[REQUEST]: Support for Blackwell GPU (50xx)
In a few words, describe your idea
Please make a version with support for pytorch 12.8
More information
Recently got a 50xx series card and we have no support for it. I'm too much of a skiddy to update the install myself. Help would be appreciated!
That's weird. Could you specify the specs of your computer, os, and RVC version?
Ah, sure. I have a 7800x3d, 5070ti, and the newest version of drivers etc. I'm running on win11, and I'm also having this issue with the latest version of the CUDA windows release. I'd venv and fix it, but again, very low level of knowledge when it comes to python, haha.
of course, the solution is to just pip install torch 128 nightly into the location, but a prebuilt installer would be better for those like myself to one click go.
Out of curiosity, I guess this issue also affects v1? v1 for many, if not all, performs better in terms of quality. Except for a few users, everyone always ends up going back to v1 for that very reason.
Yup, and the problem is that we don't have a version of torch that's an official release, and other AI related apps had to have theirs built beforehand. I expect this to be eventually resolved in like, two months.
would that manual fix for v2 work for v1 too? I feel it may be quite important for many to be able to fix v1 while it outperforms v2
Not yet. I've officially tried everything including building my own version of torchaudio for 12.8, but as of 2/25, there's still no version of torchaudio specifically that will work with cuda 12.8. someone is gonna have to build it themself, and unfortunately I am not nearly as skilled enough to do so,
Hi Folks, I'm running into the same issue here, I need to upgrade to the nightly to use sm_120 , wondering
has anyone successfully got blackwell working with this?
If so, kindly list step to do so?
thanks!
of course, the solution is to just pip install torch 128 nightly into the location, but a prebuilt installer would be better for those like myself to one click go.
pip install --pre --upgrade --no-user torch torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --target "C:\wherever_you_extracted_it\vcclient_win_cuda_2.0.76-beta\dist\main\_internal"
thanks, what about for the MMVCServerSIO ?
that command yields Errno 1101) getaddrinfor failed, could not find a version that satisfies the requirement torch, no matching distribution found for torch