Hi folks, I am running a fine-tuned version of llama 3 (not 3.1/3.2) via vLLM. I am currently unable to register simple tasks successfully. I am wondering whether this is related to the model itself not being able to call functions properly, or is there some other issue with the parser? I am using

--tool-call-parser llama3_json

in the vLLM engine args. If the LLM call is not limited, then it will go on an infinite loop. If it is given a tool, then it will hallucinate the output of the tool (here,

ls

is a wrapper for

os.listdir()

. Any input would be appreciated. Kindly thanks!