Only clear cuda cache on CUDA since it causes slowdowns on ROCm.

execution.py

@@ -189,6 +189,8 @@ class PromptExecutor:
                     self.server.send_sync("executing", { "node": None }, self.server.client_id)
 
         gc.collect()
+        if torch.cuda.is_available():
+            if torch.version.cuda: #This seems to make things worse on ROCm so I only do it for cuda
                 torch.cuda.empty_cache()
                 torch.cuda.ipc_collect()