API Reference#
- torchrl.collectors package
- Checkpointing
- Basic usage
- State-dict payload formats
- Custom components
- Reading a component without its object
- Checkpoint rotation
- Trainer integration
- Stopping at a safe boundary
- Resuming recipes
- Compatibility
- API
- Checkpoint
- CheckpointAdapter
- CheckpointError
- CheckpointLoadResult
- CheckpointOptions
- CheckpointRotation
- CheckpointFormat
- CheckpointStrictness
- DumpLoadCheckpointAdapter
- GlobalRNGState
- JSONCheckpointAdapter
- RunCheckpointer
- StateDictCheckpointAdapter
- StateDictFormat
- StopOnSignal
- resolve_checkpoint_path
- resume_config
- torchrl.data package
- Data layout: contiguous trajectories
- Trajectory boundary keys
- Trajectory boundaries: recovering episodes from storage
- The replay buffer
ndimarg and why it doesn’t multi-process well - The buffer-to-collector handoff: complete-trajectory writes
- SliceSampler: variable-length contiguous slices
- Auto-discoverability for recurrent policies
- Legacy:
split_trajectories - Narrow canonicalization for recurrent inputs
- See also
- torchrl.envs package
- LLM Interface
- torchrl.modules package
- torchrl.objectives package
- Rendering applications
- Core API
- render_policy
- make_render_env
- load_render_policy
- collect_render_rollouts
- write_render_artifact
- import_from_string
- call_with_supported_kwargs
- load_checkpoint
- save_render_checkpoint
- checkpoint_hash
- infer_state_dict
- parse_nested_key
- key_to_string
- normalize_policy
- add_step_counter
- seed_env
- normalize_env
- write_mujoco_wasm_viewer
- display_mujoco_wasm_viewer
- send_mujoco_wasm_qpos
- play_mujoco_wasm_trajectory
- extract_qpos_trajectory
- Configuration and results
- Backends
- Lower-level helpers
- Core API
- Service Registry
- Designing Training Applications with Services
- Scoped backend defaults
- Owners and clients
- Placement does not define communication
- Choosing a payload transport
- Preserving domain APIs
- Completion and failure semantics
- Lifecycle belongs to the owner
- Integrations accept owners when they can
- Environments are execution resources, not shared services
- Discovery is optional
- Design compromises
- Runnable examples
- Distributed transport implementation notes
- torchrl.trainers package
- torchrl._utils package
- Vision-Language-Action (VLA)
- TorchRL Configuration System
- Quick Start with a Simple Example
- Configuration Categories and Groups
- More Complex Example: Parallel Environment with Transforms
- Getting Available Options
- Complete Training Example
- Running Experiments
- Configuration Store Implementation Details
- Available Configuration Classes
- Base Classes
- Environment Configurations
- Environment Library Configurations
- Model and Network Configurations
- Transform Configurations
- Data Collection Configurations
- Replay Buffer and Storage Configurations
- Training and Optimization Configurations
- Trainer Hook Configurations
- Logging Configurations
- Checkpoint Configurations
- Resuming a recipe
- Creating Custom Configurations
- Best Practices
- Supported Algorithms
- Profiling collectors and envs
- Glossary