use marlin on a remote GPU machine ? third party providers ?
hi. I'm trying to use Marlin on my macbook air m4 2025, but I could not use fast GPU/MPS acceleration in this setup. It fell back to a slower Torch path, and PyTorch reported the MPS backend is unavailable for this macOS/PyTorch path. is there a way to use marlin on a remote GPU machine?
Hey! Quick heads up: MLX/MPS on Apple Silicon isn't something we support officially yet, so the slow Torch fallback you're hitting is expected on the M4.
That said, someone in the community did convert Marlin to MLX and it runs on Apple Silicon: https://huggingface.co/junwatu/Marlin-2B-MLX-8bit
One thing to know since it's not ours: use the hybrid mode from the model card (HF transformers prepares the input, MLX generates). The pure-MLX mode is a touch faster but its timestamps come out compressed, which matters a lot for Marlin since it's all about time-ranged events and grounding. Hybrid is about 28s/video and gives correct timestamps.
If you want the rock-solid reference setup, a remote NVIDIA GPU running the model from the HF page (https://huggingface.co/NemoStation/Marlin-2B) is still the ground truth. But for poking around locally on your Mac, that MLX build should get you going.
Let me know how it goes!