参加比赛，有幸有一台双T4机器使用，本来以为装上飞桨是小菜一碟，谁知过程非常曲折。

这台机器对我来说有挑战的地方有：

1 系统是centos7 ，而我以前接触AI框架都是ubuntu

2 cuda版本是10.2 而飞桨支持到10.1

3 双T4卡 

首先把cuda降级到10.1，然后通过晕头转向的操作（包括加路径，建链接等），飞桨倒是装上了，但是后来手残，装其它框架的时候动了cuda里的一些东西，导致飞桨崩了，并且我在重装cuda10.1之后，无论怎么晕头转向的操作，飞桨都报错。具体表现就是静态图ok，动态图报错：Error: Cannot load cudnn shared library. Cannot invoke method cudnnGetVersion at (/paddle/paddle/fluid/platform/dynload/cudnn.cc:63)

又重装了cuda10.0 ，还是同样报错。

最后只好又装了cuda9.2，这回报错变成了：

Error: Failed to find dynamic library: libnccl.so ( libnccl.so: cannot open shared object file: No such file or directory )

 Please specify its path correctly using following ways:

 Method. set environment variable LD_LIBRARY_PATH on Linux or DYLD_LIBRARY_PATH on Mac OS.

 For instance, issue command: export LD_LIBRARY_PATH=...

 Note: After Mac OS 10.11, using the DYLD_LIBRARY_PATH is impossible unless System Integrity Protection (SIP) is disabled. at (/paddle/paddle/fluid/platform/dynload/dynamic_loader.cc:177)

Your Paddle Fluid is installed successfully ONLY for SINGLE GPU or CPU!

 Let's start deep Learning with Paddle Fluid now

现在的问题就变成了装libnccl啦。到nvidia网站https://developer.nvidia.com/nccl，按照安装手册，装好之后，再运行测试：

python -c "import paddle; paddle.fluid.install_check.run_check()"

这回测试通过了，双卡飞桨装好了！

Running Verify Fluid Program ...

W0909 06:08:51.656754 20353 device_context.cc:252] Please NOTE: device: 0, CUDA Capability: 75, Driver API Version: 10.1, Runtime API Version: 10.0

W0909 06:08:51.660171 20353 device_context.cc:260] device: 0, cuDNN Version: 7.6.

Your Paddle Fluid works well on SINGLE GPU or CPU.

W0909 06:08:56.232347 20353 fuse_all_reduce_op_pass.cc:74] Find all_reduce operators: 2. To make the speed faster, some all_reduce ops are fused during training, after fusion, the number of all_reduce ops is 1.

Your Paddle Fluid works well on MUTIPLE GPU or CPU.

Your Paddle Fluid is installed successfully! Let's start deep Learning with Paddle Fluid now

飞桨在第一次安装之后，就没有再变动，后面只是在安装/调试cuda 。 话说，cuda这个东西，好是好，但是对非专业人士真的不太友好。