Upload app.py with huggingface_hub

app.py

@@ -6,38 +6,33 @@ import torch
 import tempfile
 import traceback
 import uvicorn
+import gc
 from fastapi import FastAPI, Request
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import HTMLResponse
 
-# --- [v160] 🚀 HERO STABILITY ENGINE (FINAL MISSION) ---
-print(f"--- [v160] 📡 BOOTING ENGINE ---")
+# --- [v161] 🚀 ULTIMATE GPU ENGINE (STABILITY MISSION) ---
+print(f"--- [v161] 📡 BOOTING ULTIMATE ENGINE ---")
 
 # 🛠️ CRITICAL: TORCHAUDIO MONKEYPATCH 🛠️
-# This bypasses the 'torchcodec' ModuleNotFoundError on Hugging Face Spaces.
 import torchaudio
 import soundfile as sf
-import numpy as np
-
 def HeroLoad(filepath, **kwargs):
-    """Robust alternative to torchaudio.load using soundfile."""
     try:
         data, samplerate = sf.read(filepath)
-        # Convert to float32 and ensure shape is (channels, samples)
         if len(data.shape) == 1:
             data = data.reshape(1, -1)
         else:
-            data = data.T # (samples, channels) -> (channels, samples)
+            data = data.T
         return torch.from_numpy(data).float(), samplerate
     except Exception as e:
-        print(f"--- [v159] ❌ PATCHED LOAD FAILED: {e} ---")
-        # Fallback to original if soundfile fails (unlikely)
+        print(f"--- [v161] ❌ PATCHED LOAD FAILED: {e} ---")
         return torchaudio.load_orig(filepath, **kwargs)
 
 if not hasattr(torchaudio, 'load_orig'):
     torchaudio.load_orig = torchaudio.load
     torchaudio.load = HeroLoad
-    print("--- [v160] 🩹 TORCHAUDIO HERO PATCH APPLIED (soundfile) ---")
+    print("--- [v161] 🩹 TORCHAUDIO HERO PATCH APPLIED ---")
 
 from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor
 from TTS.api import TTS
@@ -67,75 +62,111 @@ os.environ["PYTHONWARNINGS"] = "ignore"
 app = FastAPI()
 app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
 
-MODELS = {"stt": None, "tts": None}
+MODELS = {"stt": None, "tts": None, "current_gpu": "auto"}
 
-def load_tts_cpu():
-    global MODELS
-    if MODELS.get("tts") is None:
-        print("--- [v160] 📥 LOADING XTTS V2 (CPU MODE) ---")
-        # CPU loading is 100% stable on ZeroGPU H200
-        MODELS["tts"] = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cpu")
-        print("--- [v160] ✅ XTTS READY (CPU) ---")
+def select_best_gpu():
+    """Logic to select the best available GPU (Switch)."""
+    if torch.cuda.is_available():
+        count = torch.cuda.device_count()
+        print(f"--- [v161] 🖥️ DETECTED {count} GPUs ---")
+        # In ZeroGPU, we usually have 1 MIG instance assigned.
+        return "cuda:0"
+    return "cpu"
 
-@spaces.GPU(duration=60)
-def gpu_stt_base(temp_path, lang):
+@spaces.GPU(duration=120)
+def gpu_stt_large(temp_path, lang):
     global MODELS
+    device = select_best_gpu() if MODELS["current_gpu"] == "auto" else MODELS["current_gpu"]
+    
     if MODELS.get("stt") is None:
-        print("--- [v160] 📥 LOADING WHISPER (Base) ON GPU ---")
-        model_id = "openai/whisper-base"
-        model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch.float32).to("cuda")
+        print(f"--- [v161] 📥 LOADING WHISPER (Large-v3-Turbo) ON {device} ---")
+        model_id = "openai/whisper-large-v3-turbo"
+        model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch.float16 if "cuda" in device else torch.float32).to(device)
         processor = AutoProcessor.from_pretrained(model_id)
         MODELS["stt"] = pipeline(
             "automatic-speech-recognition",
             model=model,
             tokenizer=processor.tokenizer,
             feature_extractor=processor.feature_extractor,
-            device="cuda"
+            device=device
         )
-    print(f"--- [v160] 🎙️ RUNNING WHISPER INFERENCE ---")
-    res = MODELS["stt"](temp_path, generate_kwargs={"language": lang if lang and len(lang) <= 3 else None})
+    
+    print(f"--- [v161] 🎙️ WHISPER INFERENCE (Temp=0, TS=True) ---")
+    # Added return_timestamps=True to handle > 30s audio
+    res = MODELS["stt"](
+        temp_path, 
+        generate_kwargs={
+            "language": lang if lang and len(lang) <= 3 else None,
+            "temperature": 0.0,
+            "return_timestamps": True
+        }
+    )
     return res["text"].strip()
 
+@spaces.GPU(duration=120)
+def gpu_tts_inference(text, mapped_lang, speaker_path):
+    global MODELS
+    device = "cuda" # Always use CUDA inside the decorator
+    
+    if MODELS.get("tts") is None:
+        print(f"--- [v161] 📥 LOADING XTTS V2 ON {device} ---")
+        # We load directly to GPU inside the decorator for max stability
+        MODELS["tts"] = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)
+    
+    # Ensure model is on GPU (sometimes ZeroGPU logic moves it back to CPU)
+    if hasattr(MODELS["tts"], "to"):
+        MODELS["tts"].to(device)
+        
+    print(f"--- [v161] 🔊 XTTS INFERENCE ON GPU ---")
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as out_f:
+        out_p = out_f.name
+    
+    MODELS["tts"].tts_to_file(text=text, language=mapped_lang, file_path=out_p, speaker_wav=speaker_path)
+    
+    with open(out_p, "rb") as f:
+        audio_b64 = base64.b64encode(f.read()).decode()
+    
+    if os.path.exists(out_p):
+        os.unlink(out_p)
+        
+    return audio_b64
+
 async def handle_process(request: Request):
     t1 = time.time()
     try:
         data = await request.json()
         action = data.get("action")
-        if action == "health": return {"status": "awake", "v": "160"}
+        if action == "health": return {"status": "awake", "v": "161"}
         
-        print(f"--- [v160] 🛠️ {action.upper()} REQUESTED ---")
+        print(f"--- [v161] 🛠️ {action.upper()} REQUESTED ---")
         
-        # 🟢 STT PATH
         stt_text = ""
+        # 🟢 STT
         if action in ["stt", "s2st"]:
             audio_b64 = data.get("file")
             if not audio_b64: return {"error": "Missing audio file"}
-            
             audio_bytes = base64.b64decode(audio_b64)
             with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
-                f.write(audio_bytes)
-                temp_path = f.name
+                f.write(audio_bytes); temp_path = f.name
             try:
-                stt_text = gpu_stt_base(temp_path, data.get("lang"))
-                print(f"--- [v160] 🎙️ TRANSCRIPT: {stt_text} ---")
+                stt_text = gpu_stt_large(temp_path, data.get("lang"))
+                print(f"--- [v161] 🎙️ TRANSCRIPT: {stt_text[:100]}... ---")
             finally:
                 if os.path.exists(temp_path): os.unlink(temp_path)
-            
             if action == "stt": return {"text": stt_text}
 
-        # 🔵 TTS PATH
+        # 🔵 TTS
         if action in ["tts", "s2st"]:
             text = (data.get("text") if action == "tts" else stt_text).strip()
             if not text: return {"error": "Input text is empty"}
             
-            trans_text = text
             target = data.get("target_lang") or data.get("lang") or "en"
-            
+            trans_text = text
             if action == "s2st":
-                print(f"--- [v160] 🌏 TRANSLATING TO {target} ---")
+                print(f"--- [v161] 🌏 TRANSLATING TO {target} ---")
                 trans_text = GoogleTranslator(source='auto', target=target).translate(stt_text)
                 text = trans_text
-                print(f"--- [v160] 📝 TRANSLATED: {text} ---")
+                print(f"--- [v161] 📝 TRANSLATED: {text[:100]}... ---")
 
             XTTS_MAP = {"en": "en", "de": "de", "fr": "fr", "es": "es", "it": "it", "pl": "pl", "pt": "pt", "tr": "tr", "ru": "ru", "nl": "nl", "cs": "cs", "ar": "ar", "hu": "hu", "ko": "ko", "hi": "hi", "zh": "zh-cn"}
             clean_lang = target.split('-')[0].lower()
@@ -143,51 +174,50 @@ async def handle_process(request: Request):
             
             if not mapped_lang:
                 if HAS_CHATTERBOX:
-                    print(f"--- [v160] 📦 FALLBACK: CHATTERBOX FOR {clean_lang} ---")
                     audio_bytes = chatterbox_utils.run_chatterbox_inference(text, clean_lang)
                     audio_b64 = base64.b64encode(audio_bytes).decode()
                 else: return {"error": f"Lang {clean_lang} unsupported"}
             else:
-                load_tts_cpu()
                 speaker_wav_b64 = data.get("speaker_wav")
                 speaker_path = None
                 if speaker_wav_b64:
                     with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
-                        f.write(base64.b64decode(speaker_wav_b64))
-                        speaker_path = f.name
+                        f.write(base64.b64decode(speaker_wav_b64)); speaker_path = f.name
                 else:
                     speaker_path = "default_speaker.wav"
                     if not os.path.exists(speaker_path): speaker_path = None
                 
                 try:
-                    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as out_f:
-                        out_p = out_f.name
-                    print(f"--- [v160] 🔊 XTTS INFERENCE (CPU) ---")
-                    MODELS["tts"].tts_to_file(text=text, language=mapped_lang, file_path=out_p, speaker_wav=speaker_path)
-                    with open(out_p, "rb") as f:
-                        audio_b64 = base64.b64encode(f.read()).decode()
+                    # RUN ON GPU
+                    audio_b64 = gpu_tts_inference(text, mapped_lang, speaker_path)
                 finally:
                     if speaker_wav_b64 and speaker_path and os.path.exists(speaker_path): os.unlink(speaker_path)
-                    if 'out_p' in locals() and os.path.exists(out_p): os.unlink(out_p)
             
             if action == "tts": return {"audio": audio_b64}
             return {"text": stt_text, "translated": trans_text, "audio": audio_b64}
 
     except Exception as e:
-        print(f"❌ [v160] HERO ERROR: {traceback.format_exc()}")
+        print(f"❌ [v161] ULTIMATE ERROR: {traceback.format_exc()}")
         return {"error": str(e)}
     finally:
-        print(f"--- [v160] ✨ DONE ({time.time()-t1:.1f}s) ---")
+        print(f"--- [v161] ✨ DONE ({time.time()-t1:.1f}s) ---")
 
 @app.post("/process")
 @app.post("/api/v1/process")
 async def api_process(request: Request): return await handle_process(request)
 
 @app.get("/health")
-def health(): return {"status": "ok", "v": "160", "mode": "HERO_STABLE", "gpu": torch.cuda.is_available()}
+def health(): 
+    return {
+        "status": "ok", 
+        "v": "161", 
+        "mode": "ULTIMATE_GPU", 
+        "gpu_available": torch.cuda.is_available(),
+        "gpu_count": torch.cuda.device_count() if torch.cuda.is_available() else 0
+    }
 
 @app.get("/", response_class=HTMLResponse)
-def root(): return "<h1>🚀 AI Engine v160 (HERO STABLE)</h1>"
+def root(): return "<h1>🚀 AI Engine v161 (ULTIMATE GPU)</h1>"
 
 if __name__ == "__main__":
     uvicorn.run(app, host="0.0.0.0", port=7860)