agent-platform release v4.90.1
Added
modelServing.fastLinks: the nodes joined by a fast link (RDMA), published as the discovery ConfigMap’s spec.fastLinks (giantswarm/model-manager#190). model-manager splits one model across the nodes of one fast link — one vLLM instance, tensor parallel over the link — and refuses a split across nodes without one; the portal recommends the split where a fast link hosts the model. Per entry: name, nodes (two or more), the networks every pod of a split joins (NetworkAttachmentDefinition names), the extra resources it requests (the RDMA device plugin’s) and the env that points NCCL at the link’s devices. Empty, the default, renders no key: the discovery document of an installation without fast links is unchanged. With a fast link declared, a split’s leader and workers — the pods of a LeaderWorkerSet — reach each other on every port (network policy <release>-model-serving-multi-node); a single-node model pod does not. make verify-fast-links.
Removed
- The PolicyException
model-serving-fast-links (giantswarm/giantswarm#37941). model-manager 1.4.2 runs a split’s leader and workers with the single-node model pod’s security context (giantswarm/model-manager#193): no IPC_LOCK, SYS_RAWIO or NET_RAW, and NCCL over the fast link needs none of them. Nothing requests the capabilities any more, so no pod may add them. make verify-fast-links fails if the exception renders.