[Distributed] Add enable_expert_parallel arg (#14305)

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>

vllm/config.py

@@ -754,7 +754,7 @@ class ModelConfig:
                 " must be divisible by tensor parallel size "
                 f"({tensor_parallel_size}).")
 
-        if envs.VLLM_TEST_ENABLE_EP:
+        if parallel_config.enable_expert_parallel:
             self._verify_with_expert_parallelism()
 
         pipeline_parallel_size = parallel_config.pipeline_parallel_size
@@ -1334,6 +1334,7 @@ class ParallelConfig:
     # IP of the data parallel master.
     data_parallel_master_ip: str = "127.0.0.1"
     data_parallel_master_port: int = 29500  # Port of the data parallel master.
+    enable_expert_parallel: bool = False  # Use EP instead of TP for MoE layers.
 
     # Maximum number of multiple batches
     # when load model sequentially. To avoid RAM OOM when using tensor