EfficientEdit is an innovative inference acceleration technique that leverages edit-oriented speculative decoding to accelerate code editing tasks. Demonstrating remarkable performance gains, the method achieves up to 8× speedup on Qwen2.5-Coder-32-Instruct and 13× acceleration on DeepSeek-Coder-33B-Instruct models.
cd EfficientEdit
conda create --name efficientedit
conda activate efficientedit
pip install requirementsCreate your conda environment and notice python>=3.10
cd EfficientEdit/data/CanItEdit
conda create --name canitedit python==3.10.16
conda activate canitedit
pip install requirements
cd EfficientEdit/data/CodeIF-Bench
conda create --name codeif
conda activate codeif
pip install requirements
The inference script is inference.sh. Please fill in the required root directory.
python inference_example.py \
--data_file $DATA_DIR/test.jsonl \
--output_file $RESULTS_DIR/example.jsonl \
--target_model $MODEL_DIR/target_model \
--draft_model $MODEL_DIR/draft_model
The fine-tuning script is train.sh. Please fill in the required root directory.
DATA_PATH=""
OUTPUT_PATH=""
MODEL_PATH=""
export CUDA_VISIBLE_DEVICES=1,2
deepspeed --num_gpus 2 --master_port 6001 mask_lora.py \
--model_name_or_path $MODEL_PATH \
--data_path $DATA_PATH \
--output_dir $OUTPUT_PATH \
--num_train_epochs 1 \
--model_max_length 2048 \
--per_device_train_batch_size 4 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 4 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 50 \
--save_total_limit 20 \
--learning_rate 2e-5 \
--warmup_steps 10 \
--logging_steps 1 \
--lr_scheduler_type "cosine" \
--gradient_checkpointing True \
--report_to "wandb" \
--deepspeed ./configs/ds_config_zero3.json \
--bf16 True
You can download the lora checkpoints at link.
You can run data/CanItEdit/test.py to compute the Pass@k of CanItEdit.
You can run data/CodeIF-Bench/L1_test/test.py to compute the Pass@k of L1.
You can run data/CodeIF-Bench/L2_test/run_pass_k.sh to compute the Pass@k of L2. The script is shown below.
ROOT=/EfficientEdit/data/CodeIF-Bench/L2_test
python run_pass@k.py \
--output_file $ROOT/result.jsonl \
--log_file $ROOT/log.jsonl \
--source_code_root $ROOT/Source_Code \
--data_file $ROOT/data.jsonl \
--n 1 \
--k 1The arguments are explained as follows.
output_file: the model's predictions.log_file: the output file that stores the evaluation results.source_code_root: the path of repositories. The original repositories can be downloaded from link.data_file: the metadata file.n: number of completions per task, e.g.,1k:the k value in Pass@k, e.g.,1