Broken Tool Calling?
Tool calling doesn't seem to work properly, maybe there is a template issue? Has anyone got this working?
Hey @And1mon I personnaly have no issue with thinking, it is reliably working. However, I can also observe no tool calling is working, at least with the llama-server webui (for this test, I used --tools all and tried to make it use get_datetime) or openwebui (trying to use get_current_timestamp).
In both case, it correctly sees the list of tools available, but can't call them correctly and output for example name="get_current_timestamp">
`
Example:
[
{
"type": "reasoning",
"id": "r_817333d3d4a7488e8b79a7ae",
"status": "completed",
"start_tag": "<think>",
"end_tag": "</think>",
"attributes": {
"type": "reasoning_content"
},
"content": [
{
"type": "output_text",
"text": "Okay, the user is asking me to demonstrate the get_current_timestamp tool. First, I need to check if I have the right function. Looking back at the tools provided, there's get_current_timestamp with no parameters, and calculate_timestamp with parameters days_ago, weeks_ago, etc.\n\nThe user wants to demonstrate using get_current_timestamp, so I should call that function. The parameters are all null in get_current_timestamp, but the description says it gets the current Unix timestamp in seconds. So I need to make sure to call it without any arguments.\n\nWait, the parameters for get_current_timestamp are listed as {\"properties\": {}, \"type\": \"object\"} so no parameters. So the tool call should be name \"get_current_timestamp\" with empty arguments.\n\nThe user said \"Can you use get_current_timestamp to demonstrate?\" so I should generate a tool call for that function. Since the instructions say to return the tool call in JSON within tool_call tags.\n\nI need to make sure I don't miss any steps. The user might want to see the actual timestamp, but since they didn't provide a time range, maybe the tool is called correctly by default. The calculate_timestamp tool can also be used, but the user is asking specifically about demonstrating get_current_timestamp.\n\nSo I should output the tool call for get_current_timestamp. Let me check the example structure. The example shows \"function name arguments\". Since get_current_timestamp has no parameters, the arguments should be empty object.\n\nSo the correct tool call is {\"name\": \"get_current_timestamp\", \"arguments\": {}}\n\nWait, in the parameters, the get_current_timestamp has \"properties\": {}, \"type\": \"object\", so no parameters. So the arguments should be an empty object.\n\nI should format the tool call as:\n\n{\"name\": \"get_current_timestamp\", \"arguments\": {}}\n\nYes, that's correct.\n\nI need to make sure not to call any other tools. The user only asked about demonstrating get_current_timestamp, so only that function.\n\nSo the response should be to generate that tool call.\n"
}
],
"summary": null,
"started_at": 1779800689.0132089,
"ended_at": 1779800696.767792,
"duration": 7
},
{
"type": "message",
"id": "msg_224367e09dd3464ea549ed1b",
"status": "completed",
"role": "assistant",
"content": [
{
"type": "output_text",
"text": "name=\"get_current_timestamp\">"
}
]
}
]
Yeah the thinking issue was user error on my side, however tool calling appears broken. I think it outputs a tool calling format llama.cpp cannot interpret.
Tool Calling
For tool / function calling, SGLang is the recommended backend. MiniCPM5-1B emits XML-style tool calls and SGLang's built-in minicpm5 parser converts them to OpenAI-compatible tool_calls natively:
python -m sglang.launch_server --model-path openbmb/MiniCPM5-1B --port 30000
--tool-call-parser minicpm5 # or: --tool-call-parser auto
Thanks for raising this!
For llama.cpp, chat completions are supported today. For tool calling, we recommend SGLang or vLLM, which handle MiniCPM5’s tool call format natively (e.g. SGLang with --tool-call-parser minicpm5).