From a8693df504404ea0288e480165143d8a216f4220 Mon Sep 17 00:00:00 2001
From: Kunshang Ji <kunshang.ji@intel.com>
Date: Fri, 4 Sep 2026 00:48:45 -0700
Subject: [PATCH] [SpecDecode]Fix spec decode warmup device selection (#55245)

Signed-off-by: Kunshang Ji <jikunshang95@gmail.com>
Co-authored-by: GitHub Copilot <copilot@github.com>
---
 vllm/model_executor/warmup/spec_decode_rejection_warmup.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/vllm/model_executor/warmup/spec_decode_rejection_warmup.py b/vllm/model_executor/warmup/spec_decode_rejection_warmup.py
index 78e12b6ff8..8cedbfc4db 100644
--- a/vllm/model_executor/warmup/spec_decode_rejection_warmup.py
+++ b/vllm/model_executor/warmup/spec_decode_rejection_warmup.py
@@ -44,7 +44,7 @@ def spec_decode_rejection_warmup(worker: Worker) -> None:
     use_block_verification = rejection_method == "block"
     use_synthetic = rejection_method == "synthetic"
 
-    device = torch.device("cuda")
+    device = worker.device
     num_reqs = 1
     tokens_per_req = num_spec + 1
     num_logits = num_reqs * tokens_per_req