Is it real to run this by 8Gb VRAM ?

[LuchunPen]

I'm trying to run by1070 8Gb, but have only this:

CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 8.00 GiB total capacity; 7.33 GiB already allocated; 0 bytes free; 7.36 GiB reserved in total by PyTorch)

Could you explayn is it real to run and what I have to do ?

[Acephalia]

You can run it. Go to generation.py in the Bark folder and set SUNO_USE_SMALL_MODELS to True. This should let you download smaller models and have no issue with generation with this turned on.

[dtheb]

@Acephalia Any idea how does it affect the results?

[Acephalia]

In my opinion it is currently only usable for a little bit of fun and experimentation.

I don’t know if there is any quality drop between the full model and smaller models but the samples provided are cherry picked and you are not going to get that kind of output without a lot of work and generations. Haven’t seen anyone else say otherwise yet either.

The prompting emotions etc are very finicky and 1/50 times don’t work as intended. I haven’t been able to get a completely consistent voice in a single generation either.

But remember it’s still pretty amazing and barely in its infancy, and if we look at how fast things like stable diffusion etc have developed it shouldn’t be too long before Bark evolves into something pretty substantial.

I found this fork it is the best GUI option currently imho. Easy to setup small models, model folder can be selected and overall nice and fast. So if you are still having trouble with this one, I’d give the C0untFloyd one a go.

[gkucsko]

gonna close since a bit off topic for the issue