LLM workloads on open-source Ray (pinned to 2.57) — OpenAI-compatible serving with ray.serve.llm (vLLM-backed LLMConfig + build_openai_app) and batch inference with ray.data.llm (build_processor). Corrects the stale defaults a model produces — the archived ray-llm repo and its YAML configs, hand-rolled vLLM engines inside plain Serve deployments, the removed build_llm_processor name, deprecated boolean stage flags, top-level LLMServer/LLMRouter imports, free-form accelerator strings, one-deployment-per-LoRA-adapter designs — with the 2.57 idioms (stage configs, placement_group_config over hand-rolled PGs, deployment_config autoscaling, dynamic LoRA multiplexing, prefix-cache-affinity routing, the full OpenAI endpoint surface). Use when writing, reviewing, or productionizing LLM serving or batch inference on Ray. Classic-ML Ray (Train/Tune/Data/Serve/clusters) lives in the sibling ray skill.
Production Ray (open-source, pinned to 2.57) for classic-ML workloads from training to serving — Ray Train, Tune, Data, Serve, Core, and cluster deployment on KubeRay. Corrects the older-corpus defaults a model reaches for (ray.air session reporting, Trainer-inside-Tuner, tune.run, map_batches concurrency=, DatasetPipeline/to_torch, max_concurrent_queries, RayServeHandle + ray.get, Deployment.deploy, ray.state, ray.get-in-a-loop) with the 2.57 idioms that replaced them (Train V2 defaults, driver-function tuning, compute strategies, streaming datasets, DeploymentHandle/DeploymentResponse, serve build/deploy, ray.util.state, KubeRay CRDs and Jobs API). Use when writing, reviewing, or productionizing Python code that touches Ray distributed training, data pipelines, hyperparameter tuning, model serving, or Ray cluster operations. LLM serving/batch-inference on Ray lives in the sibling ray-llm skill.
MLflow 3 (open-source, pinned to 3.15) for classic-ML MLOps — logging and registering models, promoting versions across dev/staging/prod, standing up a tracking server, evaluating with gates, and serving. Corrects the MLflow 2-era defaults a model reaches for (artifact_path, registry stages and get_latest_versions, top-level mlflow.evaluate with baseline_model, runs-URI registration, pickle serialization, mlruns file stores, MLServer serving) with the MLflow 3 idioms that replaced them (named LoggedModels, aliases and copy_model_version, models.evaluate plus validate_evaluation_results, skops/torch.export defaults, database backends, the FastAPI scoring server). Use when writing, reviewing, or migrating Python code that touches MLflow tracking, the model registry, evaluation, or serving.
