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 <./hugging_face_format.html>`_).

# 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)

.. seealso::
   
   See the complete example `here <https://github.com/intel-analytics/ipex-llm/tree/main/python/llm/example/CPU/Native-Models>`_

1.2 KiB Raw Blame History

Native Format

1.2 KiB

Raw Blame History