Bump CUDA from 11.2 to 11.8

[weiji14]

The CUDA 11.8 migration across conda-forge is practically complete (see https://conda-forge.org/status/#cuda118), so we can start updating to a newer version of CUDA!

This should add support for the new NVIDIA Ada Lovelace and Hopper generation GPUs that requires compute capability 8.9 or 9.0 (see https://docs.nvidia.com/cuda/archive/11.8.0/hopper-compatibility-guide/index.html#verifying-hopper-compatibility-using-cuda-11-8).

Note:

• There are actually CUDA 12.0 builds on conda-forge already, but it's probably good to have some docker images with CUDA 11.8 first, and transition to CUDA 12 later.
• CUDA 11.x has good forward and backwards compatibilty (see https://docs.nvidia.com/deploy/cuda-compatibility/index.html#cuda-intro), and as long as folks are using CUDA driver 450.36.06+, it should be ok.

Changes in this PR:

• Update Pytorch, Torchvision and Tensorflow to use CUDA 11.8 builds
• Update minimum pin on Tensorflow from >=2.9.1 to >=2.14.0 because lower versions <2.13.1 only has CUDA 11.2 on conda-forge.

References:

• https://developer.nvidia.com/blog/cuda-toolkit-11-8-new-features-revealed/
• Bump minimal CUDA version to CUDA 11.8? conda-forge/conda-forge.github.io#1981
• Add CUDA 11.8 migrator conda-forge/conda-forge-pinning-feedstock#4834

[pangeo-bot]

/condalock
Automatically locking new conda environment, building, and testing images...

[github-actions]

👈 Try on Mybinder.org!

[weiji14]

/condalock

[scottyhq]

Apparently we've been on 11.8 for some time, so perhaps this * syntax isn't really pinning anything?

pangeo-docker-images/pytorch-notebook/packages.txt

Lines 114 to 115 in d2651ff

	cuda-version==11.8
	cudatoolkit==11.8.0

[weiji14]

That's cudatoolkit. CUDA 11.2 and 11.8 does have this forward/backward compatibility thing, but if you look at the conda-lock.yml file, the Pytorch build is actually the CUDA 11.2 one:

pangeo-docker-images/pytorch-notebook/conda-lock.yml

Line 9696 in d2651ff

url: https://conda.anaconda.org/conda-forge/linux-64/pytorch-2.1.0-cuda112py311ha0492fd_300.conda

Compared to Pytorch compiled with CUDA 11.2, the one compiled with CUDA 11.8 enables compute capability 8.9, as seen at https://github.com/conda-forge/pytorch-cpu-feedstock/blob/7c7a57b7515eaeda67d3879b56b68466f38f0b0d/recipe/build_pytorch.sh#L144-L153.

[scottyhq]

looks like tensorflow image size dropped a bit, but pytorch keeps inflating :) would be nice to get these below 10GB if possible one of these days

pangeo/ml-notebook 
10.8GB ->  10GB

pangeo/pytorch-notebook   
13.7GB -> 13.9GB

[weiji14]

Yes, things should get smaller! This is because conda-forge has removed the need for a large cudatoolkit package in CUDA 12 (see conda-forge/conda-forge.github.io#1963) by breaking it into smaller components that are installed as needed. So hopefully that can cut down a few hundred megabytes 🤞