Nightly Releases

Dynamo publishes nightly builds from main. Nightlies let you try the latest features and backend upgrades before they land in a stable release. This page covers what nightly publishes, how to install it, and which backend versions the current and recent nightlies ship.

Recent Nightlies

What Gets Published

Every night, the Nightly CI pipeline builds main and publishes:

• Runtime container images (CUDA 13): vllm-runtime-nightly, sglang-runtime-nightly, and tensorrtllm-runtime-nightly to NGC, each with an Elastic Fabric Adapter (EFA) variant under a -efa tag suffix.
• Component container images: kubernetes-operator-nightly, dynamo-planner-nightly, and dynamo-frontend-nightly to NGC.
• Python wheels: ai-dynamo, ai-dynamo-runtime, and kvbm to the NVIDIA prerelease index at pypi.nvidia.com.
• Helm chart: dynamo-platform-nightly to the NGC Helm registry, kept separate from the stable dynamo-platform chart.

The runtime images and the wheels gate the release: if any of them fails to build, that night publishes nothing. The component images, the EFA variants, and the Helm chart stage fail-soft, so a flake in one of them skips that artifact for the night without holding back the rest.

Nightly does not publish Rust crates — for those, use a stable or pre-release build.

Installing Nightly Containers

Nightly images live in their own -nightly NGC repositories so they cannot be pulled accidentally in place of a stable image. Every nightly build pushes an immutable YYYYMMDD-<shortsha> tag and moves the latest tag to that build.

# Most recent nightly
docker pull nvcr.io/nvidia/ai-dynamo/vllm-runtime-nightly:latest
docker pull nvcr.io/nvidia/ai-dynamo/sglang-runtime-nightly:latest
docker pull nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime-nightly:latest
# Pin one nightly build
docker pull nvcr.io/nvidia/ai-dynamo/vllm-runtime-nightly:20260830-4c7e981
# EFA variant, floating or pinned
docker pull nvcr.io/nvidia/ai-dynamo/vllm-runtime-nightly:latest-efa
docker pull nvcr.io/nvidia/ai-dynamo/vllm-runtime-nightly:20260830-4c7e981-efa

Pin the dated tag for anything you need to reproduce later: latest moves every night, so the image behind it changes underneath you. The runtime repositories also carry a nightly tag, but it stopped tracking main in July 2026 — use latest or a dated tag instead. The component repositories publish the same dated tags and a latest float, without the nightly alias or the EFA variants.

Installing Nightly Wheels

Nightly wheels are published to the NVIDIA prerelease index at pypi.nvidia.com, not the public PyPI. They are Linux manylinux builds for the Python versions in Compatibility; install on a supported Linux host or inside a Linux container. Nightly versions follow PEP 440 dev versioning, X.Y.Z.devYYYYMMDD.

# Latest nightly (uv)
uv pip install --pre --extra-index-url https://pypi.nvidia.com/ ai-dynamo
# Latest nightly (pip)
pip install --pre --extra-index-url https://pypi.nvidia.com/ ai-dynamo
# Pin a specific nightly wheel
uv pip install --pre --extra-index-url https://pypi.nvidia.com/ "ai-dynamo[vllm]==1.5.0.dev20260831"

Backend extras such as ai-dynamo[vllm] and ai-dynamo[sglang] use the same flags. For TensorRT-LLM, use the nightly container rather than a PyPI extra.

Backend Versions

Nightlies track main, so the backend versions they ship change as main advances. To find which nightly or stable build ships a given backend version, and get the exact pull or install command, use the build selector in the Kubernetes Quickstart.

To confirm the exact versions a specific nightly shipped, read them from the pulled image:

docker run --rm nvcr.io/nvidia/ai-dynamo/vllm-runtime-nightly:latest pip show vllm

See Also

• Release Artifacts — stable and pre-release artifact inventory
• Compatibility — hardware, platform, CUDA, and driver support
• Model Early Access Builds — model-specific pre-release container builds