Merge branch 'develop' of https://github.com/bjjwwang/serving into develop

core/configure/proto/multi_lang_general_model_service.proto

@@ -59,7 +59,7 @@ message SimpleResponse { required int32 err_code = 1; }
 
 message GetClientConfigRequest {}
 
-message GetClientConfigResponse { repeated string client_config_str_list = 1; }
+message GetClientConfigResponse { required string client_config_str = 1; }
 
 service MultiLangGeneralModelService {
   rpc Inference(InferenceRequest) returns (InferenceResponse) {}

doc/COMPILE.md

@@ -153,7 +153,7 @@ cmake -DPYTHON_INCLUDE_DIR=$PYTHON_INCLUDE_DIR/ \
     -DPYTHON_LIBRARIES=$PYTHON_LIBRARIES \
     -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
     -DOPENCV_DIR=${OPENCV_DIR} \
-    -DWITH_OPENCV=ON
+    -DWITH_OPENCV=ON \
     -DSERVER=ON ..
 make -j10
 ```

doc/COMPILE_CN.md

@@ -152,7 +152,7 @@ cmake -DPYTHON_INCLUDE_DIR=$PYTHON_INCLUDE_DIR/ \
     -DPYTHON_LIBRARIES=$PYTHON_LIBRARIES \
     -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
     -DOPENCV_DIR=${OPENCV_DIR} \
-    -DWITH_OPENCV=ON
+    -DWITH_OPENCV=ON \
     -DSERVER=ON ..
 make -j10
 ```

doc/PADDLE_SERVING_ON_KUBERNETES.md

@@ -25,10 +25,10 @@ kubectl apply -f https://bit.ly/kong-ingress-dbless
 在`tools/generate_runtime_docker.sh`文件下，它的使用方式如下
 
 ```bash
-bash tool/generate_runtime_docker.sh --env cuda10.1 --python 2.7 --serving 0.5.0 --paddle 2.0.0 --name serving_runtime:cuda10.1-py27
+bash tool/generate_runtime_docker.sh --env cuda10.1 --python 3.6 --serving 0.6.0 --paddle 2.0.1 --name serving_runtime:cuda10.1-py36
 ```
 
-会生成 cuda10.1，python 2.7，serving版本0.5.0 还有 paddle版本2.0.0的运行镜像。如果有其他疑问，可以执行下列语句得到帮助信息。
+会生成 cuda10.1，python 3.6，serving版本0.6.0 还有 paddle版本2.0.1的运行镜像。如果有其他疑问，可以执行下列语句得到帮助信息。
 
 ```
 bash tools/generate_runtime_docker.sh --help
@@ -39,7 +39,7 @@ bash tools/generate_runtime_docker.sh --help
 - paddle-serving-server， paddle-serving-client，paddle-serving-app，paddlepaddle，具体版本可以在tools/runtime.dockerfile当中查看，同时，如果有定制化的需求，也可以在该文件中进行定制化。
 - paddle-serving-server 二进制可执行程序
 
-也就是说，运行镜像在生成之后，我们只需要将我们运行的代码（如果有）和模型搬运到镜像中就可以。生成后的镜像名为`paddle_serving:cuda10.2-py37`
+也就是说，运行镜像在生成之后，我们只需要将我们运行的代码（如果有）和模型搬运到镜像中就可以。生成后的镜像名为`paddle_serving:cuda10.2-py36`
 
 ### 添加您的代码和模型 
 
@@ -50,8 +50,8 @@ bash tools/generate_runtime_docker.sh --help
 对于pipeline模式，我们需要确保模型和程序文件、配置文件等各种依赖都能够在镜像中运行。因此可以在`/home/project`下存放我们的执行文件时，我们以`Serving/python/example/pipeline/ocr`为例，这是OCR文字识别任务。
 
 ```bash
-#假设您已经拥有Serving运行镜像，假设镜像名为paddle_serving:cuda10.2-py37
-docker run --rm -dit --name pipeline_serving_demo paddle_serving:cuda10.2-py37 bash
+#假设您已经拥有Serving运行镜像，假设镜像名为paddle_serving:cuda10.2-py36
+docker run --rm -dit --name pipeline_serving_demo paddle_serving:cuda10.2-py36 bash
 cd Serving/python/example/pipeline/ocr
 # get models
 python -m paddle_serving_app.package --get_model ocr_rec
@@ -71,7 +71,7 @@ docker commit pipeline_serving_demo ocr_serving:latest
 ```
 docker exec -it pipeline_serving_demo bash
 cd /home/ocr
-python3.7 web_service.py 
+python3.6 web_service.py 
 ```
 
 进入容器到工程目录之后，剩下的操作和调试代码的工作是类似的。
@@ -83,8 +83,8 @@ python3.7 web_service.py
 web service模式本质上和pipeline模式类似，因此我们以`Serving/python/examples/bert`为例
 
 ```bash
-#假设您已经拥有Serving运行镜像，假设镜像名为registry.baidubce.com/paddlepaddle/serving:0.6.0-cuda10.2-py37
-docker run --rm -dit --name webservice_serving_demo registry.baidubce.com/paddlepaddle/serving:0.6.0-cpu-py27 bash
+#假设您已经拥有Serving运行镜像，假设镜像名为registry.baidubce.com/paddlepaddle/serving:0.6.0-cuda10.2-py36
+docker run --rm -dit --name webservice_serving_demo registry.baidubce.com/paddlepaddle/serving:0.6.0-cpu-py36 bash
 cd Serving/python/examples/bert
 ### download model 
 wget https://paddle-serving.bj.bcebos.com/paddle_hub_models/text/SemanticModel/bert_chinese_L-12_H-768_A-12.tar.gz
@@ -102,7 +102,7 @@ docker commit webservice_serving_demo bert_serving:latest
 ```bash
 docker exec -it webservice_serving_demo bash
 cd /home/bert
-python3.7 bert_web_service.py 9292
+python3.6 bert_web_service.py bert_seq128_model 9292
 ```
 
 进入容器到工程目录之后，剩下的操作和调试代码的工作是类似的。
@@ -118,14 +118,15 @@ kubenetes集群操作需要`kubectl`去操纵yaml文件。我们这里给出了
 - pipeline ocr示例 
 
 ```bash
-sh tools/generate_k8s_yamls.sh  --app_name ocr --image_name registry.baidubce.com/paddlepaddle/serving:k8s-pipeline-demo --workdir /home/ocr --command "python2.7 web_service.py" --port 9999
+sh tools/generate_k8s_yamls.sh  --app_name ocr --image_name registry.baidubce.com/paddlepaddle/serving:k8s-pipeline-demo --workdir /home/ocr --command "python3.6 web_service.py" --port 9999
 ```
 
 - web service bert示例
 
 ```bash
-sh tools/generate_k8s_yamls.sh  --app_name bert --image_name registry.baidubce.com/paddlepaddle/serving:k8s-web-demo --workdir /home/bert --command "python2.7 bert_web_service.py 9292" --port 9292
+sh tools/generate_k8s_yamls.sh  --app_name bert --image_name registry.baidubce.com/paddlepaddle/serving:k8s-web-demo --workdir /home/bert --command "python3.6 bert_web_service.py bert_seq128_model 9292" --port 9292
 ```
+**需要注意的是，app_name需要同URL的函数名相同。例如示例中bert的访问URL是`https://127.0.0.1:9292/bert/prediction`，那么app_name应为bert。**
 
 接下来我们会看到有两个yaml文件，分别是`k8s_serving.yaml`和 k8s_ingress.yaml`.
 
@@ -174,7 +175,7 @@ spec:
         workingDir: /home/ocr
         name: ocr
         command: ['/bin/bash', '-c']
-        args: ["python3.7 web_service.py"]
+        args: ["python3.6 bert_web_service.py bert_seq128_model 9292"]
         env:
           - name: NODE_NAME
             valueFrom:
@@ -216,7 +217,8 @@ spec:
 最终我们执行就可以启动相关容器和API网关。
 
 ```
-kubectl apply -f k8s_serving.yaml k8s_ingress.yaml
+kubectl apply -f k8s_serving.yaml
+kubectl apply -f k8s_ingress.yaml
 ```
 
 输入

python/examples/bert/README_CN.md

@@ -94,4 +94,7 @@ curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"words": "hello"}]
 bash benchmark.sh bert_seq128_model bert_seq128_client
 ```
 性能测试的日志文件为profile_log_bert_seq128_model
+
 如需修改性能测试用例的参数，请修改benchmark.sh中的配置信息。
+
+注意:bert_seq128_model和bert_seq128_client路径后不要加'/'符号,示例需要在GPU机器上运行。

python/examples/bert/benchmark.sh

@@ -17,27 +17,30 @@ sleep 5
 
 #warm up
 $PYTHONROOT/bin/python3 benchmark.py --thread 4 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
-echo -e "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
+echo -e "import psutil\nimport time\nwhile True:\n\tcpu_res = psutil.cpu_percent()\n\twith open('cpu.txt', 'a+') as f:\n\t\tf.write(f'{cpu_res}\\\n')\n\ttime.sleep(0.1)" > cpu.py
 for thread_num in 1 4 8 16
 do
 for batch_size in 1 4 16 64
 do
     job_bt=`date '+%Y%m%d%H%M%S'`
-    nvidia-smi --id=0 --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
+    nvidia-smi --id=0 --query-compute-apps=used_memory --format=csv -lms 100 > gpu_memory_use.log 2>&1 &
     nvidia-smi --id=0 --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
+    rm -rf cpu.txt
+    $PYTHONROOT/bin/python3 cpu.py &
     gpu_memory_pid=$!
     $PYTHONROOT/bin/python3 benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
-    kill ${gpu_memory_pid}
-    kill `ps -ef|grep used_memory|awk '{print $2}'`
+    kill `ps -ef|grep used_memory|awk '{print $2}'` > /dev/null
+    kill `ps -ef|grep utilization.gpu|awk '{print $2}'` > /dev/null
+    kill `ps -ef|grep cpu.py|awk '{print $2}'` > /dev/null
     echo "model_name:" $1
     echo "thread_num:" $thread_num
     echo "batch_size:" $batch_size
     echo "=================Done===================="
     echo "model_name:$1" >> profile_log_$1
     echo "batch_size:$batch_size" >> profile_log_$1
-    $PYTHONROOT/bin/python3 cpu_utilization.py >> profile_log_$1
     job_et=`date '+%Y%m%d%H%M%S'`
-    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY:", max}' gpu_use.log >> profile_log_$1
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "CPU_UTILIZATION:", max}' cpu.txt >> profile_log_$1
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY:", max}' gpu_memory_use.log >> profile_log_$1
     awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "GPU_UTILIZATION:", max}' gpu_utilization.log >> profile_log_$1
     rm -rf gpu_use.log gpu_utilization.log
     $PYTHONROOT/bin/python3 ../util/show_profile.py profile $thread_num >> profile_log_$1

python/examples/fit_a_line/README_CN.md

@@ -49,4 +49,7 @@ curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"x": [0.0137, -0.1
 bash benchmark.sh uci_housing_model uci_housing_client
 ```
 性能测试的日志文件为profile_log_uci_housing_model
+
 如需修改性能测试用例的参数，请修改benchmark.sh中的配置信息。
+
+注意:uci_housing_model和uci_housing_client路径后不要加'/'符号,示例需要在GPU机器上运行。

python/examples/fit_a_line/benchmark.py

@@ -30,6 +30,7 @@ def single_func(idx, resource):
             paddle.dataset.uci_housing.train(), buf_size=500),
         batch_size=1)
     total_number = sum(1 for _ in train_reader())
+    latency_list = []
 
     if args.request == "rpc":
         client = Client()
@@ -37,9 +38,12 @@ def single_func(idx, resource):
         client.connect([args.endpoint])
         start = time.time()
         for data in train_reader():
+            l_start = time.time()
             fetch_map = client.predict(feed={"x": data[0][0]}, fetch=["price"])
+            l_end = time.time()
+            latency_list.append(l_end * 1000 - l_start * 1000)
         end = time.time()
-        return [[end - start], [total_number]]
+        return [[end - start], latency_list, [total_number]]
     elif args.request == "http":
         train_reader = paddle.batch(
             paddle.reader.shuffle(
@@ -47,11 +51,14 @@ def single_func(idx, resource):
             batch_size=1)
         start = time.time()
         for data in train_reader():
+            l_start = time.time()
             r = requests.post(
                 'http://{}/uci/prediction'.format(args.endpoint),
                 data={"x": data[0]})
+            l_end = time.time()
+            latency_list.append(l_end * 1000 - l_start * 1000)
         end = time.time()
-        return [[end - start], [total_number]]
+        return [[end - start], latency_list, [total_number]]
 
 
 start = time.time()

python/examples/fit_a_line/benchmark.sh

+rm profile_log*
+export CUDA_VISIBLE_DEVICES=0,1
+export FLAGS_profile_server=1
+export FLAGS_profile_client=1
+export FLAGS_serving_latency=1
+
+gpu_id=0
+#save cpu and gpu utilization log
+if [ -d utilization ];then
+    rm -rf utilization
+else
+    mkdir utilization
+fi
+#start server
+$PYTHONROOT/bin/python3 -m paddle_serving_server.serve --model $1 --port 9292 --thread 4 --gpu_ids 0,1 --mem_optim  --ir_optim >  elog  2>&1 &
+sleep 5
+
+#warm up
+$PYTHONROOT/bin/python3 benchmark.py --thread 4 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+echo -e "import psutil\nimport time\nwhile True:\n\tcpu_res = psutil.cpu_percent()\n\twith open('cpu.txt', 'a+') as f:\n\t\tf.write(f'{cpu_res}\\\n')\n\ttime.sleep(0.1)" > cpu.py
+for thread_num in 1 4 8 16
+do
+for batch_size in 1 4 16 64
+do
+    job_bt=`date '+%Y%m%d%H%M%S'`
+    nvidia-smi --id=0 --query-compute-apps=used_memory --format=csv -lms 100 > gpu_memory_use.log 2>&1 &
+    nvidia-smi --id=0 --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
+    rm -rf cpu.txt
+    $PYTHONROOT/bin/python3 cpu.py &
+    gpu_memory_pid=$!
+    $PYTHONROOT/bin/python3 benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
+    kill `ps -ef|grep used_memory|awk '{print $2}'` > /dev/null
+    kill `ps -ef|grep utilization.gpu|awk '{print $2}'` > /dev/null
+    kill `ps -ef|grep cpu.py|awk '{print $2}'` > /dev/null
+    echo "model_name:" $1
+    echo "thread_num:" $thread_num
+    echo "batch_size:" $batch_size
+    echo "=================Done===================="
+    echo "model_name:$1" >> profile_log_$1
+    echo "batch_size:$batch_size" >> profile_log_$1
+    job_et=`date '+%Y%m%d%H%M%S'`
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "CPU_UTILIZATION:", max}' cpu.txt >> profile_log_$1
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY:", max}' gpu_memory_use.log >> profile_log_$1
+    awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "GPU_UTILIZATION:", max}' gpu_utilization.log >> profile_log_$1
+    rm -rf gpu_use.log gpu_utilization.log
+    $PYTHONROOT/bin/python3 ../util/show_profile.py profile $thread_num >> profile_log_$1
+    tail -n 8 profile >> profile_log_$1
+    echo "" >> profile_log_$1
+done
+done
+
+#Divided log
+awk 'BEGIN{RS="\n\n"}{i++}{print > "bert_log_"i}' profile_log_$1
+mkdir bert_log && mv bert_log_* bert_log
+ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9

python/paddle_serving_client/client.py

@@ -554,15 +554,8 @@ class MultiLangClient(object):
         get_client_config_req = multi_lang_general_model_service_pb2.GetClientConfigRequest(
         )
         resp = self.stub_.GetClientConfig(get_client_config_req)
-        model_config_path_list = resp.client_config_str_list
-        file_path_list = []
-        for single_model_config in model_config_path_list:
-            if os.path.isdir(single_model_config):
-                file_path_list.append("{}/serving_server_conf.prototxt".format(
-                    single_model_config))
-            elif os.path.isfile(single_model_config):
-                file_path_list.append(single_model_config)
-        self._parse_model_config(file_path_list)
+        model_config_str = resp.client_config_str
+        self._parse_model_config(model_config_str)
 
     def _flatten_list(self, nested_list):
         for item in nested_list:
@@ -572,23 +565,10 @@ class MultiLangClient(object):
             else:
                 yield item
 
-    def _parse_model_config(self, model_config_path_list):
-        if isinstance(model_config_path_list, str):
-            model_config_path_list = [model_config_path_list]
-        elif isinstance(model_config_path_list, list):
-            pass
-
-        file_path_list = []
-        for single_model_config in model_config_path_list:
-            if os.path.isdir(single_model_config):
-                file_path_list.append("{}/serving_client_conf.prototxt".format(
-                    single_model_config))
-            elif os.path.isfile(single_model_config):
-                file_path_list.append(single_model_config)
+    def _parse_model_config(self, model_config_str):
         model_conf = m_config.GeneralModelConfig()
-        f = open(file_path_list[0], 'r')
-        model_conf = google.protobuf.text_format.Merge(
-            str(f.read()), model_conf)
+        model_conf = google.protobuf.text_format.Merge(model_config_str,
+                                                       model_conf)
         self.feed_names_ = [var.alias_name for var in model_conf.feed_var]
         self.feed_types_ = {}
         self.feed_shapes_ = {}
@@ -598,11 +578,6 @@ class MultiLangClient(object):
             self.feed_shapes_[var.alias_name] = var.shape
             if var.is_lod_tensor:
                 self.lod_tensor_set_.add(var.alias_name)
-        if len(file_path_list) > 1:
-            model_conf = m_config.GeneralModelConfig()
-            f = open(file_path_list[-1], 'r')
-            model_conf = google.protobuf.text_format.Merge(
-                str(f.read()), model_conf)
         self.fetch_names_ = [var.alias_name for var in model_conf.fetch_var]
         self.fetch_types_ = {}
         for i, var in enumerate(model_conf.fetch_var):

python/paddle_serving_server/rpc_service.py

@@ -198,5 +198,14 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
         #model_config_path_list is list right now.
         #dict should be added when graphMaker is used.
         resp = multi_lang_general_model_service_pb2.GetClientConfigResponse()
-        resp.client_config_str_list[:] = self.model_config_path_list
+        model_config_str = []
+        for single_model_config in self.model_config_path_list:
+            if os.path.isdir(single_model_config):
+                with open("{}/serving_server_conf.prototxt".format(
+                        single_model_config)) as f:
+                    model_config_str.append(str(f.read()))
+            elif os.path.isfile(single_model_config):
+                with open(single_model_config) as f:
+                    model_config_str.append(str(f.read()))
+        resp.client_config_str = model_config_str[0]
         return resp

tools/Dockerfile.cuda10.1-cudnn7.devel

@@ -104,7 +104,7 @@ ENV PATH=usr/local/go/bin:/root/go/bin:${PATH}
 
 # Downgrade TensorRT 
 COPY tools/dockerfiles/build_scripts /build_scripts
-RUN bash /build_scripts/install_trt.sh 
+RUN bash /build_scripts/install_trt.sh cuda10.1  
 RUN rm -rf /build_scripts
 
 # git credential to skip password typing
@@ -132,9 +132,9 @@ RUN wget https://paddle-ci.gz.bcebos.com/ccache-3.7.9.tar.gz && \
     make -j8 && make install && \
     ln -s /usr/local/ccache-3.7.9/bin/ccache /usr/local/bin/ccache
 
-RUN python3.8 -m pip install --upgrade pip requests && \
-    python3.7 -m pip install --upgrade pip requests && \
-    python3.6 -m pip install --upgrade pip requests 
+RUN python3.8 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.7 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.6 -m pip install --upgrade pip==21.1.1 requests 
 
 RUN wget https://paddle-serving.bj.bcebos.com/others/centos_ssl.tar && \
     tar xf centos_ssl.tar && rm -rf centos_ssl.tar && \

tools/Dockerfile.cuda10.2-cudnn8.devel

@@ -104,7 +104,7 @@ ENV PATH=usr/local/go/bin:/root/go/bin:${PATH}
 
 # Downgrade TensorRT 
 COPY tools/dockerfiles/build_scripts /build_scripts
-RUN bash /build_scripts/install_trt.sh 
+RUN bash /build_scripts/install_trt.sh cuda10.2 
 RUN rm -rf /build_scripts
 
 # git credential to skip password typing
@@ -132,9 +132,9 @@ RUN wget https://paddle-ci.gz.bcebos.com/ccache-3.7.9.tar.gz && \
     make -j8 && make install && \
     ln -s /usr/local/ccache-3.7.9/bin/ccache /usr/local/bin/ccache
 
-RUN python3.8 -m pip install --upgrade pip requests && \
-    python3.7 -m pip install --upgrade pip requests && \
-    python3.6 -m pip install --upgrade pip requests 
+RUN python3.8 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.7 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.6 -m pip install --upgrade pip==21.1.1 requests 
 
 RUN wget https://paddle-serving.bj.bcebos.com/others/centos_ssl.tar && \
     tar xf centos_ssl.tar && rm -rf centos_ssl.tar && \

tools/Dockerfile.cuda11.2-cudnn8.devel → tools/Dockerfile.cuda11-cudnn8.devel

 # A image for building paddle binaries
 # Use cuda devel base image for both cpu and gpu environment
 # When you modify it, please be aware of cudnn-runtime version
-FROM nvidia/cuda:11.2.0-cudnn8-devel-ubuntu16.04
+FROM nvidia/cuda:11.0.3-cudnn8-devel-ubuntu16.04
 MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
 
 # ENV variables
@@ -104,7 +104,7 @@ ENV PATH=usr/local/go/bin:/root/go/bin:${PATH}
 
 # Downgrade TensorRT 
 COPY tools/dockerfiles/build_scripts /build_scripts
-RUN bash /build_scripts/install_trt.sh 
+RUN bash /build_scripts/install_trt.sh cuda11 
 RUN rm -rf /build_scripts
 
 # git credential to skip password typing
@@ -132,9 +132,9 @@ RUN wget https://paddle-ci.gz.bcebos.com/ccache-3.7.9.tar.gz && \
     make -j8 && make install && \
     ln -s /usr/local/ccache-3.7.9/bin/ccache /usr/local/bin/ccache
 
-RUN python3.8 -m pip install --upgrade pip requests && \
-    python3.7 -m pip install --upgrade pip requests && \
-    python3.6 -m pip install --upgrade pip requests 
+RUN python3.8 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.7 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.6 -m pip install --upgrade pip==21.1.1 requests 
 
 RUN wget https://paddle-serving.bj.bcebos.com/others/centos_ssl.tar && \
     tar xf centos_ssl.tar && rm -rf centos_ssl.tar && \

tools/Dockerfile.devel

@@ -132,9 +132,9 @@ RUN wget https://paddle-ci.gz.bcebos.com/ccache-3.7.9.tar.gz && \
     make -j8 && make install && \
     ln -s /usr/local/ccache-3.7.9/bin/ccache /usr/local/bin/ccache
 
-RUN python3.8 -m pip install --upgrade pip requests && \
-    python3.7 -m pip install --upgrade pip requests && \
-    python3.6 -m pip install --upgrade pip requests 
+RUN python3.8 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.7 -m pip install --upgrade pip==21.1.1 requests && \
+    python3.6 -m pip install --upgrade pip==21.1.1 requests 
 
 RUN wget https://paddle-serving.bj.bcebos.com/others/centos_ssl.tar && \
     tar xf centos_ssl.tar && rm -rf centos_ssl.tar && \

tools/Dockerfile.runtime_template

@@ -28,12 +28,12 @@ WORKDIR /home
 # install whl and bin
 WORKDIR /home
       COPY tools/dockerfiles/build_scripts /build_scripts
-      RUN bash /build_scripts/install_whl.sh 0.5.0 2.0.0 <<run_env>> <<python_version>> && rm -rf /build_scripts
+      RUN bash /build_scripts/install_whl.sh <<serving_version>> <<paddle_version>> <<run_env>> <<python_version>> && rm -rf /build_scripts
 
 # install tensorrt
 WORKDIR /home
       COPY tools/dockerfiles/build_scripts /build_scripts
-      RUN bash /build_scripts/install_trt.sh && rm -rf /build_scripts
+      RUN bash /build_scripts/install_trt.sh <<run_env>> && rm -rf /build_scripts
 
 # install go
 RUN wget -qO- https://dl.google.com/go/go1.14.linux-amd64.tar.gz | \

tools/dockerfiles/build_scripts/install_trt.sh

@@ -14,31 +14,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-VERSION=$(nvcc --version | grep release | grep -oEi "release ([0-9]+)\.([0-9])"| sed "s/release //")
-
-if [[ "$VERSION" == "10.1" ]];then
+VERSION=$1
+if [[ "$VERSION" == "cuda10.1" ]];then
   wget -q https://paddle-ci.gz.bcebos.com/TRT/TensorRT6-cuda10.1-cudnn7.tar.gz --no-check-certificate
   tar -zxf TensorRT6-cuda10.1-cudnn7.tar.gz -C /usr/local
   cp -rf /usr/local/TensorRT6-cuda10.1-cudnn7/include/* /usr/include/ && cp -rf /usr/local/TensorRT6-cuda10.1-cudnn7/lib/* /usr/lib/
+  echo "cuda10.1 trt install ==============>>>>>>>>>>>>"
   rm TensorRT6-cuda10.1-cudnn7.tar.gz
-elif [[ "$VERSION" == "11.0" ]];then
+elif [[ "$VERSION" == "cuda11" ]];then
   wget -q https://paddle-ci.cdn.bcebos.com/TRT/TensorRT-7.1.3.4.Ubuntu-16.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz --no-check-certificate
   tar -zxf TensorRT-7.1.3.4.Ubuntu-16.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz -C /usr/local
   cp -rf /usr/local/TensorRT-7.1.3.4/include/* /usr/include/ && cp -rf /usr/local/TensorRT-7.1.3.4/lib/* /usr/lib/
   rm TensorRT-7.1.3.4.Ubuntu-16.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz
-elif [[ "$VERSION" == "10.2" ]];then
+elif [[ "$VERSION" == "cuda10.2" ]];then
   wget https://paddle-ci.gz.bcebos.com/TRT/TensorRT7-cuda10.2-cudnn8.tar.gz --no-check-certificate
   tar -zxf TensorRT7-cuda10.2-cudnn8.tar.gz -C /usr/local
   cp -rf /usr/local/TensorRT-7.1.3.4/include/* /usr/include/ && cp -rf /usr/local/TensorRT-7.1.3.4/lib/* /usr/lib/
   rm TensorRT7-cuda10.2-cudnn8.tar.gz
-elif [[ "$VERSION" == "10.0" ]];then
-  wget -q https://paddle-ci.gz.bcebos.com/TRT/TensorRT6-cuda10.0-cudnn7.tar.gz --no-check-certificate
-  tar -zxf TensorRT6-cuda10.0-cudnn7.tar.gz -C /usr/local
-  cp -rf /usr/local/TensorRT6-cuda10.0-cudnn7/include/* /usr/include/ && cp -rf /usr/local/TensorRT6-cuda10.0-cudnn7/lib/* /usr/lib/
-  rm TensorRT6-cuda10.0-cudnn7.tar.gz
-elif [[ "$VERSION" == "9.0" ]];then
-  wget -q https://paddle-ci.gz.bcebos.com/TRT/TensorRT6-cuda9.0-cudnn7.tar.gz --no-check-certificate
-  tar -zxf TensorRT6-cuda9.0-cudnn7.tar.gz -C /usr/local
-  cp -rf /usr/local/TensorRT6-cuda9.0-cudnn7/include/* /usr/include/ && cp -rf /usr/local/TensorRT6-cuda9.0-cudnn7/lib/* /usr/lib/
-  rm TensorRT6-cuda9.0-cudnn7.tar.gz
 fi

tools/dockerfiles/build_scripts/install_whl.sh

@@ -40,6 +40,9 @@ if [[ $SERVING_VERSION == "0.5.0" ]]; then
     elif [[ "$RUN_ENV" == "cuda10.2" ]];then
         server_release="paddle-serving-server-gpu==$SERVING_VERSION.post102"
         serving_bin="https://paddle-serving.bj.bcebos.com/bin/serving-gpu-102-${SERVING_VERSION}.tar.gz"
+    elif [[ "$RUN_ENV" == "cuda11" ]];then
+        server_release="paddle-serving-server-gpu==$SERVING_VERSION.post11"
+        serving_bin="https://paddle-serving.bj.bcebos.com/bin/serving-gpu-11-${SERVING_VERSION}.tar.gz"
     fi
     client_release="paddle-serving-client==$SERVING_VERSION"
     app_release="paddle-serving-app==0.3.1"

tools/generate_k8s_yamls.sh

@@ -60,8 +60,8 @@ function run
   echo "named arg: command: $start_command"
   echo "named arg: port: $port"
   
-  sed -e "s/<< APP_NAME >>/$app_name/g" -e "s/<< IMAGE_NAME >>/$(echo $image_name | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< WORKDIR >>/$(echo $workdir | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< COMMAND >>/\"$start_command\"/g" -e "s/<< PORT >>/$port/g"  tools/k8s_serving.yaml_template > k8s_serving.yaml
-  sed -e "s/<< APP_NAME >>/$app_name/g" -e "s/<< IMAGE_NAME >>/$(echo $image_name | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< WORKDIR >>/$(echo $workdir | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< COMMAND >>/\"$start_command\"/g" -e "s/<< PORT >>/$port/g"  tools/k8s_ingress.yaml_template > k8s_ingress.yaml
+  sed -e "s/<< APP_NAME >>/$app_name/g" -e "s/<< IMAGE_NAME >>/$(echo $image_name | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< WORKDIR >>/$(echo $workdir | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< COMMAND >>/\"$(echo $start_command | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')\"/g" -e "s/<< PORT >>/$port/g"  tools/k8s_serving.yaml_template > k8s_serving.yaml
+  sed -e "s/<< APP_NAME >>/$app_name/g" -e "s/<< IMAGE_NAME >>/$(echo $image_name | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< WORKDIR >>/$(echo $workdir | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')/g" -e "s/<< COMMAND >>/\"$(echo $start_command | sed -e 's/\\/\\\\/g; s/\//\\\//g; s/&/\\\&/g')\"/g" -e "s/<< PORT >>/$port/g"  tools/k8s_ingress.yaml_template > k8s_ingress.yaml
   echo "check k8s_serving.yaml and k8s_ingress.yaml please." 
 }
 

tools/generate_runtime_docker.sh

@@ -66,6 +66,8 @@ function run
       base_image="nvidia\/cuda:10.1-cudnn7-runtime-ubuntu16.04"
   elif [ $env == "cuda10.2" ]; then
       base_image="nvidia\/cuda:10.2-cudnn8-runtime-ubuntu16.04"
+  elif [ $env == "cuda11" ]; then
+      base_image="nvidia\/cuda:11.0.3-cudnn8-runtime-ubuntu16.04"
   fi
   echo "base image: $base_image"
   echo "named arg: python: $python"

tools/k8s_serving.yaml_template

@@ -34,6 +34,7 @@ spec:
       containers:
       - image: << IMAGE_NAME >>
         name: << APP_NAME >>
+        imagePullPolicy: Always
         ports:
         - containerPort: << PORT >>
         workingDir: << WORKDIR >>
@@ -41,6 +42,8 @@ spec:
         command: ['/bin/bash', '-c']
         args: [<< COMMAND >>] 
         env:
+          - name: SERVING_BIN
+            value: "/usr/local/serving_bin/serving"
           - name: NODE_NAME
             valueFrom:
               fieldRef: