try to figure out how ZeroGPU works

Architectures/ControllabilityGAN/wgan/wgan_qc.py

@@ -245,7 +245,10 @@ class WassersteinGanQuadraticCost(torch.nn.Module):
         latent_samples = latent_samples.to(self.device)
         if nograd:
             with torch.no_grad():
-                generated_data = self.G(latent_samples.to("cpu"), return_intermediate=return_intermediate)
+                if isinstance(self.G, torch.nn.parallel.DataParallel):
+                    generated_data = self.G.module(latent_samples.to("cpu"), return_intermediate=return_intermediate)
+                else:
+                    generated_data = self.G(latent_samples.to("cpu"), return_intermediate=return_intermediate)
         else:
             generated_data = self.G(latent_samples)
         self.G.train()