inference_config:
  impl_type: "inline"
  inline_config:
    checkpoint_type: "pytorch"
    checkpoint_dir: /home/cyni/local/llama-3
    tokenizer_path: /home/cyni/local/llama-3/cl_toplang_128k
    model_parallel_size: 1
    max_seq_len: 2048
    max_batch_size: 1