change script

app.py

@@ -41,25 +41,22 @@ ASSETS_DIRS = (
 # ────────────────────────────────────────────────────────
 #   CONFIGURATION
 # ────────────────────────────────────────────────────────
-MODEL_ID  = os.getenv("MODEL_ID", "unsloth/Qwen3-30B-A3B")
-USE_INT8  = os.getenv("USE_INT8", "0") == "1"    # set env var to 1 for 8‑bit
-THINKING  = os.getenv("DEFAULT_THINKING", "1") == "1"  # default reasoning mode
+MODEL_ID = os.getenv("MODEL_ID", "Qwen/Qwen3-8B")
+DTYPE    = torch.bfloat16                           # or torch.float16
 
-DTYPE     = torch.bfloat16  # Hopper bf16 fast‑path
-bnb_cfg   = BitsAndBytesConfig(load_in_8bit=True) if USE_INT8 else None
-
-print(f"🔄 Loading {MODEL_ID} ({'8‑bit' if USE_INT8 else 'bf16'}) …")
+print(f"🔄 Loading {MODEL_ID} (dtype = {DTYPE}) …")
 
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID,
-    device_map="auto",
-    torch_dtype=(None if USE_INT8 else DTYPE),
-    quantization_config=bnb_cfg,
+    torch_dtype=DTYPE,
     trust_remote_code=True,
-).eval()
-DEVICE = model.device
-print("✅ Model ready.")
+)
+model.to("cuda" if torch.cuda.is_available() else "cpu")  # single-device move
+
+DEVICE = next(model.parameters()).device
+print(f"✅ Model ready on {DEVICE}.")
 
 for d in ASSETS_DIRS:
     os.makedirs(d, exist_ok=True)