Qwen3 Coder 30B 的本地函数调用不靠谱,以及如何解决此问题

At least for the 30B model (in FP8), I found the native function calling with the default chat template extremely unreliable, because the model would frequently omit the initial <tool_call> tag on tool calls, especially when the tool call was attempted after a textual response. I tried a lot of variations to the function calling instructions in the chat template, until I arrived at something that *actually* works reliably: