sgl-deep-ep

sgl-deep-ep is SGLang's binary distribution of DeepEP. The distribution name is sgl-deep-ep, while the Python import remains compatible with DeepEP:

import deep_ep

Installation

The PyPI package targets CUDA 13 and installs with:

pip install sgl-deep-ep

Each wheel is specific to a CPython version, CPU architecture, and CUDA major. The package checks these constraints when deep_ep is imported.

Host prerequisites

The wheel contains DeepEP's Python code and CUDA extension. It does not and cannot provision the host kernel or RDMA stack. Every host must provide:

• a compatible NVIDIA driver and CUDA runtime;
• the RDMA/NVLink environment required by the selected DeepEP transport.

At import time, the package checks the supported platform, PyTorch and CUDA versions, and CUDA driver/device availability. Failures raise an actionable ImportError. Transport-specific prerequisites are validated when DeepEP initializes the selected transport.

CUDA 13 wheels use DeepEP v2 and its NCCL Gin transport. Release and runtime images pin nvidia-nccl-cu13==2.30.7. NVSHMEM remains a build and runtime dependency because the current v2 extension still links the legacy DeepEP methods.

Wheel release workflow

Wheels are built and published by SGLang's Release sgl-deep-ep workflow. Implementation and packaging are maintained together on sgl-deepep. Before starting a release, merge both kinds of changes into that branch:

• CUDA 13, x86_64 and aarch64: sgl-deepep (DeepEP v2)

Run the workflow manually with:

• version: the public package version, without a leading v;
• deepep-ref: the branch, tag, or commit containing both the implementation and sgl_deep_ep/ (normally sgl-deepep; use a commit for a reproducible release).

For each selected CUDA target, the workflow uploads CUDA-tagged wheels to the cu130 index in sgl-project/whl. CUDA 13.0 wheels are also published to PyPI as sgl-deep-ep, without a CUDA suffix in the package name. Publishing requires the GH_PAT_FOR_WHL_RELEASE and SGL_DEEP_EP_PYPI_TOKEN repository secrets.

After the workflow completes, check that every expected Python and architecture artifact was published, install the appropriate wheel in a clean environment, and verify both import deep_ep and a multi-GPU DeepEP or SGLang workload.

Building from a checkout

Initialize the submodules and install the build dependencies, including the target CUDA PyTorch build, NCCL 2.30.7, NVSHMEM, ninja, setuptools, and wheel. The release container in SGLang's docker/sgl-deep-ep.Dockerfile provides these dependencies for manylinux builds. Then run:

git clone --recursive --branch sgl-deepep https://github.com/sgl-project/DeepEP.git
cd DeepEP
SGL_DEEP_EP_VERSION=0.1.3 bash sgl_deep_ep/build_sgl_deep_ep.sh dist 13.0 x86_64

The arguments are the output directory, CUDA version (13.0 or 13.4), and CPU architecture (x86_64 or aarch64). Use the CUDA version and architecture of the build environment. The script reads implementation and packaging from its own checkout, stages a temporary copy, and leaves the source unchanged. SGL_DEEP_EP_VERSION overrides sgl_deep_ep/VERSION; the wheel version includes the corresponding +cu130 or +cu134 suffix. PYTHON_BIN selects the build interpreter and MAX_JOBS controls parallelism.

This produces a native Linux wheel. The SGLang release build script applies auditwheel repair for manylinux and creates the public-version PyPI wheel. The CUDA 13.4 image also uses this entry point; the release workflow currently publishes CUDA 13.0 wheels only.