biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[core][distributed] custom allreduce when pp size > 1 (#6117)

Browse Source
This commit is contained in:
youkaichao
2024-07-03 14:41:32 -07:00
committed by GitHub
parent 47f0954af0
commit 3c6325f0fc
2 changed files with 17 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
vllm/config.py
View File

@@ -723,17 +723,11 @@ class ParallelConfig:
if self.distributed_executor_backend == "ray":
from vllm.executor import ray_utils
ray_utils.assert_ray_available()
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.")
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.")
if self.ray_workers_use_nsight and (
not self.distributed_executor_backend == "ray"):
raise ValueError("Unable to use nsight profiling unless workers "