Tracking updates for CUDA 11.3 and 11.4

[h-vetinari]

CUDA 11.3 was released a while ago. Opening this issue to track roll-out, using the pattern from #47

11.3 / 11.4

• Add CUDA 11.3 / 11.4 #60
• Add cuda 11.3 / 11.4 nvcc-feedstock#69
• add cuda 11.3 / 11.4 docker-images#185
• Windows: add cuda 11.3 / 11.4 / 11.5 conda-forge-ci-setup-feedstock#154
• Migrator

Also ref to the docs for adding a new version.

CC @jaimergp

[jaimergp]

I am working on this now! 💪

[jaimergp]

Ok, that's it for now, sans the migrator.

We need to wait on cudatoolkit, which will enable docker-images, which will enable nvcc.

The Windows part should be self-sufficient.

[isuruf]

We should get #48 done before starting the migrator. Now that we have big packages like pytorch which needs manual compilations, it's a waste of developer time to build them for 11.3 when only packages that depend on nvrtc need to be rebuilt.

[jaimergp]

Fair enough! I'll comment on the issue then.

[h-vetinari]

Might as well make this a double-issue like #47 was: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

[jaimergp]

I'll add the relevant metadata but we still need to decide whether we split or not (#48).

[jakirkham]

NVIDIA Docker images were uploaded end of last week. Jaime updated the Docker PR to include 11.4. Those builds seem to work except for the fact that cudatoolkit packages are not available yet

[h-vetinari]

What's the status here not that conda-forge/nvcc-feedstock#71 has been merged?

[jakirkham]

Oh this is done. We now use CUDA Enhanced Compatibility to build with 11.2 and then allow users to install any cudatoolkit version 11.2 or greater (excluding the next major version)

[jakirkham]

Would like to do an announcement/blogpost, but wanted to complete aarch64 and ppc64le support first ( conda-forge/conda-forge-pinning-feedstock#2001 )

[h-vetinari]

OK, so we don't need newer nvcc builds in conda-forge (I get that it's possible to build for 11.4 with 11.2)? If so, should this issue be closed?

Would you be able to comment on the PTX-vs-ABI situation (and possibility to opt out of Enhanced Compatibility) in conda-forge/nvcc-feedstock#71?

[jakirkham]

Yeah exactly. Sure :)

That's already discussed earlier in that PR.

Not sure what is meant by opt-out. Maybe a new issue would be better for discussing other things?