Deprecate CUDA 10.2 docker images

[msaroufim]

🚀 The feature

CUDA 10.2 is getting deprecated with the 1.13 release of PyTorch so we should start making our official releases use CUDA 11.6 or 11.8 https://pytorch.org/blog/PyTorch-1.13-release/#introduction-of-cuda-116-and-117-and-deprecation-of-cuda-102-and-113

Motivation, pitch

n

Alternatives

No response

Additional context

No response

[Lauler]

@msaroufim As a random user coming here to see why your containers are still built with 10.2, I concur...

I expected your newest official torchserve docker containers to already be built with CUDA 11. Spent a couple of hours debugging why GPU inference wouldn't work, convinced that my handler.py was buggy. Turned out torchserve:latest-gpu and torchserve:0.6.1-gpu are built with CUDA 10.2 :/

(RTX 3090 not supported in CUDA 10.2)

[msaroufim]

We updating cuda to 11 in the next release sometime next week

[mayankdebnath]

We updating cuda to 11 in the next release sometime next week

Any updates about this? Eagerly waiting for the Docker images to support A100.

[msaroufim]

Release happening Monday