Add exponential retry logic for gemini models (#764)

This is to avoid the following error with long context on Gemini models
due to insufficient quota:

```
Error: 429 Resource exhausted. Please try again later. Please refer to https://cloud.google.com/vertex-ai/generative-ai/docs/quotas#error-code-429 for more details.
```

This approach uses exponential backoff retries when encountering a
ResourceExhausted error.

---------

Co-authored-by: Huanzhi (Hans) Mao <[email protected]>

Author: gabrielibagon

berkeley-function-call-leaderboard/bfcl/model_handler/proprietary_model/gemini.py

@@ -1,3 +1,4 @@
+import logging
 import os
 
 import vertexai
@@ -13,6 +14,13 @@
     func_doc_language_specific_pre_processing,
     system_prompt_pre_processing_chat_model,
 )
+from google.api_core.exceptions import ResourceExhausted
+from tenacity import (
+    retry,
+    retry_if_exception_type,
+    stop_after_attempt,
+    wait_random_exponential,
+)
 from vertexai.generative_models import (
     Content,
     FunctionDeclaration,
@@ -22,6 +30,7 @@
     Tool,
 )
 
+logging.basicConfig(level=logging.INFO)
 
 class GeminiHandler(BaseHandler):
     def __init__(self, model_name, temperature) -> None:
@@ -69,6 +78,18 @@ def decode_execute(self, result):
                     )
             return func_call_list
 
+    @retry(
+        wait=wait_random_exponential(min=6, max=120),
+        stop=stop_after_attempt(10),
+        retry=retry_if_exception_type(ResourceExhausted),
+        before_sleep=lambda retry_state: print(
+            f"Attempt {retry_state.attempt_number} failed. Sleeping for {float(round(retry_state.next_action.sleep, 2))} seconds before retrying..."
+            f"Error: {retry_state.outcome.exception()}"
+        ),
+    )
+    def generate_with_backoff(self, client, **kwargs):
+        return client.generate_content(**kwargs)
+
     #### FC methods ####
 
     def _query_FC(self, inference_data: dict):
@@ -100,21 +121,17 @@ def _query_FC(self, inference_data: dict):
                 self.model_name.replace("-FC", ""),
                 system_instruction=inference_data["system_prompt"],
             )
-            api_response = client.generate_content(
-                contents=inference_data["message"],
-                generation_config=GenerationConfig(
-                    temperature=self.temperature,
-                ),
-                tools=tools if len(tools) > 0 else None,
-            )
         else:
-            api_response = self.client.generate_content(
-                contents=inference_data["message"],
-                generation_config=GenerationConfig(
-                    temperature=self.temperature,
-                ),
-                tools=tools if len(tools) > 0 else None,
-            )
+            client = self.client
+
+        api_response = self.generate_with_backoff(
+            client=client,
+            contents=inference_data["message"],
+            generation_config=GenerationConfig(
+                temperature=self.temperature,
+            ),
+            tools=tools if len(tools) > 0 else None,
+        )
         return api_response
 
     def _pre_query_processing_FC(self, inference_data: dict, test_entry: dict) -> dict:
@@ -237,19 +254,15 @@ def _query_prompting(self, inference_data: dict):
                 self.model_name.replace("-FC", ""),
                 system_instruction=inference_data["system_prompt"],
             )
-            api_response = client.generate_content(
-                contents=inference_data["message"],
-                generation_config=GenerationConfig(
-                    temperature=self.temperature,
-                ),
-            )
         else:
-            api_response = self.client.generate_content(
-                contents=inference_data["message"],
-                generation_config=GenerationConfig(
-                    temperature=self.temperature,
-                ),
-            )
+            client = self.client
+        api_response = self.generate_with_backoff(
+            client=client,
+            contents=inference_data["message"],
+            generation_config=GenerationConfig(
+                temperature=self.temperature,
+            ),
+        )
         return api_response
 
     def _pre_query_processing_prompting(self, test_entry: dict) -> dict:
@@ -275,13 +288,6 @@ def _pre_query_processing_prompting(self, test_entry: dict) -> dict:
             return {"message": []}
 
     def _parse_query_response_prompting(self, api_response: any) -> dict:
-        # Note: Same issue as with mentioned in `_parse_query_response_FC` method
-        # According to the Vertex AI documentation, `api_response.text` should be enough.
-        # However, under the hood, it is calling `api_response.candidates[0].content.parts[0].text` which is causing the issue
-        """TypeError: argument of type 'Part' is not iterable"""
-        # So again, we need to directly access the `api_response.candidates[0].content.parts[0]._raw_part.text` attribute to get the text content of the part
-        # This is a workaround for this bug, until the bug is fixed
-
         if len(api_response.candidates[0].content.parts) > 0:
             model_responses = api_response.text
         else:

berkeley-function-call-leaderboard/pyproject.toml

@@ -30,6 +30,7 @@ dependencies = [
     "tabulate>=0.9.0",
     "google-cloud-aiplatform==1.72.0",
     "mpmath==1.3.0",
+    "tenacity==9.0.0"
 ]
 
 [project.scripts]