[Core][Distributed] Refactor ipc buffer init in CustomAllreduce (#10030)

Signed-off-by: Hanzhi Zhou <hanzhi713@gmail.com>
This commit is contained in:
Hanzhi Zhou
2024-11-06 23:50:47 -08:00
committed by GitHub
parent d7263a1bb8
commit 6192e9b8fe
9 changed files with 218 additions and 260 deletions
+12 -17
View File
@@ -912,20 +912,16 @@ def get_max_shared_memory_per_block_device_attribute(device: int) -> int:
# custom ar
def init_custom_ar(meta: torch.Tensor, rank_data: torch.Tensor,
handles: List[str], offsets: List[int], rank: int,
full_nvlink: bool) -> int:
return torch.ops._C_custom_ar.init_custom_ar(meta, rank_data, handles,
offsets, rank, full_nvlink)
def init_custom_ar(ipc_tensors: List[torch.Tensor], rank_data: torch.Tensor,
rank: int, full_nvlink: bool) -> int:
return torch.ops._C_custom_ar.init_custom_ar(ipc_tensors, rank_data, rank,
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)
def all_reduce_unreg(fa: int, inp: torch.Tensor, reg_buffer: torch.Tensor,
out: torch.Tensor) -> None:
torch.ops._C_custom_ar.all_reduce_unreg(fa, inp, reg_buffer, out)
def all_reduce(fa: int, inp: torch.Tensor, out: torch.Tensor, reg_buffer: int,
reg_buffer_sz_bytes: int) -> None:
torch.ops._C_custom_ar.all_reduce(fa, inp, out, reg_buffer,
reg_buffer_sz_bytes)
def dispose(fa: int) -> None:
@@ -936,16 +932,15 @@ def meta_size() -> int:
return torch.ops._C_custom_ar.meta_size()
def register_buffer(fa: int, t: torch.Tensor, handles: List[str],
offsets: List[int]) -> None:
return torch.ops._C_custom_ar.register_buffer(fa, t, handles, offsets)
def register_buffer(fa: int, ipc_tensors: List[int]) -> None:
return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[str], List[int]]:
def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)
def register_graph_buffers(fa: int, handles: List[str],
def register_graph_buffers(fa: int, handles: List[List[int]],
offsets: List[List[int]]) -> None:
torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)