What happened?
docs/content/models/model-cache.md ("Storage prerequisites") lists two auto-provisioned sources:
The cache PVC needs a ReadWriteMany (RWX) StorageClass on the workload cluster. What the platform admin must set up depends on the cloud:
- GKE and EKS: auto-provisioned. Nothing for the admin to do.
- Existing: the admin sets up a
ReadWriteMany StorageClass on the cluster.
docs/content/platform/inference-cluster.md ("Cache storage") lists four, plus the Vultr caveat:
GKE (Filestore Enterprise), EKS (EFS), AKS (Azure Files), and Nebius (shared filesystem): auto-provisioned. Those classes are fixed; nothing for the admin to do.Vultr: none. VKE's built-in RWX class (Vultr File System) isn't usable on GPU nodes [...]
To be fair to the page: model-cache.md does link onward, immediately after its list, with "How storage is provided on each cluster source, and how to bring your own backend, is covered in Register a Cluster". So this is a summary pointing at the canonical list, not a contradiction, and I nearly didn't file it.
The reason I think it's still worth a small fix: the short list is written as an enumeration rather than as an example, so it reads as exhaustive. Someone landing on the ModelCache page to answer "can I use a cache on AKS?" gets a list of supported clouds that doesn't include AKS, and the answer is yes. The Vultr caveat, which is the one case where the answer is genuinely no, is also only on the other page.
Suggestion, since you already link to the canonical section: drop the per-cloud list from model-cache.md entirely and keep the requirement plus the link. That leaves one source of truth and nothing to drift.
Happy to send that PR if you want it. Closing this as working-as-intended is a perfectly reasonable outcome too.
How can we reproduce it?
Read the two sections side by side:
What environment did it happen in?
Docs only, main as of 2026-09-09.
What happened?
docs/content/models/model-cache.md("Storage prerequisites") lists two auto-provisioned sources:docs/content/platform/inference-cluster.md("Cache storage") lists four, plus the Vultr caveat:To be fair to the page:
model-cache.mddoes link onward, immediately after its list, with "How storage is provided on each cluster source, and how to bring your own backend, is covered in Register a Cluster". So this is a summary pointing at the canonical list, not a contradiction, and I nearly didn't file it.The reason I think it's still worth a small fix: the short list is written as an enumeration rather than as an example, so it reads as exhaustive. Someone landing on the
ModelCachepage to answer "can I use a cache on AKS?" gets a list of supported clouds that doesn't include AKS, and the answer is yes. The Vultr caveat, which is the one case where the answer is genuinely no, is also only on the other page.Suggestion, since you already link to the canonical section: drop the per-cloud list from
model-cache.mdentirely and keep the requirement plus the link. That leaves one source of truth and nothing to drift.Happy to send that PR if you want it. Closing this as working-as-intended is a perfectly reasonable outcome too.
How can we reproduce it?
Read the two sections side by side:
What environment did it happen in?
Docs only,
mainas of 2026-09-09.