diff --git a/generate.py b/generate.py
index 8446d11..ffc3954 100644
--- a/generate.py
+++ b/generate.py
@@ -169,9 +169,8 @@ def generate(
             draft_model.setup_caches(max_batch_size=1, max_seq_length=max_seq_length)
 
     # create an empty tensor of the expected final shape and fill in the current tokens
-    empty = torch.empty(T_new, dtype=dtype, device=device)
-    empty[:T] = prompt
-    seq = empty
+    seq = torch.empty(T_new, dtype=dtype, device=device)
+    seq[:T] = prompt
     input_pos = torch.arange(0, T, device=device)
 
     next_token = prefill(model, prompt.view(1, -1), input_pos, **sampling_kwargs)