fix app.py

app.py

@@ -108,6 +108,7 @@ def load_backend(model_path: str, device_str: str, attn_implementation: str):
     )
     if hasattr(processor, "audio_tokenizer"):
         processor.audio_tokenizer = processor.audio_tokenizer.to(device)
+        processor.audio_tokenizer.eval()
 
     model_kwargs = {
         "trust_remote_code": True,
@@ -558,21 +559,7 @@ def build_demo(args: argparse.Namespace):
         )
 
         run_btn.click(
-            fn=lambda text, reference_audio, mode_with_reference, duration_control_enabled, duration_tokens, temperature, top_p, top_k, repetition_penalty, max_new_tokens: run_inference(
-                text=text,
-                reference_audio=reference_audio,
-                mode_with_reference=mode_with_reference,
-                duration_control_enabled=duration_control_enabled,
-                duration_tokens=duration_tokens,
-                temperature=temperature,
-                top_p=top_p,
-                top_k=top_k,
-                repetition_penalty=repetition_penalty,
-                model_path=args.model_path,
-                device=args.device,
-                attn_implementation=args.attn_implementation,
-                max_new_tokens=max_new_tokens,
-            ),
+            fn=run_inference,
             inputs=[
                 text,
                 reference_audio,
@@ -583,6 +570,9 @@ def build_demo(args: argparse.Namespace):
                 top_p,
                 top_k,
                 repetition_penalty,
+                gr.State(args.model_path),
+                gr.State(args.device),
+                gr.State(args.attn_implementation),
                 max_new_tokens,
             ],
             outputs=[output_audio, status],
@@ -617,19 +607,6 @@ def parse_port(value: str | None, default: int) -> int:
         return default
 
 
-def build_default_args() -> argparse.Namespace:
-    return resolve_runtime_attn(
-        argparse.Namespace(
-            model_path=MODEL_PATH,
-            device="cuda:0",
-            attn_implementation=DEFAULT_ATTN_IMPLEMENTATION,
-            host=os.getenv("GRADIO_SERVER_NAME", "0.0.0.0"),
-            port=parse_port(os.getenv("GRADIO_SERVER_PORT", os.getenv("PORT")), 7860),
-            share=False,
-        )
-    )
-
-
 def main():
     parser = argparse.ArgumentParser(description="MossTTS Gradio Demo")
     parser.add_argument("--model_path", type=str, default=MODEL_PATH)
@@ -680,12 +657,5 @@ def main():
     )
 
 
-# Expose a module-level demo for Gradio hot-reload/Spaces launcher.
-demo = build_demo(build_default_args())
-
-
 if __name__ == "__main__":
-    if os.getenv("GRADIO_HOT_RELOAD"):
-        print("[Startup] GRADIO_HOT_RELOAD detected. Skipping explicit launch().", flush=True)
-    else:
-        main()
+    main()