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

Native Format

You may also convert Hugging Face Transformers models into native INT4 format for maximum performance as follows.

Note

Currently only llama/bloom/gptneox/starcoder/chatglm model families are supported; you may use the corresponding API to load the converted model. (For other models, you can use the Hugging Face transformers format as described here)

# convert the model
from ipex_llm import llm_convert
ipex_llm_path = llm_convert(model='/path/to/model/',
                            outfile='/path/to/output/',
                            outtype='int4',
                            model_family="llama")

# load the converted model
# switch to ChatGLMForCausalLM/GptneoxForCausalLM/BloomForCausalLM/StarcoderForCausalLM to load other models
from ipex_llm.transformers import LlamaForCausalLM
llm = LlamaForCausalLM.from_pretrained("/path/to/output/model.bin", native=True, ...)

# run the converted model
input_ids = llm.tokenize(prompt)
output_ids = llm.generate(input_ids, ...)
output = llm.batch_decode(output_ids)

Note

See the complete example here