sb3-extra-buffers
  • Home

User guide

  • Quickstart
    • Installation
    • Choosing compression algorithm
    • Choosing dtypes
    • Compressed rollout buffers
    • Full PPO Atari example
    • Compressed replay buffers
    • JIT warm-up
  • Supported Compression Algorithms
    • Implemented compression methods
    • Compression levels
    • Shorthand method names
    • Optional backends
  • Usage Notes
    • Input types
    • Compression algorithms
    • Buffer dtype settings
    • Recording buffers
    • Training utilities
    • Testing
    • Experimental helpers
  • Benchmarks
    • Benchmark setup
    • Summary
    • Results
    • Training validation
  • Training speed
    • Benchmark setup
      • Hardware
      • Libraries
    • PPO on PongNoFrameskip-v4
    • DQN on MsPacmanNoFrameskip-v4
    • See also
  • Validation
    • Training setup
    • Evaluation results for example training scripts
  • Citing This Project
  • API Reference
    • Compressed Buffers
      • CompressedRolloutBuffer
        • CompressedRolloutBuffer.observations
        • CompressedRolloutBuffer.actions
        • CompressedRolloutBuffer.rewards
        • CompressedRolloutBuffer.advantages
        • CompressedRolloutBuffer.returns
        • CompressedRolloutBuffer.episode_starts
        • CompressedRolloutBuffer.log_probs
        • CompressedRolloutBuffer.values
        • CompressedRolloutBuffer.reset()
        • CompressedRolloutBuffer.add()
        • CompressedRolloutBuffer.get()
        • CompressedRolloutBuffer.reconstruct_obs()
      • CompressedDictRolloutBuffer
        • CompressedDictRolloutBuffer.observation_space
        • CompressedDictRolloutBuffer.obs_shape
        • CompressedDictRolloutBuffer.observations
        • CompressedDictRolloutBuffer.reset()
        • CompressedDictRolloutBuffer.add()
        • CompressedDictRolloutBuffer.get()
      • CompressedReplayBuffer
        • CompressedReplayBuffer.observations
        • CompressedReplayBuffer.next_observations
        • CompressedReplayBuffer.actions
        • CompressedReplayBuffer.rewards
        • CompressedReplayBuffer.dones
        • CompressedReplayBuffer.timeouts
        • CompressedReplayBuffer.add()
        • CompressedReplayBuffer.reconstruct_obs()
        • CompressedReplayBuffer.reconstruct_nextobs()
      • CompressedDictReplayBuffer
        • CompressedDictReplayBuffer.observation_space
        • CompressedDictReplayBuffer.obs_shape
        • CompressedDictReplayBuffer.observations
        • CompressedDictReplayBuffer.add()
      • CompressedArray
      • DummyCls
      • find_smallest_dtype()
      • has_igzip()
      • has_numba()
      • init_jit()
      • find_buffer_dtypes()
      • Core classes
      • Helpers
      • Implementation modules
        • init_jit()
        • find_buffer_dtypes()
        • BaseCompressedBuffer
        • DummyCls
        • CompressedReplayBuffer
        • CompressedDictReplayBuffer
        • CompressedRolloutBuffer
        • CompressedDictRolloutBuffer
        • CompressedArray
        • find_optimal_shape()
        • find_smallest_dtype()
        • has_numba()
        • has_igzip()
        • has_zstd()
        • has_lz4()
    • Recording Buffers
      • BaseRecordBuffer
        • BaseRecordBuffer.frames
        • BaseRecordBuffer.features
        • BaseRecordBuffer.rewards
        • BaseRecordBuffer.actions
      • RecordBuffer
        • RecordBuffer.add()
        • RecordBuffer.add_filled()
      • FramelessRecordBuffer
        • FramelessRecordBuffer.add()
      • DummyRecordBuffer
        • DummyRecordBuffer.add()
      • Classes
      • Implementation modules
        • RecordBuffer
        • DummyRecordBuffer
        • FramelessRecordBuffer
        • DummyArray
        • BaseRecordBuffer
    • Training Utilities
      • Atari environments
        • make_env()
      • Evaluation
        • process_outcome()
        • eval_model()
      • Buffer warm-up
        • warm_up()
    • Vectorized Buffer Helpers
      • DummyVecRolloutBuffer
        • DummyVecRolloutBuffer.full
      • Implementation modules
        • DummyVecRolloutBuffer
    • GPU Buffers
      • Overview
        • GpuReplayBuffer
        • GpuRolloutBuffer
        • RawBuffer
        • SharedRawHeap
        • SlotMetadata
        • BaseGpuBuffer
        • find_gpu_buffer_dtypes()
        • has_zstd()
      • Metadata
        • SlotMetadata
        • arr_config_length()
      • Raw storage
        • write_at()
        • read_at()
        • RawBuffer
        • SharedRawHeap
      • Observation stores
        • DenseObservationStore
        • RawObservationStore
        • create_observation_store()
      • Base helpers
        • find_gpu_buffer_dtypes()
        • BaseGpuBuffer
      • Replay buffer
        • GpuReplayBuffer
      • Rollout buffer
        • GpuRolloutBuffer
      • Utilities
        • find_smallest_dtype()
        • torch_dtype_element_size()
        • estimate_max_slot_bytes()
        • estimate_total_heap_bytes()
        • numpy_dtype_to_torch()
        • torch_dtype_to_numpy()
      • Compression methods
        • GpuCompressionMethods
        • has_zstd()
      • Usage
      • Heap layout and compaction
  • Changelog
    • 0.5.0 — 2026-05-27
      • Feature
      • Documentation
      • Breaking Changes
    • 0.4.4 — 2026-05-27
      • Changes by Human
      • By Copilot
    • 0.4.3 — 2025-08-06
    • 0.4.1 — 2025-07-23
      • Benchmark for Compressed Buffers (on MsPacmanNoFrameskip-v4)
    • 0.4.0 — 2025-07-22
    • 0.3.2 — 2025-07-21
    • 0.3.1 — 2025-07-20
    • 0.2.3 — 2025-07-20
    • 0.2.2 — 2025-07-19
sb3-extra-buffers
  • Search


© Copyright 2026, Hugo Huang.

Built with Sphinx using a theme provided by Read the Docs.