fix(langchain): correctly initialize huggingface models in init_chat_model

libs/langchain/langchain/chat_models/base.py

@@ -415,10 +415,29 @@ def _init_chat_model_helper(
 
         return ChatMistralAI(model=model, **kwargs)  # type: ignore[call-arg,unused-ignore]
     if model_provider == "huggingface":
-        _check_pkg("langchain_huggingface")
-        from langchain_huggingface import ChatHuggingFace
+        try:
+            from langchain_huggingface.chat_models import ChatHuggingFace
+            from langchain_huggingface.llms import HuggingFacePipeline
+        except ImportError as e:
+            import_error_msg = "Please install langchain-huggingface to use HuggingFace models."
+            raise ImportError(import_error_msg) from e
+
+        # The 'task' kwarg is required by from_model_id but not the base constructor.
+        # We pop it from kwargs to avoid the Pydantic 'extra_forbidden' error.
+        task = kwargs.pop("task", None)
+        if not task:
+            task_error_msg = "The 'task' keyword argument is required for HuggingFace models."
+            raise ValueError(task_error_msg)
+
+        # Initialize the base LLM pipeline with the model and arguments
+        llm = HuggingFacePipeline.from_model_id(
+            model_id=model,
+            task=task,
+            **kwargs,  # Pass remaining kwargs like `device`
+        )
 
-        return ChatHuggingFace(model_id=model, **kwargs)
+        # Pass the initialized LLM to the chat wrapper
+        return ChatHuggingFace(llm=llm)
     if model_provider == "groq":
         _check_pkg("langchain_groq")
         from langchain_groq import ChatGroq
@@ -957,4 +976,4 @@ def with_structured_output(
         schema: Union[dict, type[BaseModel]],
         **kwargs: Any,
     ) -> Runnable[LanguageModelInput, Union[dict, BaseModel]]:
-        return self.__getattr__("with_structured_output")(schema, **kwargs)
+        return self.__getattr__("with_structured_output")(schema, **kwargs)

libs/langchain/tests/unit_tests/chat_models/test_base.py

@@ -6,6 +6,7 @@
 from langchain_core.language_models import BaseChatModel
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.runnables import RunnableConfig, RunnableSequence
+from langchain_huggingface.chat_models import ChatHuggingFace
 from pydantic import SecretStr
 
 from langchain.chat_models.base import __all__, init_chat_model
@@ -289,3 +290,16 @@ def test_configurable_with_default() -> None:
     prompt = ChatPromptTemplate.from_messages([("system", "foo")])
     chain = prompt | model_with_config
     assert isinstance(chain, RunnableSequence)
+
+
+def test_init_chat_model_huggingface() -> None:
+    """Test that init_chat_model works with huggingface."""
+    model_name = "google-bert/bert-base-uncased"
+
+    llm = init_chat_model(
+        model=model_name,
+        model_provider="huggingface",
+        task="text-generation",
+    )
+    assert isinstance(llm, ChatHuggingFace)
+    assert llm.llm.model_id == model_name