Goal
Deliver a V0 LLM platform on LiftWing that volunteers can experiment with during Wikimania. Success means at least one general-purpose model is reliably served and accessible (via both the public API and Lift Wing Studio), with enough documentation for someone to go from zero to a successful request, plus a performance baseline and basic operational dashboards.
Scope
- Deploy Qwen3-14B FP8 (including FP8 KV cache)
- Deploy Qwen3-27B FP8 (including FP8 KV cache)
- Establish benchmarking for the V0 platform (vllm bench serve; locust as a stretch goal)
- Create basic Grafana dashboards for GPU and vLLM metrics
- Publish V0 LLM documentation on Wikitech
Current state
- Qwen 3.6-27B already deployed (llm-qwen36-27b, T425680).
- Qwen3-14B already on LiftWing (bf16), exposed via REST Gateway, available in LiftWing Studio.
- LiftWingLLM rate-limit policy live for llm-* services (T426749).
- vLLM metrics are emitted by all LLM inference services (T431136); an initial observability dashboard is already in progress (T429597#12094788).
Non-goals
- Multi-GPU inference (e.g. tensor parallelism, distributed inference)
- Advanced tooling (e.g. web search, RAG)
Documentation lives at: https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing/Large_Language_Models