Enable 2x faster inference with unsloth

sia/llm_engine/qwq_llm_engine.py

@@ -43,6 +43,9 @@ class QwQLlmEngine(LlmEngine):
             gpu_memory_utilization = 0.8, # Reduce if out of memory
             tokenizer = self._tokenizer,
         )
+
+        # enable unsloth optimizations
+        FastLanguageModel.for_inference(self._model)
         
         # Create inference pipeline
         self._pipeline = pipeline(