If you’re running conda in the native (osx-arm64) mode you won’t be able to install some old packages because they’re not compiled for that architecture.
You could try taking advantage of Rosetta to emulate osx-arm64 but the performance will be worse
As for osx-arm64: yeah, you get what you pay for. You’ll be stuck with running old pythons under rosetta. 3.6 is EOL, so no amount of asking is going to get you any support, and it’s highly unlikely 3.7 will be made available.
I’d say try to get your tutorial working under 3.8, maybe make a PR.
More generally:
I generally can’t recommend the jupyter metapackage. It hauls along a few extra 100mb of complex dependencies (e.g. qt) which might be confounding things. And remember: none of these things have been tested natively on osx-arm64 because there are no free CI M1 assets for open source maintainers to test against.
Further, start with mambaforge, and never install anything in the base environment it gives you. Use mamba, as it halts faster (either with a working env, or with more useful errors).
pytorch is a beast to build, and the pytorch channels is one of the few I can recommend over conda-forge.
For each new environment, create and check in an environment.yml.
name: my-py36-jupyter-env
channels:
- conda-forge
- nodefaults
dependencies:
- python >=3.6,<3.7 # or whatever
- jupyterlab >=3,<4
- pytorch
# or you can try with the `{channel}::` prefix
# - pytorch::pytorch