Merge pull request #77 from jkawamoto/grammar-verbosity

Print grammar only if verbose is True

Author: Maximilian-Winter

src/llama_cpp_agent/providers/llama_cpp_python.py

@@ -151,7 +151,7 @@ def create_completion(
             if grammar in self.grammar_cache:
                 grammar = self.grammar_cache[grammar]
             else:
-                self.grammar_cache[grammar] = LlamaGrammar.from_string(grammar)
+                self.grammar_cache[grammar] = LlamaGrammar.from_string(grammar, verbose=self.llama_model.verbose)
                 grammar = self.grammar_cache[grammar]
 
         settings_dictionary = deepcopy(settings.as_dict())
@@ -179,7 +179,7 @@ def create_chat_completion(
             if grammar in self.grammar_cache:
                 grammar = self.grammar_cache[grammar]
             else:
-                self.grammar_cache[grammar] = LlamaGrammar.from_string(grammar)
+                self.grammar_cache[grammar] = LlamaGrammar.from_string(grammar, verbose=self.llama_model.verbose)
                 grammar = self.grammar_cache[grammar]
         settings_dictionary = deepcopy(settings.as_dict())
         settings_dictionary["max_tokens"] = settings_dictionary.pop("n_predict")