-
-
Notifications
You must be signed in to change notification settings - Fork 4.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
add nvidia/cutlass #17303
add nvidia/cutlass #17303
Conversation
Hi! This is the friendly automated conda-forge-linting service. I wanted to let you know that I linted all conda-recipes in your PR ( Here's what I've got... For recipes/cutlass:
|
Hi! This is the friendly automated conda-forge-linting service. I wanted to let you know that I linted all conda-recipes in your PR ( Here's what I've got... For recipes/cutlass:
|
Hi! This is the friendly automated conda-forge-linting service. I just wanted to let you know that I linted all conda-recipes in your PR ( |
Ah, the tests seem to need a GPU, which the CI doesn't have...
|
Hi! This is the friendly automated conda-forge-linting service. I was trying to look for recipes to lint for you, but couldn't find any. |
@conda-forge/core, there's a bug with the linter here. As of 23186f9, there's definitely a recipe under |
So, one thing is that things seem to get installed into The other thing is that I saw some rpath-manipulation that we might want to patch out (esp the
|
Hi! This is the friendly automated conda-forge-linting service. I just wanted to let you know that I linted all conda-recipes in your PR ( |
OK, this now builds 🥳 (Magic variable turned out to be I'm still a bit worried about the following though. Could someone either confirm or deny that we should patch this @conda-forge/cudatoolkit @jaimergp @xhochy?
|
I thought conda-build would prune unnecessary rpaths using patchelf/lief, so that we don't have to worry? |
Ah, that's amazing if correct. I wasn't aware of that. |
Well, assuming the rpaths set by the package get cleaned up correctly by patchelf/lief, then this is ready for review @conda-forge/staged-recipes |
Sorry, forgot to remove the "WIP: " from the title... |
@kkraus14 |
Thanks @kkraus14! Anyone from @conda-forge/staged-recipes wants to have a look as well? 🙃 |
Oh! So we hit another compiler bug?! |
Looks like it! But curiously, only on CUDA 11.1 & 11.2, not on CUDA 11.0 (ignore the current CI failure, this is running into an incompatible GCC-version, because I was too lazy to do all the cuda-version-specific cbc magic; previous runs show that CUDA 11.0 passed with GCC 9). |
d7d724d
to
4f4db20
Compare
I tried copying the content of conda-forge/conda-forge-pinning-feedstock#2010 into cbc.yaml, but it didn't work. Would be happy to hear how I can set GCC=10 for CUDA 11.1/11.2, but leave GCC 9 for CUDA 11.0. |
Any other difference between staged-recipes and a "normal" feedstock I'm overlooking here? From what I can tell, using parts of the pinning-cbc verbatim should always work (but doesn't here)? |
Could it be that it's due to staged-recipes not using conda-smithy, and hence not knowing about In that case, I won't get the recipe to pass here fully (but should be fine on the feedstock... 🤞) --- CUDA 11.0 needs GCC 9, and CUDA 11.1/11.2 (in this recipe) needs GCC 10. The only other way I see is to try patch the logic for CF_CUDA_ENABLED into If I'm right with this summary, this would also mean that staged-recipes in its current state will not be able to reflect conda-forge/conda-forge-pinning-feedstock#2010 once merged. CC @conda-forge/core (for conda-smithy) |
Nope. The version is determined by the config files. For eg: https://github.com/conda-forge/staged-recipes/blob/main/.ci_support/linux64_cuda102.yaml#L7 which we can edit after conda-forge-pinning PR is merged. |
Ah, makes sense, just that it's hard-coded and can't be overridden by cbc. Well, lucky that conda-forge/conda-forge-pinning-feedstock#2010 seems to be nearing completion, otherwise this one would be stuck 😅 |
@conda-forge/staged-recipes, this is finally ready! :) I also rebased for a clean commit history so that this doesn't have to be squash-merged. Still, if necessary, I can break out c4589d0 into a separate PR. |
Gentle ping @conda-forge/staged-recipes 🙃 |
Co-authored-by: Daniel Ching <carterbox@users.noreply.github.com>
@kerrmudgeon I wanted to ask if you could provide any information on the binary compatibility policies of cutlass? I.e. would a library built against 2.8.0 be able to load 2.9.0 at runtime (assuming matching CUDA version)? |
I'm going to merge this PR. You can relax the run_export in your new feedstock if needed. |
Thanks for review and merge @carterbox! |
Package https://github.com/NVIDIA/cutlass, with the goal of packaging https://github.com/openai/triton and eventually https://github.com/facebookresearch/xformers