ray
Use Ray for distributed Python applications, clusters, Data pipelines, Train/Tune experiments, Serve deployments, and RLlib workloads.
SKILL.mdskills/repositories/repo-skills/ray/SKILL.md
GitHubRay
Use this repo skill when a task is about Ray, ray, Ray Core, Ray clusters, Ray Data, Ray Train, Ray Tune, Ray Serve, or RLlib. Ray is a distributed Python runtime plus AI libraries for scalable data processing, training, tuning, serving, and reinforcement learning.
Start Here
Confirm the user’s workflow and install surface. Prefer narrow extras such as
ray[default],ray[data],ray[train],ray[tune],ray[serve], orray[rllib]; avoidray[all]unless the user explicitly needs many unrelated Ray libraries.Run the root environment checker when package state is unclear:
python scripts/check_ray_environment.py --help python scripts/check_ray_environment.py --require core --require data --jsonRoute to the focused sub-skill before writing detailed code, commands, configs, or troubleshooting steps.
For cluster-mutating commands such as
ray start,ray stop,ray up,ray down,ray job submit, orserve deploy, confirm the target cluster/address and side effects first.
Route Map
| User task | Read |
|---|---|
Use ray.remote, tasks, actors, ObjectRef, ray.get, ray.wait, resource annotations, object store behavior, or Core anti-patterns |
sub-skills/core-runtime/SKILL.md |
Operate clusters with ray CLI, Jobs, dashboard, runtime environments, ray status, ray list, logs, state APIs, or OOM observability |
sub-skills/cluster-ops/SKILL.md |
Build ray.data pipelines, load local/cloud data, transform rows/batches, write datasets, tune blocks/concurrency, or debug Data schemas/OOM |
sub-skills/data-pipelines/SKILL.md |
Use Ray Train or Tune for distributed training, checkpoints, storage, Tuner, TuneConfig, schedulers, resources, or ResultGrid |
sub-skills/train-tune/SKILL.md |
Build, run, deploy, lint, update, or troubleshoot Ray Serve apps, @serve.deployment, Serve YAML, serve run, or serve deploy |
sub-skills/serve-deployments/SKILL.md |
Configure RLlib AlgorithmConfig/PPOConfig, Gymnasium envs, EnvRunners, Learners, RLlib checkpoints, or RLlib Tune sweeps |
sub-skills/rllib-workloads/SKILL.md |
Cross-Workflow Rules
- Use
core-runtimefor Python runtime semantics even when a higher-level library starts Ray under the hood. - Use
cluster-opsfor Jobs, dashboard, state, logs, and runtime environment packaging around any Ray application. - Use
data-pipelinesfor dataset construction before handing training/tuning totrain-tuneor model serving toserve-deployments. - Use
train-tunefor generic Tune search spaces, schedulers, resources, result analysis, and storage; userllib-workloadsonly for RLlib-specific config/env/algorithm work. - Use
serve-deploymentsfor Serve application configs and lifecycle, but route generic cluster startup, dashboard, and node visibility back tocluster-ops.
Installation And Import Checks
Ray’s Python package is distributed as ray and imports as ray. The source baseline for this skill supports Python >=3.10.
pip install ray
python - <<'PY'
import ray
print(ray.__version__)
PY
Choose extras by workflow:
pip install "ray[default]" # cluster dashboard, jobs, state APIs
pip install "ray[data]" # Ray Data readers/transforms
pip install "ray[train]" # Ray Train
pip install "ray[tune]" # Ray Tune
pip install "ray[serve]" # Ray Serve
pip install "ray[rllib]" torch # RLlib with common PyTorch backend
Shared References
- Read
references/troubleshooting.mdfor cross-cutting install/import, optional extras, Python version, cluster command safety, and workflow routing failures. - Read
references/repo-provenance.mdwhen checking whether this skill is stale relative to a Ray checkout or package version. references/repo-routing-metadata.jsonis structured import metadata consumed byrepo-skills-routerduring managed import.
Bundled Root Script
scripts/check_ray_environment.pychecks package metadata, selected imports, optional CLI commands, and common workflow extras without starting a Ray cluster.
Out Of Default Scope
This skill is Python-focused. Java, C++, LLM extra-extra workflows, heavy benchmarks, release engineering, CI/build infrastructure, cloud credential provisioning, Kubernetes operator deployment authoring, and GPU/service-specific notebooks are recorded as gaps unless the user explicitly asks for those areas and provides the needed environment or platform constraints.