Update app.py

app.py

@@ -1,29 +1,27 @@
 import spaces, os
 import gradio as gr
+from kokoro import KPipeline
 
 from demo.infer import LiveCCDemoInfer
 
-with gr.Blocks() as demo:
-
-    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'
-        }
+model_path = 'chenjoya/LiveCC-7B-Instruct'
 
-        @spaces.GPU
-        def __init__(self, model_path: str = 'chenjoya/LiveCC-7B-Instruct'):
-            self.infer = LiveCCDemoInfer(model_path)
-            from kokoro import KPipeline
-            self.audio_pipeline = KPipeline(lang_code='a')
+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'
+    }
 
-        @spaces.GPU
-        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)
+    def __init__(self, infer, audio_pipeline):
+        self.infer = infer
+        self.audio_pipeline = audio_pipeline
 
-    gradio_backend = GradioBackend() 
+    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)
+            
+with gr.Blocks() as demo:
 
     gr.Markdown("## LiveCC Real-Time Commentary and Conversation - Gradio Demo")
     gr.Markdown("### [LiveCC: Learning Video LLM with Streaming Speech Transcription at Scale (CVPR 2025)](https://showlab.github.io/livecc/)")
@@ -60,8 +58,12 @@ with gr.Blocks() as demo:
             with gr.Row():
                 gr_radio_mode = gr.Radio(label="Select Mode", choices=["Real-Time Commentary", "Conversation"], elem_id="gr_radio_mode", value='Real-Time Commentary', interactive=True) 
 
+            @spaces.GPU
             def gr_chatinterface_fn(message, history, state, video_path, mode):
                 state['video_path'] = video_path
+                infer = LiveCCDemoInfer(model_path)
+                audio_pipeline = KPipeline(lang_code='a')
+                gradio_backend = GradioBackend(infer, audio_pipeline)
                 response, state = gradio_backend(query=message, state=state, mode=mode)
                 return response, state
             def gr_chatinterface_chatbot_clear_fn():