[HanFa]

Summary

The Dynamo page last changed substantively on 2026-08-19 (#63). This PR updates it to match CentML/platform main at a53e4be2b.

• Scale to zero (CentML/platform#4784, CentML/platform#4801):
  • Min Replicas accepts 0.
  • Concurrency is required only when Max Replicas exceeds both 1 and Min Replicas.
  • Cooldown Period defaults to 1800 s and also sets how long the Workers stay idle before scaling to zero.
  • A new subsection covers the cold-start window. Requests fail, with HTTP 404, until a Worker is ready. Any authenticated request to the endpoint wakes the Workers or keeps them running.
• Disaggregated serving in the console (CentML/platform#4224):
  • The Hardware step's serving-topology selector and the Prefill and Decode pool settings.
  • Edit and Clone for both topologies.
  • The Serving Topology panel on the Detail page.
  • This removes the "API-only" and "aggregated-only Edit/Clone" statements.
• KV transfer policy (CentML/platform#4532, #4533, #4584, #4585): the transport section is now split into two subsections.
  • allow_tcp keeps the existing UCX/TCP text and the B200 validation.
  • require_gpu_direct_rdma:
    • works only on AWS EKS with EFA devices managed through DRA;
    • gives each worker one EFA device and keeps all workers in the same availability zone;
    • has no TCP fallback and uses the EFA runtime;
    • reserves the FI_* and NIXL_* environment variables.
• API example:
  • Uses the canonical per-pool min_replicas/max_replicas instead of the deprecated replicas (CentML/platform#4403).
  • Adds communications.kv_transfer.policy and a note on runtime_version.
• Advanced step:
  • Dynamo Runtime Version (CentML/platform#4477, #4425, #4602).
  • Priority class.
  • What extra arguments and environment variables apply to, and which names are reserved.
  • A tip on DYN_ROUTER_MODE, which defaults to kv.
• Storage volumes in the Hardware step.
• Monitoring:
  • The pod-count wording is now observed pods ready.
  • The Detail tab's contents are updated.
  • The Usage tab's inference and KV transfer charts are listed (CentML/platform#4657).
  • In-place scaling applies to aggregated Workers only.
• Removed: "NVIDIA CCluster evaluates GPU quota against Max Replicas". POST/PUT /deployments/dynamo has no quota check, and Dynamo pods run in KAI's default-queue.
• What's next: the Frontend card now points at the v1.5.1 Dynamo docs. The old URL redirects to the unreleased dev docs.
• Screenshots: the Advanced and Hardware step screenshots now show the current console.
• Managing Deployments:
  • For Dynamo, rollout settings and in-place scaling apply to aggregated deployments only.
  • The Logs section lists Prefill and Decode pods.

Section 6 (Call the endpoint) is unchanged, so this PR merges cleanly with #84, which fixes the 404 KServe link there. I checked with git merge-tree against #84's head.

Before merging

• Python SDK. The published centml 0.7.1 pins platform-api-python-client 4.31.0 (2026-09-23). In that client, DynamoWorkerPoolConfig.min_replicas is still ge=1, so it cannot read a Dynamo deployment with Min Replicas 0. CentML/platform#4784 asks for the SDK to be published before scale to zero is announced. Either publish it first or add an SDK note to this page.
• Live wake-up check. Confirm a live 1→0→1 cycle with the ingress wake trigger. CentML/platform#4801 merged before that cycle had been exercised on a shared cluster.
• Screenshots. dynamo-advanced-configuration.png and dynamo-hardware-scaling.png were re-captured from the current console, downscaled to 1516 px wide, and given updated alt text. The Endpoint screenshot is unchanged.

Validation

I ran the CI checks locally with the pinned Fern CLI 5.168.0:

• fern check: 0 errors. The one warning is the missing-redirects check, which needs FERN_TOKEN.
• fern docs md check: all 20 MDX files valid.
• fern docs broken-links --strict: all checks passed.
• python3 tools/check-docs-mdx docs/: OK.
• python3 tools/check-docs-paths --allow '../docs/' docs fern: OK.
• git diff --check: clean.
• fern docs dev: broken-links does not check in-page anchors, so I rendered the page. The new anchors (#aggregated-workers, #scale-to-zero, #disaggregated-serving, #understand-kv-transfer-transport) match heading ids, and both tables render.
• Both new v1.5.1 Dynamo links return 200.

Context

Each statement was checked against CentML/platform main: the API schemas and validators, the dynamo_deployment chart, and the console. CentML/platform#4784 lists this docs update as a follow-up.

For #84: the versioned KServe page, https://docs.nvidia.com/dynamo/v1.5.1/knowledge-base/modular-components/frontend/kserve-grpc-frontend, matches the default runtime. /dynamo/components/frontend redirects to the Frontend overview in the dev docs instead.

Two open platform PRs will need another docs update when they merge:

• CentML/platform#4844 runs two Frontend replicas by default, adjustable per deployment.
• CentML/platform#4862 makes pause scale to zero, and paused endpoints return 503.

🤖 Generated with Claude Code

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-85.docs.buildwithfern.com/ccluster