llm-d 1.0.0 Graduates Its Router Configuration API to v1, Aligns Component Versions and Deprecates Its ROCm Image

Overview

The llm-d project, which The Machine Herald has previously covered, published v1.0.0 on October 20, 2026, according to its GitHub release page. Everything below comes from the project’s own release notes, which are the only source cited here.

What the Release Notes Say

The release notes group the release under four themes:

• Multi-tenant quality of service. Priority- and fairness-aware flow control with controlled load shedding, so that latency-sensitive requests and bursty agent workloads can share a fleet. Flow control configuration is consolidated under a single flowControl key.
• Zero-downtime operations. Router high availability with a hot standby and safe rolling updates, plus a DisaggregatedSet resource that manages prefill and decode replicas as one logical unit, with coordinated rollouts and lockstep upgrades.
• A stable foundation. The router configuration API graduates to v1, and the Responses API supports previous_response_id. A topology-aware accelerator guide based on DRA and DRANET is included.
• Agent-aware serving. A session-state-producer plugin and session affinity keep multi-turn conversations close to their cached context.

The notes also link to a tracking issue for the release goals, adding that “not all of that was accomplished.” That tracking issue lists “Target release date: October 19, 2026”, one day before the tag. In the issue’s checklists as fetched by the Herald, items such as active-active endpoint-picker work and inference request fail-over and migration are unchecked, as is a WideEP upstream vLLM image on GKE; the notes do not say which unchecked items correspond to the goals that were not accomplished.

Versioning and Deprecations

According to the release notes, the maintainers cited community confusion about version alignment and used the move to v1.0.0 to align all components across the organization, with the stated goal of enforcing this for major and minor versions going forward. In the core component table, the endpoint picker, disaggregation sidecar, inference simulator, benchmark, autoscaling and several other components are listed at v1.0.0; the endpoint picker’s previous version is listed as v0.11.0.

The notes also state: “We are deprecating building our own inferencing images.” The ghcr.io/llm-d/llm-d-rocm image build is deprecated; the notes say it was previously kept only for ROCm prefill/decode disaggregation support and that the project will eventually move to upstream vLLM images. As a stopgap, the release cherry-picks the official vLLM v0.32.0-rc.2 ROCm image into the project’s registry to guard against image garbage collection.

The release also lists three new ecosystem components: modelexpress from the ai-dynamo organization, which the notes say llm-d chose to integrate with for peer-to-peer loading of model weights, plus llm-d-api-extensions and llm-d-pd-utils, the latter described as providing preflight diagnostics, GPU topology discovery and network benchmarks.

Engine Versions Listed

The release tables pin several model-server engines. The core component table lists vllm-project/vllm at v0.31.0 (“varies per accelerator”), up from v0.26.0, while the model-server image table lists the vllm/vllm-openai image at v0.30.0, the ROCm and CPU images at v0.31.0, and SGLang at v0.5.22, up from v0.5.20. The TensorRT-LLM image is listed at 1.3.0rc29. The notes do not explain the difference between the two vLLM figures.

Upgrade Constraint

The notes add a requirement for disaggregated deployments: the prefill/decode guide now needs LeaderWorkerSet v0.11.1 or newer with the DisaggregatedSet API enabled, via --set enableDisaggregatedSet=true with Helm. The notes highly recommend upgrading to v1.0.0, the version listed for LeaderWorkerSet in the infrastructure table, and say almost all guides will break without v0.11.1 or newer.

What Is Not in the Release Notes

The release notes’ theme summaries do not report performance figures for the flow-control or DisaggregatedSet features, and this article does not cite any outside assessment of them. The themes above are the project’s own description of what it shipped.