pip install fails on Mac with Apple M1 chips

[dimiboeckaerts]

Hey ColabFold team! Just wanted to let you know that I haven't been able to locally install ColabFold on my Apple MacBook Air with M1 chip due to incompatibilities. The error I get is the following:

ERROR: Could not find a version that satisfies the requirement tensorflow-cpu<3.0.0,>=2.6.0 (from colabfold[alphafold]) (from versions: none) ERROR: No matching distribution found for tensorflow-cpu<3.0.0,>=2.6.0

I do however have TensorFlow installed, specifically for M1 chips (cfr https://caffeinedev.medium.com/how-to-install-tensorflow-on-m1-mac-8e9b91d93706). Is there a way around this?

Thanks!

[YoshitakaMo]

Hi @dimiboeckaerts, I'm now developing an installer for Macbook with M1 chip. ColabFold worked on my M1 Mac finally. I will test the installer again and release the installer if there are no issues. But it's 10 times slower than Linux+NVIDIA GPU.

[dimiboeckaerts]

That's great! I have checked out the localcolabfold repo but saw that indeed Mac M1 was currently unsupported. The question I still have though is why it is not possible to run ColabFold with the GPU enabled version of TensorFlow on Mac? As I mentioned, I have TensorFlow set up for my Mac M1, complete with GPU support and Apple Metal integration. Should this be possible, in theory?

[YoshitakaMo]

I updated my repo. Please use the installer for Mac M1 and report if some errors occurred!

The question I still have though is why it is not possible to run ColabFold with the GPU enabled version of TensorFlow on Mac?

I know tenforflow-metal for Mac M1, but I think it has no effect for ColabFold (and AlphaFold) because the structure prediction largely depends on JAX and jaxlib, not tensorflow. I also tried tensorflow-gpu instead of tensorflow-cpu to check the calculation time on my Linux, but the time didn't decrease much.

[dimiboeckaerts]

OK thank you Yoshitaka!