fix nccl version (#8540)

* fix nccl version

* enable nccl test

Author: QiJune

Dockerfile

@@ -22,7 +22,8 @@ COPY ./paddle/scripts/docker/root/ /root/
 
 RUN apt-get update && \
     apt-get install -y \
-    git python-pip python-dev openssh-server bison libnccl-dev \
+    git python-pip python-dev openssh-server bison \
+    libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
     python-matplotlib gcc-4.8 g++-4.8 \

paddle/fluid/platform/nccl_test.cu

@@ -129,9 +129,6 @@ TEST(NCCL, all_reduce) {
 }  // namespace paddle
 
 int main(int argc, char** argv) {
-  // FIXME(tonyyang-svail):
-  //   Due to the driver issue on our CI, disable for now
-  return 0;
   dev_count = paddle::platform::GetCUDADeviceCount();
   if (dev_count <= 1) {
     LOG(WARNING)

paddle/scripts/docker/build.sh

@@ -171,7 +171,7 @@ EOF
 EOF
 
     if [[ ${WITH_GPU} == "ON"  ]]; then
-        NCCL_DEPS="apt-get install -y libnccl-dev &&"
+        NCCL_DEPS="apt-get install -y libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 &&"
     else
         NCCL_DEPS=""
     fi