Support for RTX 50XX (sm120)

[CLG68]

Hi,

are you planning to update Chai-1 to support sm120 GPU (RTX 50XX) so the latest Pytorch and Cuda (12.8)?

Thx

[arogozhnikov]

chai is compatible with multiple torch versions. I don't expect problems with fresh CUDA and 5090.

[CLG68]

Hi,

sm120 (Blackwell GPU) are supported by the nightly build of Pytorch and Cuda 12.8

I'm a bit new at this so could you validate how it needs to be done, please?

1. download Chai-1's archive from GitHub
2. Create/activate an environment chai1 with python 3.12
3. pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
   (do I need torchvision & torchaudio?)
4. edit: requirements.in and remove or modify:
   torch>=2.3.1,<2.7
5. pip install -e .

I upgraded my GPU to an sm120 and suddenly nothing works... the sm90 cards with at least 16gb are now either impossible to find or infinitely expensive.

[arogozhnikov]

1. pip install from github (see command in readme)
2. pip install torch==... (whatever version you use, even if it isn't compatible). chai-lab + torch 2.7 works fine, if you still observe any problems - that's likely your particular setup is broken