perf(hybrid): prevent expensive array slicing when cache is disabled

Added a `max_checkpoints > 0` check to the `finally` block of the generation loop.

Previously, even though the underlying C++ state extraction was bypassed, the Python layer was still executing `self._input_ids[:self.n_tokens].tolist()`. For long contexts, slicing and converting this massive array to a Python list caused unnecessary CPU overhead and garbage collection (GC) pressure. This intercept acts as a double-layer isolation, ensuring absolute zero memory allocation and zero overhead for hybrid models running in single-turn mode.

Author: JamePeng

llama_cpp/llama.py

@@ -1396,7 +1396,11 @@ def adapter(token_data_array: llama_cpp.llama_token_data_array):
                             ]
                         )
         finally:
-            if self.is_hybrid and self._hybrid_cache_mgr is not None:
+            if (
+                self.is_hybrid
+                and self._hybrid_cache_mgr is not None
+                and self._hybrid_cache_mgr.max_checkpoints > 0
+            ):
                 current_history = self._input_ids[:self.n_tokens].tolist()
 
                 self._hybrid_cache_mgr.save_checkpoint(