Optimize model execution with CUDA graph (#1926)

Co-authored-by: Chen Shen <scv119@gmail.com>
Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
This commit is contained in:
Woosuk Kwon
2023-12-16 21:12:08 -08:00
committed by GitHub
co-authored by Chen Shen Antoni Baum
parent eed74a558f
commit 37ca558103
34 changed files with 557 additions and 254 deletions
@@ -1,8 +1,10 @@
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,
)
@@ -15,8 +17,12 @@ def tensor_model_parallel_all_reduce(input_):
if get_tensor_model_parallel_world_size() == 1:
return input_
# All-reduce.
torch.distributed.all_reduce(input_,
group=get_tensor_model_parallel_group())
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())
return input_