Bugfix: only eval new tokens

Author: abetlen

Diff for: llama_cpp/llama.py

@@ -280,6 +280,7 @@ def generate(
             if self.verbose:
                 print("generate cache hit", file=sys.stderr)
             reset = False
+            tokens = tokens[len(self.tokens) :]
         ###
         if reset:
             self.reset()