readme

README.md

@@ -254,13 +254,22 @@ cd ./examples/benchmark/peft/paddle
 RCCL_NCHANNELS=8 HSA_FORCE_FINE_GRAIN_PCIE=1 python3 -m paddle.distributed.launch --gpus "0,1,2,3,4,5,6,7" benchmark.py --model_name_or_path facebook/llama-13b --english --train_data_size 1000  --intokens --intokens_length 1024  --num_train_epochs 1 --per_device_train_batch_size 2 --gradient_accumulation_steps 2 --evaluation_strategy no --save_strategy no  --fp16 --fp16_opt_level O2 --recompute --tensor_parallel_degree 8 --logging_steps 50 --output_dir outputs
 ```
 
-### 推理benchmark
+### 推理benchmark 1
 
 ```
 cd ./examples/benchmark/peft/paddle
 python3  inference_benchmark.py   --model_name_or_path facebook/llama-13b --dtype float16 --do_forward --do_generate
 ```
 
+### 推理benchmark 2（换用[PaddleNLP-develop](https://github.com/PaddlePaddle/PaddleNLP/tree/28158b9735837495e6c73f848925e1d72b821863)）
+
+```
+pip3 uninstall paddlenlp
+cd ./llm
+PYTHONPATH=../:$PYTHONPATH \
+python3 predictor.py --model_name_or_path facebook/llama-13b --dtype float16 --src_length 300 --max_length 100 --output_file "infer.json" --batch_size 1 --benchmark
+```
+
 ### LAMBADA推理评估
 
 ```