Update README.md

README.md

@@ -42,6 +42,9 @@ print(texts[0])
 
 > The first `generate()` call is slower due to `torch.compile` building optimized kernels. Subsequent calls are much faster.
 
+We already use PagedInference which is quite fast for most interactive tasks, but for large-scale deployment, check the Deployment section 
+which provides a vLLM backend.
+
 ## Categories
 
 By default, category is `"plain"` (general text extraction). You can specify a category to use a task-specific prompt:
@@ -96,6 +99,9 @@ for det in results[0]:
 }
 ```
 
+## Deployment
+TODO: explain how to set up the vLLM server.
+
 ## Citation