diff --git a/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py b/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py
index ca9c8ca41..87fadae8c 100644
--- a/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py
+++ b/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py
@@ -1122,7 +1122,7 @@ class Qwen3TTSTalkerForConditionalGeneration(nn.Module):
         )
         # Prefer GPU for encoder if available; otherwise keep CPU.
         dev = next(self.parameters()).device
-        if getattr(dev, "type", None) == "cuda":
+        if dev.type != "cpu":
             try:
                 tok.model.to(dev)
                 tok.device = dev