Changes and Releases
Updates on Giant Swarm workload cluster releases, apps, UI improvements and documentation changes.
Fixed
- (model-serving) Triton’s kernel cache lives on the cache claim beside vLLM’s compiled graphs — TRITON_CACHE_DIR=/mnt/vllm-cache/triton on the runtime container in #573 by @teemow
Full Changelog: https://github.com/giantswarm/agent-platform/compare/v4.40.0...v4.40.1
Added
- (model-serving) The slice’s llm-d images come from the llm-d-fast/ prefix (re-layered, zstd llm-d-cuda) and the pre-pull names that image (#568) in #569 by @teemow
Full Changelog: https://github.com/giantswarm/agent-platform/compare/v4.39.0...v4.40.0
Added
- (model-serving) The cache claim’s defaults are 100Gi at 500 MiB/s / 3000 IOPS — the shipped presets need < 40 GiB and the read is loader-bound at ≈ 300 MB/s; the default class is named with a digest of its parameters and an existing claim keeps its class and size (#570) in #571 by @teemow
Full Changelog: https://github.com/giantswarm/agent-platform/compare/v4.38.3...v4.39.0
Fixed
- (model-serving) Qwen3-8-flash-next-nvfp4 memory limit 118Gi — B12X keeps the weights as mlocked process memory the cgroup charges in #567 by @teemow
Full Changelog: https://github.com/giantswarm/agent-platform/compare/v4.38.2...v4.38.3
Fixed
- (model-serving) The serving namespace is kept on uninstall whatever the cache switch says — modelServing.namespace.keep (default true), so a slice composed without the cache no longer deletes a cache claim an earlier slice left there (#565) in #566 by @teemow
Full Changelog: https://github.com/giantswarm/agent-platform/compare/v4.38.1...v4.38.2
Added
- Built-in static clients
mcpCapi and mcpPrometheus (oidc.staticClients.mcpCapi, oidc.staticClients.mcpPrometheus) with the shape and rules of mcpKubernetes: clientID, redirectURI, trustedPeers, and exactly one of clientSecret and clientSecretRef: {name, key}; both are trusted peers of dex-k8s-authenticator. Values under these keys were accepted by the schema and ignored by the templates before: an installation that carries them gets the two clients with this version, and one that carries a clientID without a secret fails the render naming the client. Installations without them render unchanged.
Fixed
- (mcpserver) A remote server’s spec.timeout bounds every operation, tool calls included in #1282 by @teemow
Full Changelog: https://github.com/giantswarm/muster/compare/v5.27.1...v5.27.2
Added
- Static clients read their secret from a Kubernetes Secret:
secretRef: {name, key} in an extraStaticClients entry, clientSecretRef: {name, key} in the pre-defined clients (gitopsui, muster, mcpKubernetes, dexK8SAuthenticator), in place of the inline secret. The chart sets DEX_CLIENT_SECRET_<ID> on the dex container from the referenced key and names it in the client’s secretEnv, so a client is added by a new Secret and a plaintext list entry. Exactly one of the inline secret and the reference per client: both, or neither on a client that is not public, fails the render naming the client. Inline clients render unchanged.
Changed
- A pre-defined static client with a
clientID but no secret (gitopsui, muster, mcpKubernetes) fails the render naming the client instead of being left out of the configuration silently.
Changed
- (golden) Drop the holds GOLDEN_REF now carries — every one only narrowed the comparison, and three of them compared a shape no installation gets in #531 by @fiunchinho
Full Changelog: https://github.com/giantswarm/agent-platform/compare/v4.38.0...v4.38.1
Added
- (bot-prs) A classification is a way into the queue, and an empty column is dropped in #2463 by @QuentinBisson
Full Changelog: https://github.com/giantswarm/backstage/compare/v2.29.1...v2.30.0