Repository navigation
Studio: save a model's run settings without loading it - #214
danielhanchen wants to merge 1 commit into
Conversation
a4ca3ab to
3d6dcfc
Compare
|
@codex security review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3d6dcfc8a9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
3d6dcfc to
0a58dc3
Compare
|
@codex review |
1 similar comment
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0a58dc343f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| // settings the load path strips or has already captured. Forget stores nothing, so | ||
| // broken saved arguments must not lock it. | ||
| disabled={ | ||
| stagedMetadataPending || |
There was a problem hiding this comment.
Wait for GGUF classification before saving
The Save gate does not actually wait for classification in all cases: stagedMetadataPending is false while metadata is unresolved unless the draft has manual GPU placement, selected GPUs, or batch settings. If a user changes another diffusion-unsupported field such as Reasoning Budget, Tensor Parallelism, or Vision while a slow GGUF metadata request is pending, Save remains enabled; commitDraft consequently treats the model as non-diffusion and persists that field locally and to the API override instead of applying withoutUnsupportedDiffusionSettings. Gate Save on unresolved classifiedIsDiffusion itself, rather than this narrower metadata predicate.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Not changing this here: Load on main has the same stagedMetadataPending gate and runs the same persistConfig path, so an unclassified draft is already persisted by Load today; Save matches it. Gating on an unresolved classification instead would leave Save disabled for good when the metadata fetch never resolves.

Fixes #10168.
The run settings panel only let you save settings as part of loading the model. If you wanted to set up a model ahead of time, you had to load it every time.
This adds a Save settings button next to Load model. It stores the settings for that model, including the copy the API uses for its own loads, and leaves the model unloaded. The footer now wraps in the picker so the "Remember for this model" label is not cut off.
Tested in the browser against a fresh install of main and this branch. On main the button does not exist and nothing is saved. On this branch clicking it shows "Settings saved.", the server has the new settings, and the model stays unloaded.
Before / after
Run settings for an unloaded on-device GGUF (
gemma-3-270m-it-GGUFUD-Q4_K_XL), merge base7617fc398vs head332ed632e. Before, the footer only offers Load model and Reset. After, Save settings sits between them; clicking it shows "Settings saved.", stores the config, and no model is loaded.Mirror of unslothai/unsloth#10216 by @NilayYadav, opened for review only.
Do not merge here. Fixes agreed on this mirror must be applied to the source PR.
Source description digest at mirror time:
bad119a97e93