Refactor run_pipeline and update_metrics methods to support inference on HF ZeroGPU

interface.py

@@ -1,5 +1,6 @@
 import time
 import uuid
+from functools import partial
 
 import gradio as gr
 import spaces
@@ -9,6 +10,34 @@ from characters import CHARACTERS
 from pipeline import SingingDialoguePipeline
 
 
+@spaces.GPU(duration=120)
+def run_pipeline(audio_path, interface):
+    if not audio_path:
+        return gr.update(value=None), gr.update(value=None)
+    tmp_file = f"audio_{int(time.time())}_{uuid.uuid4().hex[:8]}.wav"
+    results = interface.pipeline.run(
+        audio_path,
+        interface.svs_model_map[interface.current_svs_model]["lang"],
+        interface.character_info[interface.current_character].prompt,
+        interface.current_voice,
+        output_audio_path=tmp_file,
+    )
+    formatted_logs = f"ASR: {results['asr_text']}\nLLM: {results['llm_text']}"
+    return gr.update(value=formatted_logs), gr.update(
+        value=results["output_audio_path"]
+    )
+
+
+@spaces.GPU(duration=120)
+def update_metrics(audio_path, interface):
+    if not audio_path or not interface.results:
+        return gr.update(value="")
+    results = interface.pipeline.evaluate(audio_path, **interface.results)
+    results.update(interface.results.get("metrics", {}))
+    formatted_metrics = "\n".join([f"{k}: {v}" for k, v in results.items()])
+    return gr.update(value=formatted_metrics)
+
+
 class GradioInterface:
     def __init__(self, options_config: str, default_config: str):
         self.options = self.load_config(options_config)
@@ -148,12 +177,12 @@ class GradioInterface:
                     fn=self.update_voice, inputs=voice_radio, outputs=voice_radio
                 )
                 mic_input.change(
-                    fn=self.run_pipeline,
+                    fn=partial(run_pipeline, interface=self),
                     inputs=mic_input,
                     outputs=[interaction_log, audio_output],
                 )
                 metrics_button.click(
-                    fn=self.update_metrics,
+                    fn=partial(update_metrics, interface=self),
                     inputs=audio_output,
                     outputs=[metrics_output],
                 )
@@ -161,6 +190,7 @@ class GradioInterface:
             return demo
         except Exception as e:
             import traceback
+
             print(traceback.format_exc())
             return gr.Blocks()
 
@@ -212,29 +242,3 @@ class GradioInterface:
     def update_voice(self, voice):
         self.current_voice = self.svs_model_map[self.current_svs_model]["voices"][voice]
         return gr.update(value=voice)
-
-    @spaces.GPU
-    def run_pipeline(self, audio_path):
-        if not audio_path:
-            return gr.update(value=None), gr.update(value=None)
-        tmp_file = f"audio_{int(time.time())}_{uuid.uuid4().hex[:8]}.wav"
-        self.results = self.pipeline.run(
-            audio_path,
-            self.svs_model_map[self.current_svs_model]["lang"],
-            self.character_info[self.current_character].prompt,
-            self.current_voice,
-            output_audio_path=tmp_file,
-        )
-        formatted_logs = f"ASR: {self.results['asr_text']}\nLLM: {self.results['llm_text']}"
-        return gr.update(value=formatted_logs), gr.update(
-            value=self.results["output_audio_path"]
-        )
-
-    @spaces.GPU
-    def update_metrics(self, audio_path):
-        if not audio_path or not self.results:
-            return gr.update(value="")
-        results = self.pipeline.evaluate(audio_path, **self.results)
-        results.update(self.results.get("metrics", {}))
-        formatted_metrics = "\n".join([f"{k}: {v}" for k, v in results.items()])
-        return gr.update(value=formatted_metrics)