diff --git a/vllm_omni/deploy/qwen3_tts.yaml b/vllm_omni/deploy/qwen3_tts.yaml
index 55bac5833..69085fc5a 100644
--- a/vllm_omni/deploy/qwen3_tts.yaml
+++ b/vllm_omni/deploy/qwen3_tts.yaml
@@ -10,8 +10,7 @@
 #     codec prefill length (Q * num_frames) exceeds the 32k default.
 #   * Stage 0 does not set enforce_eager — talker runs cudagraph by default.
 #     Stage 1 defaults to false so Code2Wav can use its inner cudagraph;
-#     set true to disable that path in eager mode. NPU platform flips stage 0 to true where
-#     cudagraph is not yet verified.
+#     set true to disable that path in eager mode.
 async_chunk: true
 
 connectors:
@@ -93,4 +92,4 @@ platforms:
   npu:
     stages:
       - stage_id: 0
-        enforce_eager: true
+        enforce_eager: false
diff --git a/vllm_omni/platforms/npu/worker/npu_model_runner.py b/vllm_omni/platforms/npu/worker/npu_model_runner.py
index 5a50f523d..3b9b4937a 100644
--- a/vllm_omni/platforms/npu/worker/npu_model_runner.py
+++ b/vllm_omni/platforms/npu/worker/npu_model_runner.py
@@ -219,7 +219,10 @@ class OmniNPUModelRunner(OmniGPUModelRunner, NPUModelRunner):
         ):
             # Make sure padding doesn't exceed max_num_tokens
             assert num_tokens_padded <= self.max_num_tokens
-            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder or self.enable_prompt_embeds:
+            if getattr(getattr(self, "model", None), "has_preprocess", False):
+                input_ids = self.input_ids.gpu[:num_tokens_padded]
+                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
+            elif self.supports_mm_inputs and not self.model_config.is_encoder_decoder or self.enable_prompt_embeds:
                 input_ids = None
                 inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
             else: