Ray LLM

Library-reference skill for LLM workloads on open-source Ray — 13 rules across 5 categories covering ray.serve.llm (OpenAI-compatible, vLLM-backed serving) and ray.data.llm (batch inference). This surface churned faster than any other part of Ray — a standalone repo was absorbed and archived, entry points were renamed, and config shapes restructured — so the examples a model learned from mostly no longer run. Each rule names the wrong default it corrects; there is no rule for things a capable model already gets right.

Scope is the LLM-specific layer. Generic Serve/Data/cluster decisions (deployment lifecycle, autoscaling semantics, KubeRay) are the sibling ray skill — the two compose.

Pinned to ray 2.57.0 (ray[llm] extra, which pins its matching vLLM). API claims were verified against the unpacked 2.57.0 wheel and the installed package source, and every config example in the rules was constructed under CPU-only pydantic validation (including the traps, which fail exactly as described); engine/GPU runtime behavior is source-verified only — no model was actually served.

When to Apply

• Standing up or reviewing an OpenAI-compatible LLM serving deployment on Ray
• Writing batch LLM inference over datasets — summarization, embedding, scoring at scale
• Sizing or placing multi-GPU models — tensor/pipeline parallelism, accelerator selection
• Scaling LLM deployments — replica autoscaling, ingress sizing, request routing
• Serving families of LoRA fine-tunes of a shared base model
• Migrating code that uses the archived ray-llm repo, hand-rolled vLLM engines, or pre-2.5x ray.data.llm names

Rule Categories

Quick Reference

1. Serving Setup

• serve-builtin-not-handrolled — LLMConfig + build_openai_app; the ray-llm repo is archived, hand-rolled engines re-implement less
• serve-model-id-vs-source — model_id is the client-facing name; model_source is where weights live
• serve-ray-llm-extra-pins-vllm — ray[llm] pins its exact vLLM; don't mix independently chosen versions
• serve-deprecated-server-router — LLMServer/LLMRouter moved; the ingress class is OpenAiIngress (exact casing)

2. Batch Inference

• batch-build-processor-renamed — build_llm_processor is removed; build_processor + vLLMEngineProcessorConfig
• batch-stage-configs-not-flags — boolean apply_chat_template/tokenize/detokenize gave way to stage configs
• batch-concurrency-and-alternatives — accelerator_type=None means CPU; (min, max) concurrency; HTTP/Serve processors

3. Placement & Accelerators

• place-engine-builds-pg — the engine builds its TP×PP placement group; override strategy, don't hand-roll
• place-accelerator-type-validated — canonical accelerator constants; "A10" normalizes, most typos raise

4. Autoscaling & Routing

• scale-autoscaling-deployment-config — standard Serve autoscaling nested in deployment_config; replicas are whole engines
• scale-prefix-cache-routing — PrefixCacheAffinityRouter keeps same-prefix requests on warm KV caches

5. LoRA & API Surface

• api-lora-dynamic-multiplexing — dynamic adapter loading from cloud storage, not one deployment per adapter
• api-endpoint-surface-cpu-validation — embeddings/transcription/score/tokenize are served too; configs validate without GPUs

How to Use

Read a reference file when its decision comes up. Each rule names the wrong default it corrects, then shows the canonical way (with an incorrect/correct contrast only where the wrong way is a real trap).

• Section definitions — category structure
• Rule template — for adding new rules
• AGENTS.md — auto-built table of contents across all rules

Related Skills

• ray — the sibling rule pack for classic-ML Ray (Train, Tune, Data, Serve, Core, KubeRay production topology); generic Serve and cluster decisions live there

Reference Files