Improve the output precision of embedding models (#19092)

vllm/model_executor/models/bert.py

@@ -414,10 +414,15 @@ class BertEmbeddingModel(nn.Module, SupportsV0Only, SupportsQuant):
         intermediate_tensors: Optional[IntermediateTensors] = None,
         inputs_embeds: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
-        return self.model(input_ids=input_ids,
-                          position_ids=positions,
-                          inputs_embeds=inputs_embeds,
-                          intermediate_tensors=intermediate_tensors)
+        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
 
     def pooler(
         self,