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 15, 09:45 UTC
[Frontend] deprecate `--device` arg (#18399)
Signed-off-by: Kebe <mail@kebe7jun.com>
Kebe committed
May 21, 2025 at 16:21 UTC
5d7f5452044c2a21dcd75fd8d4266a25052fe43b
3 files changed
+9
-4
docs/source/getting_started/installation/gpu/xpu.inc.md
-1
@@ -66,7 +66,6 @@ XPU platform supports **tensor parallel** inference/serving and also supports **
66
python -m vllm.entrypoints.openai.api_server \
67
--model=facebook/opt-13b \
68
--dtype=bfloat16 \
69
- --device=xpu \
69
--max_model_len=1024 \
70
--distributed-executor-backend=ray \
71
--pipeline-parallel-size=2 \
vllm/config.py
+5
-1
@@ -2201,7 +2201,11 @@ class DeviceConfig:
2201
"""Configuration for the device to use for vLLM execution."""
2202
2203
device: Union[Device, torch.device] = "auto"
2204
- """Device type for vLLM execution."""
2204
+ """Device type for vLLM execution.
2205
+ This parameter is deprecated and will be
2206
+ removed in a future release.
2207
+ It will now be set automatically based
2208
+ on the current platform."""
2209
device_type: str = field(init=False)
2210
"""Device type from the current platform. This is set in
2211
`__post_init__`."""
vllm/engine/arg_utils.py
+4
-2
@@ -737,7 +737,9 @@ class EngineArgs:
737
title="DeviceConfig",
738
description=DeviceConfig.__doc__,
739
)
740
- device_group.add_argument("--device", **device_kwargs["device"])
740
+ device_group.add_argument("--device",
741
+ **device_kwargs["device"],
742
+ deprecated=True)
743
744
# Speculative arguments
745
speculative_group = parser.add_argument_group(
@@ -977,7 +979,7 @@ class EngineArgs:
979
from vllm.platforms import current_platform
980
current_platform.pre_register_and_update()
981
980
- device_config = DeviceConfig(device=self.device)
982
+ device_config = DeviceConfig(device=current_platform.device_type)
983
model_config = self.create_model_config()
984
985
# * If VLLM_USE_V1 is unset, we enable V1 for "supported features"