Hello,
I've encountered some strange behavior in rope.py when using the perception models with Torch 2.9.1+cu130, CUDA 13.0, and Python 3.11.14 in a container.
I made a small change to line 340 to check that the "freq" attribute is correctly defined:
if self.freq is not None: self.freq = self.freq.to(device)
I don't know if this is optimal, but it works for me.
Hello,
I've encountered some strange behavior in rope.py when using the perception models with Torch 2.9.1+cu130, CUDA 13.0, and Python 3.11.14 in a container.
I made a small change to line 340 to check that the "freq" attribute is correctly defined:
if self.freq is not None: self.freq = self.freq.to(device)
I don't know if this is optimal, but it works for me.