目的是云端算法中执行LSTM部分计算过程的加速，即用cu文件编译出so，用此so中的LSTM类或函数替代tf.LSTMCell进行运算。
整个项目见Github，流程见博客，博主也刚入门cuda，欢迎留言探讨~

使用自定义操作提供TensorFlow模型

TensorFlow预先构建了一个广泛的操作库和操作内核（实现），可针对不同的硬件类型（CPU，GPU等）进行微调。这些操作自动链接到TensorFlow Serving ModelServer二进制文件，无需用户进行额外的工作。但是，有两个用例需要用户在ops中显式链接到ModelServer：

您已经编写了自己的自定义操作（例如，使用本指南）
您正在使用TensorFlow未附带的已实现的操作系统

注意：从2.0版开始，TensorFlow不再分发contrib模块; 如果您使用contrib ops提供TensorFlow程序，请使用本指南明确地将这些操作链接到ModelServer。

无论您是否实现了操作，为了使用自定义操作来提供模型，您都需要访问操作系统的源代码。本指南将指导您完成使用源以使自定义操作可用于服务的步骤。有关自定义操作的实现的指导，请参阅 tensorflow。

先决条件：已经编写了自定义操作并注册到tensorflow op。

将op源复制到Serving项目中

tensorflow_serving文件夹下创建以cuda_lstm_forward命名的文件夹

然后同时把"00_lstm.cu", “00_lstm.so” , “cuda_lstm_forward.h”, “cuda_lstm_forward.cc”，"cuda_lstm_forward.so,"即所有依赖项放到当前文件夹下

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-VNiHKFHe-1571281039666)(/Users/zhipeng/Library/Application Support/typora-user-images/image-20190923101116942.png)]

为op构建静态库

在cuda_lstm_forward的文件夹，您会看到一个生成共享对象文件（.so）的目标，您可以将其加载到python中以创建和训练模型。但是，TensorFlow服务在构建时静态链接操作，并且需要一个.a文件。因此，需要创建一个生成此文件的构建规则 tensorflow_serving/cuda_lstm_forward/BUILD：

package(
    default_visibility = [
        "//tensorflow_serving:internal",
    ],
    features = ["-layering_check"],
)



cc_library(
    name = "cuda_lstm_forward.so",
    visibility = ["//visibility:public"],
    srcs = [
            "cuda_lstm_forward.cc",
            #"cuda_lstm_forward.h",
            "lib00_lstm.so",
            #"00_lstm.cu.cc"
           ] ,
    copts = ["-std=c++11"],
    deps = ["@org_tensorflow//tensorflow/core:framework_headers_lib",
            "@org_tensorflow//tensorflow/core/util/ctc",
            "@org_tensorflow//third_party/eigen3",
    ],
    alwayslink=1,
)

bazel BUILD规则参考：bazel C/C++ Rules

使用op链接构建ModelServer

要为使用自定义操作的模型提供服务，您必须使用链接的操作构建ModelServer二进制文件。具体来说，您将cuda_lstm_forward上面创建的构建目标添加到ModelServer的BUILD文件中。

编辑tensorflow_serving/model_servers/BUILD以添加目标中SUPPORTED_TENSORFLOW_OPS包含的自定义op构建server_lib目标：

找到“SUPPORTED_TENSORFLOW_OPS”，做如下修改：

SUPPORTED_TENSORFLOW_OPS = [
    "@org_tensorflow//tensorflow/contrib:contrib_kernels",
    "@org_tensorflow//tensorflow/contrib:contrib_ops_op_lib",
    #Added this line
    #"//tensorflow_serving/fsmn_forward:fsmn_forward.so",
    "//tensorflow_serving/cuda_lstm_forward:cuda_lstm_forward.so"  
]

然后使用tensorflow_serving上层目录下的build_tf.sh编译ModelServer：

#FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04
export TF_CUDA_VERSION=8.0
export TF_CUDNN_VERSION=6
TF_SERVING_COMMIT=tags/1.5.0
#TF_COMMIT=tags/ais_v0.0.1
BAZEL_VERSION=0.15.0

export  TF_NEED_CUDA=1
export  TF_NEED_S3=1
export  TF_CUDA_COMPUTE_CAPABILITIES="3.5,5.2,6.1"
export  TF_NEED_GCP=1
export  TF_NEED_JEMALLOC=0
export  TF_NEED_HDFS=0
export  TF_NEED_OPENCL=0
#export  TF_NEED_MKL=1
export  TF_NEED_VERBS=0
export  TF_NEED_MPI=0
export  TF_NEED_GDR=0
export  TF_ENABLE_XLA=0
export  TF_CUDA_CLANG=0
export  TF_NEED_OPENCL_SYCL=0
export  CUDA_TOOLKIT_PATH=/usr/local/cuda
export  CUDNN_INSTALL_PATH=/usr/local/cuda
#export  MKL_INSTALL_PATH=/opt/intel/mkl
export  GCC_HOST_COMPILER_PATH=/usr/bin/gcc
export  PYTHON_BIN_PATH=/usr/bin/python
export  PYTHON_LIB_PATH=/usr/lib/python2.7/site-packages/
export  CC_OPT_FLAGS="-march=native"

if [ ! -d "./tensorflow" ]; then
        git clone https://gitlab.spetechcular.com/core/tensorflow.git
fi

if [ ! -d "./build_out" ]; then
    mkdir ./build_out
fi

#git checkout $TF_SERVING_COMMIT
cd ./tensorflow && \
#git checkout $TF_COMMIT
TF_SET_ANDROID_WORKSPACE= ./configure

cd ..
    #bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=cuda -k --verbose_failures  --crosstool_top=@local_config_cuda//crosstool:toolchain tensorflow_serving/model_servers:tensorflow_model_server

    #bazel build -c opt --copt=-mavx --copt=-mfpmath=both --copt=-msse4.2 --config=cuda -k --verbose_failures --crosstool_top=@local_config_cuda//crosstool:toolchain --spawn_strategy=standalone tensorflow_serving/model_servers:tensorflow_model_server
    bazel build -c opt --config=cuda -k --verbose_failures --crosstool_top=@local_config_cuda//crosstool:toolchain tensorflow_serving/model_servers:tensorflow_model_server

提供包含您的自定义操作的模型

把savemodel拷贝进/home/public/tfs_sever_gpu/tfs_models/cudalstm/ 重命名为1，需要用命令行改mv savedmodel/ 1
把依赖项lib00_lstm.so拷贝到/home/public/tfs_sever_gpu/cuda_so/
把编译好的/home/public/serving_gpu_15addopjiami/bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server拷贝到/home/public/tfs_sever_gpu/bin/下，拷贝时无权限用chmod 777 ，用ll查看更新时间确保拷贝成功
ps -ef |grep tensorflow查看端口号
在/home/public/tfs_sever_gpu/下 sh run13.sh开服务器，不需要激活环境和本机无关
在/home/public/tfs_sever_gpu/model_transfer/下python lstmctc.py and python cudalstm.py，环境source activate /home/pz853/anaconda3/envs/py2，这里的环境缺很多包grpc, tensorflow-serving-apiGPUpip第二个后少abs，并且无可逆注意下
多并发即同样目录下写个shell脚本用 & 多次跑，测到1 2 4 8 16 32 64 128

#!/home/public/tfs_sever_gpu/run.sh

basepath=$(dirname $(readlink -f $0))

# ipcpath 按需修改
#ipcpath=unix:${basepath}/f
ipcpath=10.12.8.26:9001

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${basepath}/cuda_so

# del model.cfg if exist
if [ -f "${basepath}/cfg/model.cfg" ]; then
    rm "${basepath}/cfg/model.cfg"
fi

# prepare model.cfg file
model_cfg=${basepath}/cfg/model.cfg
touch "$model_cfg"
echo  "model_config_list: {" >> $model_cfg
for dir in `ls ${basepath}/tfs_models`
do
    if [ -d "${basepath}/tfs_models/${dir}" ]; then
        echo "  config: {" >> $model_cfg
        echo "    name: \"${dir}\"," >> $model_cfg
        echo "    ## must be absolute path" >> $model_cfg
        echo "    base_path: \"${basepath}/tfs_models/${dir}\"," >> $model_cfg
        echo "    model_platform: \"tensorflow\"" >> $model_cfg
        echo "    model_version_policy: {" >> $model_cfg
        echo "      all: {}" >> $model_cfg
        echo "    }" >> $model_cfg
        echo "  }" >> $model_cfg
    fi
done
echo "}" >> $model_cfg

if [ ! -f "$model_cfg" ]; then
    echo "error: $model_cfg not exist"
    exit
fi

platform_cfg=${basepath}/cfg/platform.cfg
if [ ! -f "$platform_cfg" ]; then
    echo "error: $platform_cfg not exist"
    exit
fi

tfs_bin=${basepath}/bin/tensorflow_model_server_addop

#cpuinfo=`cat /proc/cpuinfo |grep flags | sed -n '1p' |grep avx -c`
#if [ $cpuinfo -gt 0 ]; then
#    tfs_bin=${basepath}/bin/tensorflow_model_server_xla
#fi
#cpuinfo=`cat /proc/cpuinfo |grep flags | sed -n '1p' |grep avx2 -c`
#if [ $cpuinfo -gt 0 ]; then
#    tfs_bin=${basepath}/bin/tensorflow_model_server_sse
#fi
#echo $tfs_bin 

#tfs_bin=${basepath}/bin/tensorflow_model_server

while true;
do
    #CUDA_VISIBLE_DEVICES=0 exec ${tfs_bin}  --ipcpath=10.12.8.26:9002 --model_config_file=${model_cfg} --platform_config_file=${platform_cfg}
    CUDA_VISIBLE_DEVICES=0 exec ${tfs_bin}  --port=9001 --model_config_file=${model_cfg} --platform_config_file=${platform_cfg}
done

原文链接：https://blog.csdn.net/qq_43208303/article/details/102601555