API Reference

SkyRL

• Backends — AbstractBackend, JaxBackend, SkyRLTrainBackend
• Tinker Engine — Orchestration engine for RL training
• Types — Request/response types for the Tinker API
• TX Models — ModelConfig, ModelForCausalLM, CausalLMOutput
• Entrypoint — BasePPOExp, EvalOnlyEntrypoint
• Configuration — Configuration dataclasses for SkyRL training
• Environment Variables — Configuration via environment variables
• SkyRL-Train Backend:
  • Data Interface — TensorBatch, TrainingInput, GeneratorInput, GeneratorOutput
  • Generator — GeneratorInterface, InferenceEngineInterface
  • Trainer — RayPPOTrainer, Dispatch, Worker, PPORayActorGroup
  • Algorithm Registry — Advantage estimators, policy loss registries

SkyRL-Gym

• Environment — Env, BaseTextEnv, EnvStepOutput
• Tools — ToolGroup, tool