[CI] Add macos build test #3994
Conversation
| ~/.triton/nvidia | ||
| ~/.triton/pybind11 | ||
| ~/.triton/json | ||
| key: ${{ runner.os }}-${{ runner.arch }}-llvm-${{ steps.cache-key.outputs.llvm }}-nvidia-${{ steps.cache-key.outputs.nvidia }}-pybind11-${{ steps.cache-key.outputs.pybind11 }} |
There was a problem hiding this comment.
Instead of copy-pasting sections like this one, can we use the YAML anchors? Replace this section with
- *cache-build-dependencies-step
(Also it seems that we're caching ~/.triton/json here but I don't see it in the CUDA runner. That's part of the reason to use the anchors, so that code that's supposed to be the same is the same.)
There was a problem hiding this comment.
No problem, I'll do that. Just trying to make the build runner pass first.
There was a problem hiding this comment.
ty ty
Sorry that there's some unused cruft in here. I'm happy to remove it if after you're done with your changes if @ThomasRaoux is ok with that.
| @@ -155,7 +155,7 @@ jobs: | |||
| runs-on: ${{ matrix.runner }} | |||
| timeout-minutes: 30 | |||
| env: | |||
| CMAKE_BUILD_TYPE: "MIN_SIZE_REL" | |||
| CMAKE_BUILD_TYPE: "Debug" | |||
There was a problem hiding this comment.
Just to over-communicate, I would not expect "Debug" or "MinSizeRel" is the fastest build; I'd expect -O1 is probably faster than both.
But also seeing as Debug seems to be the same speed as Release (15m) maybe something else is causing the slowness...
|
github is not loading the logs for the most recent build so I wasn't able to investigate this myself, but it may be worth investigating if you haven't how much of the 12m30s build step is due to downloading dependencies versus actually building Triton. IIRC downloading the CUDA and nvidia dependencies was slow for Linux, so we started caching them. But I'm not sure if the cache works here yet. |
|
Hi @jlebar, it's ready for re-review. :) |
| brew install ccache | ||
| brew install llvm |
There was a problem hiding this comment.
| brew install ccache | |
| brew install llvm | |
| brew install ccache llvm |
| TRITON_BUILD_WITH_O1: "true" | ||
| # macos-latest has a 3-core M1 vcpu | ||
| # https://docs.github.com/en/actions/using-github-hosted-runners/about-github-hosted-runners/about-github-hosted-runners#standard-github-hosted-runners-for-public-repositories | ||
| MAX_JOBS: 3 |
There was a problem hiding this comment.
I don't think this is quite the right comment. ninja -j will read the number of CPUs and launch f(ncpu) jobs, I think 2*ncpu.
You're restricting to 3 jobs and because I presume you found it's faster? Probably due to limited RAM?
There was a problem hiding this comment.
Yes. Due to limited RAM.
I'd like to launch ncpu instead of 2*cpu because M1 doesn't have hyperthreading.
There was a problem hiding this comment.
Right, just suggest updating the comment to explain the reason, which isn't really that M1 only has 3 cores.
(Also usually build systems will do 2ncpu even if there's no hyperthreading, because sometimes jobs are I/O bound. Indeed I think they probably do 2nvcpu, i.e. if M1 had hyperthreading it would by default launch 12 jobs.)
| python3 -m pip install cython setuptools wheel cmake==3.24 ninja pytest-xdist lit | ||
| - name: Install Triton | ||
| env: | ||
| TRITON_BUILD_WITH_CCACHE: "true" |
There was a problem hiding this comment.
I don't see us caching the ccache directory, in which case ccache is just overhead.
There was a problem hiding this comment.
Sorry, forgot the caching step...
| @@ -35,7 +35,8 @@ endif() | |||
|
|
|||
| # Check if the platform is MacOS | |||
| if(APPLE) | |||
| set(PROTON_PYTHON_LDFLAGS "-undefined dynamic_lookup -flto") | |||
| set(CMAKE_SHARED_LIBRARY_SUFFIX ".so") | |||
| set(PROTON_PYTHON_LDFLAGS "-undefined dynamic_lookup") | |||
There was a problem hiding this comment.
| set(PROTON_PYTHON_LDFLAGS "-undefined dynamic_lookup") | |
| # Other platforms build with -flto, but we found that this adds significant overhead to our macos CI without providing a major benefit. | |
| set(PROTON_PYTHON_LDFLAGS "-undefined dynamic_lookup") |
|
LGTM! |
No description provided.