Upload app.py with huggingface_hub

app.py

@@ -18,7 +18,7 @@ import logging
 from threading import Thread, Lock
 from huggingface_hub import snapshot_download
 
-# 🛡️ 1. SILENCE LOGS & WARNINGS (v104: Extreme Stability & Precedence)
+# 🛡️ 1. SILENCE LOGS & WARNINGS (v105: CUDA Stability Protocol)
 logging.getLogger("transformers").setLevel(logging.ERROR)
 logging.getLogger("TTS").setLevel(logging.ERROR)
 logging.getLogger("onnxruntime").setLevel(logging.ERROR)
@@ -81,8 +81,8 @@ except ImportError:
             if f is None: return lambda x: x
             return f
 
-# FORCE BUILD TRIGGER: 13:35:00 Jan 21 2026
-# v104: FastAPI-First Precedence. Fix 405 error. ZeroGPU Unified.
+# FORCE BUILD TRIGGER: 14:00:00 Jan 21 2026
+# v105: CUDA Stability Fix. int8_float16. Device Indexing.
 
 os.environ["COQUI_TOS_AGREED"] = "1"
 MODELS = {"stt": None, "translate": None, "tts": None, "denoiser": None}
@@ -91,7 +91,7 @@ WARMUP_STATUS = {"complete": False, "in_progress": False}
 WARMUP_LOCK = Lock()
 
 def activate_gpu_models(action):
-    """v104: Stability Milestone Activation"""
+    """v105: MISSION-CRITICAL GPU ACTIVATION"""
     global MODELS, WARMUP_STATUS
     local_only = WARMUP_STATUS["complete"]
     
@@ -100,19 +100,22 @@ def activate_gpu_models(action):
         try: stt_on_gpu = MODELS["stt"] is not None and MODELS["stt"].model.device == "cuda"
         except: pass
         if not stt_on_gpu:
-            print(f"🎙️ [v104] Activating Whisper (GPU)...")
+            print(f"🎙️ [v105] Activating Whisper (GPU: int8_float16 Stability)...")
             try:
                 if MODELS["stt"]: del MODELS["stt"]
                 gc.collect(); torch.cuda.empty_cache()
+                time.sleep(0.5) # Let CUDA driver settle
+                # v105: int8_float16 is the most stable for CTranslate2 on H200 MIGs
                 MODELS["stt"] = WhisperModel(
                     "large-v3", 
                     device="cuda", 
-                    compute_type="float16", 
+                    device_index=0,
+                    compute_type="int8_float16", 
                     num_workers=1,
                     local_files_only=local_only
                 )
             except Exception as e:
-                print(f"⚠️ GPU Init failed: {e}")
+                print(f"⚠️ CUDA v105 Stability Init failed: {e}")
                 MODELS["stt"] = WhisperModel("large-v3", device="cpu", compute_type="int8", local_files_only=True)
 
     if action in ["tts", "s2st"]:
@@ -122,7 +125,7 @@ def activate_gpu_models(action):
             tts_on_gpu = "cuda" in curr
         except: pass
         if MODELS["tts"] is None or not tts_on_gpu:
-            print(f"🔊 [v104] Activating XTTS-v2 (GPU)...")
+            print(f"🔊 [v105] Activating XTTS-v2 (GPU)...")
             try:
                 if MODELS["tts"] is None:
                     MODELS["tts"] = TTS(model_name="tts_models/multilingual/multi-dataset/xtts_v2", gpu=True)
@@ -137,9 +140,9 @@ def activate_gpu_models(action):
     if MODELS["translate"] is None: MODELS["translate"] = "active"
 
 def release_gpu_models():
-    """v104: Resident RAM Protocol"""
+    """v105: Safe Resident Cleanup"""
     global MODELS
-    print("🧹 [v104] Releasing GPU resources.")
+    print("🧹 [v105] Releasing GPU resources.")
     try:
         if MODELS["stt"] and MODELS["stt"].model.device == "cuda":
             del MODELS["stt"]
@@ -151,20 +154,21 @@ def release_gpu_models():
     except: pass
     gc.collect()
     if torch.cuda.is_available(): torch.cuda.empty_cache()
+    time.sleep(0.5) # Driver buffer
 
 def warmup_task():
-    """Silent Warmup (v104)"""
+    """Silent Pre-loading (v105)"""
     global WARMUP_STATUS
     with WARMUP_LOCK:
         if WARMUP_STATUS["complete"] or WARMUP_STATUS["in_progress"]: return
         WARMUP_STATUS["in_progress"] = True
-    print("\n🔥 --- V104: PRECEDENCE WARMUP STARTED ---")
+    print("\n🔥 --- V105: STABILITY WARMUP STARTED ---")
     try:
         MODELS["stt"] = WhisperModel("large-v3", device="cpu", compute_type="int8")
         MODELS["tts"] = TTS(model_name="tts_models/multilingual/multi-dataset/xtts_v2", gpu=False)
         chatterbox_utils.warmup_chatterbox()
         WARMUP_STATUS["complete"] = True
-        print(f"✅ --- SYSTEM READY: v104 --- \n")
+        print(f"✅ --- SYSTEM READY: v105 --- \n")
     except: pass
     finally: WARMUP_STATUS["in_progress"] = False
 
@@ -216,7 +220,7 @@ def _tts_logic(text, lang, speaker_wav_b64):
 def core_process(request_dict):
     action = request_dict.get("action")
     t1 = time.time()
-    print(f"--- [v104] 🚀 GPU SESSION: {action} ---")
+    print(f"--- [v105] 🚀 GPU SESSION: {action} ---")
     activate_gpu_models(action)
     try:
         if action == "stt": res = _stt_logic(request_dict)
@@ -229,7 +233,7 @@ def core_process(request_dict):
             res = {"text": stt_res.get("text"), "translated": translated, "audio": tts_res.get("audio")}
         else: res = {"error": f"Unknown action: {action}"}
     finally:
-        print(f"--- [v104] ✨ SUCCESS: {action} ({time.time()-t1:.2f}s) ---")
+        print(f"--- [v105] ✨ SUCCESS: {action} ({time.time()-t1:.2f}s) ---")
         release_gpu_models()
     return res
 
@@ -247,7 +251,6 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
-# 🚀 STEP 2: REGISTER ROUTES ON PRIMARY APP (High Precedence)
 @app.post("/api/v1/process")
 async def api_process(request: Request):
     try:
@@ -258,7 +261,7 @@ async def api_process(request: Request):
     except Exception as e: return {"error": str(e)}
 
 @app.get("/health")
-def health(): return {"status": "ok", "warm": WARMUP_STATUS["complete"], "v": "104"}
+def health(): return {"status": "ok", "warm": WARMUP_STATUS["complete"], "v": "105"}
 
 @app.post("/api/v1/clear_cache")
 async def clear_cache_api():
@@ -267,7 +270,7 @@ async def clear_cache_api():
         return {"status": "success"}
     except: return {"status": "error"}
 
-# 🚀 STEP 3: DEFINE GRADIO INTERFACE
+# 🚀 STEP 2: DEFINE GRADIO INTERFACE
 def gradio_fn(req_json):
     try: return json.dumps(core_process(json.loads(req_json)))
     except Exception as e: return json.dumps({"error": str(e)})
@@ -276,15 +279,12 @@ demo = gr.Interface(
     fn=gradio_fn, 
     inputs="text", 
     outputs="text", 
-    title="🚀 AI Engine v104"
+    title="🚀 AI Engine v105"
 )
 
-# 🚀 STEP 4: MOUNT GRADIO ONTO FASTAPI (Low Precedence)
-# This way, our /api/v1/process route is checked BEFORE Gradio handles it.
+# 🚀 STEP 3: MOUNT GRADIO ONTO FASTAPI
 app = gr.mount_gradio_app(app, demo, path="/")
 
-# 🚀 STEP 5: FINAL LAUNCH
 if __name__ == "__main__":
-    print("🚀 [v104] Starting Unified Server on Port 7860...")
-    # Using uvicorn on the unified app instance.
+    print("🚀 [v105] Starting Unified Server on Port 7860...")
     uvicorn.run(app, host="0.0.0.0", port=7860, log_level="error")