[core] platform agnostic executor via collective_rpc (#11256)

Signed-off-by: youkaichao <youkaichao@gmail.com>

vllm/config.py

@@ -1294,8 +1294,11 @@ class ParallelConfig:
             from vllm.executor import ray_utils
             backend = "mp"
             ray_found = ray_utils.ray_is_available()
-            if (current_platform.is_cuda()
-                    and cuda_device_count_stateless() < self.world_size):
+            if current_platform.is_neuron():
+                # neuron uses single process to control multiple devices
+                backend = "uni"
+            elif (current_platform.is_cuda()
+                  and cuda_device_count_stateless() < self.world_size):
                 if not ray_found:
                     raise ValueError("Unable to load Ray which is "
                                      "required for multi-node inference, "
@@ -1328,13 +1331,14 @@ class ParallelConfig:
         from vllm.executor.executor_base import ExecutorBase
         from vllm.platforms import current_platform
         if self.distributed_executor_backend not in (
-                "ray", "mp", None) and not (isinstance(
+                "ray", "mp", "uni", None) and not (isinstance(
                     self.distributed_executor_backend, type) and issubclass(
                         self.distributed_executor_backend, ExecutorBase)):
             raise ValueError(
                 "Unrecognized distributed executor backend "
                 f"{self.distributed_executor_backend}. Supported "
-                "values are 'ray', 'mp' or custom ExecutorBase subclass.")
+                "values are 'ray', 'mp' 'uni', or custom ExecutorBase"
+                " subclass.")
         if self.use_ray:
             from vllm.executor import ray_utils
             ray_utils.assert_ray_available()