Update app.py

app.py

@@ -1,3 +1,4 @@
+import spaces  # 추가
 import gradio as gr
 import os
 import asyncio
@@ -55,7 +56,12 @@ except:
 
 # MeloTTS imports (for local mode)
 try:
-    os.system("python -m unidic download")
+    # unidic 다운로드를 조건부로 처리
+    if not os.path.exists("/usr/local/lib/python3.10/site-packages/unidic"):
+        try:
+            os.system("python -m unidic download")
+        except:
+            pass
     from melo.api import TTS as MeloTTS
     MELO_AVAILABLE = True
 except:
@@ -92,6 +98,7 @@ class UnifiedAudioConverter:
         """Initialize API mode with Together API (now fallback)"""
         self.llm_client = OpenAI(api_key=api_key, base_url="https://api.together.xyz/v1")
         
+    @spaces.GPU(duration=120)
     def initialize_local_mode(self):
         """Initialize new local mode with Llama CPP"""
         if not LLAMA_CPP_AVAILABLE:
@@ -126,6 +133,7 @@ class UnifiedAudioConverter:
                 print(f"Failed to initialize local LLM: {e}")
                 raise RuntimeError(f"Failed to initialize local LLM: {e}")
 
+    @spaces.GPU(duration=60)
     def initialize_legacy_local_mode(self):
         """Initialize legacy local mode with Hugging Face model (fallback)"""
         if self.legacy_local_model is None:
@@ -168,6 +176,7 @@ class UnifiedAudioConverter:
         if not os.path.exists("cli/inference.py"):
             print("Warning: Spark-TTS CLI not found. Please clone the Spark-TTS repository.")
 
+    @spaces.GPU(duration=60)
     def initialize_melo_tts(self):
         """Initialize MeloTTS models"""        
         if MELO_AVAILABLE and self.melo_models is None:
@@ -240,6 +249,7 @@ class UnifiedAudioConverter:
             {"role": "user", "content": self._build_prompt(text, language)}
         ]
 
+    @spaces.GPU(duration=120)
     def extract_conversation_local(self, text: str, language: str = "English", progress=None) -> Dict:
         """Extract conversation using new local LLM (primary method)"""
         try:
@@ -294,6 +304,7 @@ class UnifiedAudioConverter:
             print(f"Local LLM failed: {e}, falling back to legacy local method")
             return self.extract_conversation_legacy_local(text, language, progress)
 
+    @spaces.GPU(duration=120)
     def extract_conversation_legacy_local(self, text: str, language: str = "English", progress=None) -> Dict:
         """Extract conversation using legacy local model (fallback)"""
         try:
@@ -465,6 +476,7 @@ class UnifiedAudioConverter:
 
         return tmp_path
 
+    @spaces.GPU(duration=60)
     def text_to_speech_spark(self, conversation_json: Dict, language: str = "English", progress=None) -> Tuple[str, str]:
         """Convert text to speech using Spark TTS CLI"""
         if not SPARK_AVAILABLE or not self.spark_model_dir:
@@ -557,6 +569,7 @@ class UnifiedAudioConverter:
         except Exception as e:
             raise RuntimeError(f"Failed to convert text to speech with Spark TTS: {e}")
 
+    @spaces.GPU(duration=60)
     def text_to_speech_melo(self, conversation_json: Dict, progress=None) -> Tuple[str, str]:
         """Convert text to speech using MeloTTS"""
         if not MELO_AVAILABLE or not self.melo_models:
@@ -751,6 +764,19 @@ def update_tts_engine_for_korean(language):
         )
 
 
+# 모델 초기화 (앱 시작 시)
+if LLAMA_CPP_AVAILABLE:
+    try:
+        model_path = hf_hub_download(
+            repo_id=converter.config.local_model_repo,
+            filename=converter.config.local_model_name,
+            local_dir="./models"
+        )
+        print(f"Model downloaded to: {model_path}")
+    except Exception as e:
+        print(f"Failed to download model at startup: {e}")
+
+
 # Gradio Interface
 with gr.Blocks(theme='soft', title="URL to Podcast Converter") as demo:
     gr.Markdown("# 🎙️ URL to Podcast Converter")