---

name: qwen-vl description: "Route Qwen-VL multimodal inference, serving, finetuning, and official evaluation workflows." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Qwen-VL

This generated repo skill is the router for Qwen-VL-family workflows. Start here when you are not yet sure whether the request is about direct inference, serving, finetuning, or evaluation. The detailed workflow notes live in the sub-skills and bundled references.

Read first

• Repository provenance to see the source snapshot, dirty-state note, and evidence paths used to build this skill.
• Model and workflow overview for the model-family map and the fastest route to the right sub-skill.
• Installation for the runtime package groups and a safe smoke check command.
• Troubleshooting for cross-cutting install, dependency, and backend issues.
• Runtime smoke helper when you want a quick import or CUDA probe before launching a larger workflow.

Route by intent

• Use sub-skills/inference/SKILL.md for direct chat, grounding, box rendering, generation, and quantization-aware loading.
• Use sub-skills/serving/SKILL.md for the Gradio demo or the OpenAI-compatible FastAPI service.
• Use sub-skills/finetuning/SKILL.md for full finetuning, LoRA, Q-LoRA, and conversation-data validation.
• Use sub-skills/evaluation/SKILL.md for captioning, VQA, grounding, ScienceQA, MMBench, SEED-Bench, MME reference notes, and TouchStone reference notes.

What this skill covers

• Qwen-VL, Qwen-VL-Chat, and Qwen-VL-Chat-Int4 checkpoint selection.
• Multimodal prompts with one or more images and optional grounding markup.
• Local service launch planning without auto-starting listeners.
• Adapter training and checkpoint/adapter preparation.
• Official benchmark conversion, inference, scoring, and submission formatting.

Operating notes

1. Use trust_remote_code=True unless you have an audited local copy of the custom model code.
2. Use Qwen/Qwen-VL-Chat for assistant-style chat and Qwen/Qwen-VL for base-model generation.
3. Use the bundled sub-skill scripts and references rather than depending on the original repository checkout.
4. Keep service launch, training, and benchmark execution separate unless the user explicitly asks for a multi-step workflow.
5. If a request mentions box, grounding, ref, or quad, start with inference; if it mentions port, share, or OpenAI-compatible, start with serving; if it mentions LoRA, Q-LoRA, or DeepSpeed, start with finetuning; if it mentions dataset names or benchmark outputs, start with evaluation.

Quick start

• For a quick environment sanity check, run python scripts/runtime_smoke.py --check-cuda.
• For a quick command plan, open the sub-skill that matches the user request and then read the bundled reference linked there.
• For a small source-free orientation, use references/model-and-workflow-overview.md before opening the deeper workflow docs.

Not included here

• Long API tables, benchmark file layouts, and command catalogs live in the sub-skill references.
• Review artifacts, usability test cases, and verification notes live under skills/tests/qwen-vl/, not in this runtime skill.