Upload app.py

app.py

@@ -33,8 +33,9 @@ except Exception:
     spaces = _SpacesFallback()
 
 
-DEFAULT_MODEL_ID = "openbmb/MiniCPM4.1-8B"
+DEFAULT_MODEL_ID = "Alphaplasti/ToneBridge-MiniCPM4.1-8B"
 MODEL_ID = os.getenv("MODEL_ID", DEFAULT_MODEL_ID).strip() or DEFAULT_MODEL_ID
+HF_TOKEN = (os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN") or "").strip() or None
 DEFAULT_TTS_MODEL_ID = "openbmb/VoxCPM2"
 TTS_MODEL_ID = os.getenv("TTS_MODEL_ID", DEFAULT_TTS_MODEL_ID).strip() or DEFAULT_TTS_MODEL_ID
 ENABLE_SERVER_TTS = os.getenv("ENABLE_SERVER_TTS", "true").strip().lower() in {"1", "true", "yes", "y"}
@@ -279,12 +280,13 @@ def load_model():
 
     try:
         cuda_available = torch.cuda.is_available()
-        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True, token=HF_TOKEN)
         load_kwargs = {
             "torch_dtype": "auto",
             "device_map": "auto",
             "trust_remote_code": True,
             "low_cpu_mem_usage": True,
+            "token": HF_TOKEN,
         }
         if LOAD_IN_4BIT and cuda_available:
             load_kwargs["quantization_config"] = BitsAndBytesConfig(