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 11, 00:00 UTC
[XPU] Remove distributed_executor_backend check (#30760)
Signed-off-by: sihao.li <sihao.li@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
sihao_li committed
Dec 24, 2025 at 13:34 UTC
471ddb99a059bc32486d566e369887a7bc618c14
1 file changed
-27
vllm/platforms/xpu.py
-27
@@ -7,7 +7,6 @@ from typing import TYPE_CHECKING, Optional
7
8
import torch
9
10
-import vllm.envs as envs
10
from vllm.attention.backends.registry import AttentionBackendEnum
11
from vllm.logger import init_logger
12
@@ -168,32 +167,6 @@ class XPUPlatform(Platform):
167
if vllm_config.kv_transfer_config is not None:
168
vllm_config.kv_transfer_config.enable_permute_local_kv = True
169
171
- if parallel_config.distributed_executor_backend is None:
172
- if parallel_config.world_size > 1:
173
- parallel_config.distributed_executor_backend = "ray"
174
- else:
175
- parallel_config.distributed_executor_backend = "uni"
176
- elif parallel_config.distributed_executor_backend == "mp":
177
- # FIXME(kunshang):
178
- # spawn needs calling `if __name__ == '__main__':`
179
- # fork is not supported for xpu start new process.
180
- if envs.VLLM_WORKER_MULTIPROC_METHOD != "spawn":
181
- os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
182
- logger.warning(
183
- "Please use spawn as start method if you want to use mp."
184
- )
185
- elif (
186
- parallel_config.distributed_executor_backend != "ray"
187
- and parallel_config.distributed_executor_backend != "uni"
188
- and parallel_config.distributed_executor_backend != "external_launcher"
189
- ):
190
- logger.warning(
191
- "%s is not supported on XPU, fallback to ray distributed"
192
- " executor backend.",
193
- parallel_config.distributed_executor_backend,
194
- )
195
- parallel_config.distributed_executor_backend = "ray"
196
-
170
if model_config and model_config.use_mla:
171
logger.info(
172
"MLA is enabled on a non-GPU platform; forcing chunked "