compile with nccl2 #8196
compile with nccl2 #8196
Conversation
Dockerfile
Outdated
| @@ -22,7 +22,7 @@ COPY ./paddle/scripts/docker/root/ /root/ | |||
|
|
|||
| RUN apt-get update && \ | |||
| apt-get install -y \ | |||
| git python-pip python-dev openssh-server bison libnccl-dev \ | |||
There was a problem hiding this comment.
Should also update tools/manylinux1/Dockerfile.x64 for release builds.
There was a problem hiding this comment.
libnccl is not at the nvidia/cuda:8.0-cudnn5-devel-ubuntu16.04. So I added it back. In this case, we don't need to update tools/manylinux1/Dockerfile.x64. :)
paddle/platform/CMakeLists.txt
Outdated
| @@ -1,5 +1,5 @@ | |||
| if(WITH_GPU) | |||
| cc_library(enforce SRCS enforce.cc DEPS nccl) | |||
| cc_library(enforce SRCS enforce.cc DEPS) | |||
There was a problem hiding this comment.
I vaguely remember that NVIDIA guy had upgraded these things with CUDA9, cudnn7, NCCL2.
You teach me that we need to call group_start before call ALLReduce NCCL calls. Is it right?
There was a problem hiding this comment.
Thank you for reminding me. I've added group_start and group_end in this PR.
|
Yang Yang seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account. You have signed the CLA already but the status is still pending? Let us recheck it. |
fix #8195