llm-d v0.10.0 Deprecates Custom CUDA and AWS Images for Upstream vLLM

The llm-d project released v0.10.0 on September 29, 2026, with a focus on operational hardening and production readiness. The release deprecates the project's own CUDA and AWS container images in favor of upstream vLLM images, and introduces a new image build and signing workflow that separates development builds from production artifacts.

Image Build Refactor and Signing

The release team changed how container images are published. Instead of maintaining separate dev and regular images under different names on GitHub Container Registry, llm-d now uses quay.io as a staging registry and ghcr.io as the production registry, with identical image names across both. Images built from code merged to main or from official releases go to ghcr.io and are signed with cosign for provenance. Images built from open pull requests go to quay.io and are intentionally unsigned, reflecting that PR builds are a courtesy for developers and their integrity cannot be guaranteed by the community.

This signing policy currently applies only to the inference server images produced in the llm-d/llm-d repository. The router image and sidecar image are not included, and the project may extend the approach if the community approves.

Deprecations and Upstream Migration

The release deprecates the ghcr.io/llm-d/llm-d-cuda image in favor of docker.io/vllm/vllm-openai, and the ghcr.io/llm-d/llm-d-aws image in favor of public.ecr.aws/deep-learning-containers/vllm. The llm-d-kv-cache repository moved into the llm-d-router repository, and the llmd-fs-connector was integrated upstream into vLLM as the OffloadingConnector, so it no longer requires separate installation. The workload-variant-autoscaler repository was renamed to llm-d-autoscaling, and its guides are marked as deprecated.

The release also updates the base vLLM image from v0.26.0 to v0.30.0, with corresponding updates for vLLM Omni, TPU, XPU, ROCm, and CPU variants. SGLang moved to v0.5.20, and TRT-LLM to 1.3.0rc28. The llm-d CPU image remains at v0.10.0, and a new llm-d-xpu-sglang image was added.

Component and Infrastructure Updates

The release includes version bumps across several components. The router endpoint picker moved to v0.11.0, the disagg sidecar to v0.11.0, the inference simulator to v0.11.2, and the batch gateway to v0.6.0. The async component is at v0.10.0, and the autoscaling Helm chart and image are at v0.9.0. The Gateway API CRDs remain at v1.5.1, and the Gateway API inference extension CRDs at v1.5.0. AgentGateway moved from v1.1.0 to v1.5.0, and the Leader Worker Set and Disaggregated Set are now at 0.11.0.

The project published a release matrix tracking tested guides, noting that some tests could not run due to infrastructure constraints. The release notes also list numerous fixes, including corrections to nightly CI lanes, XPU image handling, CPU builds for NIXL, and documentation updates for flow control and observability.