Pf buildbot #552
Pf buildbot #552
Conversation
|
There's also a patch needed to the llvm-zorg repo, which controls invocations of the buildbots from lab.llvm.org. In buildbot terms, this patch adds our buildbot to the configuration of the master, while the changes to llvm-project-mlir are for our buildbot worker. There are three hunks in the patch: one tells the command we want, a cmake of check-mlir with the ROCM flags on; one adds an email notification for failures; one adds our worker to the list. |
| ARGS+=" -e BUILDBOT_MASTER=${BUILDBOT_MASTER}" | ||
| fi | ||
|
|
||
| # Proxy needed on lockhart hosts. |
There was a problem hiding this comment.
This script is just like the one in llvm-zorg up to this point, where we set up a proxy for lockhart hosts (so the worker can read from github.com) and then run the docker image with access to the GPU and with useful options.
There was a problem hiding this comment.
Note to self for tomorrow: pass on how to talk to GitHub from Lockhart tomorrow
| RUN chmod a+rx /home/buildbot/run.sh | ||
|
|
||
| USER buildbot | ||
| ENV WORKER_NAME="mlir-rocm-mi200" |
There was a problem hiding this comment.
Mostly the same as the one I'm adapting from llvm-zorg, except for name changes.
|
|
||
| # Set up buildbot host and maintainer info. | ||
| mkdir -p "${WORKER_NAME}/info/" | ||
| echo "dl.mlse.buildbot@amd.com" > "${WORKER_NAME}/info/admin" |
There was a problem hiding this comment.
Email distribution list in case of problems.
| @@ -0,0 +1 @@ | |||
| rocmmlir22 | |||
There was a problem hiding this comment.
Password for the worker account. Will be sent by separate email to the buildbot-master administrator when we're ready to start.
mlir/utils/buildbot/build-run.sh
Outdated
| runproxy='' | ||
| fi | ||
|
|
||
| docker build ${buildproxy} -t "${IMAGE_NAME}:latest" . |
There was a problem hiding this comment.
Do we want to go ahead and make the proxy vars all caps to match the rest of the style?
There was a problem hiding this comment.
I thought at first that that was only for the ones passed to the environment, but no, it's all of them, so I updated.
Gather the buildbot setup files under mlir/utils/buildbot so they don't get lost. They're adapted from files in llvm-zorg/buildbot/google/docker, but don't quite fit there.
Start the buildbot on the intended worker host with
build-run.sh is in this review. It will build a docker image based on rocm/mlir:rocm5.0-latest using the Dockerfile here, culminating in running run.sh to start the buildbot worker in a docker container. The worker will connect to the master and listen for commands; the worker password is in the file named token. The container will persist, keeping its data in a couple of docker volumes, and will try to restart if the system is rebooted.