From 87873c7e912c5db5819b3995023c26896dc8d413 Mon Sep 17 00:00:00 2001
From: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
Date: Wed, 26 Nov 2025 18:07:33 -0800
Subject: [PATCH] [https://nvbugs/5698581][fix] Init draft tokens for CUDA
 graph dummy request

Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
---
 tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py b/tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
index d924f2ea457c..b8e2754a9cb3 100644
--- a/tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
+++ b/tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
@@ -389,6 +389,7 @@ def _get_padded_batch(self, batch: ScheduledRequests,
             if spec_res_mgr:
                 spec_res_mgr.add_dummy_requests([CUDA_GRAPH_DUMMY_REQUEST_ID])
 
+        self.padding_dummy_request.py_draft_tokens = [0] * runtime_draft_len
         batch.generation_requests.extend([self.padding_dummy_request] *
                                          padding_size)
         return padding_size