Fix issue with regular torch version.

comfy/model_management.py

@@ -390,7 +390,11 @@ def unet_inital_load_device(parameters, dtype):
         return torch_dev
 
     cpu_dev = torch.device("cpu")
-    model_size = dtype.itemsize * parameters
+    dtype_size = 4
+    if dtype == torch.float16 or dtype == torch.bfloat16:
+        dtype_size = 2
+
+    model_size = dtype_size * parameters
 
     mem_dev = get_free_memory(torch_dev)
     mem_cpu = get_free_memory(cpu_dev)