First commit

vllm/executor/multiproc_xpu_executor.py

@@ -0,0 +1,26 @@
+import vllm.envs as envs
+from vllm.executor.multiproc_gpu_executor import (
+    MultiprocessingGPUExecutor, MultiprocessingGPUExecutorAsync)
+from vllm.executor.xpu_executor import XPUExecutor
+from vllm.logger import init_logger
+from vllm.utils import make_async
+
+logger = init_logger(__name__)
+
+
+class MultiprocessingXPUExecutor(MultiprocessingGPUExecutor, XPUExecutor):
+    """Python multiprocessing-based multi-XPU executor"""
+
+    def _check_executor_parameters(self):
+        mp_method = envs.VLLM_WORKER_MULTIPROC_METHOD
+        if mp_method != "spawn":
+            raise RuntimeError(
+                "XPU multiprocess executor only support spawn as mp method")
+
+
+class MultiprocessingXPUExecutorAsync(MultiprocessingXPUExecutor,
+                                      MultiprocessingGPUExecutorAsync):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.driver_exec_model = make_async(self.driver_worker.execute_model)