| SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd ) | |
| ROOT_DIR=$(dirname $SCRIPT_DIR) | |
| export TORCHINDUCTOR_CACHE_DIR=$ROOT_DIR/cache/compiled_kernels | |
| # train eagle3 for qwq-32b | |
| NUM_GPUS=${1:-4} | |
| TP_SIZE=${2:-4} | |
| BUILD_DATASET_NUM_PROC=${BUILD_DATASET_NUM_PROC:-64} | |
| torchrun \ | |
| --standalone \ | |
| --nproc_per_node $NUM_GPUS \ | |
| $ROOT_DIR/scripts/train_eagle3.py \ | |
| --target-model-path Qwen/QwQ-32B \ | |
| --draft-model-config $ROOT_DIR/configs/qwq-32B-eagle3.json \ | |
| --train-data-path $ROOT_DIR/cache/dataset/sharegpt_train.jsonl \ | |
| --build-dataset-num-proc $BUILD_DATASET_NUM_PROC \ | |
| --output-dir $ROOT_DIR/outputs/qwq-32b-eagle3-sharegpt \ | |
| --num-epochs 10 \ | |
| --batch-size 1 \ | |
| --learning-rate 1e-4 \ | |
| --max-length 4096 \ | |
| --chat-template qwen \ | |
| --cache-dir $ROOT_DIR/cache \ | |
| --embedding-key model.embed_tokens.weight \ | |
| --tp-size $TP_SIZE \ | |
| --target-model-backend sglang | |