reduce position embeddings

app.py

@@ -30,7 +30,8 @@ model = AutoModelForCausalLM.from_pretrained(
     quantization_config=quantization_config,  # Apply quantization
     # device_map="auto",  # Automatically map to available devices
     torch_dtype=torch.bfloat16,
-    token=HF_TOKEN
+    token=HF_TOKEN,
+    max_position_embeddings=8192  # Reduce context window to 8k tokens (from 128k)
 )
 
 @spaces.GPU