fix: incorrect HuggingFace initialization in init_chat_model

[bhardwajjvaibhav]

Summary

This PR fixes a bug in the HuggingFace backend of init_chat_model where all kwargs were passed directly to HuggingFacePipeline.from_model_id().

This caused validation errors such as:

Validation Error

ValidationError: 1 validation error for ChatHuggingFace
llm
  Field required [type=missing]

Root Cause

The previous implementation passed parameters like temperature, max_tokens, timeout, etc. to HuggingFacePipeline, which does not accept these arguments.

As a result, Pydantic validation inside ChatHuggingFace failed.

What this PR fixes

1. Adds default task for decoder-only models

If no task is provided, we set:

🛠️ Added Default Task for Decoder-Only Models

kwargs["task"] = "text-generation"

2. Filters arguments to only those supported by HuggingFacePipeline

Allowed params:

{"task", "model_kwargs", "device"}

Other parameters (temperature, max_tokens, etc.) are ignored for the pipeline and used only by LangChain.

Before

if model_provider == "huggingface":
    _check_pkg("langchain_huggingface")
    from langchain_huggingface import ChatHuggingFace, HuggingFacePipeline

    llm = HuggingFacePipeline.from_model_id(model_id=model, **kwargs)
    return ChatHuggingFace(llm=llm)

This sent all kwargs → leading to ValidationError.

After (Fixed)

if model_provider == "huggingface":
    _check_pkg("langchain_huggingface")
    from langchain_huggingface import ChatHuggingFace, HuggingFacePipeline

    # Add default task for decoder-only models
    if "task" not in kwargs:
        kwargs["task"] = "text-generation"

    # Filter only parameters allowed by HuggingFacePipeline
    pipeline_allowed_params = {"task", "model_kwargs", "device"}
    pipeline_kwargs = {k: v for k, v in kwargs.items() if k in pipeline_allowed_params}

    llm = HuggingFacePipeline.from_model_id(
        model_id=model,
        **pipeline_kwargs
    )

    return ChatHuggingFace(llm=llm)

Example Code That Now Works

from langchain.chat_models import init_chat_model

llm = init_chat_model(
    model="microsoft/Phi-3-mini-4k-instruct",
    model_provider="huggingface",
    temperature=0,
    max_tokens=1024,
    timeout=None,
    max_retries=2,
)

Test Case:

from langchain.langchain_classic.chat_models import init_chat_model

def test_hf_init_chat_model_initializes_successfully():
    llm = init_chat_model(
        model="microsoft/Phi-3-mini-4k-instruct",
        model_provider="huggingface",
        temperature=0,
        max_tokens=1024,
        timeout=None,
        max_retries=2,
    )

    assert llm is not None

Result

The HuggingFace initialization now correctly separates pipeline parameters and LangChain LLM settings, preventing validation errors.