Skip to content

fix(gallery): place context options at the root - #12590

Merged
mudler merged 1 commit into
masterfrom
fix/gallery-context-placement
Oct 9, 2026
Merged

mudler merged 1 commit into
masterfrom
fix/gallery-context-placement

Conversation

@mudler-agent

Copy link
Copy Markdown
Collaborator

Description

Closes #12528.

Gallery installation of qwen3-vl-4b-thinking uses gallery/qwen3.yaml. That template placed context_size, f16, and mmap under parameters. The installer preserved that nesting, but the model loader ignored those fields and applied the global context default.

  • Move the Qwen3 template options to the model root, preserving their values.
  • Correct context overrides for Supra2 and both Shieldstral entries.
  • Add checks against the real template and gallery index, including generated configuration for the reported Qwen3 model.
  • Document repair of already-installed YAML. Template updates do not rewrite existing installations.

Related: #12529 documents placement and precedence; this PR fixes the gallery producer. No runtime precedence changes or automatic migration are included.

Notes for Reviewers

The new regression tests failed with three intended placement failures before the data corrections. After the fix, 41 selected specs and the standard package tests passed. The verification command was independently rerun successfully:

go test ./core/gallery -run Test -ginkgo.focus='gallery config placement|gallery variant lint helpers|gallery/index.yaml variant invariants|Inference defaults persisted at install|InstallModelFromGallery with an empty base config' -count=1

git diff --check passed. Independent code review found no blockers. The installation regression uses placeholder weight files and does not download models.

Full builds, the complete test suite, and hardware inference were not run. Activating the existing explicit context values can change KV-cache memory use. Moving mmap: true to the root also makes that declared setting effective; Intel hardware behavior was not tested.

Signed commits

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

The Qwen3 template nests context_size, f16, and mmap under parameters,
where the typed model configuration ignores them. Move these options to
the root and correct three model-specific context overrides.

Test the real gallery data and generated qwen3-vl-4b-thinking YAML.
Document manual repair because template updates do not rewrite installed
configurations.

Fixes #12528

Assisted-by: nib:gpt-6-astra
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
@mudler
mudler merged commit 8559589 into master Oct 9, 2026
71 checks passed
@mudler
mudler deleted the fix/gallery-context-placement branch October 9, 2026 08:16
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Model-specific context_size ignored, global Default Context Size always applied

2 participants