Fixes

modeling_functionary.py

@@ -48,7 +48,9 @@ class FunctionaryForCausalLM(LlamaForCausalLM):
         negative_prompt_attention_mask: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Union[GenerateOutput, torch.LongTensor]:
-        
+
+        tokenizer = kwargs.pop("tokenizer", None)  # Pull this out first, we use it to parse raw output
+
         results = self.generate(
             inputs=inputs,
             generation_config=generation_config,
@@ -63,7 +65,6 @@ class FunctionaryForCausalLM(LlamaForCausalLM):
             **kwargs,
         )
 
-        tokenizer = kwargs.pop("tokenizer", None)  # Pull this out first, we use it to parse raw output
         input_ids = kwargs.pop("input_ids")
         function_call_token = "<|reserved_special_token_249|>"