fix: revert layer offload iteration

Author: Rypo

OmniGen/transformer.py

@@ -33,13 +33,15 @@ def prefetch_layer(self, layer_idx: int, device: torch.device):
         "Starts prefetching the next layer cache"
         with torch.cuda.stream(self.prefetch_stream):
             # Prefetch next layer tensors to GPU
-            self.layers[layer_idx] = self.layers[layer_idx].to(device, non_blocking=True)
+            for name, param in self.layers[layer_idx].named_parameters():
+                param.data = param.data.to(device, non_blocking=True)
 
     def evict_previous_layer(self, layer_idx: int):
         "Moves the previous layer cache to the CPU"
         prev_layer_idx = layer_idx - 1
-        self.layers[prev_layer_idx] = self.layers[prev_layer_idx].to("cpu")
-            
+        for name, param in self.layers[prev_layer_idx].named_parameters():
+            param.data = param.data.to("cpu")
+
     def get_offload_layer(self, layer_idx: int, device: torch.device):
         # init stream
         if not hasattr(self, "prefetch_stream"):