diff --git a/comfy/model_management.py b/comfy/model_management.py
index 731fb584..15800a8e 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -1121,9 +1121,8 @@ def soft_empty_cache(force=False):
     elif is_ascend_npu():
         torch.npu.empty_cache()
     elif torch.cuda.is_available():
-        if force or is_nvidia(): #This seems to make things worse on ROCm so I only do it for cuda
-            torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
 
 def unload_all_models():
     free_memory(1e30, get_torch_device())