I am unable to setup the environment as explained in the README.md as it installs a version of PyTorch that is incompatible with my CUDA.
NVIDIA GeForce RTX 5080 with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_70 sm_75 sm_80 sm_86.
However, I was able to solve with two minor changes:
patch_t=128 instead of 160, otherwise, the positional embeddings have shape (1, 10, 128) instead of (1, 8, 128) as in the checkpoint.- in
SRDTrans/MainFrame.py, SqueezeLayer::forward, I changed before_down = self.conv_net(x) to before_down = self.conv_net(x.float())
Now the inference works fine on Python 3.12.3, torch (2.9.0+cu128).
I am unable to setup the environment as explained in the
README.mdas it installs a version of PyTorch that is incompatible with my CUDA.However, I was able to solve with two minor changes:
patch_t=128instead of160, otherwise, the positional embeddings have shape (1, 10, 128) instead of (1, 8, 128) as in the checkpoint.SRDTrans/MainFrame.py,SqueezeLayer::forward, I changedbefore_down = self.conv_net(x)tobefore_down = self.conv_net(x.float())Now the inference works fine on Python 3.12.3, torch (2.9.0+cu128).