diff --git a/nodes.py b/nodes.py
index 609716e..7957bd5 100644
--- a/nodes.py
+++ b/nodes.py
@@ -145,7 +145,7 @@ class GGUFModelPatcher(comfy.model_patcher.ModelPatcher):
             if device_to is not None:
                 out_weight = weight.to(device_to, copy=True)
             else:
-                out_weight = weight.clone()
+                out_weight = weight.to(weight.device, copy=True) # make sure patches are copied
 
             if self.patch_on_device:
                 patches = move_patch_to_cuda(patches, self.load_device)
@@ -165,6 +165,10 @@ class GGUFModelPatcher(comfy.model_patcher.ModelPatcher):
         else:
             comfy.utils.set_attr_param(self.model, key, out_weight)
 
+    def load(self, *args, force_patch_weights=False, **kwargs):
+        # always call `patch_weight_to_device` even for lowvram
+        return super().load(*args, force_patch_weights=True, **kwargs)
+
     def clone(self, *args, **kwargs):
         n = GGUFModelPatcher(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
diff --git a/tools/convert.py b/tools/convert.py
index a49ef1c..63bded9 100644
--- a/tools/convert.py
+++ b/tools/convert.py
@@ -109,6 +109,9 @@ def handle_tensors(args, writer, state_dict):
 
         if data.dtype == torch.bfloat16:
             data = data.to(torch.float32).numpy()
+        # this is so we don't break torch 2.0.X
+        elif data.dtype in [getattr(torch, "float8_e4m3fn", "_invalid"), getattr(torch, "float8_e5m2", "_invalid")]:
+            data = data.to(torch.float16).numpy()
         else:
             data = data.numpy()