Upload app.py with huggingface_hub

app.py

@@ -11,10 +11,10 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import HTMLResponse
 import uvicorn
 
-# --- [v152] 🚀 REFINED HYBRID ENGINE (CPU-STT + CPU-TRANS + GPU-TTS) ---
-print(f"--- [v152] 📡 BOOTING REFINED HYBRID ENGINE ---")
+# --- [v153] 🚀 EPHEMERAL GPU ENGINE ---
+print(f"--- [v153] 📡 BOOTING EPHEMERAL ENGINE ---")
 
-from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor
+from transformers import pipeline
 from TTS.api import TTS
 from deep_translator import GoogleTranslator
 
@@ -42,66 +42,52 @@ os.environ["PYTHONWARNINGS"] = "ignore"
 app = FastAPI()
 app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
 
-MODELS = {"stt": None, "tts": None}
+MODELS = {"stt": None}
 
 def load_stt_cpu():
-    """STT on CPU is stable and fast for Whisper Base."""
     global MODELS
     if MODELS.get("stt") is None:
-        print("--- [v152] 📥 LOADING WHISPER (Base) ON CPU ---")
-        model_id = "openai/whisper-base"
-        MODELS["stt"] = pipeline(
-            "automatic-speech-recognition",
-            model=model_id,
-            device="cpu", # Force CPU for stability
-            torch_dtype=torch.float32
-        )
-        print("--- [v152] ✅ WHISPER READY (CPU) ---")
-
-def load_tts_gpu():
-    global MODELS
-    if MODELS.get("tts") is None:
-        print("--- [v152] 📥 LOADING XTTS V2 ---")
-        # Load to CPU first, then move to CUDA inside the decorated function if needed, 
-        # or just load directly if ZeroGPU allows.
-        MODELS["tts"] = TTS("tts_models/multilingual/multi-dataset/xtts_v2")
-        print("--- [v152] ✅ XTTS READY (CPU MEMORY) ---")
-
-@spaces.GPU(duration=120)
-def gpu_tts_inference(text, mapped_lang, speaker_path):
-    """Isolated GPU inference for XTTS."""
-    global MODELS
-    if MODELS["tts"] is None:
-        load_tts_gpu()
-    
-    # Move to GPU inside the decorated scope
-    MODELS["tts"].to("cuda")
-    
-    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as out_f:
-        out_p = out_f.name
-    
+        print("--- [v153] 📥 LOADING WHISPER (Base) ON CPU ---")
+        MODELS["stt"] = pipeline("automatic-speech-recognition", model="openai/whisper-base", device="cpu")
+        print("--- [v153] ✅ WHISPER READY (CPU) ---")
+
+@spaces.GPU(duration=180)
+def ephemeral_tts(text, mapped_lang, speaker_path):
+    """Pure ephemeral loading on GPU to bypass VRAM watchdogs."""
+    print(f"--- [v153] 📥 LOADING XTTS EPOCH... ---")
+    local_tts = None
     try:
-        MODELS["tts"].tts_to_file(text=text, language=mapped_lang, file_path=out_p, speaker_wav=speaker_path)
+        local_tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda")
+        local_tts.to(torch.float32)
+        
+        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as out_f:
+            out_p = out_f.name
+        
+        print(f"--- [v153] 🔊 INFERENCE... ---")
+        local_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()
         return audio_b64
     finally:
-        # Move back to CPU to release VRAM as per ZeroGPU guidelines
-        MODELS["tts"].to("cpu")
-        if os.path.exists(out_p): os.unlink(out_p)
+        print(f"--- [v153] 🧹 CLEANUP ---")
+        if local_tts:
+            del local_tts
+        if 'out_p' in locals() and os.path.exists(out_p): os.unlink(out_p)
+        gc.collect()
         torch.cuda.empty_cache()
 
 async def handle_process(request: Request):
     try:
         data = await request.json()
         action = data.get("action")
-        if action == "health": return {"status": "awake", "v": "152"}
+        if action == "health": return {"status": "awake", "v": "153"}
         
-        print(f"--- [v152] 🛠️ ENGINE ACTION: {action} ---")
+        print(f"--- [v153] 🛠️ {action} ---")
         t1 = time.time()
         
         # 🎙️ STT (CPU)
-        stt_text = None
+        stt_text = ""
         if action in ["stt", "s2st"]:
             load_stt_cpu()
             audio_bytes = base64.b64decode(data.get("file"))
@@ -111,11 +97,12 @@ async def handle_process(request: Request):
                 lang = data.get("lang")
                 res = MODELS["stt"](temp_path, generate_kwargs={"language": lang if lang and len(lang) <= 3 else None})
                 stt_text = res["text"].strip()
-                if action == "stt": return {"text": stt_text}
             finally:
                 if os.path.exists(temp_path): os.unlink(temp_path)
+            
+            if action == "stt": return {"text": stt_text}
 
-        # 🔊 TTS (Hybrid GPU)
+        # 🔊 TTS (GPU)
         if action in ["tts", "s2st"]:
             text = (data.get("text") if action == "tts" else stt_text).strip()
             trans_text = text
@@ -147,7 +134,7 @@ async def handle_process(request: Request):
                     if not os.path.exists(speaker_path): speaker_path = None
                 
                 try:
-                    audio_b64 = gpu_tts_inference(text, mapped_lang, speaker_path)
+                    audio_b64 = ephemeral_tts(text, mapped_lang, speaker_path)
                 finally:
                     if speaker_wav_b64 and speaker_path and os.path.exists(speaker_path): os.unlink(speaker_path)
             
@@ -155,21 +142,20 @@ async def handle_process(request: Request):
             return {"text": stt_text, "translated": trans_text, "audio": audio_b64}
 
     except Exception as e:
-        print(f"❌ [v152] ERROR: {traceback.format_exc()}")
+        print(f"❌ [v153] ERROR: {traceback.format_exc()}")
         return {"error": str(e)}
     finally:
-        print(f"--- [v152] ✨ DONE ({time.time()-t1:.1f}s) ---")
+        print(f"--- [v153] ✨ 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": "152", "gpu": torch.cuda.is_available()}
+def health(): return {"status": "ok", "v": "153", "gpu": torch.cuda.is_available()}
 
 @app.get("/", response_class=HTMLResponse)
-def root(): return "<h1>🚀 AI Engine v152 (REFINED HYBRID)</h1>"
+def root(): return "<h1>🚀 AI Engine v153 (EPHEMERAL)</h1>"
 
 if __name__ == "__main__":
     uvicorn.run(app, host="0.0.0.0", port=7860)