Skip to content

[AUTOTUNER] Use hosted default FlagTune Manifest URL - #1103

Open
k6699k wants to merge 38 commits into
flagos-ai:mainfrom
HenryRenYz:backend
Open

[AUTOTUNER] Use hosted default FlagTune Manifest URL#1103
k6699k wants to merge 38 commits into
flagos-ai:mainfrom
HenryRenYz:backend

Conversation

@k6699k

@k6699k k6699k commented Sep 4, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Add the FlagOS-hosted Manifest URL as the default FLAGTUNE_MANIFEST_URL.
  • Keep explicit FLAGTUNE_MANIFEST_URL values as overrides.
  • Preserve the ability to disable remote Manifest access with an explicit empty URL.
  • Update the runtime documentation to reflect the new default behavior.
  • Add regression coverage for the default URL and keep unrelated tests offline.

HenryRenYz and others added 30 commits July 20, 2026 01:12
Reorganize FlagTune into contract, runtime, and training namespaces; update public imports and tests to use the new boundaries; and remove the obsolete migration module.

Introduce an architecture-neutral event/replay benchmark protocol with complete cache identity and audit metadata. Extend driver capabilities so NVIDIA and AMD backends can expose replay timing without leaking device-specific APIs into callers.

Update the FlagTune autotuner to consume resolved benchmarkers, preserve event-cache compatibility, and add unit coverage for protocol validation, replay fallback, and backend capability behavior.
Apply the Code-Format-Check pre-commit output to FlagTune contracts, runtime, training, and their focused tests without changing behavior.

The resulting files pass the configured Ruff, YAPF, clang-format applicability, and mypy checks.
Replace gpu_key identity with platform_key while keeping architecture as device metadata. Add flat versioned platform packages, model migration without retraining, loading, caching, validation, and tests.
A platform package covers only the operators, variants, and dtype
combinations it was built for, so a package that resolves but carries no
bundle for one identity is a statement about coverage rather than a fault.
Integration layers need to tell that apart to fall back, but it previously
raised the same IncompatibleModelError as a corrupt archive, an
incompatible FlagTune version, and an identity mismatch.

ModelBundleMissingError is raised only for that case and subclasses
IncompatibleModelError, so existing handlers keep working.

Also name the cause when download-latest narrowed a cache lookup: asking
for the Manifest's newest package while forbidding downloads reports only
that this version is uncached, which sends the reader after the network
when an older cached package was simply not eligible. Behaviour is
unchanged; that pair remains a supported offline configuration.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants