api-pages.yaml

# API Reference page configuration for Fumadocs
# This is a custom configuration format modelled after `mkdocstrings`.
- path: skyrl
  pages:
    - slug: backends
      title: Backends
      description: Backend abstractions for training and inference.
      sections:
        - heading: Abstract Backend
          description: The base class all backends implement.
          objects:
            - skyrl.backends.backend.AbstractBackend
        - heading: JAX Backend
          description: JAX-based backend for training and inference.
          objects:
            - skyrl.backends.jax.JaxBackend
        - heading: SkyRL-Train Backend
          description: Backend wrapping the SkyRL-Train training pipeline.
          objects:
            - skyrl.backends.skyrl_train_backend.SkyRLTrainBackend

    - slug: tinker-engine
      title: Tinker Engine
      description: Orchestration engine for RL training.
      sections:
        - heading: Engine
          description: The TinkerEngine orchestrates backends, models, and training loops.
          objects:
            - skyrl.tinker.engine.TinkerEngine

    - slug: types
      title: Types
      description: Request/response types for the Tinker API.
      sections:
        - heading: Enums
          description: ""
          objects:
            - skyrl.tinker.types.RequestType
            - skyrl.tinker.types.CheckpointType
        - heading: Configuration
          description: ""
          objects:
            - skyrl.tinker.types.TinkerPath
            - skyrl.tinker.types.AdamParams
            - skyrl.tinker.types.LoraConfig
            - skyrl.tinker.types.SamplingParams
            - skyrl.tinker.types.ModelMetadata
        - heading: Model Lifecycle
          description: ""
          objects:
            - skyrl.tinker.types.CreateModelInput
            - skyrl.tinker.types.CreateModelOutput
            - skyrl.tinker.types.UnloadModelInput
            - skyrl.tinker.types.UnloadModelOutput
        - heading: Forward / Backward
          description: ""
          objects:
            - skyrl.tinker.types.ModelInputChunk
            - skyrl.tinker.types.ModelInput
            - skyrl.tinker.types.TensorData
            - skyrl.tinker.types.LossFnInputs
            - skyrl.tinker.types.Datum
            - skyrl.tinker.types.ForwardBackwardInput
            - skyrl.tinker.types.ForwardBackwardOutput
        - heading: Optimization
          description: ""
          objects:
            - skyrl.tinker.types.OptimStepInput
            - skyrl.tinker.types.OptimStepOutput
        - heading: Checkpointing
          description: ""
          objects:
            - skyrl.tinker.types.SaveWeightsForSamplerInput
            - skyrl.tinker.types.SaveWeightsForSamplerOutput
            - skyrl.tinker.types.SaveWeightsInput
            - skyrl.tinker.types.SaveWeightsOutput
            - skyrl.tinker.types.LoadWeightsInput
            - skyrl.tinker.types.LoadWeightsOutput
        - heading: Sampling
          description: ""
          objects:
            - skyrl.tinker.types.SampleInput
            - skyrl.tinker.types.GeneratedSequence
            - skyrl.tinker.types.SampleOutput
        - heading: Engine Types
          description: ""
          objects:
            - skyrl.tinker.types.ErrorResponse
            - skyrl.tinker.types.EngineMetrics
            - skyrl.tinker.types.PreparedModelPassBatch
            - skyrl.tinker.types.PreparedSampleBatch

    - slug: tx-models
      title: TX Models
      description: Model loading and configuration for the JAX backend.
      sections:
        - heading: Model Configuration
          description: ""
          objects:
            - skyrl.tx.models.configs.ModelConfig
        - heading: Model Interface
          description: ""
          objects:
            - skyrl.tx.models.types.ModelForCausalLM
            - skyrl.tx.models.types.CausalLMOutput

    - slug: entrypoints
      title: Entrypoint
      description: Entrypoint API — BasePPOExp, EvalOnlyEntrypoint.
      sections:
        - heading: Training Entrypoint
          description: The main entrypoint is the `BasePPOExp` class which runs the main training loop.
          objects:
            - skyrl.train.entrypoints.main_base.BasePPOExp
        - heading: Evaluation Entrypoint
          description: The evaluation-only entrypoint is the `EvalOnlyEntrypoint` class which runs evaluation without training.
          objects:
            - skyrl.train.entrypoints.main_generate.EvalOnlyEntrypoint

    - slug: config
      title: Configuration
      description: Configuration dataclasses for SkyRL training.
      sections:
        - heading: Top-Level Config
          description: The root configuration object and helpers.
          objects:
            - skyrl.train.config.config.SkyRLTrainConfig
            - skyrl.train.config.config.make_config
            - skyrl.train.config.config.get_config_as_dict
            - skyrl.train.config.config.get_config_as_yaml_str
        - heading: Data & Model
          description: ""
          objects:
            - skyrl.train.config.config.DataConfig
            - skyrl.train.config.config.DataLoaderConfig
            - skyrl.train.config.config.ModelConfig
            - skyrl.train.config.config.SkyRLLoraConfig
        - heading: Training
          description: ""
          objects:
            - skyrl.train.config.config.TrainerConfig
            - skyrl.train.config.config.OptimizerConfig
            - skyrl.train.config.config.MixedPrecisionConfig
            - skyrl.train.config.config.FakeInt4QatConfig
            - skyrl.train.config.config.MTPConfig
        - heading: Backend Config
          description: ""
          objects:
            - skyrl.train.config.config.FSDPConfig
            - skyrl.train.config.config.MegatronConfig
            - skyrl.train.config.config.MegatronDDPConfig
            - skyrl.train.config.config.MegatronLoraConfig
            - skyrl.train.config.config.MegatronHFExportConfig
            - skyrl.train.config.config.TorchProfilerConfig
        - heading: Placement
          description: ""
          objects:
            - skyrl.train.config.config.PlacementConfig
        - heading: Policy & Algorithm
          description: ""
          objects:
            - skyrl.train.config.config.PolicyConfig
            - skyrl.train.config.config.CriticConfig
            - skyrl.train.config.config.RefConfig
            - skyrl.train.config.config.AlgorithmConfig
            - skyrl.train.config.config.KLCtrlConfig
        - heading: Algorithm Extensions
          description: ""
          objects:
            - skyrl.train.config.config.SAPOConfig
            - skyrl.train.config.config.DynamicSamplingConfig
            - skyrl.train.config.config.ClipCovConfig
            - skyrl.train.config.config.KLCovConfig
            - skyrl.train.config.config.CISPOConfig
            - skyrl.train.config.config.DPPOConfig
            - skyrl.train.config.config.OffPolicyCorrectionConfig
            - skyrl.train.config.config.FullyAsyncConfig
        - heading: Inference & Generation
          description: ""
          objects:
            - skyrl.train.config.config.SamplingParams
            - skyrl.train.config.config.ChatTemplateConfig
            - skyrl.train.config.config.InferenceEngineConfig
            - skyrl.train.config.config.DeltaWeightSyncConfig
            - skyrl.train.config.config.GeneratorConfig
        - heading: Environment
          description: ""
          objects:
            - skyrl.train.config.config.EnvironmentConfig
            - skyrl.train.config.config.SkyRLGymConfig
            - skyrl.train.config.config.GSM8kLLMJudgeEnvConfig

    - slug: callbacks
      title: Callbacks
      description: Training callbacks for the SFT and RL trainers.
      sections:
        - heading: Callback Interface
          description: Subclass `TrainingCallback` and override the events you need.
          objects:
            - skyrl.train.utils.callbacks.TrainingCallback
        - heading: Callback I/O
          description: State passed to each event and the mutable control flags callbacks can set.
          objects:
            - skyrl.train.utils.callbacks.CallbackInput
            - skyrl.train.utils.callbacks.TrainingControl
        - heading: Dispatch
          description: Fan-out handler that trainers hold internally.
          objects:
            - skyrl.train.utils.callbacks.CallbackHandler

    - slug: env-vars
      title: Environment Variables
      description: Configuration via environment variables.
      sections:
        - heading: null
          description: ""
          objects:
            - skyrl.env_vars

- path: skyrl/sft
  pages:
    - slug: sft_config
      title: Configuration
      description: Supervised Fine-Tuning configuration.
      sections:
        - heading: Configuration
          description: ""
          objects:
            - skyrl.train.config.sft_config.SFTPlacementConfig
            - skyrl.train.config.sft_config.SFTConfig
            - skyrl.train.config.sft_config.TrainOnWhat
        - heading: Config Bridge
          description: ""
          objects:
            - skyrl.train.config.sft_config.validate_sft_cfg
            - skyrl.train.config.sft_config.build_skyrl_config_for_sft
      
    - slug: sft_trainer
      title: SFT Trainer
      description: Supervised Fine-Tuning Trainer.
      sections:
        - heading: Trainer
          description: ""
          objects:
            - skyrl.train.sft_trainer.SFTTrainer
    
    - slug: sft_samplers
      title: Samplers
      description: Supervised Fine-Tuning Samplers.
      sections:
        - heading: Samplers
          description: ""
          objects:
            - skyrl.train.dataset.samplers.StatefulSequentialSampler
            - skyrl.train.dataset.samplers.DataMixingSampler
        - heading: Sampler Utilities
          description: ""
          objects:
            - skyrl.train.dataset.samplers.import_sampler_class

- path: skyrl/skyrl-train
  pages:
    - slug: trainer
      title: Trainer
      description: Trainer API — RayPPOTrainer, Dispatch, Worker APIs.
      sections:
        - heading: Trainer Class
          description: ""
          objects:
            - skyrl.train.trainer.RayPPOTrainer
        - heading: Dispatch APIs
          description: ""
          objects:
            - skyrl.backends.skyrl_train.distributed.dispatch.Dispatch
            - skyrl.backends.skyrl_train.distributed.dispatch.MeshDispatch
            - skyrl.backends.skyrl_train.distributed.dispatch.PassThroughDispatch
        - heading: Worker APIs
          description: The base worker abstraction in SkyRL.
          objects:
            - skyrl.backends.skyrl_train.workers.worker.Worker
            - skyrl.backends.skyrl_train.workers.worker.PPORayActorGroup

    - slug: utils
      title: Utils
      description: Utility classes for the SkyRL-Train backend.
      sections:
        - heading: Placement Group
          description: ""
          objects:
            - skyrl.train.utils.utils.ResolvedPlacementGroup

    - slug: data
      title: Data Interface
      description: Data Interface — TensorBatch, TrainingInput, GeneratorInput/Output.
      sections:
        - heading: Trainer APIs
          description: ""
          objects:
            - skyrl.backends.skyrl_train.training_batch.TensorBatch
            - skyrl.backends.skyrl_train.training_batch.TrainingInput
            - skyrl.backends.skyrl_train.training_batch.TrainingInputBatch
            - skyrl.backends.skyrl_train.training_batch.TrainingOutputBatch
        - heading: Generator APIs
          description: ""
          objects:
            - skyrl.train.generators.base.GeneratorInput
            - skyrl.train.generators.base.GeneratorOutput
            - skyrl.train.generators.base.MetricsOutput

    - slug: generators
      title: Generator
      description: Generator API — GeneratorInterface, InferenceEngineInterface.
      sections:
        - heading: Core APIs
          description: ""
          objects:
            - skyrl.train.generators.base.GeneratorInterface
            - skyrl.backends.skyrl_train.inference_servers.base.InferenceEngineInterface
            - skyrl.backends.skyrl_train.inference_servers.remote_inference_client.RemoteInferenceClient

    - slug: registry
      title: Algorithm Registry
      description: Algorithm Registry API — Advantage estimators, policy loss registries.
      sections:
        - heading: Base Registry Classes
          description: The registry system provides a way to register and manage custom algorithm functions across distributed Ray environments.
          objects:
            - skyrl.backends.skyrl_train.utils.ppo_utils.BaseFunctionRegistry
            - skyrl.backends.skyrl_train.utils.ppo_utils.RegistryActor
            - skyrl.backends.skyrl_train.utils.ppo_utils.sync_registries
        - heading: Advantage Estimator Registry
          description: The advantage estimator registry manages functions that compute advantages and returns.
          objects:
            - skyrl.backends.skyrl_train.utils.ppo_utils.AdvantageEstimatorRegistry
            - skyrl.backends.skyrl_train.utils.ppo_utils.AdvantageEstimator
            - skyrl.backends.skyrl_train.utils.ppo_utils.register_advantage_estimator
        - heading: Policy Loss Registry
          description: The policy loss registry manages functions that compute policy losses for PPO.
          objects:
            - skyrl.backends.skyrl_train.utils.ppo_utils.PolicyLossRegistry
            - skyrl.backends.skyrl_train.utils.ppo_utils.PolicyLossType
            - skyrl.backends.skyrl_train.utils.ppo_utils.register_policy_loss

- path: skyrl-gym
  pages:
    - slug: environment
      title: Environment
      description: Environment API — Env, BaseTextEnv, step outputs.
      sections:
        - heading: Core Classes
          description: ""
          objects:
            - skyrl_gym.core.Env
            - skyrl_gym.core.EnvStepOutput
        - heading: Text Environment
          description: ""
          objects:
            - skyrl_gym.envs.base_text_env.BaseTextEnv
            - skyrl_gym.envs.base_text_env.BaseTextEnvStepOutput

    - slug: tools
      title: Tools
      description: Tools API — ToolGroup, tool decorator.
      sections:
        - heading: ToolGroup
          description: ""
          objects:
            - skyrl_gym.tools.core.ToolGroup
        - heading: Tool Decorator
          description: ""
          objects:
            - skyrl_gym.tools.core.tool