reminder: an op changes the devices of p_gains and d_gains to default cuda:0: self.p_gains, self.d_gains = to_torch(self.p_gains), to_torch(self.d_gains)

[VictoriYang]

A kind reminder: there is an op in phc/env/tasks/humanoid.py line 1179: self.p_gains, self.d_gains = to_torch(self.p_gains), to_torch(self.d_gains) this op changes the device to "cuda:0", which could cause a bug when "cuda:0" is not the target device.

Here is a solution below: self.p_gains, self.d_gains = to_torch(self.p_gains, device=self.device), to_torch(self.d_gains, device=self.device)