Malformed `system_prompt` in `/completions` request crashes server

[crclark]

Observed behavior

If I send a string for the system prompt instead of the expected json object, the server terminates.

Desired behavior

The server responds with an HTTP 400 error and doesn't terminate.

Environment

Running the server via docker:

docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp@sha256:b4675af8c9a8b3e7019a7baf536b95c3984a9aaacd0eafce7422377a299e31f4 -m /models/Meta-Llama-3-8B.Q4_K_M.gguf --port 8000 --host 0.0.0.0 -n 512

Using this gguf quant (though I don't think it matters -- it is crashing with different ggufs I have tried).

Request

json_body=$(cat <<EOF
{
  "system_prompt": "Always reply in markdown lists.",
  "prompt": "Building a website can be done in 10 simple steps:",
  "n_predict": 128
}
EOF
)

curl --request POST \
    --url http://localhost:8000/completion \
    --header "Content-Type: application/json" \
    --data "$json_body"

Server error logs

phantom-llama-1  | terminate called after throwing an instance of 'nlohmann::json_abi_v3_11_3::detail::type_error'
phantom-llama-1  |   what():  [json.exception.type_error.306] cannot use value() with string

[ngxson]

The bug should be in void system_prompt_set(const json & sys_props) where sys_props is a string instead of an object