NovaSky

license

Apache-2.0(OSI, root LICENSE)+MIT(OSI, declared by the skyrl-gym package)

source

public(NovaSky-AI/SkyRL)

core features withheld

no — the tree at 2c54cef carries one root Apache-2.0 LICENSE and no ee/, enterprise/, commercial/ or proprietary/ path. The project is developed at the Berkeley Sky Computing Lab, and Anyscale runs the same open library on its paid managed platform rather than selling an edition of it

SkyRL is Apache-2.0 licensed, and its public repository holds the trainer, the Tinker backend, the environments and the agent layer, so it builds and runs from published source with no account. Anyscale offers SkyRL on its paid managed platform, but that is a separate service running the same open library, and nothing in SkyRL is held back for it.

• https://docs.anyscale.com/llm/fine-tuning recorded 2026-10-19

  Anyscale's documentation for its managed platform says "Anyscale supports multiple frameworks for LLM post-training, including LLaMA-Factory, SkyRL, and Ray Train" and lists a "GRPO with SkyRL" tutorial: the paid platform runs the open library.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/LICENSE recorded 2026-10-19

  The root LICENSE is the stock Apache License 2.0 text, ending with the unfilled appendix line "Copyright [yyyy] [name of copyright owner]"; no added term or condition is attached.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/pyproject.toml recorded 2026-10-19

  The root manifest builds the skyrl package, "Unified API for training and inference", from the repository's skyrl* packages, with fsdp, megatron, jax and tinker extras for its backends.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/README.md recorded 2026-10-19

  The README describes SkyRL as "a full-stack RL library" whose components skyrl, skyrl-agent and skyrl-gym live in this repository, says the work is done at the Berkeley Sky Computing Lab in collaboration with Anyscale, and mentions no enterprise, paid or license-gated edition.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/skyrl-gym/pyproject.toml recorded 2026-10-19

  The skyrl-gym manifest declares name = "skyrl-gym", version = "0.4.0" and license = { text = "MIT" }.

Adoption is the trailing month of PyPI downloads of skyrl, the package that now holds the trainer and the Tinker backend; the retired skyrl-train and skyrl-tx packages and the skyrl-gym environment library are left out. Installs from a git checkout, which the installation guide uses, and from the project's Docker images are not counted, so the figure is a lower bound on use.

• https://pypi.org/pypi/skyrl-train/json recorded 2026-10-19

  skyrl-train's newest release is 0.3.1, uploaded 2026-01-02, with no release since.

• https://pypi.org/pypi/skyrl-tx/json recorded 2026-10-19

  skyrl-tx's newest release is 0.3.0, uploaded 2026-02-09, with no release since.

• https://pypi.org/pypi/skyrl/json recorded 2026-10-19

  skyrl 0.4.0, uploaded 2026-10-01, with the summary "Unified API for training and inference", the description the repository's root manifest gives. Its fsdp and megatron extras pull skyrl[skyrl-train], which requires skyrl-gym.

• https://pypistats.org/api/packages/skyrl-gym/overall?mirrors=false recorded 2026-10-19

  Daily skyrl-gym downloads without mirrors: 7,883, 9,199, 7,455 and 6,595 on 2026-09-11 to 2026-09-14, then under 400 a day from 2026-09-15 to 2026-10-18 apart from 1,351 on 2026-10-06.

• https://pypistats.org/api/packages/skyrl-gym/recent recorded 2026-10-19

  last_month 34,949, last_week 187, last_day 22 for skyrl-gym.

• https://pypistats.org/api/packages/skyrl-train/recent recorded 2026-10-19

  last_month 61, last_week 13, last_day 1 for skyrl-train.

• https://pypistats.org/api/packages/skyrl-tx/recent recorded 2026-10-19

  last_month 27, last_week 2, last_day 0 for skyrl-tx.

• https://pypistats.org/api/packages/skyrl/recent recorded 2026-10-19

  last_month 210, last_week 26, last_day 5 for the skyrl package.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/docs/content/docs/getting-started/installation.mdx recorded 2026-10-19

  The installation guide marks "Docker (recommended)", starts a novaskyai/skyrl-train-ray image, and inside it has the user run "git clone https://github.com/novasky-ai/SkyRL.git"; that path does not go through PyPI.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/docs/content/docs/getting-started/library.mdx recorded 2026-10-19

  The library guide offers "Option A: Install from PyPI (stable release)" as skyrl[fsdp], and "Option B" from Git.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/skyrl-tx/README.md recorded 2026-10-19

  The skyrl-tx README says "SkyRL is undergoing a repo reorganization" into a single package and that the code "previously in the skyrl-tx folder can now be found in skyrl/{backends, tinker, tx, utils}".

SkyRL covers a broad RL post-training stack: PPO, GRPO and several newer policy-optimization variants over FSDP or Megatron, vLLM or SGLang rollout, fully asynchronous training, and a local implementation of the Tinker API. It sits a step below verl, which documents a wider algorithm catalog and has been used to train frontier reasoning models.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/docs/content/docs/algorithms/dapo.mdx recorded 2026-10-19

  The DAPO guide says "DAPO is supported with the fsdp and megatron backends" and walks through Clip-Higher, dynamic sampling, token-level loss and overlong reward shaping.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/README.md recorded 2026-10-19

  The README lists skyrl, skyrl-agent and skyrl-gym as SkyRL's components, links a 397B RL training guide, and announces day-0 LoRA, full fine-tuning and MXFP8 RL recipes for a 501B-A23B MoE model on a release branch.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/skyrl-train/README.md recorded 2026-10-19

  The trainer's feature list gives "PPO and GRPO", FSDP and Megatron training backends, vLLM, SGLang and OpenAI-compatible inference endpoints with weight sync, Ulysses sequence parallelism, synchronous, one-step-off and "fully async RL with in-flight weight updates", "Algorithmic support for RLOO, REINFORCE, GSPO, CISPO, SAPO" and "5D Parallelism support for MoE models" with Megatron.

• https://raw.githubusercontent.com/NovaSky-AI/SkyRL/main/skyrl-tx/README.md recorded 2026-10-19

  SkyRL tx implements a backend for the Tinker API with multi-user LoRA, "SFT & RL Support", FSDP, tensor and expert parallelism, and GPU and TPU launch commands.