diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index e700c09038e2..177542ed96c8 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -524,7 +524,7 @@ def __init__(
         self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
         self.transfer_event = torch.cuda.Event()
         self.sampled_token_ids_pinned_cpu = torch.empty(
-            (self.max_model_len, 1),
+            (self.max_num_reqs, 1),
             dtype=torch.int64,
             device="cpu",
             pin_memory=self.pin_memory,