enable all moe models for mrv2

Signed-off-by: yewentao256 <zhyanwentao@126.com>

tests/test_config.py

@@ -90,6 +90,16 @@ def test_v2_model_runner_env_tri_state(monkeypatch, env_value, expected):
             ),
             True,
         ),
+        (
+            SimpleNamespace(
+                model="Qwen/Qwen2-7B-Instruct",
+                architectures=["Qwen2ForCausalLM"],
+                runner_type="generate",
+                is_moe=False,
+                is_quantized=False,
+            ),
+            False,
+        ),
         (
             SimpleNamespace(
                 model="meta-llama/Llama-3.2-1B",
@@ -140,6 +150,16 @@ def test_v2_model_runner_env_tri_state(monkeypatch, env_value, expected):
             ),
             True,
         ),
+        (
+            SimpleNamespace(
+                model="deepseek-ai/DeepSeek-V3",
+                architectures=["DeepseekV3ForCausalLM"],
+                runner_type="generate",
+                is_moe=True,
+                is_quantized=False,
+            ),
+            True,
+        ),
         (
             SimpleNamespace(
                 model="Qwen/Qwen1.5-MoE-A2.7B",
@@ -178,7 +198,7 @@ def test_v2_model_runner_env_tri_state(monkeypatch, env_value, expected):
                 is_moe=True,
                 is_quantized=False,
             ),
-            False,
+            True,
         ),
         (
             SimpleNamespace(

vllm/config/vllm.py

@@ -555,6 +555,9 @@ class VllmConfig:
         if model_config.runner_type != "generate":
             return False
 
+        if model_config.is_moe:
+            return True
+
         architectures = getattr(model_config, "architectures", [])
         return any(
             arch in DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES for arch in architectures