xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 04:34:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

Disable custom all reduce by default (#2808)

Browse Source
This commit is contained in:
Woosuk Kwon 2024-02-08 09:58:03 -08:00 committed by GitHub
parent 65b89d16ee
commit 3711811b1d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 18 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
vllm/config.py
View File

@ -388,16 +388,26 @@ class ParallelConfig:
if self.pipeline_parallel_size > 1:
raise NotImplementedError(
"Pipeline parallelism is not supported yet.")
if is_hip():
if not self.disable_custom_all_reduce and self.world_size > 1:
if is_hip():
self.disable_custom_all_reduce = True
logger.info(
"Disabled the custom all-reduce kernel because it is not "
"supported on AMD GPUs.")
elif self.pipeline_parallel_size > 1:
self.disable_custom_all_reduce = True
logger.info(
"Disabled the custom all-reduce kernel because it is not "
"supported with pipeline parallelism.")
# FIXME(woosuk): Fix the stability issues and re-enable the custom
# all-reduce kernel.
if not self.disable_custom_all_reduce and self.world_size > 1:
self.disable_custom_all_reduce = True
logger.info(
"Disabled the custom all-reduce kernel because it is not "
"supported on AMD GPUs.")
elif self.pipeline_parallel_size > 1:
self.disable_custom_all_reduce = True
logger.info(
"Disabled the custom all-reduce kernel because it is not "
"supported with pipeline parallelism.")
"Custom all-reduce kernels are temporarily disabled due to "
"stability issues. We will re-enable them once the issues are "
"resolved.")
class SchedulerConfig: