stable-retro-turbo 1.0.1.post37

🚀 Blazing-fast Stable Retro fork with native vectorization and preprocessing 🚀

stable-retro-turbo is a Python library for reinforcement-learning developers who need faster batched rollouts from classic console games. It keeps Stable Retro's game integrations and single-environment API, and adds RetroVecEnv, a Gymnasium vector environment that steps many libretro emulators and preprocesses observations in native code.

Install the package, import your legally obtained ROMs, and use the upstream-compatible stable_retro import. The native path is useful for parallel training workloads that would otherwise spend substantial time crossing between Python wrappers and individual emulator instances.

Install

Release wheels require Python 3.14 and support macOS on Apple Silicon and Linux on x86-64.

uv venv --python 3.14
source .venv/bin/activate
uv pip install stable-retro-turbo
python -m stable_retro.import /path/to/your/roms

ROMs are not included. The importer matches supported ROMs to Stable Retro's game integrations.

Check the games available on your machine, then open one in the interactive player:

stable-retro-turbo play --list
stable-retro-turbo play nes
stable-retro-turbo play SuperMarioBros-Nes-v0 --press START
stable-retro-turbo play Breakout-Atari2600-v0 --mode 32 --difficulty A

Pass a full game ID such as SuperMarioBros-Nes-v0, or use all to open one imported game per platform. Add --show-obs to display the raw game beside its PPO-style preprocessed observation.

--press BUTTON[:COUNT] applies repeatable startup inputs using the selected game's own button names. Atari games additionally accept --mode N, which pulses the console SELECT switch N times before RESET, and --difficulty A or --difficulty B, which sets both console difficulty switches. For example, Breakout mode values 0, 4, 8, …, 44 select its twelve one-player cartridge variants. Use --state none to launch a game from its power-on state.

Use

import numpy as np
import stable_retro as retro

env = retro.RetroVecEnv(
    "SuperMarioBros-Nes-v0",
    state="Level1-1",
    num_envs=32,
    num_threads=16,
    render_mode="rgb_array",
    obs_crop=(32, 0, 0, 0),
    obs_crop_mode="mask",
    obs_resize=(84, 84),
    obs_resize_algorithm="area",
    obs_grayscale=True,
    obs_layout="chw",
    frame_skip=4,
    frame_stack=4,
    maxpool_last_two=True,
    info_filter="terminal",
)

obs, infos = env.reset(seed=123)
obs, rewards, terminations, truncations, infos = env.step(
    env.action_space.sample()
)

done = terminations | truncations
if done.any():
    obs, reset_infos = env.reset(options={"reset_mask": done})

RetroVecEnv uses Gymnasium's disabled-autoreset semantics. A finished lane keeps its terminal observation and cannot be stepped again until it is selected by a masked reset; unselected lanes keep their emulator state, RNG stream, frame stack, and sticky-action history.

Turbo Vector API v1

RetroVecEnv implements the strict Turbo Vector API v1:

• metadata["turbo_api_version"] is 1, and metadata["render_modes"] advertises rgb_array.
• Immutable capabilities and signal_schema declarations describe supported features and the dtype, shape, and reset/step availability of every signal.
• buttons, action_mode, action_preset, action_table, action_meanings, and action_table_hash expose the resolved action semantics without provider-specific probing.
• state_catalog is an immutable ordered tuple. Callers select reset states with an int32 state_indices array and inspect the read-only active indices with active_state_indices(); state sampling and lane routing remain caller-owned.
• observation_ownership and observation_buffer_depth declare the exact lifetime of returned observations. render_lane(index) renders one lane, get_images() renders all lanes, and render() renders lane zero.

When env.supports_live_snapshots is true, live positions can be captured without advancing emulation and restored into any lane of the same environment:

capture_mask = np.zeros(env.num_envs, dtype=np.bool_)
capture_mask[0] = True
captured = env.capture_snapshots(capture_mask)

restore_mask = np.zeros(env.num_envs, dtype=np.bool_)
restore_mask[3] = True
starts = [None] * env.num_envs
starts[3] = captured[0]
obs, infos = env.reset(
    options={"reset_mask": restore_mask, "snapshots": starts},
)
env.close()

Handles are reusable, session-local, and intentionally not pickleable. A single masked reset can mix snapshot starts with ordinary state_indices; infos["start_source"] distinguishes "snapshot" from "environment". Scripted scenarios and cores that cannot serialize exact state report the capability as unavailable.

The fast path also supports:

• native crop, mask, resize, grayscale, layout conversion, frame skip, frame stack, and two-frame max-pooling;
• ordered saved-state catalogs with explicit per-lane state_indices; state sampling and curriculum routing stay with the caller;
• copy-safe, safe-view, and benchmark-only unsafe-view observation ownership;
• sticky actions, random no-op starts, reward clipping, and native info filtering;
• Atari through the packaged Stella core, using the same RetroEnv and RetroVecEnv APIs.

The inherited RetroEnv API remains available for single-environment use. RetroVecEnv supports one player, image observations, and no movie recording.

Develop

git clone https://github.com/tsilva/stable-retro-turbo.git
cd stable-retro-turbo
uv sync --frozen

Source builds require Python 3.14, CMake, a C/C++ compiler, and the platform dependencies needed by the selected emulator cores.

Commands

Run these commands from the repository root:

uv run --frozen stable-retro-turbo play --list                                      # list imported games by platform
uv run --frozen python scripts/benchmark_vec_env.py --list-profiles                 # list benchmark profiles
uv run --frozen python scripts/benchmark_vec_env.py --profile supermario-level1-1  # benchmark native and classic paths
uv run --frozen --with pytest pytest tests/test_python/test_cli.py tests/test_python/test_benchmark_vec_env.py  # run quick tests
uv run --frozen --with build python -m build                                        # build source and wheel artifacts

Benchmark definitions live in scripts/benchmark_vec_env.json. Use real saved states for representative throughput measurements; State.NONE is reserved for explicit direct-ROM diagnostics and requires --allow-state-none.

Notes

• The distribution is stable-retro-turbo; the Python package is stable_retro. The upstream-compatible retro import remains available for scalar integrations, while new code should import stable_retro.
• RetroVecEnv implements Gymnasium's vector API directly. It is not a Stable-Baselines3 VecEnv, and Stable-Baselines3 is not a runtime dependency.
• A scalar reset seed expands to seed + lane_index. Seed sequences must contain one integer or None per lane.
• state_catalog preloads an ordered saved-state catalog. Select reset lanes with reset_mask and their exact catalog entries with state_indices; Turbo does not sample states.
• capture_snapshots(mask) returns lane-aligned live handles for exact same-instance continuation. The caller owns archive selection, eviction, and curriculum policy; handle.nbytes exposes approximate payload size.
• active_state_indices() returns a read-only NumPy view. Copy it when you need a stable snapshot.
• Stable Retro remains the source for inherited game, integration, and emulator documentation.
• Bundled emulator cores have their own licenses; see LICENSES.md.

Architecture

License

MIT