feat: use Gemini response metadata for token counting (#11743)

api/core/model_runtime/model_providers/google/llm/llm.py

@@ -292,8 +292,12 @@ class GoogleLargeLanguageModel(LargeLanguageModel):
                     )
                 else:
                     # calculate num tokens
-                    prompt_tokens = self.get_num_tokens(model, credentials, prompt_messages)
-                    completion_tokens = self.get_num_tokens(model, credentials, [assistant_prompt_message])
+                    if hasattr(response, "usage_metadata") and response.usage_metadata:
+                        prompt_tokens = response.usage_metadata.prompt_token_count
+                        completion_tokens = response.usage_metadata.candidates_token_count
+                    else:
+                        prompt_tokens = self.get_num_tokens(model, credentials, prompt_messages)
+                        completion_tokens = self.get_num_tokens(model, credentials, [assistant_prompt_message])
 
                     # transform usage
                     usage = self._calc_response_usage(model, credentials, prompt_tokens, completion_tokens)