[CI] Add mteb testing for rerank models (#19344)

vllm/model_executor/models/bert.py

@@ -414,15 +414,10 @@ class BertEmbeddingModel(nn.Module, SupportsV0Only, SupportsQuant):
         intermediate_tensors: Optional[IntermediateTensors] = None,
         inputs_embeds: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids=input_ids,
-                                   position_ids=positions,
-                                   inputs_embeds=inputs_embeds,
-                                   intermediate_tensors=intermediate_tensors)
-
-        # convert the embedding output to float32,
-        # otherwise precision will be lost significantly
-        hidden_states = hidden_states.to(torch.float32)
-        return hidden_states
+        return self.model(input_ids=input_ids,
+                          position_ids=positions,
+                          inputs_embeds=inputs_embeds,
+                          intermediate_tensors=intermediate_tensors)
 
     def pooler(
         self,