[Model Runner v2] Migration from v1 to v2, with Qwen and DSv2 MOE models [3/N] (#42667)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

tests/test_config.py

@@ -122,8 +122,58 @@ def test_v2_model_runner_env_tri_state(monkeypatch, env_value, expected):
         ),
         (
             SimpleNamespace(
-                model="Qwen/Qwen3-30B-A3B",
-                architectures=["Qwen3MoeForCausalLM"],
+                model="deepseek-ai/DeepSeek-V2-Lite-Chat",
+                architectures=["DeepseekV2ForCausalLM"],
+                runner_type="generate",
+                is_moe=True,
+                is_quantized=False,
+            ),
+            True,
+        ),
+        (
+            SimpleNamespace(
+                model="deepseek-ai/DeepSeek-V2-Chat",
+                architectures=["DeepseekV2ForCausalLM"],
+                runner_type="generate",
+                is_moe=True,
+                is_quantized=False,
+            ),
+            True,
+        ),
+        (
+            SimpleNamespace(
+                model="Qwen/Qwen1.5-MoE-A2.7B",
+                architectures=["Qwen2MoeForCausalLM"],
+                runner_type="generate",
+                is_moe=True,
+                is_quantized=False,
+            ),
+            True,
+        ),
+        (
+            SimpleNamespace(
+                model="Qwen/Qwen1.5-MoE-A2.7B-Chat",
+                architectures=["Qwen2MoeForCausalLM"],
+                runner_type="generate",
+                is_moe=True,
+                is_quantized=False,
+            ),
+            True,
+        ),
+        (
+            SimpleNamespace(
+                model="ibm-research/PowerMoE-3b",
+                architectures=["GraniteMoeForCausalLM"],
+                runner_type="generate",
+                is_moe=True,
+                is_quantized=False,
+            ),
+            False,
+        ),
+        (
+            SimpleNamespace(
+                model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+                architectures=["MixtralForCausalLM"],
                 runner_type="generate",
                 is_moe=True,
                 is_quantized=False,

vllm/config/vllm.py

@@ -67,9 +67,11 @@ logger = init_logger(__name__)
 
 DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES = frozenset(
     {
+        "Qwen3ForCausalLM",
+        "DeepseekV2ForCausalLM",
+        "Qwen2MoeForCausalLM",
         "LlamaForCausalLM",
         "MistralForCausalLM",
-        "Qwen3ForCausalLM",
     }
 )
 
@@ -559,13 +561,13 @@ class VllmConfig:
         if model_config.runner_type != "generate":
             return False
 
-        architectures = getattr(model_config, "architectures", [])
-        if not any(
-            arch in DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES for arch in architectures
-        ):
+        if model_config.is_quantized:
             return False
 
-        return not model_config.is_moe and not model_config.is_quantized
+        architectures = getattr(model_config, "architectures", [])
+        return any(
+            arch in DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES for arch in architectures
+        )
 
     @property
     def needs_dp_coordinator(self) -> bool:
@@ -2020,6 +2022,9 @@ class VllmConfig:
         if self.parallel_config.enable_dbo:
             unsupported.append("dual batch overlap")
 
+        if self.parallel_config.enable_elastic_ep:
+            unsupported.append("elastic expert parallelism")
+
         if model_config is not None and model_config.enable_return_routed_experts:
             # Will be added by https://github.com/vllm-project/vllm/pull/38163
             unsupported.append("routed experts capture")