What about GPU accelerated images?

[benz0li]

i.e. CUDA-based images

[benz0li]

The MAX Engine does not support NVIDIA GPUs yet: https://docs.modular.com/max/faq#what-hardware-is-currently-supported-by-max-engine

GPU accelerated 'MAX Engine'-based images, i.e.

• glcr.b-data.ch/jupyterlab/cuda/max/base
• glcr.b-data.ch/jupyterlab/cuda/max/scipy

will be published as soon as the MAX Engine supports Python 3.12 and NVIDIA GPUs.
ℹ️ Supposedly this summer: https://twitter.com/Modular/status/1785131461345157140

Cross references:

• What about 'MAX Engine'-based images? #1
• https://github.com/modularml/max/issues/149

[benz0li]

Regarding the new MAX & Mojo Community License Terms, see #1 (comment).

[benz0li]

Here they are:

• glcr.b-data.ch/jupyterlab/cuda/max/base
• glcr.b-data.ch/jupyterlab/cuda/max/scipy

[benz0li]

For models with an optimised [GPU accelerated] MAX Graph implementation, one could also use the regular images:

• glcr.b-data.ch/jupyterlab/max/base
• glcr.b-data.ch/jupyterlab/max/scipy

and run it with option --gpus '"device=all"'. PyTorch is still required, but the CPU-only version is sufficient.

Cross reference:

• https://github.com/modularml/max/issues/273