Standard-Intelligence / hertz-dev

first base model for full-duplex conversational audio
https://si.inc
Apache License 2.0
1.6k stars 102 forks source link

Support for macOS & Apple M1-M4 #5

Open ChristianWeyer opened 3 weeks ago

ChristianWeyer commented 3 weeks ago

Is this planned? Thanks!

calculating commented 2 weeks ago

planned, and we'll also give a $500 bounty if someone makes a PR we can merge.

chigkim commented 1 week ago

Pytorch nightly build now has the fix for convolutions with output_channels > 65536, and I got both inference_server and inference_client to work together on MPS after changing bfloat16 to float16. However, it seems too slow for my M3 Max 64GB to work with Torch and mps. It can only play a short burst of samples at a time, so the output audio is extremely choppy. :( Is there a way to have the model to accept longer input and generate longer output? It would increase the latency, but it might work.