Request: Create 8 GB VRAM sized model, or models for common GPU VRAM sizes, to optimize model sizes to GPU's

[Hukkinen]

Hey,

Thank you for the Whisper.

For what it's worth, I have a 8GB VRAM GPU, and it's suboptimally in between of large (10 GB) and medium (5GB model).

Also, I have 4 GB, too, so while you're at it..

In general, consider optimizing model sizes for common VRAM sizes, so people can have the best bang for their hardware.

Also, in terms of compute and energy investment (carbon footprint & climate change mitigation), it is justified to aim to get the optimal quality per compute out of a particular hardware.

Thanks!

Or what do you people think?

[phineas-pta]

Create xx GB VRAM sized model

u think creating neural networks is like just add/remove random things until they attain a certain size ? NO

consider optimizing model sizes for common VRAM sizes

it's not the authors job, there're already a lot of forks that's optimized for low-end/mid-end gpu, even no gpu

so people can have the best bang for their hardware

it's already free of charge, what bang u want more ?