add refusal

Author: lhoestq

src/huggingface_hub/_webhooks_payload.py

@@ -39,13 +39,27 @@ def model_json_schema(cls, *args, **kwargs) -> dict[str, Any]:
                 " should be installed separately. Please run `pip install --upgrade pydantic` and retry."
             )
 
+        @classmethod
+        def schema(cls, *args, **kwargs) -> dict[str, Any]:
+            raise ImportError(
+                "You must have `pydantic` installed to use `WebhookPayload`. This is an optional dependency that"
+                " should be installed separately. Please run `pip install --upgrade pydantic` and retry."
+            )
+
         @classmethod
         def model_validate_json(cls, json_data: str | bytes | bytearray, *args, **kwargs) -> "BaseModel":
             raise ImportError(
                 "You must have `pydantic` installed to use `WebhookPayload`. This is an optional dependency that"
                 " should be installed separately. Please run `pip install --upgrade pydantic` and retry."
             )
 
+        @classmethod
+        def parse_raw(cls, json_data: str | bytes | bytearray, *args, **kwargs) -> "BaseModel":
+            raise ImportError(
+                "You must have `pydantic` installed to use `WebhookPayload`. This is an optional dependency that"
+                " should be installed separately. Please run `pip install --upgrade pydantic` and retry."
+            )
+
 
 # This is an adaptation of the ReportV3 interface implemented in moon-landing. V0, V1 and V2 have been ignored as they
 # are not in used anymore. To keep in sync when format is updated in

src/huggingface_hub/inference/_client.py

@@ -878,13 +878,15 @@ def chat_completion(
         ```
         """
         if issubclass(response_format, BaseModel):
-            base_model = response_format
+            response_model = response_format
             response_format = ChatCompletionInputGrammarType(
                 type="json",
-                value=base_model.model_json_schema(),
+                value=response_model.model_json_schema()
+                if hasattr(response_model, "model_json_schema")
+                else response_model.schema(),
             )
         else:
-            base_model = None
+            response_model = None
 
         model_url = self._resolve_chat_completion_url(model)
 
@@ -922,13 +924,18 @@ def chat_completion(
             return _stream_chat_completion_response(data)  # type: ignore[arg-type]
 
         chat_completion_output = ChatCompletionOutput.parse_obj_as_instance(data)  # type: ignore[arg-type]
-        if base_model:
+        if response_model:
             for choice in chat_completion_output.choices:
                 if choice.message.content:
                     try:
-                        choice.message.parsed = base_model.model_validate_json(choice.message.content)
+                        # pydantic v2 uses model_validate_json
+                        choice.message.parsed = (
+                            response_model.model_validate_json(choice.message.content)
+                            if hasattr(response_model, "model_validate_json")
+                            else response_model.parse_raw(choice.message.content)
+                        )
                     except ValueError:
-                        pass
+                        choice.message.refusal = f"Failed to generate the response as a {response_model.__name__}"
         return chat_completion_output
 
     def _resolve_chat_completion_url(self, model: Optional[str] = None) -> str:

src/huggingface_hub/inference/_generated/_async_client.py

@@ -933,13 +933,15 @@ async def chat_completion(
         ```
         """
         if issubclass(response_format, BaseModel):
-            base_model = response_format
+            response_model = response_format
             response_format = ChatCompletionInputGrammarType(
                 type="json",
-                value=base_model.model_json_schema(),
+                value=response_model.model_json_schema()
+                if hasattr(response_model, "model_json_schema")
+                else response_model.schema(),
             )
         else:
-            base_model = None
+            response_model = None
 
         model_url = self._resolve_chat_completion_url(model)
 
@@ -977,13 +979,18 @@ async def chat_completion(
             return _async_stream_chat_completion_response(data)  # type: ignore[arg-type]
 
         chat_completion_output = ChatCompletionOutput.parse_obj_as_instance(data)  # type: ignore[arg-type]
-        if base_model:
+        if response_model:
             for choice in chat_completion_output.choices:
                 if choice.message.content:
                     try:
-                        choice.message.parsed = base_model.model_validate_json(choice.message.content)
+                        # pydantic v2 uses model_validate_json
+                        choice.message.parsed = (
+                            response_model.model_validate_json(choice.message.content)
+                            if hasattr(response_model, "model_validate_json")
+                            else response_model.parse_raw(choice.message.content)
+                        )
                     except ValueError:
-                        pass
+                        choice.message.refusal = f"Failed to generate the response as a {response_model.__name__}"
         return chat_completion_output
 
     def _resolve_chat_completion_url(self, model: Optional[str] = None) -> str:

src/huggingface_hub/inference/_generated/types/chat_completion.py

@@ -199,6 +199,7 @@ class ChatCompletionOutputMessage(BaseInferenceType):
     content: Optional[str] = None
     tool_calls: Optional[List[ChatCompletionOutputToolCall]] = None
     parsed: Optional[BaseModel] = None
+    refusal: Optional[str] = None
 
 
 @dataclass