chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Roll back to use vllm custom allreduce (#3006)

Browse Source
This commit is contained in:
Lianmin Zheng
2025-01-20 04:03:15 -08:00
committed by GitHub
parent dc1881326f
commit 89cd923581
10 changed files with 18 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
python/sglang/srt/model_executor/cuda_graph_runner.py
View File

@@ -33,7 +33,6 @@ from sglang.srt.model_executor.forward_batch_info import (
ForwardBatch,
ForwardMode,
)
from sglang.srt.utils import monkey_patch_vllm_all_gather
if TYPE_CHECKING:
from sglang.srt.model_executor.model_runner import ModelRunner
@@ -72,7 +71,6 @@ def patch_model(
try:
if enable_compile:
_to_torch(model, reverse=False, batch_size=batch_size)
monkey_patch_vllm_all_gather()
backup_ca_comm = tp_group.ca_comm
# Use custom-allreduce here.
# We found the custom allreduce is much faster than the built-in allreduce in torch,
@@ -88,7 +86,6 @@ def patch_model(
finally:
if enable_compile:
_to_torch(model, reverse=True, batch_size=batch_size)
monkey_patch_vllm_all_gather(reverse=True)
tp_group.ca_comm = backup_ca_comm