Upstream: ggml-org#25510
Malformed client-supplied tool_calls[].function.arguments in replayed chat history is treated as an internal server failure instead of a client error.
Environment: Linux aarch64, NVIDIA GB10 / CUDA 13, current upstream df03399b885831b2a1603b3abb0d8c156808e363, Qwen2.5 0.5B Instruct Q4_K_M.
llama-server -m qwen2.5-0.5b-instruct-q4_k_m.gguf -ngl 99 --jinja -c 2048 -b 256
Send POST /v1/chat/completions with a prior assistant message containing:
{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"echo","arguments":"{"}}]}Observed: HTTP 500 with Failed to parse tool call arguments as JSON. Repeating the malformed history repeats the 500.
Likely path: func_args_not_string() in common/chat.cpp converts the parse error to std::runtime_error, which the server maps to HTTP 500.
Upstream: ggml-org#25510
Malformed client-supplied
tool_calls[].function.argumentsin replayed chat history is treated as an internal server failure instead of a client error.Environment: Linux aarch64, NVIDIA GB10 / CUDA 13, current upstream
df03399b885831b2a1603b3abb0d8c156808e363, Qwen2.5 0.5B Instruct Q4_K_M.Send
POST /v1/chat/completionswith a prior assistant message containing:{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"echo","arguments":"{"}}]}Observed: HTTP 500 with
Failed to parse tool call arguments as JSON. Repeating the malformed history repeats the 500.Likely path:
func_args_not_string()incommon/chat.cppconverts the parse error tostd::runtime_error, which the server maps to HTTP 500.