Rebuild cuda118, cuda1120, and protobuf

[RaulPPelaez]

Checklist

• Used a personal fork of the feedstock to propose changes
• Bumped the build number (if the version is unchanged)
• Reset the build number to 0 (if the version changed)
• Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
• Ensured the license file is being packaged.

Continuing #191 I added the necessary CUDA arch flags to build_pytorch.sh.

[conda-forge-webservices]

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 (recipe) and found it was in an excellent condition.

[RaulPPelaez]

@conda-forge-admin, please rerender

[RaulPPelaez]

These ones were running, but they were timed out:

  [5220/5707] Building CUDA object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/ForeachUnaryOp.cu.o
  [5221/5707] Building CUDA object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/IGammaKernel.cu.o
  [5222/5707] Building CUDA object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/Im2Col.cu.o
  [5223/5707] Building CUDA object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/GridSampler.cu.o
##[error]The operation was canceled.
Finishing: Run docker build

They just take really really long it seems. Not sure what can be done in order to reduce this time.
Any advice here @jakirkham?

[hmaarrfk]

I've always been building cuda locally. the fat binaries are just really slow to compile.

[hmaarrfk]

I can try to start building this locally and see where it goes. I am not overly enthusiastic to build out 3 cuda versions (11.2, 11.8, and 12.0).....

[RaulPPelaez]

I've always been building cuda locally. the fat binaries are just really slow to compile.

I did not know that was an option hehe. I know how to build locally, but I assume the results are not something I can somehow upload here myself?

I can try to start building this locally and see where it goes. I am not overly enthusiastic to build out 3 cuda versions (11.2, 11.8, and 12.0).....

I understand it is a hassle, but torch is a really big package and I believe it is worth to provide these options. It is true though that minor CUDA versions are supposed to be compatible amongst them, but even in that case I would assume CUDA 11.8 provides some kind of performance enhancement with respect 11.2.
At least CUDA 12 should be an option to reduce the chances of dependency hell.

[hmaarrfk]

See https://github.com/conda-forge/cfep/blob/main/cfep-03.md

is 11.8 useful if we provide 12.0?

It just really is ALOT of package data (storage) and compilation time.

[hmaarrfk]

@RaulPPelaez were you able to get Cuda 12.0 to also compile? if so do you want to add it to this merge request and I can trigger a rebuild.

[jakirkham]

the biggest barrier is me starting my script.

Yeah we need to figure out a way to automate builds

[RaulPPelaez]

@RaulPPelaez were you able to get Cuda 12.0 to also compile? if so do you want to add it to this merge request and I can trigger a rebuild.

Those also timed out, the PR for CUDA 12 #193 should also be compatible with CUDA 11.8.
Not sure how the pipeline works in this case. I merged #193 here, not sure if that is enough.

[RaulPPelaez]

@conda-forge-admin, please rerender

[jakirkham]

Think what Mark is asking is if you have tried building locally (and if so whether those builds completed)

Guessing most build errors occur early (like those seen previously: #193 (comment) )

Though there are some errors that might not show up until the linking stage

Another question would be whether we see issues during testing (not sure what testing is done for the CUDA packages)

[hmaarrfk]

I have the following done canaries completed.

conda-forge-build-done-linux_64_blas_implgenericc_compiler_version11cuda_compilernvcccuda_compiler_version11.8cxx_compiler_version11numpy1.22python3.10.____cpython
conda-forge-build-done-linux_64_blas_implgenericc_compiler_version11cuda_compilernvcccuda_compiler_version11.8cxx_compiler_version11numpy1.22python3.8.____cpython
conda-forge-build-done-linux_64_blas_implgenericc_compiler_version11cuda_compilernvcccuda_compiler_version11.8cxx_compiler_version11numpy1.22python3.9.____cpython
conda-forge-build-done-linux_64_blas_implgenericc_compiler_version11cuda_compilernvcccuda_compiler_version11.8cxx_compiler_version11numpy1.23python3.11.____cpython
conda-forge-build-done-linux_64_blas_implmklc_compiler_version10cuda_compilernvcccuda_compiler_version11.2cxx_compiler_version10numpy1.22python3.10.____cpython
conda-forge-build-done-linux_64_blas_implmklc_compiler_version10cuda_compilernvcccuda_compiler_version11.2cxx_compiler_version10numpy1.23python3.11.____cpython
conda-forge-build-done-linux_64_blas_implmklc_compiler_version12cuda_compilercuda-nvcccuda_compiler_version12.0cxx_compiler_version12numpy1.22python3.8.____cpython

That said, i'm wondering if for the sake of conda-forge's storage quote, if i shoul dcancel (now that I have one build of 11.2, 11.8 and 12.0 complete) cancel these builds and work to integrate #197

[jakirkham]

@h-vetinari do you have a sense of how #197 should factor in here?

[h-vetinari]

@h-vetinari do you have a sense of how #197 should factor in here?

If it can be integrated, all the better! If not, it's not the end of the world, but should then go into the next build. With pytorch, the recent protobuf bumps seemed to go pretty smoothly though (from what I can tell, esp. compared to TF).

[hmaarrfk]

they work. builds take a full day. restarting them now with protobuf migration

[hmaarrfk]

log files

log_files.zip

[Tobias-Fischer]

Hi @hmaarrfk - have these builds actually been uploaded to the conda-forge channel? I might be blind seeing only the cpu builds.

[Tobias-Fischer]

Background: I’m trying to figure out why conda-forge/pytorch_scatter-feedstock#55 can’t find solve the environment

[hmaarrfk]

lol, maybe not.... I might just upload the next ones if that is ok?

[jakirkham]

Meaning with the changes from PR ( #199 )?

[hmaarrfk]

yes.

I mean, i'm really not excited about them either, but they are complete....

[jakirkham]

Yeah that seems better from a user standpoint (recent version + Python 3.12)

[hmaarrfk]

Ohhhh. i found the builds:

https://anaconda.org/mark.harfouche/pytorch/files
https://anaconda.org/mark.harfouche/pytorch-gpu/files

Just copied them over.

[jakirkham]

Thanks Mark! 🙏