Remove dependency on CuPy (#2152)
This commit is contained in:
@@ -1,10 +1,8 @@
|
||||
import torch
|
||||
|
||||
from vllm.model_executor.parallel_utils import cupy_utils
|
||||
from vllm.model_executor.parallel_utils.parallel_state import (
|
||||
get_tensor_model_parallel_world_size,
|
||||
get_tensor_model_parallel_group,
|
||||
is_custom_nccl_enabled_for_all_reduce,
|
||||
)
|
||||
|
||||
|
||||
@@ -17,12 +15,8 @@ def tensor_model_parallel_all_reduce(input_):
|
||||
if get_tensor_model_parallel_world_size() == 1:
|
||||
return input_
|
||||
# All-reduce.
|
||||
if is_custom_nccl_enabled_for_all_reduce():
|
||||
# TODO: support multiple parallel groups.
|
||||
cupy_utils.all_reduce(input_)
|
||||
else:
|
||||
torch.distributed.all_reduce(input_,
|
||||
group=get_tensor_model_parallel_group())
|
||||
torch.distributed.all_reduce(input_,
|
||||
group=get_tensor_model_parallel_group())
|
||||
return input_
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user