From d2f70da116257d02f34d12bf478c974ea9179dac Mon Sep 17 00:00:00 2001
From: Uranus <109661872+UranusSeven@users.noreply.github.com>
Date: Fri, 5 Jun 2026 15:43:04 +0800
Subject: [PATCH] fix: pad dummy run query_start_loc (#44603)

Signed-off-by: UranusSeven <109661872+UranusSeven@users.noreply.github.com>
---
 vllm/v1/worker/gpu_model_runner.py | 3 +++
 1 file changed, 3 insertions(+)

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