[Bugfix] Use ReplicatedLinear for SequenceClassification head (vllm-project#23836)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

Author: Isotr0py

tests/models/language/pooling/test_qwen3_reranker.py

@@ -96,8 +96,5 @@ def test_rerank_models_mteb_tp(vllm_runner,
         "tensor_parallel_size": 2,
     }
 
-    mteb_test_rerank_models(Qwen3RerankerHfRunner,
-                            vllm_runner,
-                            model_info,
-                            vllm_extra_kwargs,
-                            atol=1.2e-2)
+    mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info,
+                            vllm_extra_kwargs)

vllm/model_executor/models/adapters.py

@@ -248,7 +248,7 @@ def as_seq_cls_model(cls: _T) -> _T:
         return cls
 
     # Lazy import
-    from vllm.model_executor.layers.linear import RowParallelLinear
+    from vllm.model_executor.layers.linear import ReplicatedLinear
     from vllm.model_executor.layers.pooler import (ClassifierPooler,
                                                    DispatchPooler, Pooler,
                                                    PoolingMethod, PoolingType)
@@ -264,10 +264,9 @@ def _init_pooler(self, vllm_config: "VllmConfig", prefix: str = ""):
             config = vllm_config.model_config.hf_config
             quant_config = vllm_config.quant_config
 
-            self.score = RowParallelLinear(
+            self.score = ReplicatedLinear(
                 config.hidden_size,
                 config.num_labels,
-                input_is_parallel=False,
                 bias=False,
                 params_dtype=torch.float32,
                 quant_config=quant_config,