Skip to content

fix(ui): setup shows the same model size and name as Settings - #8431

Open
kovtcharov wants to merge 1 commit into
amd:mainfrom
kovtcharov:fix/setup-tab-model-name-and-size
Open

kovtcharov wants to merge 1 commit into
amd:mainfrom
kovtcharov:fix/setup-tab-model-name-and-size

Conversation

@kovtcharov

Copy link
Copy Markdown
Contributor

Setup listed Muse Glimmer 30B as a 19.9 GB download while Settings said 21.4 GB, and Settings › Setup read "Running Qwen3.6-35B-A3B-GGUF · local" where every other screen says "Qwen3.6 35B A3B". Setup now shows the same size and the same name as the rest of the app.

Closes #8349
Closes #8356

Test plan

🔍 Technical details
  • Lemonade's catalog lists Muse Glimmer at 19.9 (GiB); GAIA's registry has 21.36 GB (weights + vision projector + draft). The plan took the server's figure first.
  • Only models outside setup's own set (the profile's models and DEFAULT_MODEL_LADDER) switch to model_choice.download_size_gb. Setup's own models keep the server's size, so the fit checks fix(ui): leave the setup plan's own size figures alone #8243 protected are untouched; Lemonade 2026.44.1 lists those at GAIA's figures already (23.3, 5.97).
  • Not looked at in the app: verified by tests only. The "Installed: n, size" total on the Setup tab sums the same catalog sizes, so it follows.

Setup listed Muse Glimmer 30B as a 19.9 GB download where Settings said
21.4 GB, and Settings > Setup read "Running Qwen3.6-35B-A3B-GGUF · local"
where every other screen says "Qwen3.6 35B A3B". Both now match.

Lemonade lists Muse Glimmer in GiB, and the setup plan took the server's
figure first. A model outside setup's own set (the profile's models and the
default ladder) now shows GAIA's figure, as the model picker and Settings
do. Setup's own models keep the server's size, which the plan's fit checks
run on (amd#8243); Lemonade lists those in the same units as GAIA.

The Setup tab printed a local model's label, which is its id plus an
untranslated "local". It now shows the display name.

Closes amd#8349
Closes amd#8356
@github-actions github-actions Bot added the tests Test changes label Nov 17, 2026
@github-actions

Copy link
Copy Markdown
Contributor

Verdict: Approve with suggestions

Setup now shows Muse Glimmer at the same 21.4 GB that Settings shows, and Settings › Setup names the running local model "Qwen3.6 35B A3B" instead of its raw id followed by "· local". The change is narrow and correct. Setup's own models keep the server's size, so their fit checks don't change. Each fix has a test.

  • Missing evidence (not blocking): this changes a screen, but the PR has no screenshot and no captured setup-plan response. The PR also says nobody has looked at it in the app. Before merge, attach a Settings › Setup screenshot in English and Arabic, or at least the plan's Muse Glimmer entry.

Real-world evidence

N/A: this job wrote no evidence bundle. The verdict rests on static review and the PR's unit tests. Neither the Settings › Setup screen nor the setup-plan route ran here.

🔍 Technical details

Issues

🟡 No real-world evidence for a UI-visible change. This needs either a screenshot of Settings › Setup, or the /api setup-plan response showing Muse Glimmer's full_size_gb: 21.36. The PR's last Test plan box is unchecked. Per CLAUDE.md, a UI change also needs npm run test:rtl and an Arabic look at the "Running …" line.

🟢 Models outside setup's own set now have their fit judged on GAIA's registry size (src/gaia/ui/setup_plan.py:449-452). full also feeds memory_fit and remaining_download_gb. For Muse Glimmer that means 21.36 instead of Lemonade's 19.9, so the memory need grows by about 7%. That is the conservative direction, and arguably more correct, because the registry figure includes the vision projector and draft model. Still, a model close to the limit could flip from fits to doesn't-fit on this change. Models missing from Lemonade's list already took this path through the old _registry_size fallback, so nothing new is mixed. Worth one line in the PR body so nobody is surprised.

Checked, no issue

  • _registry_size is still used at lines 136 and 603, so removing its call here leaves no dead code.
  • For a remote model with no label, the fallback is displayModelName(active.model). Previously it fell through to display_name and then the raw id, so cloud names are unchanged.
  • The "Installed" total adds up the same catalog sizes, so it updates along with the per-model size.

Strengths

  • The size fix reuses model_choice.download_size_gb, the figure Settings already shows, instead of adding a third source of sizes.
  • The backend test checks both cases: a model outside setup's set switches to the registry size, and DEFAULT_MODEL_NAME keeps 5.6.
  • The frontend test checks that the GGUF and · local text is gone, not only that the new name appears.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

tests Test changes

Projects

None yet

1 participant