Issue evidence · Serve · ggml-org/llama.cpp

LFM2.5-24B-A2B emits Qwen3-Coder-style <tool_call> XML as content: calls silently lost

GitHub issue: https://github.com/ggml-org/llama.cpp/issues/30490

Not reproduced

Intermittently, LFM2.5-24B-A2B returns a tool call as <tool_call> XML in message content with empty tool_calls.

What Badgr ran

Served LiquidAI/LFM2.5-24B-A2B-GGUF Q4_0 with ghcr.io/ggml-org/llama.cpp:server-cuda (build b11610) on an RTX 4090 through Badgr with --jinja -fa on -c 16384, and sent 60 chat requests with two tools (edit, read_file) at temperature 0.7, 6 at a time.

What came back

  • 60 of 60 returned structured tool_calls; 0 had <tool_call> XML in content.

The reporter is on ROCm (gfx1100, WSL2), a 131072 context and an earlier build, and sees it only intermittently, so 60 clean runs do not rule it out.

Checked 2026-10-14 in Badgr’s local development environment.

← All issue evidence