Setup:
∙ DGX Spark (GB10, SM12.1)
∙ vLLM 0.17.2rc1 (pinned — 0.18.x breaks NVFP4 on SM12.1)
∙ --tool-call-parser qwen3_coder
∙ --enable-auto-tool-choice
∙ Open WebUI as frontend
Problem:
Tool calls don’t execute — the model outputs raw JSON instead of a proper function call that vLLM intercepts. Example output:
{ “tool”: “market-prices”, “arguments”: { “symbol”: “AAPL” } }
The model’s own thinking block shows it deciding not to use tools (“we should say we can’t”), suggesting the tool schema isn’t being properly injected or parsed.
What I know:
∙ The tool call fix for qwen3_coder was merged in vLLM 0.18.0 (per the NVIDIA improvements thread)
∙ 0.18.x currently crashes on SM12.1 with MergedColumnParallelLinear has no attribute ‘workspace’
∙ So I’m stuck on 0.17.2rc1 without the fix
Question:
Is there a backport or workaround to get tool calling working on 0.17.2rc1? Or is there a patched 0.18.x build that works on SM12.1?