Snowflake Research

Explore Snowflake's initiatives and contribution to making enterprise AI easy, efficient, and trusted through open research and innovation.

Arctic RL

Arctic RL Overview

Arctic RL is a high-throughput reinforcement learning training and inference backend designed to integrate with existing RL frameworks, providing unified GPU backends and system optimizations for accelerated post-training of large language models.

Configuration Reference

Configure the Arctic RL client with backend, communication protocol, model, and GPU resource settings through the ArcticRLClientConfig model. This page documents all available configuration options for initializing and tuning the platform.

Client API

The ArcticRL Client API provides a unified interface for reinforcement learning training, sampling, and inference operations. It supports both HTTP and Ray communication protocols with automatic job initialization and lifecycle management.

Weight Synchronization

The `WeightSyncCoordinator` manages efficient transfer of model weights from training GPUs to inference replicas using NCCL topology, enabling thread-safe coordination between training and inference clients with lazy sender setup and HTTP-orchestrated synchronization.

Framework Integrations

Arctic RL integrates into existing RL frameworks rather than replacing them, with completed integrations into SkyRL and an in-progress integration into Verl, plus additional framework integrations planned for TRL, Axolotl, unsloth, and PrimeRL.

Training Pipeline & Processors

The Arctic RL training pipeline provides a composable, two-phase architecture for forward passes: post-forward processors compute derived outputs (like logprobs from logits), and loss functions compute scalar losses and metrics. Register processors by name to make them available at runtime via batch configuration.

DeepSpeed Worker

A Ray-managed single-GPU actor that handles DeepSpeed training, inference, and weight synchronization for reinforcement learning workloads. Supports distributed training with gradient checkpointing, ZeRO optimization, and optional Liger kernel acceleration.

Ray Server & Deployment

The Arctic RL platform uses Ray to manage distributed DeepSpeed workers and ArcticInference ReplicaPools for training, sampling, and log probability computation. This page covers the server architecture, initialization, and key deployment patterns including colocated GPU resource sharing.

Overview

Contributing

Getting Started

ZoRRo Train

ZoRRo Train Overview

ZoRRo Train eliminates redundant prompt computation in RL training by automatically deduplicating identical prompts at the attention layer, reducing 80–95% of wasted token computations while maintaining mathematical correctness. This transparent optimization delivers 2–4x speedup for high-deduplication scenarios through prompt grouping, optimized attention, and automatic reconstruction.

ZoRRo Train API Reference

Complete API reference for ZoRRo Train's core classes and methods for prompt deduplication in RL training. Covers DeduplicatedActor, ZoRRoTrain, and supporting utilities.

Performance Benchmarks

ZoRRo Train achieves significant speedups through prompt deduplication, with expected 2–4x improvements under high deduplication scenarios and dependencies on hardware, sequence length, and attention implementation.

How Prompt Deduplication Works

Prompt deduplication removes redundant computation in RL training by identifying sequences with identical prompts and processing them together, eliminating up to 82% of wasted token processing while maintaining mathematical correctness for gradients.

Supported Models

ZoRRo Train currently supports Qwen3 model families including dense, MoE, and hybrid architectures, with plans to expand support to additional models.

Testing & Correctness Verification

ZoRRo Train includes comprehensive CPU and GPU tests to verify deduplication correctness, gradient accuracy, and performance. Run tests with pytest and benchmarks with dedicated scripts.

Getting Started with ZoRRo Train

ZoRRo Train is a prompt deduplication optimization that eliminates redundant computation in RL training by identifying identical prompts and processing them once. Learn how to install, configure, and use the system to achieve 2-4x speedups in typical RLHF scenarios.

Recipes

ZoRRo Inference