: torch.Tensor, max_size: int, world_size: int,
|
||||
full_nvlink: bool) -> bool:
|
||||
return torch.ops._C_custom_ar.should_custom_ar(inp, max_size, world_size,
|
||||
full_nvlink)
|
||||
|
||||
|
||||
def all_reduce_reg(fa: int, inp: torch.Tensor, out: torch.Tensor) -> None:
|
||||
torch.ops._C_custom_ar.all_reduce_reg(fa, inp, out)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user