fix(vram): persist remote probe metadata

[localai-org-maint-bot]

Description

This PR fixes #11483.

The gallery startup warmer previously rebuilt process-memory-only size and GGUF metadata caches after every restart. Successful HTTP(S) probes now persist for 24 hours under the LocalAI cache directory, so frequent restarts reuse them.

The persistent cache:

• bypasses local model files
• keeps at most 4,096 entries
• uses atomic, versioned records without storing source URLs
• rejects corrupt or incomplete records and falls back to the remote probe
• purges records when the gallery changes
• rejects stale in-flight writes that finish after invalidation

The VRAM management documentation now describes persistence, expiry, and the entry limit.

Notes for Reviewers

Verification:

• go test -count=1 ./pkg/vram
• go test -count=1 ./core/application
• go test -count=1 ./core/gallery -ginkgo.focus="refresh|cached gallery"
• go test -race -count=1 ./pkg/vram
• go vet ./pkg/vram ./core/gallery ./core/application
• git diff --check

The full core/gallery suite also ran 375 passing specs; three unrelated network-dependent specs failed because GitHub raw/gist requests return HTTP 403 in this environment.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable

[mmelvin0]

Does it make sense to either add an option to disable this entirely (at startup) or if the existing LOCALAI_AUTOLOAD_GALLERIES is set to false?

[mudler]

Does it make sense to either add an option to disable this entirely (at startup) or if the existing LOCALAI_AUTOLOAD_GALLERIES is set to false?

it does indeed make sense. @localai-org-maint-bot update the PR, incorporate the feedback, and also add a runtime toggle wired in the setting section of the webui

[localai-org-maint-bot]

Implemented and pushed in 998f40ec3.

• Added a Persist remote VRAM estimates runtime toggle under Settings > Galleries.
• Added --vram-persistent-cache / LOCALAI_VRAM_PERSISTENT_CACHE (default: true).
• autoload_galleries=false now suppresses both startup warming and persistent-cache setup.
• Runtime changes take effect immediately and keep in-memory estimation available when disk persistence is off.
• Updated the VRAM, runtime-settings, and CLI docs.

Verified with go test ./pkg/vram ./core/config, the focused Settings endpoint suite, focused UI ESLint, the inline-style gate, and a production UI build. The full endpoint package additionally needs rootless Docker for its unrelated node/Postgres specs.

[mudler]

@localai-org-maint-bot we want also a toggle to disable automatic loading and pre-warm on boot

[localai-org-maint-bot]

Implemented and pushed in 99c80d5d2.

The existing autoload_galleries control already gates both gallery loading and metadata pre-warming at boot. I renamed its Settings toggle to Load and pre-warm galleries on boot, clarified the behavior in the description and runtime-settings docs, and added a focused UI regression test.