Studio: redesign Select model dropdown to match Hub design - #6364
Conversation
Make the chat Select model picker easier to scan by reusing the Hub on-device card's visual language. - Rows now split owner/name, add a param chip, a DotTag format pill, a tabular size, and a Loaded marker on the active model. - Hub models / Fine-tuned tabs reuse the Hub's exact .hub-tab-toggle styling (selectors extended in hub.css to the selector menu). - Add a Downloaded / Recommended / Custom section toggle on the Hub tab to filter the list. - Widen the popover and nudge the scrollbar toward the edge.
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request refactors the model selector UI by replacing the standard tabs component with a custom segmented pill toggle (PillTabs) and introducing a new section-based navigation system for model picking. It also improves the row presentation for models by parsing metadata to display formatted tags, parameter counts, and owner information. The review feedback highlights an opportunity to improve the accessibility of the new PillTabs component by implementing WAI-ARIA keyboard navigation patterns, such as arrow key support and proper focus management.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
| {tabs.map((tab) => ( | ||
| <button | ||
| key={tab.value} | ||
| type="button" | ||
| role="tab" | ||
| aria-selected={value === tab.value} | ||
| onClick={() => onValueChange(tab.value)} | ||
| className={cn( | ||
| "relative z-10 inline-flex h-9 min-w-0 flex-1 items-center justify-center rounded-full px-3 text-[12.5px] transition-colors", | ||
| value === tab.value | ||
| ? "text-foreground" | ||
| : "text-muted-foreground hover:text-foreground", | ||
| )} | ||
| > | ||
| {tab.label} | ||
| </button> | ||
| ))} |
There was a problem hiding this comment.
To comply with WAI-ARIA tablist keyboard interaction patterns, only the active tab should be focusable via tabIndex={0} (with inactive tabs having tabIndex={-1}), and users should be able to navigate between tabs using the ArrowLeft and ArrowRight keys. Adding an onKeyDown handler and setting the appropriate tabIndex will significantly improve the keyboard accessibility of the segmented pill toggle.
{tabs.map((tab, index) => (
<button
key={tab.value}
type="button"
role="tab"
aria-selected={value === tab.value}
tabIndex={value === tab.value ? 0 : -1}
onClick={() => onValueChange(tab.value)}
onKeyDown={(e) => {
if (e.key === "ArrowRight" || e.key === "ArrowLeft") {
e.preventDefault();
const nextIndex = (index + (e.key === "ArrowRight" ? 1 : -1) + tabs.length) % tabs.length;
onValueChange(tabs[nextIndex].value);
(e.currentTarget.parentElement?.querySelectorAll('button[role="tab"]')[nextIndex] as HTMLElement)?.focus();
}
}}
className={cn(
"relative z-10 inline-flex h-9 min-w-0 flex-1 items-center justify-center rounded-full px-3 text-[12.5px] transition-colors",
value === tab.value
? "text-foreground"
: "text-muted-foreground hover:text-foreground",
)}
>
{tab.label}
</button>
))}
Put Downloaded / Recommended / Custom under the search bar in their own row so Hub models / Fine-tuned no longer wrap. The section toggle uses a smaller font and sizes each tab to its label instead of equal widths.
Move splitRepoLabel, classifyMetaToken, and parseMetaTokens out of pickers.tsx into row-meta.ts. No behaviour change; keeps the presentation logic free of React/DOM deps so it is easy to test in isolation.
|
Validated this end to end with isolated simulations before merging. Pure logic (52 cases, against the extracted Cross browser rendering
Static |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b612d3ccc8
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| // Section gating, ignored while searching. Downloaded stays visible while | ||
| // searching so a searched-for downloaded model doesn't disappear. | ||
| const showDownloaded = showHfSection || section === "downloaded"; | ||
| const showCustom = !showHfSection && section === "custom"; |
There was a problem hiding this comment.
When the user is on the new Custom section and types in the search box, showHfSection becomes true, which forces showCustom to false and removes all custom-folder rows from both the DOM and the roving option list. This means custom local models cannot be searched from the Custom tab; the picker instead switches to downloaded/HF results even if the query matches a custom model, despite the UI presenting the search as spanning the Hub sections.
Useful? React with 👍 / 👎.
| const [activeTab, setActiveTab] = useState<string>(() => | ||
| chatOnly ? chatOnlyTabsDefault : studioTabsDefault, | ||
| ); |
There was a problem hiding this comment.
Because activeTab is initialized only once, the selector can stay on hub even after the selected value is later found in loraModels or externalModels (both are populated asynchronously in the chat page). In that case reopening the dropdown for a persisted fine-tuned or connected model hides the currently loaded row until the user manually switches tabs, so the default should be derived again when the selected model/source list changes or when the popover opens.
Useful? React with 👍 / 👎.
| const effectiveTab = tabs.some((tab) => tab.value === activeTab) | ||
| ? activeTab | ||
| : tabs[0].value; | ||
| const [hubSection, setHubSection] = useState<HubSection>("downloaded"); |
There was a problem hiding this comment.
When the current checkpoint is a Hub model that is not in Downloaded (for example a recommended/search result before the cached lists catch up, or a restored selection that only exists in models), hubSection always starts as downloaded, so reopening the selector shows the empty/downloaded view and the new Loaded marker is absent until the user manually switches sections. The new section state needs the same selected-model derivation as the source tab.
Useful? React with 👍 / 👎.
Size the Hub models / Fine-tuned tabs to their labels (with side padding) like the section toggle, instead of stretching full width. Add a leading download, star, and folder icon to Downloaded, Recommended, and Custom.
The popover is a flex column, so the fit toggle stretched full width; add w-fit/self-start so it sizes to its content. Also hide the Fine-tuned tab when there are no fine-tuned models, defaulting to Hub models.
Local models (LM Studio, Ollama, custom folders) carry source "local" and already show in the Hub tab's Downloaded / Custom sections, so exclude them from the Fine-tuned tab and from its visibility count. Extract the tab rules into source-tabs.ts.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9113cd7f8e
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| // Ollama, custom folders) carry source "local" and live in the Hub tab's | ||
| // Downloaded / Custom sections instead. | ||
| const fineTunedModels = useMemo( | ||
| () => loraModels.filter((model) => isFineTunedSource(model.source)), |
There was a problem hiding this comment.
Filtering every source: "local" entry out of the Fine-tuned tab hides local models that the chat page still passes through loraModels (including models_dir entries and LM Studio entries outside chat-only mode). The Hub picker does not add those back: its local refresh only stores lmstudio and custom sources, and LM Studio rows are only rendered for chatOnly, so users with models under the local models directory can no longer select them from the dropdown after this change.
Useful? React with 👍 / 👎.
Show LM Studio and other local provider models in the Downloaded section in all modes (was chat-only). Put Recommended first and make it the default section. Add a little more space below the search bar.
Replace the static Recommended list (and its collapse chevron) with a sort dropdown over Unsloth's own models: Recommended, Trending, Most likes, Downloads, Recently updated. Recommended shows recently uploaded GGUF/MLX models that fit the device (hidden if they do not); the other sorts list all Unsloth models, badged but never hidden. Adds a sort option to useHfModelSearch and a pure recommended-fit helper.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4b9dd55011
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| const rows = recommendedSearch.results.filter( | ||
| (r) => !downloadedSet.has(r.id.toLowerCase()), | ||
| ); | ||
| if (recommendedSort !== "recommended") return rows; |
There was a problem hiding this comment.
In chat-only/macOS mode, switching the Recommended list to Downloads/Likes/Trending takes this early return and skips the runnable-format filter. useHfModelSearch returns all unsloth repos for those sorts, so regular safetensors/checkpoint repos that the rest of the picker still hides in chat-only become selectable and are sent to onSelect as Hub models, leaving chat-only users trying to load unsupported models; keep the chat-only GGUF/MLX filter even for alternate sorts.
Useful? React with 👍 / 👎.
| recommendedRows.map((r) => { | ||
| const id = r.id; | ||
| const info = recommendedMeta.get(id); | ||
| const isG = isKnownGgufRepo(id); |
There was a problem hiding this comment.
For a Recommended search result where Hugging Face marks r.isGguf true but the repo name does not match the -GGUF suffix, the row passes the filter via isRunnableRecommendedFormat(r.id, r.isGguf) but this check ignores that hint. Such rows render and click as plain Hub checkpoints instead of opening GgufVariantExpander, so users cannot choose/download a quantized file for tagged-only GGUF repos; carry r.isGguf into the isG decision.
Useful? React with 👍 / 👎.
| obs.disconnect(); | ||
| }; | ||
| }, [recommendedSentinel, hasMoreRecommended, recommendedPage, scrollRef]); | ||
| }, [recommendedSentinel, recommendedSearch.hasMore, recommendedSearch.fetchMore, scrollRef]); |
There was a problem hiding this comment.
When the Recommended tab has more results after the first fetchMore, the observer has already disconnected and this effect does not rerun because none of its dependencies change while hasMore stays true. That means the Recommended list can load the initial page plus one extra page, then further scrolling never fetches the remaining Hugging Face results; include a changing page/result/loading value in the dependencies or avoid permanently disconnecting.
Useful? React with 👍 / 👎.
…issing GGUF and MLX repos rarely expose safetensors metadata, so a large model with no size could pass the Recommended fit check because unknown size was treated as fitting. Parse the parameter count from the repo id, including the Gemma E series, and hide anything we still cannot size.
Thread tags and the pipeline tag through the model search results and add a pure helper that infers vision, reasoning and audio plus the architecture family, falling back to repo-name keywords when tags are absent.
Give each model row more detail and make the Hub sections easier to scan: - Show vision, reasoning and audio badges plus the architecture family tag on each row, alongside the params, format and size. - Drop the redundant unsloth/ prefix on the Recommended rows. - Rename the Recommended section tab to Unsloth and enlarge the section tabs. - Move the sort dropdown inline to the right of the tabs at a fixed width. - Add Recent, Size and Downloaded sorting to the Downloaded and Custom tabs. - Remove the header icons, pad the subheadings, and grow the list height.
- Recommended now lists the most recently created Unsloth repos. - Narrow the sort dropdown, remove its border, and truncate long labels. - Tighten the gap between the section tab icons and their labels. - Remove the architecture family tag from rows since it repeats the name.
Move the segmented pill toggle out of the model selector into its own file so the Hub picker can reuse it for a format filter without duplicating the markup.
- Re-attach the scroll observer on each loaded page so a filtered Recommended list keeps paging until the viewport fills instead of spinning forever with nothing new appearing. - Add an All / GGUF / MLX / Safetensors toggle on the Unsloth listing that filters every sort.
…ce, and fade the scroll edge Sort: default the Recommended view to Trending and add a Name option to the On Device / Custom sort. Recent now orders by last load time while Downloaded orders by file date, tracked in localStorage (model-usage.ts). Formats: show the format filter on all three tabs (Unsloth, On Device, Custom), exclude mobile GGUF builds from Recommended, and flag GGUF rows that exceed the device with the same OOM badge as safetensors. Polish: download-icon badge on already-downloaded Recommended rows, the hugeicons view stroke-rounded vision badge, Search all models placeholder, matched popover padding, and a top-edge mask fade once the list scrolls.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Repos with no <n>B token in the name (Kimi, MiniMax) had no param count and so never showed an OOM badge. Request the gguf expand field from Hugging Face and read gguf.total, so those repos get a param chip and an OOM badge when they exceed the device budget. Keep the row name full contrast when over budget (the OOM badge already signals the fit), shorten the format and sort dropdowns, narrow the popover, and rename Recently updated to Recent and All formats to All.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Add WAI-ARIA roving tabindex and Arrow Left/Right navigation to the pill toggle so only the active tab is in the tab order. Keep the chat-only GGUF/MLX filter for every Recommended sort, not just Recommended, so chat-only users do not see unrunnable checkpoints under Trending. Feed both listings' GGUF hints into repo detection so a tag-only GGUF in Recommended expands variants instead of loading as a checkpoint.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Went through the review feedback. Dispositions below. Fixed in 0952e12
Already addressed on the current head
Tracking separately
Validated the three fixes with isolated logic and cross-browser (Chromium, Firefox, WebKit) simulations: roving focus and Arrow nav, the chat-only filter across sorts, and the GGUF-hint detection. |
Search is now per section. The Unsloth tab searches the Unsloth HF listing only, On Device filters downloaded and LM Studio models by name, and Custom filters custom-folder models, each with its own empty state. MLX repos get an MLX pill mirroring the GGUF tag. Downloaded quants in the Unsloth and search lists get the same delete action as On Device. Also: revert the model name to normal weight, narrow the popover to 558px so the format and sort dropdowns sit one gap-2 from the tabs, tighten the dropdown menus to match the Projects activity Select, and make the empty On Device state name the active format filter.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Models under the local models directory (source models_dir) flow in as local models but were dropped from every list: filtered out of Fine-tuned and never re-added by the Hub picker, which kept only LM Studio and custom-folder sources. Capture them in the local refresh and render a Local models group on the On Device tab, with the same format, search, and chat-only GGUF rules as the other local groups.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 35c8ed46df
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".
| }).catch(() => { | ||
| const store = useChatRuntimeStore.getState(); | ||
| if (!store.pendingSelection) store.resetModelSettingsToLoaded(); |
There was a problem hiding this comment.
When an autoLoad pick finishes downloading but the subsequent selectModel call fails (for example OOM, gated repo, or validation error), selectModel only clears pendingSelection on success, so this catch usually sees the hidden auto-load stage still present and does nothing. That leaves an orphaned staged model/edited load knobs in the store even though the sheet was never opened; opening Run settings later shows the stale staged pick and those settings can leak into later loads. Clear or abandon the stage when the pending selection still matches this failed auto-load pick.
Useful? React with 👍 / 👎.
| const downloadedEmpty = | ||
| visibleCachedGguf.length === 0 && | ||
| visibleCachedModelRows.length === 0 && | ||
| sortedLmStudio.length === 0 && | ||
| sortedLocalDir.length === 0; |
There was a problem hiding this comment.
When the On Device tab has only fine-tuned models (or a search matches only fine-tuned rows), downloadedEmpty is still true because it ignores fineTunedRows. The empty-state block then renders “No downloaded models yet” or “No matching models on device” immediately above the actual Fine-tuned section, so users see a false empty message despite having selectable on-device models.
Useful? React with 👍 / 👎.
| // Per-model pre-load inference settings, persisted in localStorage so the load | ||
| // dialog can offer "Remember settings for <model>". | ||
|
|
||
| const KEY = "unsloth_load_settings"; |
There was a problem hiding this comment.
This new localStorage key is not included in GeneralTab’s PREFS_KEYS, so using “Reset all local preferences” leaves remembered per-model context/KV/speculative/tensor-parallel settings behind. After a reset, staging the same model can silently reapply those old load settings, which contradicts the reset action’s promise to clear local preferences.
Useful? React with 👍 / 👎.
| if entry.is_dir(): | ||
| if not entry.name.startswith("."): | ||
| queue.append(entry) | ||
| elif entry.name.lower().endswith((".gguf", ".safetensors")): |
There was a problem hiding this comment.
The new recommended-folder gate only treats .gguf and .safetensors as generic weights, but the local model scanners still support PyTorch .bin checkpoints such as pytorch_model.bin/model*.bin. When a well-known directory contains only those valid .bin weights, /recommended-folders now omits it, so the one-click folder chip disappears even though the folder has a loadable local model. Include the supported .bin weight names in this check.
Useful? React with 👍 / 👎.
- chat-page auto-load: selectModel only clears pendingSelection on success, so a failed auto-load left the hidden stage (and its edited load knobs) behind. Abandon the stage when it still matches the failed pick. - model picker: count fine-tuned rows in the On Device empty check so a fine-tuned-only tab no longer shows a false 'No models on device' message above the Fine-tuned section. - general settings: add the remembered per-model load settings key to PREFS_KEYS so 'Reset all local preferences' actually clears it. - recommended-folders: recognize PyTorch .bin weights (gated by the scanner's weight-name prefixes) so a .bin-only model folder still earns a chip; add tests.
|
Thanks, all four of these were valid and are addressed in 7b71d64:
Frontend typecheck/build/i18n and the backend folder-recommendation tests pass. |
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7b71d64d98
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".
| has_non_gguf = any(path.glob("*.safetensors")) or any(path.glob("*.bin")) | ||
| return None if has_non_gguf else "gguf" |
There was a problem hiding this comment.
When a suffixless GGUF directory also contains a non-weight .bin file such as tokenizer.bin, this treats it as having non-GGUF weights and returns None instead of "gguf". The frontend relies on this hint for LM Studio/custom folders whose names do not end in -GGUF; without it, those rows are labeled/filtered as plain local checkpoints and are sent through the wrong load path instead of expanding GGUF variants. The .bin check should mirror the weight-name gating used elsewhere rather than counting every .bin companion file.
Useful? React with 👍 / 👎.
| export const CHAT_EXPAND_QUANTIZATIONS_KEY = | ||
| "unsloth_chat_expand_quantizations"; | ||
| export const CHAT_SHOW_ALL_QUANTIZATIONS_KEY = | ||
| "unsloth_chat_show_all_quantizations"; |
There was a problem hiding this comment.
These new selector settings are persisted to localStorage, but the Reset all local preferences path only clears keys listed in GeneralTab's PREFS_KEYS, and these two keys are not included there. After a user toggles Expand quantizations or Show all quantizations, using Reset all local preferences leaves the model selector behavior changed instead of returning to defaults.
Useful? React with 👍 / 👎.
| (!hasGgufSource(selection) && !wantManagerDownload) || | ||
| (store.loadOnSelection && selection.isDownloaded) |
There was a problem hiding this comment.
When Load on selection is off, an uncached non-GGUF Hub repo now falls through this branch to stageModel so the settings sheet opens during the manager download. If a GGUF model is currently loaded, ChatSettingsPanel still derives its GGUF controls from the loaded model, so the staged safetensors/MLX pick shows stale context/KV/speculative knobs and can snapshot those values on Load; the staged pick should suppress loaded-model GGUF controls or avoid opening that sheet path for non-GGUF downloads.
Useful? React with 👍 / 👎.
| // Connected when an external model is active, else On Device when the | ||
| // user has downloads, else their last section. | ||
| setHubSection(wantsConnectedDefault ? "connected" : defaultHubSection()); |
There was a problem hiding this comment.
After moving fine-tuned models into the Hub section, reopening the selector while a fine-tuned model is active no longer routes to where that active row lives. With no persisted On Device choice (or on a fresh profile), this falls back to Recommended, so the currently loaded fine-tuned model and its delete/reload controls are hidden until the user manually switches to On Device; the old dedicated LoRA tab did this routing via studioTabsDefault.
Useful? React with 👍 / 👎.
| return sortedCachedGguf.filter((c) => | ||
| normalizeForSearch(c.repo_id).includes(q), | ||
| ); |
There was a problem hiding this comment.
When the user types a search in On Device, this branch filters cached GGUF rows only by text and stops applying formatFilter. As a result, selecting Safetensors (or MLX) and then searching still shows matching GGUF downloads, while the no-query path correctly hides them; include matchesFormatFilter in the search branch too so the format dropdown remains consistent.
Useful? React with 👍 / 👎.
…nce reset Follow-up to the codex review on the model_format/recommended-folder paths: - _dir_model_format and _scan_models_dir treated any .bin (incl. tokenizer.bin) as a non-GGUF weight, so a suffixless GGUF folder shipping a companion .bin was misclassified as a plain checkpoint and routed through the wrong load path. Factor the scanner's weight-name gating into shared _is_weight_bin / _has_non_gguf_weights helpers and use them everywhere (also in _dir_has_downloaded_model). - PREFS_KEYS was missing the new 'Select model settings' keys (load on selection, expand/show-all quantizations), so 'Reset all local preferences' left them set. - On Device cached search dropped the active format filter while a query was typed; keep matchesFormatFilter applied so the format dropdown stays consistent. Adds tests for the tokenizer.bin vs weight-.bin classification.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Triaged this round; three were addressed in b1abe33 and two are flagged for the staging/selector work. Fixed
Flagged for the staging/selector work (not changed)These are real but sit inside the actively-developed deferred-load and selector-redesign code and are UX/design calls, so leaving them to the author:
Frontend typecheck/build/i18n and the backend model_format/recommended-folder tests pass. |
for more information, see https://pre-commit.ci
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 73a3f62890
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".
| if ( | ||
| latest?.autoLoad && | ||
| (latest.ggufVariant ?? null) === (variant ?? null) | ||
| ) { | ||
| onAutoLoadRef.current?.(latest); |
There was a problem hiding this comment.
When an auto-load download is replaced by another staged pick with the same quantization, this completion handler can act on the new pendingSelection even though the callback belongs to the old repo: abandonStaged() cancels asynchronously and immediately stages the next pick, while the old listener remains until React re-renders. If the old job completes in that window, the variant-only check passes and onAutoLoad tries to load the new repo before its own download has finished. Please also verify the repo/model id associated with the completed job before auto-loading the latest pending selection.
Useful? React with 👍 / 👎.
| if m.is_file(): | ||
| return True |
There was a problem hiding this comment.
When an Ollama directory has manifest files but its blobs/ directory is empty or missing the referenced layer blobs (for example after a failed or pruned download), this returns true as soon as it sees any manifest file. _scan_ollama_dir only emits a selectable model after resolving manifest layers to existing blob files, so /recommended-folders can still show a recommended chip that opens to no models—the empty-scaffold regression this helper is intended to avoid. Please verify that at least one referenced blob exists before returning true.
Useful? React with 👍 / 👎.
| @@ -473,9 +656,10 @@ function GgufVariantExpander({ | |||
| ggufVariant: quant, | |||
| isDownloaded: isLocalPath ? true : downloaded, | |||
| expectedBytes: sizeBytes, | |||
| contextLength: nativeContext, | |||
There was a problem hiding this comment.
When a repo already has one GGUF variant downloaded, /gguf-variants returns a context_length, so this now attaches contextLength to clicks for every variant, including ones with downloaded === false. The staging effect exits on pendingHasContext before calling startDownloadRef, so choosing an undownloaded quant from that partially cached repo (especially with Load on selection enabled, where the sheet stays closed) never starts the download and never auto-loads. Only seed the context for already-local/downloaded picks, or make pendingHasContext skip only the metadata fetch, not the download.
Useful? React with 👍 / 👎.
| if (!gpuData?.available || !gpuData.devices?.length) { | ||
| // No discrete GPU (e.g. Mac): still surface system RAM so memory math | ||
| // (unified memory) has a budget to work with. | ||
| const info: GpuInfo = { ...DEFAULT_GPU, systemRamAvailableGb: ramAvailableGb }; |
There was a problem hiding this comment.
When /api/system reports no discrete GPU (for example Mac/unified-memory hosts), this branch now returns a nonzero systemRamAvailableGb while leaving available: false, but the GGUF variant fit path still gates on gpu.available/gpuGb and returns fits before considering system RAM. As a result those hosts see every quantization as fitting (and no OOM/tight ordering) even though the Recommended filter uses the RAM budget. Please either pass/use systemRamAvailableGb for no-GPU hosts or adjust the fit classifier so the new RAM value is honored.
Useful? React with 👍 / 👎.
…on no-GPU hosts - recommended-folders: only count an Ollama dir once its manifest resolves to an on-disk model blob, so a failed/pruned pull no longer surfaces an empty chip - GGUF variant click: only seed the staged contextLength for already-downloaded picks, so choosing an undownloaded quant from a partially cached repo still starts its download (the staging effect short-circuits on a known context) - device fit: classify GGUF variants against the system-RAM budget on no-GPU / unified-memory hosts instead of reporting everything as fits, and pass systemRamGb to every variant expander regardless of gpu.available
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Pushed Fixed
Flagged (not fixed here)
Scope confirmation (re: unsloth-only) The chat model dropdown and its Recommended section are unsloth-only: the picker calls the search hook with
|
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5ad8b38c7c
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".
| pinUnslothFirst: true, | ||
| keepUnsupportedTags: true, | ||
| accessToken, | ||
| enabled: online, |
There was a problem hiding this comment.
When the picker is on the On Device or Connected section, typing in the search box still enables this Hugging Face search (and the empty recommendedSearch below is also enabled), even though those results are discarded unless section === "recommended". In the On Device flow this produces unnecessary network requests/spinner state for every local-only search and breaks the intended offline/local behavior; gate the HF hooks on the Recommended section or pass an empty/inactive query for other sections.
Useful? React with 👍 / 👎.
| expectedBytes: selection.expectedBytes, | ||
| nativePathToken: selection.nativePathToken, | ||
| isGguf: selection.isGguf, | ||
| isHubRepo: wantManagerDownload || undefined, |
There was a problem hiding this comment.
When Load on selection is off and an uncached non-GGUF Hub model is selected, wantManagerDownload stages it here, but the settings sheet still computes isGguf as isLoadedGguf || pendingIsGguf. If a GGUF model is currently loaded, the staged safetensors/MLX pick opens Run settings with GGUF-only context/KV/speculative controls from the loaded model; either do not route non-GGUF picks through this staged GGUF settings path or make the sheet prefer the pending model's type while a pending selection exists.
Useful? React with 👍 / 👎.
| type="button" | ||
| onClick={(e) => { | ||
| e.stopPropagation(); | ||
| useChatRuntimeStore.getState().stageModel({ |
There was a problem hiding this comment.
When this action is rendered inside the Compare pane model selectors (they reuse ModelSelector via GeneralCompareHeader), clicking the gear for a downloaded quant still calls the single chat runtime stageModel here and opens/stages the global Run settings instead of configuring that compare pane. In Compare mode this leaves the selected left/right model unchanged and mutates the main chat pending model; hide this action outside the main selector or route it through a pane-scoped handler.
Useful? React with 👍 / 👎.
| sortLocalModels( | ||
| localDirModels.filter( | ||
| (m) => | ||
| (!chatOnly || localModelIsGguf(m)) && |
There was a problem hiding this comment.
In chat-only Mac builds this predicate keeps only GGUF entries from ./models, so a local MLX model is removed before the MLX format filter can match it, even though the new selector now treats MLX as runnable/recommendable on Mac. Users who place or download an MLX model under the local models directory cannot reselect it from On Device; include an MLX check (or a backend format hint) alongside localModelIsGguf.
Useful? React with 👍 / 👎.
…, keep local MLX on Mac - model picker: only run the Hub search hooks on the Recommended section. On Device / Connected render local data, so typing there no longer fires HF requests or a spinner and the local/offline flow is preserved - chat settings: when a pick is staged, decide the GGUF-only controls from the staged model's type, not the currently loaded model's. A staged non-GGUF Hub repo no longer inherits a loaded GGUF's context/KV/speculative controls - On Device: keep local MLX builds in ./models selectable on Mac (chat-only ran GGUF/MLX only, but the filter dropped MLX before the format toggle)
|
Pushed Fixed
Flagged (not fixed here)
|
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
PR #6364 added a branch that overrode the unsloth owner avatar with the bundled circle-logo-small.png sticker. Revert it so unsloth uploads use the live Hugging Face org avatar again, falling back to the colored initial tile. Upstream re-uploads are unaffected since they still resolve through provider logos. Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
4 participants
What
Reworks the chat Select model dropdown so it is easier to scan and lines up with the Hub page's design language. The dropdown now carries the same detail you get on the Hub, with per-tab search, per-section sorting, a format filter and clearer model metadata.
Rows
owner / nameso the model name reads as the primary text and the uploader is de-emphasized.DotTagformat pill: GGUF, MLX and Safetensors repos each get a coloured tag, adapters their own. Safetensors rows now show param, the Safetensors pill and size so their meta matches GGUF and MLX.Loadedmarker.OOMbadge alone signals the fit.Search and sections
.hub-tab-togglestyling (thehub.cssselectors are extended to also match.unsloth-model-selector-menu)../models) now show in a Local models group on On Device so they stay selectable, with the same format, search and chat-only GGUF rules as the other local groups.Unslothwhen other providers (LM Studio / Ollama) also appear.Recommended
Recommendedkeeps the most recently created GGUF and MLX models that fit the device budget (0.7 * GPU + 0.7 * RAM); the other sorts list everything with the existing fit badges.ggufexpand field from Hugging Face and readgguf.totalfor the param count. That sizes repos with no<n>Btoken in the name (Kimi, MiniMax) so they get a param chip and an OOM badge instead of slipping through unsized.unsloth/prefix is dropped on these rows.Accessibility and polish
gap-2from the tabs. The two dropdown menus match the Projects activity Select (tight padding, 14px radius). The Search Hub button matches the dropdown width and lines up with the Trending dropdown on the right, and the list scrollbar sits inside the box.No downloaded MLX models yet.Search models.Fine-tuned models
Load settings and staging
Load on selectionpreference lives in Settings, Chat (not the popover). On: Unsloth auto-picks the best settings and loads on selection. Off: picking a model opens Run settings to customize first.Select model settings (Settings, Chat)
Load on selection(default on).Expand quantizations(default off): expand every On Device GGUF model's quantizations by default, or keep them behind a click. With it on, clicking a model collapses or re-expands just that model; the collapse state is in memory only, so it resets on reload and when the setting is toggled.Show all quantizations(default on): On Device only. List every quant including ones not downloaded, or downloaded only. Recommended and other browse lists always show every quant.Both options affect On Device models only, and the copy says so.
Settings layout
Chat with Files (RAG).Picker polish
Notes
Visionbadge on the model name; Recommended keeps it inline with the variant list.role="tab"./kv-cache-estimateroute, and a quant GGUF resolver) and gates the recommended-folder chips on folders that actually hold a downloaded model. No existing route's contract changes and model loading, training and hardware detection are untouched.Testing
tsc -btype checks clean andvite buildsucceeds;i18n:checkpasses parity../modelssurfacing, GGUF sizing and OOM, the chat-only filter, format detection, row metadata) and a cross-browser Playwright harness (Chromium, Firefox, WebKit) for the rendering, roving keyboard nav, scroll fade and localStorage behaviour all pass.