Birch-san

hmm it sounds like the "custom kernel" solutions will add complexity and require a fair bit of engineering. without even being clear whether it's worth it. I dunno how expensive...

Yup, I'm okay with releasing my contribution under a permissive license. There are some bits of @AminRezaei0x443's code remaining, so good to find out their licensing desires. But I believe...

Oh, you might want to read the latest commits I added: https://github.com/Birch-san/diffusers/pull/1 I reduced the number of times the key needed to be transposed. This complicates the API slightly. Has...

> could only get worse performance okay, that's interesting. I'm skeptical. I measured the impact of pre-transpose on _unchunked_ attention a while back, and it didn't seem to change the...

okay yeah, got 12% faster when I updated to latest master. might be that `torch.linear()` optimization. https://twitter.com/Birchlabs/status/1611507344302608388

okay, you're right: kulinseth's branch is faster than master. I measured it as 10% faster (on float16). https://twitter.com/Birchlabs/status/1611525288642613248 also, the [PR](https://github.com/pytorch/pytorch/pull/89436) to which the "memory usage" [issue](https://github.com/pytorch/pytorch/issues/90464) refers… has merged...

From the [upstream](https://github.com/vaklinov/zcash-swing-wallet-ui) readme: > Limitation: The list of transactions does not show all outgoing ones (specifically outgoing Z address transactions). A corresponding issue https://github.com/zcash/zcash/issues/1438 has been opened for the...

Aye, I can believe that. Okay, probably I shouldn't make plans to play via Voobly on launch day. :)

Perhaps this is a big ask, but could you please compile this thing, sign it in Glass Echidna's name, and submit a release somewhere? (I suppose you could fork it...

ah, because debugDraw doesn't give you any control over ordering? you don't need to use debugDraw. you can actually just retain in a variable a reference to every body or...