stateset-rl-core

High-performance Rust implementations of reinforcement learning operations, with optional Python bindings via PyO3.

Features

• GAE (Generalized Advantage Estimation) - Fast, parallel GAE computation
• Advantage Computation - Group-relative advantages for GRPO/PPO training
• Reward Normalization - Welford's algorithm for numerically stable online normalization
• GSPO Support - Sequence-level importance ratios and clipping for GSPO
• PPO Surrogate - Clipped surrogate objective computation
• Parallel Processing - Automatic parallelization via Rayon

Installation

As a Python extension (recommended)

Pre-built abi3 wheels are published to PyPI on each rust-core-v* tag. Install directly or via the [rust] extra on stateset-agents:

# Standalone
pip install stateset-rl-core

# Bundled with the framework
pip install "stateset-agents[rust]"

One wheel per (OS, arch) covers Python 3.10–3.13 via the stable PyO3 ABI. Supported targets: Linux x86_64/aarch64 (manylinux2014), macOS x86_64/arm64, Windows x86_64.

As a Rust crate

[dependencies]
stateset-rl-core = "0.1"

From source (development)

cd rust_core
maturin develop --release

Usage

Rust

use stateset_rl_core::{compute_gae_internal, compute_advantages_for_group};

// Compute GAE
let rewards = vec![1.0, 0.0, 1.0, 0.0];
let values = vec![0.5, 0.5, 0.5, 0.5, 0.0]; // n+1 values for bootstrap
let advantages = compute_gae_internal(&rewards, &values, 0.99, 0.95);

// Compute group-relative advantages
let group_rewards = vec![1.0, 2.0, 3.0, 4.0];
let advantages = compute_advantages_for_group(&group_rewards, "mean", true);

Python

import numpy as np
import stateset_rl_core

# Compute GAE
rewards = np.array([1.0, 0.0, 1.0, 0.0])
values = np.array([0.5, 0.5, 0.5, 0.5, 0.0])
advantages = stateset_rl_core.compute_gae(rewards, values, gamma=0.99, gae_lambda=0.95)

# Batch GAE (parallel)
all_rewards = [np.random.randn(100) for _ in range(32)]
all_values = [np.random.randn(101) for _ in range(32)]
all_advantages = stateset_rl_core.batch_compute_gae(all_rewards, all_values)

# Group-relative advantages for GRPO
rewards_2d = np.random.randn(16, 4)  # 16 groups, 4 samples each
advantages = stateset_rl_core.compute_group_advantages(rewards_2d, "mean", normalize=True)

# Reward normalization with running stats
rewards = np.array([1.0, 2.0, 3.0])
normalized, mean, var, count = stateset_rl_core.normalize_rewards(rewards)

# GSPO importance ratios
log_probs_new = np.array([-10.0, -12.0, -11.0])
log_probs_old = np.array([-10.5, -11.5, -11.0])
seq_lengths = np.array([50, 60, 55])
ratios = stateset_rl_core.compute_gspo_importance_ratios(log_probs_new, log_probs_old, seq_lengths)

# PPO surrogate objective
ratios = np.array([1.1, 0.9, 1.05])
advantages = np.array([1.0, -1.0, 0.5])
objectives = stateset_rl_core.compute_ppo_surrogate(ratios, advantages, clip_epsilon=0.2)

API Reference

GAE Functions

• compute_gae(rewards, values, gamma=0.99, gae_lambda=0.95) - Single trajectory GAE
• batch_compute_gae(all_rewards, all_values, gamma=0.99, gae_lambda=0.95) - Parallel batch GAE

Advantage Functions

• compute_group_advantages(rewards_2d, baseline_type, normalize) - GRPO-style group advantages
  • baseline_type: "mean", "median", or "min"

Reward Functions

• normalize_rewards(rewards, running_mean=0, running_var=1, count=0, epsilon=1e-8) - Online normalization
• clip_rewards(rewards, min_val, max_val) - Reward clipping
• compute_reward_statistics(rewards) - Compute mean, std, min, max, median

Policy Gradient Functions

• compute_gspo_importance_ratios(log_probs_new, log_probs_old, sequence_lengths) - GSPO ratios
• apply_gspo_clipping(ratios, advantages, clip_left=3e-4, clip_right=4e-4) - GSPO clipping
• compute_ppo_surrogate(ratios, advantages, clip_epsilon=0.2) - PPO clipped objective

Performance

This crate is optimized for performance:

• LTO enabled - Link-time optimization for maximum speed
• Single codegen unit - Better optimization opportunities
• Rayon parallelization - Automatic multi-threading for batch operations
• Zero-copy Python interop - Minimal overhead when called from Python

Typical speedups over pure Python/NumPy: 10-100x for batch operations.

License

BUSL-1.1