Update app.py

app.py

@@ -11,20 +11,18 @@ tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
 tokenizer.pad_token = tokenizer.eos_token
 
 # Base model (CPU)
+# 1. Base model (disk offload aktif)
 model = AutoModelForCausalLM.from_pretrained(
     BASE_MODEL,
     torch_dtype=torch.float32,
-    offload_folder="offload",
-    device_map={"": "cpu"},
+    device_map="auto",
+    offload_dir="offload",
+    offload_state_dict=True,
     low_cpu_mem_usage=True
 )
 
-# Load LoRA
-model = PeftModel.from_pretrained(
-    model,
-    LORA_PATH,
-    is_trainable=False
-)
+# 2. LoRA (RAM'de kalır)
+model = PeftModel.from_pretrained(model, LORA_PATH)
 model.eval()