So, I have an old GPU - GTX780. If I correctly understand, last version CUDA Toolkit for it is 11.4.4 and version of driver is 474.30, CUDA_ARCH_LIST=“3.5”
Is it possible to build PyTorch v.2.0.0 from source or maybe other last versions?
I didn’t find MAGMA for CUDA 11.4 for building and looks like it’s not exists.
I will to try build from source on Windows 10.
It is definitely possible to build from source:
If you are not using MAGMA it is also possible to remove it from the build e.g., via the
USE_MAGMA=0 environment variable when running the build command.
Alternatively, if you are OK with using a somewhat older version to avoid building from source, it looks like
sm_35 should be supported by default in the published binaries to at least 1.8.1+, according to: GPU compute capability support for each pytorch version - #7 by moi90