feat: add libwholegraph wheel#182
Conversation
* Remove `USE_NCCL_WHEEL` arg where I was getting ahead of myself * Add notes about what things do * set tests and benchmarks to OFF * make sure name passed to `project` matches name passed to `rapids_cuda_init_architectures`
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
|
/ok to test |
5751242 to
081a3ac
Compare
|
/ok to test |
1 similar comment
|
/ok to test |
f933c99 to
a70942a
Compare
|
/ok to test |
don't need to expand over Python versions, only CUDA and ARCH
|
/ok to test |
1 similar comment
|
/ok to test |
jameslamb
left a comment
There was a problem hiding this comment.
Hope you don't mind me crashing in, but I was interested when I saw @alexbarghi-nv post that this was up. Put up some comments for your consideration.
Leaving a "comment" review so my review doesn't block merging in case you get approvals between now and the next time I'm able to look at this.
Co-authored-by: James Lamb <jaylamb20@gmail.com>
This reverts commit eed6b84.
Co-authored-by: Bradley Dice <bdice@bradleydice.com>
|
The pip devcontainer build is looking for Also you might need to update the devcontainers like we did here for cuml: rapidsai/devcontainers#442. Get that PR ready first, and we'll merge it just before we merge this one (to avoid an admin-merge cycle). |
There is an unsuffixed fallback in the I think it's failing because there are no |
That's not it. @bdice 's advice in #182 (comment) is the fix... get a |
|
devcontainers are currently broken by |
Replied here: rapidsai/devcontainers#494 (comment) |
|
All looks fine here with one required follow-up: we're shipping a ~350MB wheel but the conda package is only 18MB. The difference appears to be in NCCL. We need to stop shipping NCCL in this wheel as @divyegala recently did with RAFT and cuVS. @gforsyth Here's what I propose:
|
|
Yeah, I have a note down for the NCCL devendoring followup |
xref: rapidsai/build-planning#33 xref: rapidsai/cugraph-gnn#182 Add `libwholegraph` wheels to the build manifest. Currently pointing at my fork in `cugraph-gnn` for testing. --------- Co-authored-by: Bradley Dice <bdice@bradleydice.com>
|
/merge |
Followup to #182, related to #386 (review) Proposes using `--no-build-isolation` for `libwholegraph` wheel builds. `--no-build-isolation` preserves all of the source file paths, which should lead to better `sccache` hit rate and faster builds. All of the other `lib{something}` wheels in RAPIDS are built this way, see rapidsai/build-planning#108 Also switches `cugraph-pyg` builds to smaller runners (cpu4). That's a pure-Python package that shouldn't need many resources to build. Authors: - James Lamb (https://github.com/jameslamb) Approvers: - Bradley Dice (https://github.com/bdice) URL: #388
Contributes to rapidsai/build-planning#33
Resolves rapidsai/rapids-wheels-planning#59
This PR proposes packaging
libwholegraphas a wheel, which is then re-used bypylibwholegraph,cugraph-dgl, andcugraph-pyg.It should avoid recompiling RAFT and RMM, as those are now both handled by their respective wheels.
The
wholegraphlibrary is now compiled as a shared library.