update setup && doc for trt

6e5a02fa · MRXLT · 0e796150 · 6e5a02fa · 6e5a02fa · 6e5a02fa
5 changed file
--- a/cmake/paddlepaddle.cmake
+++ b/cmake/paddlepaddle.cmake
@@ -130,4 +130,9 @@ SET_PROPERTY(TARGET xxhash PROPERTY IMPORTED_LOCATION ${PADDLE_INSTALL_DIR}/thir
 LIST(APPEND external_project_dependencies paddle)

 LIST(APPEND paddle_depend_libs
-        xxhash nvinfer nvinfer_plugin)
+    xxhash)
+
+if(WITH_TRT)
+LIST(APPEND paddle_depend_libs
+    nvinfer nvinfer_plugin)
+endif()
--- a/doc/COMPILE.md
+++ b/doc/COMPILE.md
@@ -205,7 +205,8 @@ To compile the Paddle Serving GPU version on bare metal, you need to install the

 - CUDA
 - CuDNN
- NCCL2
+
+To compile the TensorRT version, you need to install the TensorRT library.

 Note here:

@@ -215,21 +216,12 @@ Note here:

 The following is the base library version matching relationship used by the PaddlePaddle release version for reference:

-|        |  CUDA   |          CuDNN           | NCCL2  |
-| :----: | :-----: | :----------------------: | :----: |
-| CUDA 8 | 8.0.61  | CuDNN 7.1.2 for CUDA 8.0 | 2.1.4  |
-| CUDA 9 | 9.0.176 | CuDNN 7.3.1 for CUDA 9.0 | 2.2.12 |
+|          |  CUDA   |          CuDNN           | TensorRT |
+| :----:   | :-----: | :----------------------: | :----:   |
+| post9    |  9.0    | CuDNN 7.3.1 for CUDA 9.0 |          |
+| post10   |  10.0   | CuDNN 7.5.1 for CUDA 10.0|          |
+| trt      |  10.1   | CuDNN 7.5.1 for CUDA 10.1| 6.0.1.5  |

 ### How to make the compiler detect the CuDNN library

 Download the corresponding CUDNN version from NVIDIA developer official website and decompressing it, add `-DCUDNN_ROOT` to cmake command, to specify the path of CUDNN.
-
-### How to make the compiler detect the nccl library
-
-After downloading the corresponding version of the nccl2 library from the NVIDIA developer official website and decompressing it, add the following environment variables (take nccl2.1.4 as an example):
-
-```shell
-export C_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$C_INCLUDE_PATH
-export CPLUS_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$CPLUS_INCLUDE_PATH
-export LD_LIBRARY_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/lib/:$LD_LIBRARY_PATH
-```
--- a/doc/COMPILE_CN.md
+++ b/doc/COMPILE_CN.md
@@ -202,7 +202,8 @@ Paddle Serving通过PaddlePaddle预测库支持在GPU上做预测。WITH_GPU选

 - CUDA
 - CuDNN
- NCCL2
+
+编译TensorRT版本，需要安装TensorRT库。

 这里要注意的是：

@@ -211,21 +212,12 @@ Paddle Serving通过PaddlePaddle预测库支持在GPU上做预测。WITH_GPU选

 以下是PaddlePaddle发布版本所使用的基础库版本匹配关系，供参考：

-|        |  CUDA   |          CuDNN           | NCCL2  |
-| :----: | :-----: | :----------------------: | :----: |
-| CUDA 8 | 8.0.61  | CuDNN 7.1.2 for CUDA 8.0 | 2.1.4  |
-| CUDA 9 | 9.0.176 | CuDNN 7.3.1 for CUDA 9.0 | 2.2.12 |
+|          |  CUDA   |          CuDNN           | TensorRT |
+| :----:   | :-----: | :----------------------: | :----:   |
+| post9    |  9.0    | CuDNN 7.3.1 for CUDA 9.0 |          |
+| post10   |  10.0   | CuDNN 7.5.1 for CUDA 10.0|          |
+| trt      |  10.1   | CuDNN 7.5.1 for CUDA 10.1| 6.0.1.5  |

 ### 如何让Paddle Serving编译系统探测到CuDNN库

-从NVIDIA developer官网下载对应版本CuDNN并在本地解压后，在cmake编译命令中增加`-DCUDNN_ROOT`参数，指定CuDNN库所在路径。
-
-### 如何让Paddle Serving编译系统探测到nccl库
-
-从NVIDIA developer官网下载对应版本nccl2库并解压后，增加如下环境变量 (以nccl2.1.4为例)：
-
-```shell
-export C_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$C_INCLUDE_PATH
-export CPLUS_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$CPLUS_INCLUDE_PATH
-export LD_LIBRARY_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/lib/:$LD_LIBRARY_PATH
-```
+从NVIDIA developer官网下载对应版本CuDNN并在本地解压后，在cmake编译命令中增加`-DCUDNN_LIBRARY`参数，指定CuDNN库所在路径。
--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -80,6 +80,16 @@ if (SERVER)
            COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
            DEPENDS ${SERVING_SERVER_CORE} server_config_py_proto ${PY_FILES})
        add_custom_target(paddle_python ALL DEPENDS ${PADDLE_SERVING_BINARY_DIR}/.timestamp)
+    elseif(WITH_TRT)
+        add_custom_command(
+            OUTPUT ${PADDLE_SERVING_BINARY_DIR}/.timestamp
+            COMMAND cp -r
+            ${CMAKE_CURRENT_SOURCE_DIR}/paddle_serving_server_gpu/ ${PADDLE_SERVING_BINARY_DIR}/python/
+            COMMAND env ${py_env} ${PYTHON_EXECUTABLE} gen_version.py
+            "server_gpu" trt
+            COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
+            DEPENDS ${SERVING_SERVER_CORE} server_config_py_proto ${PY_FILES})
+        add_custom_target(paddle_python ALL DEPENDS ${PADDLE_SERVING_BINARY_DIR}/.timestamp)
    else()
        add_custom_command(
            OUTPUT ${PADDLE_SERVING_BINARY_DIR}/.timestamp

--- a/python/setup.py.server_gpu.in
+++ b/python/setup.py.server_gpu.in
@@ -19,11 +19,13 @@ from __future__ import print_function
 from setuptools import setup, Distribution, Extension
 from setuptools import find_packages
 from setuptools import setup
-from paddle_serving_server_gpu.version import serving_server_version
+from paddle_serving_server_gpu.version import serving_server_version, cuda_version
 import util

-max_version, mid_version, min_version = util.python_version()
+if cuda_version != "trt":
+    cuda_version = "post" + cuda_version

+max_version, mid_version, min_version = util.python_version()
 # gen pipeline proto code
 util.gen_pipeline_code("paddle_serving_server_gpu")

@@ -56,7 +58,7 @@ package_data={'paddle_serving_server_gpu': ['pipeline/gateway/libproxy_server.so

 setup(
    name='paddle-serving-server-gpu',
-    version=serving_server_version.replace('-', '') + '.post@CUDA_VERSION_MAJOR@',
+    version=serving_server_version.replace('-', '') + "." + cuda_version,
    description=
    ('Paddle Serving Package for saved model with PaddlePaddle'),
    url='https://github.com/PaddlePaddle/Serving',