Hugging Face Local Pipelines

Hugging Face models can be run locally through the HuggingFacePipeline class.

The Hugging Face Model Hub hosts over 120k models, 20k datasets, and 50k demo apps (Spaces), all open source and publicly available, in an online platform where people can easily collaborate and build ML together.

These can be called from LangChain either through this local pipeline wrapper or by calling their hosted inference endpoints through the HuggingFaceHub class.

To use, you should have the transformers python package installed, as well as pytorch. You can also install xformer for a more memory-efficient attention implementation.

%pip install --upgrade --quiet  transformers --quiet

Model Loading

Models can be loaded by specifying the model parameters using the from_model_id method.

from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline

hf = HuggingFacePipeline.from_model_id(
    model_id="gpt2",
    task="text-generation",
    pipeline_kwargs={"max_new_tokens": 10},
)

They can also be loaded by passing in an existing transformers pipeline directly

from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "gpt2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=10)
hf = HuggingFacePipeline(pipeline=pipe)

Create Chain

With the model loaded into memory, you can compose it with a prompt to form a chain.

from langchain_core.prompts import PromptTemplate

template = """Question: {question}

Answer: Let's think step by step."""
prompt = PromptTemplate.from_template(template)

chain = prompt | hf

question = "What is electroencephalography?"

print(chain.invoke({"question": question}))

GPU Inference

When running on a machine with GPU, you can specify the device=n parameter to put the model on the specified device. Defaults to -1 for CPU inference.

If you have multiple-GPUs and/or the model is too large for a single GPU, you can specify device_map="auto", which requires and uses the Accelerate library to automatically determine how to load the model weights.

Note: both device and device_map should not be specified together and can lead to unexpected behavior.

gpu_llm = HuggingFacePipeline.from_model_id(
    model_id="gpt2",
    task="text-generation",
    device=0,  # replace with device_map="auto" to use the accelerate library.
    pipeline_kwargs={"max_new_tokens": 10},
)

gpu_chain = prompt | gpu_llm

question = "What is electroencephalography?"

print(gpu_chain.invoke({"question": question}))

Batch GPU Inference

If running on a device with GPU, you can also run inference on the GPU in batch mode.

gpu_llm = HuggingFacePipeline.from_model_id(
    model_id="bigscience/bloom-1b7",
    task="text-generation",
    device=0,  # -1 for CPU
    batch_size=2,  # adjust as needed based on GPU map and model size.
    model_kwargs={"temperature": 0, "max_length": 64},
)

gpu_chain = prompt | gpu_llm.bind(stop=["\n\n"])

questions = []
for i in range(4):
    questions.append({"question": f"What is the number {i} in french?"})

answers = gpu_chain.batch(questions)
for answer in answers:
    print(answer)

Inference with OpenVINO backend

To deploy a model with OpenVINO, you can specify the backend="openvino" parameter to trigger OpenVINO as backend inference framework.

If you have an Intel GPU, you can specify model_kwargs={"device": "GPU"} to run inference on it.

%pip install --upgrade-strategy eager "optimum[openvino,nncf]" --quiet

ov_config = {"PERFORMANCE_HINT": "LATENCY", "NUM_STREAMS": "1", "CACHE_DIR": ""}

ov_llm = HuggingFacePipeline.from_model_id(
    model_id="gpt2",
    task="text-generation",
    backend="openvino",
    model_kwargs={"device": "CPU", "ov_config": ov_config},
    pipeline_kwargs={"max_new_tokens": 10},
)

ov_chain = prompt | ov_llm

question = "What is electroencephalography?"

print(ov_chain.invoke({"question": question}))

Inference with local OpenVINO model

It is possible to export your model to the OpenVINO IR format with the CLI, and load the model from local folder.

!optimum-cli export openvino --model gpt2 ov_model_dir

It is recommended to apply 8 or 4-bit weight quantization to reduce inference latency and model footprint using --weight-format:

!optimum-cli export openvino --model gpt2  --weight-format int8 ov_model_dir # for 8-bit quantization

!optimum-cli export openvino --model gpt2  --weight-format int4 ov_model_dir # for 4-bit quantization

ov_llm = HuggingFacePipeline.from_model_id(
    model_id="ov_model_dir",
    task="text-generation",
    backend="openvino",
    model_kwargs={"device": "CPU", "ov_config": ov_config},
    pipeline_kwargs={"max_new_tokens": 10},
)

ov_chain = prompt | ov_llm

question = "What is electroencephalography?"

print(ov_chain.invoke({"question": question}))

You can get additional inference speed improvement with Dynamic Quantization of activations and KV-cache quantization. These options can be enabled with ov_config as follows:

ov_config = {
    "KV_CACHE_PRECISION": "u8",
    "DYNAMIC_QUANTIZATION_GROUP_SIZE": "32",
    "PERFORMANCE_HINT": "LATENCY",
    "NUM_STREAMS": "1",
    "CACHE_DIR": "",
}

For more information refer to OpenVINO LLM guide and OpenVINO Local Pipelines notebook.