mlx-audio icon indicating copy to clipboard operation
mlx-audio copied to clipboard

Orpheus in Swift example - Layer shapes

Open BenLumenDigital opened this issue 10 months ago • 2 comments

I'm trying to get Orpheus working in Swift here: https://github.com/Blaizzy/mlx-audio/pull/84

But I can't understand some layer shape stuff.

There is an embedding layer with hidden_size of 3072.

But when I run the model, I see a shape of 384.

And I can't see any way that this should be projected... what am I missing here!?

BenLumenDigital avatar Apr 14 '25 10:04 BenLumenDigital

It's difficult to debug this without seeing the code -- do you have a branch somewhere with it?

lucasnewman avatar Apr 21 '25 15:04 lucasnewman

There is a PR here from @BenLumenDigital

Blaizzy avatar Apr 21 '25 15:04 Blaizzy