[ModelRunner v2] Enable MRV2 for pooling models by default (#48290)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>

tests/models/language/pooling/test_colbert.py

@@ -323,8 +323,8 @@ class TestColbertSharedEngine:
     [
         pytest.param("bert", True, id="bert-v2"),
         pytest.param("modernbert", True, id="modernbert-v2"),
-        pytest.param("jina", False, id="jina-v1"),
-        pytest.param("lfm2", False, id="lfm2-v1"),
+        pytest.param("jina", True, id="jina-v2"),
+        pytest.param("lfm2", True, id="lfm2-v2"),
     ],
 )
 def test_colbert_hf_comparison(vllm_runner, monkeypatch, backend, use_v2):

tests/test_config.py

@@ -344,13 +344,36 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
         ),
         (
             SimpleNamespace(
-                model="Qwen/Qwen3-Embedding-0.6B",
-                architectures=["Qwen3ForCausalLM"],
+                model="sentence-transformers/all-MiniLM-L6-v2",
+                architectures=["BertModel"],
                 runner_type="pooling",
+                is_multimodal_model=False,
                 is_moe=False,
                 is_quantized=False,
             ),
-            False,
+            True,
+        ),
+        (
+            SimpleNamespace(
+                model="Qwen/Qwen3-Embedding-0.6B",
+                architectures=["Qwen3ForCausalLM"],
+                runner_type="pooling",
+                is_multimodal_model=False,
+                is_moe=False,
+                is_quantized=False,
+            ),
+            True,
+        ),
+        (
+            SimpleNamespace(
+                model="TomoroAI/tomoro-colqwen3-embed-4b",
+                architectures=["ColQwen3"],
+                runner_type="pooling",
+                is_multimodal_model=True,
+                is_moe=False,
+                is_quantized=False,
+            ),
+            True,
         ),
     ],
 )

vllm/config/vllm.py

@@ -676,9 +676,6 @@ class VllmConfig:
         if model_config is None:
             return False
 
-        if model_config.runner_type != "generate":
-            return False
-
         architectures = getattr(model_config, "architectures", [])
         default_architectures = default_v2_model_runner_architectures()
         is_default_v2_architecture = any(