Laurent Mazare

An initial version of metavoice is now available, #1717 , you can give this a shot with [this example](https://github.com/huggingface/candle/tree/main/candle-examples/examples/metavoice). Please let us know how it goes, note that speaker embeddings...

Yeah it takes a bit of time to get the generation back, maybe we should have some progress bar or some other way to know that the process is not...

Interesting, the memory usage of candle during backprop is certainly not optimal. Could you maybe measure the memory footprint after doing the forward pass but before doing the backward pass?...

Also could you give it a try with #1243 ? This should only improve memory consumption on the backprop part avoiding that all the intermediary values get stored until the...

> When I try with [#1243](https://github.com/huggingface/candle/pull/1243), I can successfully set the batch size back to 16. Thank you very much. That seems to have solved my problem. But I would...

I've merged #1243 as it seems to fix things for you. @laptou would be great to have your experience with this too! I wouldn't expect the new behavior to be...

Thanks for trying this out, that's very interesting. Do you have a sense of why the pytorch model wouldn't have to retain the intermediary values? It could be that we...

Thanks, hard to tell by just looking at the code as the model seems pretty involved. A way to optimize this a bit would probably be to use `heaptrack` as...

The idea is that non-resolved tokens are actually accumulated, the decoder (`TokenOutputStream`) is stateful as decoding some tokens can only be done when knowing the following tokens so it's expected...

"weight is invalid in distribution" usually means that the model generated a nan, so somehow one of the metal kernel probably did not work in the same way on iphone...