m1_huggingface_diffusers_demo VS conda

The trick is that you have to deactivate the virtual environment and then resource it after adding Jupyter to that virtual environment.

Most shells cache executable paths, so the path for jupyter will be the global path, not the one for your virtual environment. This is unfortunately not at all obvious and leads to very hard to track down bugs that seem to disappear and reappear if you aren't familiar with the issue.

I have a recipe here which always works: https://github.com/nlothian/m1_huggingface_diffusers_demo#se...

If you don't have requirements.txt then do this: `pip3 install jupyter` for that line, then `deactivate` and `source ./venv/bin/activate`.

This is how I did it on an M1 in September: https://github.com/nlothian/m1_huggingface_diffusers_demo

I think it probably needs updating now, but it should give you something to start with.

On my M1 MAx with 32 GB I'm getting 1.5 iterations/second (ie, ~30 seconds for the standard 50 iterations) using this example: https://github.com/nlothian/m1_huggingface_diffusers_demo

Out of interest I've been running a bunch of the huggingface version of StableDiffusion using the M1 accelerated branch on my M1 Max[1]. I'm getting 1.54 it/s compared to 2.0 it/s for a Nvidia T4 Tesla on Google Collab.

T4 Tesla gets 21,691 queries/second for for ResNet, compared to 81,292 q/s for the new H100, 41,893 q/s for the A100 and 6164 q/s for the new Jetson.

So you can expect maybe 15,000 q/s on a M1 Max. But some tests seem to indicate a lot less[2] - not sure what is happening there.

[1] Setup like this: https://github.com/nlothian/m1_huggingface_diffusers_demo

[2] https://tlkh.dev/benchmarking-the-apple-m1-max#heading-resne...

Python's venv module is officially recommended for creating virtual environments since Python 3.5 comes packaged with your Python installation. While there still are additional older tools available, such as conda and virtualenv, if you are new to virtual environments, it is best to use venv now.

conda's dependency solver solves a harder problem than pip's. This quote alludes to it "Conda will never be as fast as pip, so long as we're doing real environment solves and pip satisfies itself only for the current operation." (from https://github.com/conda/conda/issues/7239). Thus mamba was created to improve performance and now conda is bringing in that performance boost.

The other challenge I ran into sometimes was that if I was running a lower memory/storage Codespace instance, when I tried to use Conda from the command line to modify environments, the process would be killed after a few seconds. This turns out to be related to some performance issues Conda has that make it consume a lot of memory when trying to work with the conda-forge installation channel. You can always then just increase the size of the Codespace your are working with (just go to your Codespaces list and use the triple dots to change the settings for a Codespace).

It's worth noting that [ana]conda isn't even fully compatible yet with 3.11 (you can use it to create 3.11 environments--and you really should rather than waiting on relying on the system python--but conda itself can only run on 3.10.

It took some time but as great Christmas present Miniconda was finally released with Python 3.10!

The September update broke it.

According to this this issue is high on their priority list (whatever that means).