Ross Wightman

@stas-sl thanks, I'll try taking out the vmap in patch interpolation with your approach and replace if it matches. For the position embeddings, you are correct, this is not particularly...

pos embed is updated and tested in #2518 ... haven't had a chance to work on the pos embed yet

@stas-sl I've got an impl of both in the PR now. The patch embed resizing is a win on simplicity and performance. In practical terms the grid_sample would appear to...

@stas-sl hah, right that was a bit silly, I just focused on the fn w/ set interface but was no need to do that

Okay, fixed yesterdays quick tunnel vision impl, altered the interface and leverage patch_coords as I should have

> Looks like I also messed up yesterday during my PR review - I added comments on some lines and assumed that was enough, but I just realized they probably...

Hello, I came here by way of #110080 looking for a device type agnostic API. TLDR this is nice, but why aren't all CPU and accelerators just 'devices', more in...

@guangyey having written quite a few train/val scripts I feel it is possible to support many different types of devices in one scripts fairly cleanly. When I refer to capabilities...

I guess my main point is I feel life would be easier if we stopped thinking about it in relation to device type, but instead focus on capabilities. Still having...

usages of `from timm.models.registry import register_model` that exist in this lib should be changed to `from timm.models import register_model` it was deprecated 2 years ago