[Model] Pooling model activation supports per request control by PoolingParams (#20538)

Signed-off-by: wang.yuqi <noooop@126.com>

vllm/model_executor/models/config.py

@@ -44,6 +44,15 @@ class GteNewModelConfig(VerifyAndUpdateConfig):
         }
 
 
+class JambaForSequenceClassificationConfig(VerifyAndUpdateConfig):
+
+    @staticmethod
+    def verify_and_update_config(vllm_config: "VllmConfig") -> None:
+        pooler_config = vllm_config.model_config.pooler_config
+        if pooler_config.activation is None:
+            pooler_config.activation = False
+
+
 class JinaRobertaModelConfig(VerifyAndUpdateConfig):
 
     @staticmethod
@@ -155,6 +164,26 @@ class NomicBertModelConfig(VerifyAndUpdateConfig):
             vllm_config.recalculate_max_model_len(max_model_len)
 
 
+class Qwen2ForProcessRewardModelConfig(VerifyAndUpdateConfig):
+
+    @staticmethod
+    def verify_and_update_config(vllm_config: "VllmConfig") -> None:
+        pooler_config = vllm_config.model_config.pooler_config
+
+        if pooler_config.step_tag_id is None:
+            pooler_config.step_tag_id = 151651
+
+
+class Qwen2ForRewardModelConfig(VerifyAndUpdateConfig):
+
+    @staticmethod
+    def verify_and_update_config(vllm_config: "VllmConfig") -> None:
+        pooler_config = vllm_config.model_config.pooler_config
+
+        if pooler_config.softmax is None:
+            pooler_config.softmax = False
+
+
 class Qwen3ForSequenceClassificationConfig(VerifyAndUpdateConfig):
 
     @staticmethod
@@ -309,8 +338,11 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
     "GteModel": SnowflakeGteNewModelConfig,
     "GteNewModel": GteNewModelConfig,
     "NomicBertModel": NomicBertModelConfig,
+    "Qwen2ForProcessRewardModel": Qwen2ForProcessRewardModelConfig,
+    "Qwen2ForRewardModel": Qwen2ForRewardModelConfig,
     "Qwen3ForSequenceClassification": Qwen3ForSequenceClassificationConfig,
     "XLMRobertaModel": JinaRobertaModelConfig,
     "JinaVLForRanking": JinaVLForSequenceClassificationConfig,
+    "JambaForSequenceClassification": JambaForSequenceClassificationConfig,
     "GraniteMoeHybridForCausalLM": GraniteMoeHybridModelConfig,
 }

vllm/model_executor/models/jamba.py

@@ -593,7 +593,5 @@ class JambaForSequenceClassification(JambaForCausalLM):
                 pooler_config,
                 classifier=self.score,
                 default_pooling_type=PoolingType.LAST,
-                default_normalize=False,
-                default_softmax=False,
             ),
         })

vllm/model_executor/models/jina_vl.py

@@ -90,15 +90,12 @@ class JinaVLForSequenceClassification(Qwen2VLForConditionalGeneration,
                          prefix=maybe_prefix(prefix, "qwen2_vl"))
         config = vllm_config.model_config.hf_config
         pooler_config = vllm_config.model_config.pooler_config
+        assert pooler_config is not None
 
         # logit bias for sigmoid normalization
         self.LOGIT_BIAS = 2.65
 
         self.score = JinaVLScorer(config)
-
-        pooler_config = vllm_config.model_config.pooler_config
-        assert pooler_config is not None
-
         self.pooler = DispatchPooler({
             "encode":
             Pooler.for_encode(pooler_config),

vllm/model_executor/models/qwen2_rm.py

@@ -117,8 +117,5 @@ class Qwen2ForProcessRewardModel(Qwen2RewardBaseModel):
             Pooler.for_encode(
                 pooler_config,
                 default_pooling_type=PoolingType.STEP,
-                default_normalize=False,
-                default_softmax=True,
-                default_step_tag_id=151651,
             )
         })