Sometimes I have to compile packages on a non-CUDA machine and later use that compilation on a CUDA machine.
This PR unconditionally builds the CUDA extension when FORCE_CUDA=1 is set.
Also I tried to give helpful warnings when users stumble upon the same error; also, we error out early if the user does not have a GPU but tries using fused Adan.
Finally, I took the opportunity to also include an increase in the version string here. That way, when people pip install -U ...Adan.git, they get fused Adan without having to force-reinstall the package (otherwise pip does not see that there's a new feature available).
Sometimes I have to compile packages on a non-CUDA machine and later use that compilation on a CUDA machine. This PR unconditionally builds the CUDA extension when
FORCE_CUDA=1is set.Also I tried to give helpful warnings when users stumble upon the same error; also, we error out early if the user does not have a GPU but tries using fused Adan.
Finally, I took the opportunity to also include an increase in the version string here. That way, when people
pip install -U ...Adan.git, they get fused Adan without having to force-reinstall the package (otherwisepipdoes not see that there's a new feature available).