feat: Add additional parameters for GenerationConfig

- `presence_penalty`
- `frequency_penalty`
- `response_mime_type`

PiperOrigin-RevId: 627067043

Author: holtskinner

tests/system/vertexai/test_generative_models.py

@@ -17,6 +17,7 @@
 # pylint: disable=protected-access, g-multiple-import
 """System tests for generative models."""
 
+import json
 import pytest
 
 # Google imports
@@ -30,6 +31,7 @@
 
 GEMINI_MODEL_NAME = "gemini-1.0-pro-002"
 GEMINI_VISION_MODEL_NAME = "gemini-1.0-pro-vision"
+GEMINI_15_MODEL_NAME = "gemini-1.5-pro-preview-0409"
 
 
 # A dummy function for function calling
@@ -150,6 +152,31 @@ def test_generate_content_with_parameters(self):
         )
         assert response.text
 
+    def test_generate_content_with_gemini_15_parameters(self):
+        model = generative_models.GenerativeModel(GEMINI_15_MODEL_NAME)
+        response = model.generate_content(
+            contents="Why is sky blue? Respond in JSON Format.",
+            generation_config=generative_models.GenerationConfig(
+                temperature=0,
+                top_p=0.95,
+                top_k=20,
+                candidate_count=1,
+                max_output_tokens=100,
+                stop_sequences=["STOP!"],
+                presence_penalty=0.0,
+                frequency_penalty=0.0,
+                response_mime_type="application/json",
+            ),
+            safety_settings={
+                generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
+                generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_ONLY_HIGH,
+                generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
+                generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_NONE,
+            },
+        )
+        assert response.text
+        assert json.loads(response.text)
+
     def test_generate_content_from_list_of_content_dict(self):
         model = generative_models.GenerativeModel(GEMINI_MODEL_NAME)
         response = model.generate_content(

tests/unit/vertexai/test_generative_models.py

@@ -404,6 +404,8 @@ def test_generate_content(self, generative_models: generative_models):
                 candidate_count=1,
                 max_output_tokens=200,
                 stop_sequences=["\n\n\n"],
+                presence_penalty=0.0,
+                frequency_penalty=0.0,
             ),
             safety_settings=[
                 generative_models.SafetySetting(
@@ -420,6 +422,33 @@ def test_generate_content(self, generative_models: generative_models):
         )
         assert response2.text
 
+        model3 = generative_models.GenerativeModel("gemini-1.5-pro-preview-0409")
+        response3 = model3.generate_content(
+            "Why is sky blue? Respond in JSON.",
+            generation_config=generative_models.GenerationConfig(
+                temperature=0.2,
+                top_p=0.9,
+                top_k=20,
+                candidate_count=1,
+                max_output_tokens=200,
+                stop_sequences=["\n\n\n"],
+                response_mime_type="application/json",
+            ),
+            safety_settings=[
+                generative_models.SafetySetting(
+                    category=generative_models.SafetySetting.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
+                    threshold=generative_models.SafetySetting.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
+                    method=generative_models.SafetySetting.HarmBlockMethod.SEVERITY,
+                ),
+                generative_models.SafetySetting(
+                    category=generative_models.SafetySetting.HarmCategory.HARM_CATEGORY_HATE_SPEECH,
+                    threshold=generative_models.SafetySetting.HarmBlockThreshold.BLOCK_ONLY_HIGH,
+                    method=generative_models.SafetySetting.HarmBlockMethod.PROBABILITY,
+                ),
+            ],
+        )
+        assert response3.text
+
     @mock.patch.object(
         target=prediction_service.PredictionServiceClient,
         attribute="stream_generate_content",

vertexai/generative_models/_generative_models.py

@@ -1189,6 +1189,9 @@ def __init__(
         candidate_count: Optional[int] = None,
         max_output_tokens: Optional[int] = None,
         stop_sequences: Optional[List[str]] = None,
+        presence_penalty: Optional[float] = None,
+        frequency_penalty: Optional[float] = None,
+        response_mime_type: Optional[str] = None,
     ):
         r"""Constructs a GenerationConfig object.
 
@@ -1199,6 +1202,18 @@ def __init__(
             candidate_count: Number of candidates to generate.
             max_output_tokens: The maximum number of output tokens to generate per message.
             stop_sequences: A list of stop sequences.
+            presence_penalty: Positive values penalize tokens that have appeared in the generated text,
+                thus increasing the possibility of generating more diversed topics. Range: [-2.0, 2.0]
+            frequency_penalty: Positive values penalize tokens that repeatedly appear in the generated
+                text, thus decreasing the possibility of repeating the same content. Range: [-2.0, 2.0]
+            response_mime_type: Output response mimetype of the generated
+                candidate text. Supported mimetypes:
+
+                -  ``text/plain``: (default) Text output.
+                -  ``application/json``: JSON response in the candidates.
+
+                The model needs to be prompted to output the appropriate
+                response type, otherwise the behavior is undefined.
 
         Usage:
             ```
@@ -1222,6 +1237,9 @@ def __init__(
             candidate_count=candidate_count,
             max_output_tokens=max_output_tokens,
             stop_sequences=stop_sequences,
+            presence_penalty=presence_penalty,
+            frequency_penalty=frequency_penalty,
+            response_mime_type=response_mime_type,
         )
 
     @classmethod
@@ -1650,7 +1668,7 @@ def prompt_feedback(
 
     @property
     def usage_metadata(
-        self
+        self,
     ) -> gapic_prediction_service_types.GenerateContentResponse.UsageMetadata:
         return self._raw_response.usage_metadata