Refactor model loading to remove device_map and manually move model to device

app.py

@@ -66,13 +66,13 @@ snapshot_download(
 print("Model files downloaded successfully")
 
 print("Loading main model...")
-# Load model and tokenizer with token
+# Load model and tokenizer with token (removed device_map)
 model = AutoModelForCausalLM.from_pretrained(
     model_name, 
     torch_dtype=torch.bfloat16,
-    token=hf_token,
-    device_map="auto"
+    token=hf_token
 )
+model = model.to(device)  # Move to device manually
 
 print("Loading tokenizer...")
 tokenizer = AutoTokenizer.from_pretrained(