[misc][distributed] error on invalid state (#6092)

vllm/executor/multiproc_gpu_executor.py

@@ -10,6 +10,7 @@ from vllm.executor.multiproc_worker_utils import (ProcessWorkerWrapper,
 from vllm.logger import init_logger
 from vllm.sequence import ExecuteModelRequest, SamplerOutput
 from vllm.utils import (cuda_device_count_stateless,
+                        error_on_invalid_device_count_status,
                         get_distributed_init_method, get_open_port,
                         get_vllm_instance_id, make_async,
                         update_environment_variables)
@@ -39,6 +40,8 @@ class MultiprocessingGPUExecutor(DistributedGPUExecutor):
         assert world_size <= cuda_device_count_stateless(), (
             "please set tensor_parallel_size to less than max local gpu count")
 
+        error_on_invalid_device_count_status()
+
         # Multiprocessing-based executor does not support multi-node setting.
         # Since it only works for single node, we can use the loopback address
         # 127.0.0.1 for communication.

vllm/executor/ray_gpu_executor.py

@@ -11,7 +11,8 @@ from vllm.executor.distributed_gpu_executor import (  # yapf: disable
 from vllm.executor.ray_utils import RayWorkerWrapper, ray
 from vllm.logger import init_logger
 from vllm.sequence import ExecuteModelRequest, SamplerOutput
-from vllm.utils import (get_distributed_init_method, get_ip, get_open_port,
+from vllm.utils import (error_on_invalid_device_count_status,
+                        get_distributed_init_method, get_ip, get_open_port,
                         get_vllm_instance_id, make_async)
 
 if ray is not None:
@@ -175,6 +176,8 @@ class RayGPUExecutor(DistributedGPUExecutor):
         distributed_init_method = get_distributed_init_method(
             driver_ip, get_open_port())
 
+        error_on_invalid_device_count_status()
+
         # Initialize the actual workers inside worker wrapper.
         init_worker_all_kwargs = [
             self._get_worker_kwargs(