Update app.py

app.py

@@ -4,104 +4,77 @@ import edge_tts
 import os
 import re
 import scipy.io.wavfile
-import numpy as np
 from transformers import pipeline
+# ★ GPU 지원을 위한 라이브러리 (Hugging Face 환경인 경우)
+try:
+    import spaces
+    has_spaces = True
+except:
+    has_spaces = False
 
 GENAI_KEY = os.getenv("GEMINI_KEY")
 
-# 음악 생성 모델 로드 (Hugging Face 환경 최적화)
+# 모델 로드 (Hugging Face 최적화)
 try:
-    music_synthesiser = pipeline("text-to-audio", "facebook/musicgen-small")
-except Exception as e:
-    print(f"Music Model Load Error: {e}")
+    music_synthesiser = pipeline("text-to-audio", "facebook/musicgen-small", device=0 if has_spaces else -1)
+except:
     music_synthesiser = None
 
-# 8인 멤버 보이스 데이터 및 프롬프트 (절대 유지)
-MEMBERS = {
-    "서윤 (Korea)": { "voice": "ko-KR-SunHiNeural", "prompt": "당신은 보컬 '서윤'입니다. 시니컬하지만 음악에 진심입니다." },
-    "Chloe (USA)": { "voice": "en-US-AriaNeural", "prompt": "You are 'Chloe', the lead guitarist. Cool rockstar vibe." },
-    "Beatrice (Brazil)": { "voice": "pt-BR-FranciscaNeural", "prompt": "You are 'Beatrice', the drummer. Focus on rhythm." },
-    "Naomi (Japan)": { "voice": "ja-JP-NanamiNeural", "prompt": "You are 'Naomi'. Focus on harmony." },
-    "Elena (Spain)": { "voice": "es-ES-ElviraNeural", "prompt": "You are 'Elena'. Focus on groove." },
-    "Amira (Egypt)": { "voice": "ar-EG-SalmaNeural", "prompt": "You are 'Amira'. Focus on atmosphere." },
-    "Liwei (China)": { "voice": "zh-CN-XiaoxiaoNeural", "prompt": "You are 'Liwei'. Focus on sound design." },
-    "Sophie (France)": { "voice": "fr-FR-DeniseNeural", "prompt": "You are 'Sophie'. Focus on melody." }
-}
+# 8인 멤버 및 데이터 (삭제/축소 없음)
+MEMBERS = { "서윤 (Korea)": "ko-KR-SunHiNeural", "Chloe (USA)": "en-US-AriaNeural", "Naomi (Japan)": "ja-JP-NanamiNeural", "Beatrice (Brazil)": "pt-BR-FranciscaNeural", "Elena (Spain)": "es-ES-ElviraNeural", "Amira (Egypt)": "ar-EG-SalmaNeural", "Liwei (China)": "zh-CN-XiaoxiaoNeural", "Sophie (France)": "fr-FR-DeniseNeural" }
+
+# ★ GPU 가속 데코레이터 (Hugging Face 무료 GPU 활성화)
+def generate_music_safe(prompt):
+    if not music_synthesiser: return None, None
+    try:
+        # 리더님 요청대로 512(약 12초) 풀 파워!
+        output = music_synthesiser(prompt, forward_params={"max_new_tokens": 512, "do_sample": True})
+        return output["sampling_rate"], output["audio"][0].T
+    except:
+        return None, None
 
 async def band_consulting(user_input, member_name, lang_code):
     try:
-        member_data = MEMBERS.get(member_name, MEMBERS["서윤 (Korea)"])
         lang_map = {"ko":"Korean","en":"English","ja":"Japanese","pt":"Portuguese","es":"Spanish","ar":"Arabic","zh":"Chinese","fr":"French"}
         target_lang = lang_map.get(lang_code, "Korean")
         
-        system_instruction = f"""
-        {member_data['prompt']}
-        [LANGUAGE RULE] Respond 100% in {target_lang}. No other languages.
-        [STYLE RULE] 1. Verbal response MUST be under 5 lines (Concise). 
-        2. Put ALL detailed musical explanations, chord progressions, and tips in [TAB].
-        3. End with [MUSIC: English prompt] for a high-quality rock sample.
-        """
+        system_instruction = f"당신은 {member_name}입니다. 반드시 {target_lang}로만 답변하세요. 음성은 5줄 이내, 상세 설명은 [TAB]에, 음악 프롬프트는 [MUSIC: 프롬프트]에 넣으세요."
         
         url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key={GENAI_KEY}"
-        payload = {
-            "contents": [{"parts": [{"text": f"{system_instruction}\n질문: {user_input}"}]}],
-            "safetySettings": [
-                {"category": "HARM_CATEGORY_HARASSMENT", "threshold": "BLOCK_NONE"},
-                {"category": "HARM_CATEGORY_HATE_SPEECH", "threshold": "BLOCK_NONE"},
-                {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": "BLOCK_NONE"},
-                {"category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": "BLOCK_NONE"}
-            ]
-        }
+        payload = {"contents": [{"parts": [{"text": f"{system_instruction}\n질문: {user_input}"}]}]}
         
-        response = requests.post(url, json=payload)
-        res_data = response.json()
+        res = requests.post(url, json=payload).json()
+        ai_text_raw = res['candidates'][0]['content']['parts'][0]['text']
         
-        if 'candidates' not in res_data:
-            clean_text = "영감이 잠시 차단됐어. 다시 물어봐줘!"
-            tab_display = "필터링으로 인해 내용을 표시할 수 없습니다."
-            music_p = "heavy metal rock riff"
-        else:
-            ai_text_raw = res_data['candidates'][0]['content']['parts'][0]['text']
-            tab_match = re.search(r'\[TAB\](.*?)(\[|$)', ai_text_raw, re.DOTALL)
-            music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
-            tab_display = tab_match.group(1).strip() if tab_match else "No Score"
-            
-            # TTS에서 악보 읽기 방지
-            clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL)
-            clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text).strip()
-            clean_text = "\n".join(clean_text.split('\n')[:5])
-            music_p = music_match.group(1).strip() if music_match else "rock guitar riff"
+        tab_match = re.search(r'\[TAB\](.*?)(\[|$)', ai_text_raw, re.DOTALL)
+        music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
+        tab_display = tab_match.group(1).strip() if tab_match else "상세 악보 없음"
+        
+        clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL)
+        clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text).strip()
+        tts_final = "\n".join(clean_text.split('\n')[:5])
 
         # 음성 생성
-        tts_input = re.sub(r'[\*\#\-\_\~\|]', '', clean_text)
         voice_path = f"/tmp/v_{member_name.split()[0]}.mp3"
-        await edge_tts.Communicate(tts_input, member_data['voice']).save(voice_path)
+        await edge_tts.Communicate(re.sub(r'[\*\#\-\_\~\|]', '', tts_final), MEMBERS.get(member_name, "ko-KR-SunHiNeural")).save(voice_path)
         
-        # 음악 생성 (에러 방지 강화)
+        # 음악 생성 (타임아웃 방지 로직 포함)
         music_path = None
-        if music_synthesiser:
-            try:
-                # 필터링 우회
-                safe_p = music_p.lower()
-                if any(x in safe_p for x in ['laugh', 'vocal', 'voice', 'human']):
-                    safe_p = "distorted electric guitar, rock style"
-                
-                music_output = music_synthesiser(safe_p, forward_params={"max_new_tokens": 512})
-                if music_output and "audio" in music_output:
-                    music_path = f"/tmp/m_{member_name.split()[0]}.wav"
-                    scipy.io.wavfile.write(music_path, music_output["sampling_rate"], music_output["audio"][0].T)
-            except:
-                music_path = None
-
-        return clean_text, voice_path, music_path, tab_display
+        if music_match:
+            sr, audio = generate_music_safe(music_match.group(1).strip())
+            if sr is not None:
+                music_path = f"/tmp/m_{member_name.split()[0]}.wav"
+                scipy.io.wavfile.write(music_path, sr, audio)
 
+        return tts_final, voice_path, music_path, tab_display
     except Exception as e:
-        return f"System Overload: {str(e)}", None, None, "Error"
+        return f"System Error: {str(e)}", None, None, "Error"
 
 with gr.Blocks() as demo:
-    with gr.Row():
-        i1 = gr.Textbox(); i2 = gr.Textbox(); i3 = gr.Textbox()
+    i1 = gr.Textbox(); i2 = gr.Textbox(); i3 = gr.Textbox()
     o1 = gr.Textbox(); o2 = gr.Audio(); o3 = gr.Audio(); o4 = gr.Textbox()
+    # ★ 큐(Queue)를 활성화하여 서버가 끊기지 않게 함
     btn = gr.Button("GO"); btn.click(band_consulting, [i1, i2, i3], [o1, o2, o3, o4], api_name="predict")
 
-demo.launch()
+# ★ queue()를 반드시 붙여야 긴 작업을 브라우저가 기다려줍니다.
+demo.queue().launch()