Using SkyRL as a Library

To use SkyRL as a dependency in your project, configure your pyproject.toml with uv. You can install from either PyPI (stable releases) or Git (latest development version).

Some additional dependencies that depend on custom indexes/wheels, such as flash-attn and flashinfer-jit-cache. Below is an example configuration using uv:

[project]
dependencies = [
    # Option A: Install from PyPI (stable release)
    "skyrl[fsdp]",
    # Option B: Install from Git (latest)
    # "skyrl[fsdp]",

    "flashinfer-jit-cache==0.6.13",
    "flash-attn==2.8.3+cu.13.0.torch.2.11"
]

[tool.uv.sources]
flashinfer-jit-cache = { index = "flashinfer-cu130" }
flash-attn = { index = "astral-cu130" }
# Only needed for Option B (Git install):
# skyrl-gym = { git = "https://github.com/NovaSky-AI/SkyRL", subdirectory = "skyrl-gym" }
# skyrl = { git = "https://github.com/NovaSky-AI/SkyRL" }

[[tool.uv.index]]
name = "flashinfer-cu130"
url = "https://flashinfer.ai/whl/cu130"
explicit = true

[[tool.uv.index]]
name = "astral-cu130"
url = "https://wheels.astral.sh/simple/cu130/"
explicit = true

Option A (PyPI): Use skyrl[fsdp]. No additional sources needed.

Option B (Git): Uncomment the skyrl-gym and skyrl sources to install from the latest commit.

SkyRL is built against CUDA 13 and torch 2.11, so both indexes above are the cu130 variants. flash-attn comes from Astral's prebuilt GPU wheel index, which stamps each wheel with the (CUDA, torch) pair it was built against as a PEP 440 local version — the pin has to name that build exactly. Keep it at 2.8.3, not 2.8.3.post1: Transformer Engine only accepts flash-attn up to 2.8.3.