diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 261995f4b01..8c726be337e 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -5787,6 +5787,9 @@ class GPUModelRunner(
                     num_scheduled_tokens, self.query_pos.np
                 )
                 self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
+                self.query_start_loc.np[num_reqs + 1 : num_reqs_padded + 1].fill(
+                    cum_num_tokens[-1]
+                )
                 self.query_start_loc.copy_to_gpu()
 
                 # Sync block table CPU->GPU so cleared rows from