diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index e350988456f1..7324d41310ba 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -2321,11 +2321,19 @@ def _bookkeeping_sync(
                 sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
             else:
                 sampled_ids = valid_sampled_token_ids[req_idx]
+
+            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
+
+            if cu_num_accepted_tokens is not None:
+                cu_num_accepted_tokens.append(
+                    cu_num_accepted_tokens[-1] + num_sampled_ids
+                )
+
             if not sampled_ids:
                 continue
 
             start_idx = self.input_batch.num_tokens_no_spec[req_idx]
-            end_idx = start_idx + len(sampled_ids)
+            end_idx = start_idx + num_sampled_ids
             assert end_idx <= self.max_model_len, (
                 "Sampled token IDs exceed the max model length. "
                 f"Total number of tokens: {end_idx} > max_model_len: "
@@ -2341,11 +2349,6 @@ def _bookkeeping_sync(
             req_state = self.requests[req_id]
             req_state.output_token_ids.extend(sampled_ids)
 
-            if cu_num_accepted_tokens is not None:
-                cu_num_accepted_tokens.append(
-                    cu_num_accepted_tokens[-1] + len(sampled_ids)
-                )
-
         logprobs_lists = (
             logprobs_tensors.tolists(cu_num_accepted_tokens)
             if not self.use_async_scheduling and logprobs_tensors is not None