MLC-MiniCPM

中文版

Based on MLC-LLM, we run MiniCPM and MiniCPM-V on android devices.

Android APK

0. Install APK

1. Accept camera & photo permission: the permission are for MiniCPM-V which can process multimodel input (text + image)

   

2. Download model: (1) Press the download button (2) Wait for the progress bar to fill up (3) Start chat

   Caution：currently the two model can not be downloaded simultaneously due to the bug in downloading.

   

3. Chat with MiniCPM: (1) Wait for model initialization until "Ready to chat" pop up. (2) Type and send question

4. Chat with MiniCPM-V: (1) Wait for model initialization until "Ready to chat" pop up. (2) Upload image (3) Wait until "process image done" show up (4) Type and send question

   Note：image process may take some time.

   

5. Demo:

   

Note that the models run on android are quantized to 4-bit and may lose some performance. The non-quantized models can be found here.

Prepare Enviroment

Follow https://llm.mlc.ai/docs/deploy/android.html to prepare requirements.

For the Compile PyTorch Models from HuggingFace session, use our github repo and conduct the following instructions to install our modified version of mlc_chat.

mkdir -p build && cd build
# generate build configuration
python3 ../cmake/gen_cmake_config.py && cd ..
# build `mlc_chat_cli`
cd build && cmake .. && cmake --build . --parallel $(nproc) && cd ..
# install
cd python && pip install -e . && cd ..

Compile Model

put huggingface downloaded model checkpoint into dist/models.

For MiniCPM

MODEL_NAME=MiniCPM
QUANTIZATION=q4f16_1
MODEL_TYPE=minicpm
mlc_chat convert_weight --model-type ${MODEL_TYPE} ./dist/models/${MODEL_NAME}-hf/ --quantization $QUANTIZATION -o dist/$MODEL_NAME/
mlc_chat gen_config --model-type ${MODEL_TYPE} ./dist/models/${MODEL_NAME}-hf/ --quantization $QUANTIZATION --conv-template LM --sliding-window-size 768 -o dist/${MODEL_NAME}/
mlc_chat compile --model-type ${MODEL_TYPE} dist/${MODEL_NAME}/mlc-chat-config.json --device android -o ./dist/libs/${MODEL_NAME}-android.tar
cd ./android/library
./prepare_libs.sh
cd -

For MiniCPM-V (vision version)

MODEL_NAME=MiniCPM-V
QUANTIZATION=q4f16_1
MODEL_TYPE=minicpm_v
mlc_chat convert_weight --model-type ${MODEL_TYPE} ./dist/models/${MODEL_NAME}-hf/ --quantization $QUANTIZATION -o dist/$MODEL_NAME/
mlc_chat gen_config --model-type ${MODEL_TYPE} ./dist/models/${MODEL_NAME}-hf/ --quantization $QUANTIZATION --conv-template LM --sliding-window-size 768 -o dist/${MODEL_NAME}/
mlc_chat compile --model-type ${MODEL_TYPE} dist/${MODEL_NAME}/mlc-chat-config.json --device android -o ./dist/libs/${MODEL_NAME}-android.tar
cd ./android/library
./prepare_libs.sh
cd -

--sliding-window-size are set only for mobile phones to limit memory usage, and can be set smaller or larger base on your phone.

Build Android App

Go to android/ and use Android Studio to build the app. (Follow https://llm.mlc.ai/docs/deploy/android.html)