search

substratusai / images

Official Substratus Container Images
1 stars 0 forks source link

docker images are currently incompatible with m1/m2 mac platform - unable to use locally #14

Closed brandonjbjelland closed 1 year ago

brandonjbjelland commented 1 year ago

This is looking like a yak shave so I'll come back to it but running anything locally on my Mac isn't supported atm. Why?:

  1. we don't publish an arm compatible image. Simple enough substitute platforms: "linux/amd64,linux/arm64" in our github workflow

  2. We'll have a container image that expects local CUDA libraries. No mac ships with a nvidia GPU - those libs wont exist.

    qemu-x86_64: Could not open '/lib64/ld-linux-x86-64.so.2': No such file or directory

Some paths forward:

  1. encourage only local (non-docker) development for mac users - I think we can do better than this
  2. create a separate base image not based on CUDA - I didn't like this thought initially but since we parameterize FROM anyway, we could pass in a more universally-compatible base image for local build and load use cases.
brandonjbjelland commented 1 year ago

https://github.com/substratusai/images/pull/15

brandonjbjelland commented 1 year ago

Closing as won't do - the story for running docker with GPUs on macs is poor. Nvidia has a stranglehold here and Apple Metal Performance Shaders (MSP) is currently not supported: https://github.com/sonoisa/arm64-docker-pytorch-tensorflow/issues/1

The answer here is to:

  1. use substratus for notebook dev
  2. develop locally using a local kernel, not in docker