Currently, mlc-serve takes prompts in two paths: request.mesasges and request.debug_options.prompt.
https://github.com/octoml/mlc-llm/blob/batch-serving/serve/mlc_serve/engine/engine_common.py#L43

The expected behavior is not clear until you deep-dive the codebase.
Better way can be have a flag to disable the application of chat template.