inference/README.md

Inference code for DeepSeek models

First convert huggingface model weight files to the format of this project.

export EXPERTS=256
export MP=8
export CONFIG=config_w4a16.json
python convert_w4a16.py --hf-ckpt-path ${HF_CKPT_PATH} --save-path ${SAVE_PATH} --n-experts ${EXPERTS} --model-parallel ${MP}

Then chat with DeepSeek model at will!

torchrun --nproc-per-node ${MP} generate.py --ckpt-path ${SAVE_PATH} --config ${CONFIG} --interactive

Or batch inference from file.

torchrun --nproc-per-node ${MP} generate.py --ckpt-path ${SAVE_PATH} --config ${CONFIG} --input-file ${FILE}

Or multi nodes inference.

torchrun --nnodes ${NODES} --nproc-per-node $((MP / NODES)) --node-rank $RANK --master-addr $ADDR generate.py --ckpt-path ${SAVE_PATH} --config ${CONFIG} --input-file ${FILE}