Kacper Michajłow

Ok, fair, it can wait. I was using pytorch wheels from AMD https://repo.radeon.com/rocm/manylinux/rocm-rel-6.4.1/ to validate those changes.

> pytorch rocm 6.4 is still nightly so I can't merge some of these changes as is or it's going to break people using the stable pytorch which is on...

> > > pytorch rocm 6.4 is still nightly so I can't merge some of these changes as is or it's going to break people using the stable pytorch which...

Well, yes. It still uses lots of memory, but it's less than before and previously fallback to tailed vae would take ages because of conversions, this seems to work more...

Rebased. Pytorch 2.8.0 with support for ROCm 6.4 has been released too.

Rebased after recent changes, some changes were already upstreamed.

This is probably effect of bt.1886 implementation in libplacebo ``` c const float lb = powf(csp_min, 1/2.4f); const float lw = powf(csp_max, 1/2.4f); const float a = powf(lw - lb,...

> there's a behavioral quirk with libplacebo where this option controls all colorspace linearization, including the linearization that occurs before sigmoidal resampling. I don't remember if this was ever fixed...

> When no hint is provided When there is no hint, there is no hint. mpv/libplacebo output is expected to be configured to matching (already configured) display colorspace in this...

Can you provide more information, log? There were changes recently in this area.