[1/N] Elastic EP Milestone 2 (#34861)

Signed-off-by: Yongji Wu <wuyongji317@gmail.com>
Signed-off-by: Itay Alroy <ialroy@nvidia.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Ron Tourgeman <rtourgeman@nvidia.com>
Co-authored-by: Yongji Wu <wuyongji317@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Ron Tourgeman <rtourgeman@nvidia.com>
This commit is contained in:
Itay Alroy
2026-02-28 04:46:42 +00:00
committed by GitHub
co-authored by Yongji Wu Tyler Michael Smith Ron Tourgeman
parent 90805ff464
commit dea268336f
53 changed files with 3612 additions and 1015 deletions
@@ -312,10 +312,19 @@ class PyNcclCommunicator:
)
if stream is None:
stream = current_stream()
if tensor.dtype in [
torch.float8_e5m2,
torch.float8_e4m3fn,
torch.float8_e4m3fnuz,
torch.float8_e5m2fnuz,
]:
nccl_dtype = ncclDataTypeEnum.from_torch(torch.uint8)
else:
nccl_dtype = ncclDataTypeEnum.from_torch(tensor.dtype)
self.nccl.ncclSend(
buffer_type(tensor.data_ptr()),
tensor.numel(),
ncclDataTypeEnum.from_torch(tensor.dtype),
nccl_dtype,
dst,
self.comm,
cudaStream_t(stream.cuda_stream),
@@ -330,10 +339,19 @@ class PyNcclCommunicator:
)
if stream is None:
stream = current_stream()
if tensor.dtype in [
torch.float8_e5m2,
torch.float8_e4m3fn,
torch.float8_e4m3fnuz,
torch.float8_e5m2fnuz,
]:
nccl_dtype = ncclDataTypeEnum.from_torch(torch.uint8)
else:
nccl_dtype = ncclDataTypeEnum.from_torch(tensor.dtype)
self.nccl.ncclRecv(
buffer_type(tensor.data_ptr()),
tensor.numel(),
ncclDataTypeEnum.from_torch(tensor.dtype),
nccl_dtype,
src,
self.comm,
cudaStream_t(stream.cuda_stream),
@@ -384,3 +402,17 @@ class PyNcclCommunicator:
def deregister_comm_window(self, window):
return self.nccl.ncclCommWindowDeregister(self.comm, window)
def batch_isend_irecv(self, p2p_ops: list, stream=None):
if self.disabled:
return
if stream is None:
stream = current_stream()
self.group_start()
for op in p2p_ops:
if op.op is torch.distributed.isend:
self.send(op.tensor, op.group_peer, stream)
elif op.op is torch.distributed.irecv:
self.recv(op.tensor, op.group_peer, stream)
self.group_end()