Project Scope & Roadmap

Project Scope & Roadmap

Arctic Platform is a framework for simplifying and accelerating post-training for large language models (LLMs). It addresses challenges in current post-training frameworks through modularity across training and inference components, simplified code structures, and integrated pipelines for creating and cleaning synthetic data.

Current Scope

Arctic Platform covers the full post-training stack for LLMs behind a small, composable API. The codebase is being built out incrementally.

Available Today

Arctic Reinforcement Learning — A high-throughput RL training and inference backend that plugs into existing RL frameworks rather than replacing them. The framework keeps ownership of the training loop, rollouts, rewards, and advantage estimation, while Arctic Platform provides the compute engines:

• Training engine — A DeepSpeed engine for forward/backward passes and optimizer steps
• Log-prob / reference engine — A forward-only DeepSpeed engine for reference and old log-prob computation
• Sampling engine — A vLLM engine with ArcticInference for fast rollouts

These engines are orchestrated over Ray, can be colocated on shared GPUs via fractional Ray resources or split across separate GPUs, and keep the sampler in sync with the trainer through NCCL or CUDA-IPC weight transfer.

ZoRRo Train — A prompt-deduplication optimization that removes redundant prompt computation during RL training. In RL training (PPO/GRPO), the same prompt is sampled many times, meaning 80–95% of tokens in long-sequence batches are redundant prompt tokens. ZoRRo Train detects sequences sharing a prompt, runs the model once over deduplicated sequences, and transparently reconstructs per-response logprobs and entropy, achieving substantial memory savings and throughput improvements while remaining mathematically equivalent to naive forward/backward computation.

ZoRRo Inference — Forest Cascade Attention (FCA) for efficient rollout steps that eliminates redundant memory accesses through grouping. During RL rollouts, many sequences are generated from the same prompt. FCA discovers groups of requests sharing a KV-cache prefix, splits attention calls into a grouped pass over shared prefix blocks and per-request passes over unique suffix blocks, then reduces the results with rigorous weighting. This reads each shared prefix block once per group instead of once per request.

Coming Next

• Additional trainers (SFT/distillation)
• Synthetic data generation and cleaning pipelines
• Tighter inference integration

Framework Integrations

Integrated

• SkyRL — Fully integrated. Documentation

Integration Complete but Not Yet Merged

• Verl — Reference integration that drives Arctic RL from its PPO/GRPO trainer. End-to-end recipes are available in recipes/rl/verl/, including a simple single-GPU GSM8K starter, Txt2SQL, and long-context QA. PR #6422

Upcoming Integrations

Multiple additional framework integrations are in progress, including TRL, Axolotl, unsloth, PrimeRL, and others.

Getting Started

To begin training a model with Arctic Platform:

1. Install the package
2. Follow the recipes

Installation

From PyPI

pip install arctic-platform

From Source (Git)

For the latest development version:

git clone https://github.com/Snowflake-AI-Research/Arctic-Platform.git
cd Arctic-Platform
pip install -e .[rl]

Example Usage

To set up Arctic RL on a single 8-GPU node:

from arctic_platform.rl import ArcticRLClientConfig, create_arctic_rl_client

config = ArcticRLClientConfig(
    model_name="Qwen/Qwen3-4B",
    comm_protocol="ray",        # or "http"
    training_gpus=8,
    sampling_gpus=8,
    log_prob_gpus=0,
    colocate=True,
)
client = create_arctic_rl_client(config)

An RL framework integrates this module by constructing a client and driving standard operations: generate, forward/backward, optimizer step, sync_weights, and wake/sleep for memory management.