A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 07:00 UTC
[Refactor][MISC] del redundant code in ParallelConfig.postinit (#10614)
Signed-off-by: MengqingCao <cmq0113@163.com>
Mengqing Cao committed
Nov 25, 2024 at 13:14 UTC
7ea3cd7c3e9fa1db06cdf8ad1973237b061b7d64
1 file changed
+5
-10
vllm/config.py
+5
-10
@@ -998,20 +998,15 @@ class ParallelConfig:
998
raise ValueError(f"worker-use-ray can't be used with "
999
f"distributed executor backend "
1000
f"'{self.distributed_executor_backend}'.")
1001
-
1002
- if current_platform.is_tpu() and self.world_size > 1:
1003
- if self.distributed_executor_backend is None:
1004
- self.distributed_executor_backend = "ray"
1005
- if self.distributed_executor_backend != "ray":
1006
- raise ValueError(
1007
- "TPU backend only supports Ray for distributed inference.")
1008
-
1009
- if current_platform.is_hpu() and self.world_size > 1:
1001
+ ray_only_devices = ["tpu", "hpu"]
1002
+ if (current_platform.device_type in ray_only_devices
1003
+ and self.world_size > 1):
1004
if self.distributed_executor_backend is None:
1005
self.distributed_executor_backend = "ray"
1006
if self.distributed_executor_backend != "ray":
1007
raise ValueError(
1014
- "HPU backend only supports Ray for distributed inference.")
1008
+ f"{current_platform.device_type.upper()} backend only "
1009
+ "supports Ray for distributed inference.")
1010
1011
if self.distributed_executor_backend is None and self.world_size > 1:
1012
# We use multiprocessing by default if world_size fits on the