Arlo Phoenix

> Please @arlo-phoenix or @fxmarty create an RFC issue of the same format as the Apple one to centralize discussions, as well as decisions and tracking of open work. I'll...

> https://github.com/TimDettmers/bitsandbytes/blob/a13af4542f5a163d8e9eacb0b8b2a5b26d9e2b15/Makefile#L19-L23 > > You have logic bug there, if `ROCM_HOME` is empty, it will skip `ROCM_TARGET` check and will try to compile anyway (will fail later due to missing...

> FWIW, Arch Linux's ROCM doesn't seem to be distributing hipblaslt (yet), so I am also getting missing hipblaslt.h errors on 5.6 there. It doesn't seem to be a straightforward...

> Hi, is there any update on this? Would love to have this merged! Since there seems to now be an official plan to extend support to multiple platforms /...

> @arlo-phoenix rocm 6.0 docker has arrived https://hub.docker.com/r/rocm/pytorch/tags Thanks for the info! Only tested the basic stuff and will probably only further test after the holidays, but it still compiles,...

@purefire > Hello, any news about 6.0 update? Status quo is still > Thanks for the info! Only tested the basic stuff and will probably only further test after the...

> @arlo-phoenix if you need a wave64 device to test with, i have a mi25 im not using sitting on a shelf, i could give it to you for free....

> Thank you so much for this contribution. I am sorry that it took so long to reply and look at this. We are currently working on integrating different devices...

> Note that for hipblaslt you need at least gfx90a not gfx908, i do have access to mi100s but that wont help here since thats gfx908. The matrix instructions in...

@IMbackK my latest commit tries to adjust the kQuantizeBlockwise function to work with wavefront 64 (just only use 1 load, etc. per thread instead of 2). Didn't really test much,...