Update app.py

app.py

@@ -1,22 +1,24 @@
 import spaces, os
 import gradio as gr
-from kokoro import KPipeline
-from qwen_vl_utils import process_vision_info
 
+from kokoro import KPipeline
 from demo.infer import LiveCCDemoInfer
 
-model_path = 'chenjoya/LiveCC-7B-Instruct'
-
-def _init_infer():
-    # create a singleton LiveCCDemoInfer inside GPU
-    import torch
-    from kokoro import KPipeline
-    from demo.infer import LiveCCDemoInfer
-    infer = LiveCCDemoInfer(model_path, device='cuda')
-    return infer
+class GradioBackend:
+    waiting_video_response = 'Waiting for video input...'
+    not_found_video_response = 'Video does not exist...'
+    mode2api = {
+        'Real-Time Commentary': 'live_cc',
+        'Conversation': 'video_qa'
+    }
+    def __init__(self, model_path: str = 'chenjoya/LiveCC-7B-Instruct'):
+        self.infer = LiveCCDemoInfer(model_path)
+        self.audio_pipeline = KPipeline(lang_code='a')
+    
+    def __call__(self, query: str = None, state: dict = {}, mode: str = 'Real-Time Commentary', **kwargs):
+        return getattr(self.infer, self.mode2api[mode])(query=query, state=state, **kwargs)
 
-# We'll keep a module-global placeholder
-infer = None
+gradio_backend = None
             
 with gr.Blocks() as demo:
     gr.Markdown("## LiveCC Conversation and Real-Time Commentary - Gradio Demo")
@@ -57,15 +59,15 @@ with gr.Blocks() as demo:
 
             @spaces.GPU
             def gr_chatinterface_fn(message, history, state, video_path, mode):
-                global infer
+                global gradio_backend
                 yield '(initializing model, thanks for waiting...)', state
-                if infer is None:
-                    infer = _init_infer()
+                if gradio_backend is None:
+                    gradio_backend = GradioBackend()
                 state['video_path'] = video_path
                 yield '(finished initialization, responding...)', state
                 if mode != 'Conversation':
                     yield 'waiting video input...', state
-                response, _ = infer.video_qa(query=message, state=state)
+                response, _ = gradio_backend.video_qa(query=message, state=state)
                 yield response, {}                
                 
             def gr_chatinterface_chatbot_clear_fn():