diff --git a/examples/code_generation/codegen/README.md b/examples/code_generation/codegen/README.md
index ee842be55e3a..cf6497caa750 100644
--- a/examples/code_generation/codegen/README.md
+++ b/examples/code_generation/codegen/README.md
@@ -106,7 +106,7 @@ python codegen_server.py
##### 配置参数说明
在codegen_server.py中配置如下参数：
-- `model_name_or_path`：模型名，默认为 "Salesforce/codegen-2B-mono"
+- `model_name_or_path`：模型名，默认为 "Salesforce/codegen-350M-mono"
- `device`：运行设备，默认为"gpu"
- `temperature`：解码参数temperature，默认为0.5
- `top_k`：解码参数top_k，默认为10
@@ -114,7 +114,7 @@ python codegen_server.py
- `repetition_penalty`：解码重复惩罚项，默认为1.0
- `min_length`：生成的最小长度，默认为0
- `max_length`：生成的最大长度，默认为16
-- `decode_strategy`：解码策略，默认为"sampling"
+- `decode_strategy`：解码策略，默认为"greedy_search"
- `load_state_as_np`：以numpy格式加载模型参数，可节省显存，默认为True
- `use_faster`：是否使用Fastergeneration，可加速推理，默认为True
- `use_fp16_decoding`：是否使用fp16推理，可节省显存和加速推理，默认为True
@@ -165,7 +165,16 @@ print(result)
- 如果使用FasterGeneration，需要设置[codegen_server.py](#配置参数说明)中`use_faster=True`，第一次推理会涉及到编译，会耗费一些时间。FasterGeneration的环境依赖参考[这里](https://github.com/PaddlePaddle/PaddleNLP/blob/develop/paddlenlp/ops/README.md#%E4%BD%BF%E7%94%A8%E7%8E%AF%E5%A2%83%E8%AF%B4%E6%98%8E)。
- 如果要使用自己训练好的模型，可以设置[codegen_server.py](#配置参数说明)中`model_name_or_path`为本地模型路径。
- 如果要从本地访问服务器，上述的`127.0.0.1`需要换成服务器的对外IP。
-
+- 如果出现下方的提示和报错，则说明FasterGeneration没有启动成功，需要定位下失败的原因。或者也可设置`use_faster=False`，不启动FasterGeneration加速，但推理速度会较慢。
+```shell
+ FasterGeneration is not available, and the original version would be used instead.
+```
+```shell
+ RuntimeError: (NotFound) There are no kernels which are registered in the unsqueeze2 operator.
+ [Hint: Expected kernels_iter != all_op_kernels.end(), but received kernels_iter == all_op_kernels.end().] (at /home/Paddle/paddle/fluid/imperative/prepared_operator.cc:341)
+ [operator < unsqueeze2 > error]
+```
+- 本代码也支持插件[fauxpilot](https://marketplace.visualstudio.com/items?itemName=Venthe.fauxpilot)，感谢[@linonetwo](https://github.com/linonetwo)测试。`settings.json`中配置"fauxpilot.server": "http://服务器ip:8978/v1/engines"
## 训练定制
@@ -307,3 +316,4 @@ hello_world()
## References
- Nijkamp, Erik, et al. "A conversational paradigm for program synthesis." arXiv preprint arXiv:2203.13474 (2022).
- [https://github.com/features/copilot/](https://github.com/features/copilot/)
+- [https://github.com/AndPuQing/Papilot](https://github.com/AndPuQing/Papilot)
diff --git a/examples/code_generation/codegen/requirements.txt b/examples/code_generation/codegen/requirements.txt
index 37e5ae958c12..ae00f4799fa1 100644
--- a/examples/code_generation/codegen/requirements.txt
+++ b/examples/code_generation/codegen/requirements.txt
@@ -3,4 +3,5 @@ pydantic==1.9.1
python-dotenv==0.20.0
sse_starlette==0.10.3
uvicorn==0.17.6
-openai==0.8.0
\ No newline at end of file
+openai==0.8.0
+regex==2022.6.2
\ No newline at end of file
diff --git a/faster_generation/README.md b/faster_generation/README.md
index 58cb2a6f4f99..dc156550f72e 100644
--- a/faster_generation/README.md
+++ b/faster_generation/README.md
@@ -43,25 +43,25 @@ FasterGeneration的高性能解码相比原版generate方法加速明显，并
- torch version 1.10.0+cu113
- transformers version 4.12.5
-**BART** (bart-base, batch_size=4, max_length=32)
+### **BART** (bart-base, batch_size=4, max_length=32)
-**GPT** (gpt2, batch_size=4, max_length=32)
+### **GPT** (gpt2, batch_size=4, max_length=32)
-**OPT** (opt, batch_size=4, max_length=32)
+### **OPT** (opt, batch_size=4, max_length=32)
-**CodeGen:**
+### **CodeGen:**
* 环境和超参
- Platform: Tesla V100-SXM2-32GB
- CUDA 10.1