ipex-llm/docs/mddocs/Overview/llm.md

IPEX-LLM in 5 minutes

You can use IPEX-LLM to run any Hugging Face Transformers PyTorch model. It automatically optimizes and accelerates LLMs using low-precision (INT4/INT5/INT8) techniques, modern hardware accelerations and latest software optimizations.

Hugging Face transformers-based applications can run on IPEX-LLM with one-line code change, and you'll immediately observe significant speedup^[1].

Here, let's take a relatively small LLM model, i.e open_llama_3b_v2, and IPEX-LLM INT4 optimizations as an example.

Load a Pretrained Model

Simply use one-line transformers-style API in ipex-llm to load open_llama_3b_v2 with INT4 optimization (by specifying load_in_4bit=True) as follows:

from ipex_llm.transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path="openlm-research/open_llama_3b_v2",
                                             load_in_4bit=True)

Tip

open_llama_3b_v2 is a pretrained large language model hosted on Hugging Face. openlm-research/open_llama_3b_v2 is its Hugging Face model id. from_pretrained will automatically download the model from Hugging Face to a local cache path (e.g. ~/.cache/huggingface), load the model, and converted it to ipex-llm INT4 format.

It may take a long time to download the model using API. You can also download the model yourself, and set pretrained_model_name_or_path to the local path of the downloaded model. This way, from_pretrained will load and convert directly from local path without download.

Load Tokenizer

You also need a tokenizer for inference. Just use the official transformers API to load LlamaTokenizer:

from transformers import LlamaTokenizer

tokenizer = LlamaTokenizer.from_pretrained(pretrained_model_name_or_path="openlm-research/open_llama_3b_v2")

Run LLM

Now you can do model inference exactly the same way as using official transformers API:

import torch

with torch.inference_mode():
    prompt = 'Q: What is CPU?\nA:'
    
    # tokenize the input prompt from string to token ids
    input_ids = tokenizer.encode(prompt, return_tensors="pt")

    # predict the next tokens (maximum 32) based on the input token ids
    output = model.generate(input_ids,
                            max_new_tokens=32)

    # decode the predicted token ids to output string
    output_str = tokenizer.decode(output[0], skip_special_tokens=True)
    
    print(output_str)

---

^{[1] Performance varies by use, configuration and other factors. ipex-llm may not optimize to the same degree for non-Intel products. Learn more at www.Intel.com/PerformanceIndex.}