ipex-llm/python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek_v3

Deepseek V3/R1

In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on Deepseek-V3/R1 models.

Currently only BF16 models (unsloth/DeepSeek-V3-bf16 and unsloth/DeepSeek-R1-BF16) are validated. It may need some config modifications and walkaround to run the official models.

0. Requirements

To run these examples with IPEX-LLM, we have some recommended requirements for your machine, please refer to here for more information.

Example: Predict Tokens using generate() API

In the example generate.py, we show a basic use case for a Deepseek model to predict the next N tokens using generate() API, with IPEX-LLM INT4 optimizations.

1. Install

We suggest using conda to manage environment:

On Linux:

conda create -n llm python=3.11 # recommend to use Python 3.11
conda activate llm

# install the latest ipex-llm nightly build with 'all' option
pip install --pre --upgrade ipex-llm[all] --extra-index-url https://download.pytorch.org/whl/cpu

On Windows:

conda create -n llm python=3.11
conda activate llm

pip install --pre --upgrade ipex-llm[all]

pip install transformers==4.48.3
pip install trl==0.12.0

2. Run

python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT

Arguments info:

• --repo-id-or-model-path REPO_ID_OR_MODEL_PATH: argument defining the huggingface repo id for the Deepseek model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be 'unsloth/DeepSeek-R1-BF16'.
• --prompt PROMPT: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be 'What is AI?'.
• --n-predict N_PREDICT: argument defining the max number of tokens to predict. It is default to be 32.

Note

: When loading the model in 4-bit, IPEX-LLM converts linear layers in the model into INT4 format. In theory, a XB model saved in 16-bit will requires approximately 2X GB of memory for loading, and ~0.5X GB memory for further inference. For 671B model, ~1.3 TB memory is needed during the load procedure.

2.1 Client

On client Windows machine, it is recommended to run directly with full utilization of all cores:

python ./generate.py 

2.2 Server

For optimal performance on server, it is recommended to set several environment variables (refer to here for more information), and run the example with all the physical cores of a single socket.

E.g. on Linux,

# set IPEX-LLM env variables
source ipex-llm-init

# e.g. for a server with 48 cores per socket
export OMP_NUM_THREADS=48
numactl -C 0-47 -m 0 python ./generate.py