Look up variant-qualified concrete model paths

When the settings UI saves a multi-quant GGUF whose target.id is a concrete local/snapshot path, it stores the override as path:VARIANT. Auto-switch resolves that model to a public override_id plus the concrete target_id, but this loop checks only override_id:VARIANT, override_id, and the unqualified target_id. Consequently the saved settings are silently ignored for local folders and models in non-active Hugging Face caches; include target_id:variant in the lookup.

Useful? React with 👍 / 👎.

Preserve the partial-download state when loading settings

When settings are opened for a partial inventory row, target.meta.isDownloaded is deliberately set to false, but this load call hardcodes it to true. The model runtime therefore treats the operation as a cached load, skips download-progress polling, and shows “Starting model” while the remaining weights are still downloading; pass through target.meta.isDownloaded so partial downloads retain their progress and phase reporting.

Useful? React with 👍 / 👎.

Preserve bare-repo launch flags when creating quant overrides

When a pre-existing override stores llama_extra_args under the bare repo ID, the first settings save for a GGUF now uses repo:VARIANT; this exact-key lookup finds no flags, creates a nonempty variant entry without them, and auto-switch then prefers that entry over the bare fallback. Since the new settings UI cannot display or resend launch flags, saving any per-quant setting silently disables the user's legacy flags; carry them over from the bare-repo entry when the qualified entry does not yet exist.

Useful? React with 👍 / 👎.

Build the displayed API URL from the reachable backend base

In the packaged Tauri app the backend is reached through getApiBase()/the dynamically selected localhost port, while window.location.origin is the Tauri webview origin rather than the API server. The monitor therefore displays and copies an unusable tauri://…/v1 URL (and can also miss the fallback port); derive this value from the platform server URL or getApiBase() on desktop, as the existing Agents settings tab does.

Useful? React with 👍 / 👎.

Gate the loaded-quant preference by model identity

When another GGUF model is currently loaded, this selection still prefers its active quant for the row being opened. Because common quants such as Q4_K_M appear in many repositories, opening model B's settings while model A uses Q4_K_M silently targets B's Q4_K_M instead of B's configured default, so subsequent saves and loads affect an unintended variant; only prefer activeGgufVariant when activeCheckpoint matches this row.

Useful? React with 👍 / 👎.

Revalidate saved GPU IDs before auto-switch loads

These IDs are validated only as integers in a broad numeric range, so an override saved before a GPU removal, visibility-mask change, or move to another host retains IDs that are no longer available. On the next API auto-switch, _validate_requested_gpu_ids calls resolve_requested_gpu_ids and turns that stale field into a 400, failing the entire model load rather than dropping only the unusable setting as intended; reconcile the stored IDs with the current backend devices at load time and omit the field when invalid.

Useful? React with 👍 / 👎.

Fully validate saved Vulkan GPU ordinals

On a Vulkan backend, resolve_requested_gpu_ids(..., is_vulkan=True) only checks for duplicate or negative ordinals, so a saved ordinal from another host still returns True here. _resolve_gguf_gpu_ids_for_request later probes the actual Vulkan devices and raises a 400 when that ordinal is absent, causing the auto-switch load to fail instead of dropping the stale field. The fresh evidence beyond the existing stale-GPU finding is this mismatch between the new reconciliation helper and the later Vulkan-specific probe; use the same device probe before retaining the override.

Useful? React with 👍 / 👎.

Distinguish saving defaults from forgetting the model

When the user keeps “Remember for this model” enabled but the effective configuration is entirely default, toApiOverride(config) returns {} while config !== null, so this sends only model_id. The backend explicitly interprets that payload as removal, which deletes any pre-existing llama_extra_args even though the UI cannot display or edit them; the fresh unhandled path beyond the earlier launch-flag finding is a non-null default save, rather than a save containing another field. Preserve a distinct save/remove signal so only the explicit config === null path clears those flags.

Useful? React with 👍 / 👎.

Migrate existing per-model settings to the server

For users upgrading with entries already stored in unsloth_model_configs, the only repo-wide call to syncModelOverride is this action handler, so none of those existing settings reach the new backend override map until the user opens every model and explicitly saves or loads it again. An API auto-switch immediately after upgrade therefore loads those models with app defaults despite the UI still showing them as remembered; add a one-time reconciliation of existing local entries when the authenticated frontend starts or when the new feature is initialized.

Useful? React with 👍 / 👎.

Stop polling after automatic monitor opening is disabled

When autoOpen is false and the overlay is closed, this effect still starts an immediate request and continues calling /api/inference/monitor every five seconds indefinitely. Since the panel cannot open or display those results in that state, opting out leaves every open Studio window generating permanent authentication, network, and backend polling load; short-circuit the effect when both !autoOpen and !isOpen, and include autoOpen in its dependencies so toggling the preference actually stops or restarts the loop.

Useful? React with 👍 / 👎.

Preserve original model-key casing during backfill

Existing v2 localStorage keys are produced by modelStorageKey after normalizing both repo IDs and GGUF variants to lowercase, and listPerModelConfigs returns those normalized values, so this call persists keys such as org/model:q4_k_m. Auto-switch performs exact dictionary lookups using resolver casing such as Org/Model:Q4_K_M, meaning the migrated entry is never found and upgraded users still load with defaults. This is fresh evidence beyond the prior missing-migration comment: the newly added migration writes keys that cannot match; preserve the original identifiers or normalize both sides of the server lookup.

Useful? React with 👍 / 👎.

Retry detail revisions skipped during an in-flight fetch

When the selected request advances to a new revision while its previous detail fetch is still running, requestDetail rejects this call through its in-flight guard, but lastFetchedRef has already recorded the new revision. If that race occurs on the terminal update, the older fetch can return a running or otherwise stale payload, updated_at never advances again, and this effect permanently skips the terminal detail, leaving only the truncated row preview visible. Record the revision only when a fetch actually starts or queue a retry after the in-flight request settles.

Useful? React with 👍 / 👎.

Ignore stale asynchronous settings opens

Opening settings for a variant-required GGUF waits for listGgufVariants, but there is no request identity or cancellation check before committing this target. If the user opens row A and then row B while A's lookup is still pending, whichever lookup finishes last overwrites settingsTarget; the page can therefore display, save, or load settings for A after the user most recently selected B. Track the latest requested row and discard results from older lookups before calling setSettingsTarget.

Useful? React with 👍 / 👎.

Accept loader-valid local paths as override keys

Local inventory targets use their concrete filesystem path as model_id, optionally with a quant suffix, but this new 512-character limit is stricter than LoadRequest.model_path, which has no corresponding bound, and valid filesystem paths can be substantially longer. For such a model the local settings save succeeds, while the best-effort server sync receives a 422 and only logs a warning, so the UI continues to show the configuration as remembered even though API auto-switch loads it with defaults. Remove this arbitrary limit or align it with the actual loader and request-size constraints.

Useful? React with 👍 / 👎.

Honor remove even when config fields are present

When a client sets remove: true while reusing a payload that still contains fields such as max_seq_length, is_removal only disables the launch-argument carry-over; these fields are still passed to set_model_override, so the entry is re-created and the endpoint returns success without removing it. Treat an explicit removal as an unconditional empty override rather than forwarding the remaining payload.

Useful? React with 👍 / 👎.

Reject Vulkan diffusion GPU pins in the preflight

When a DiffusionGemma override with saved gpu_ids is replayed on a Vulkan host, this preflight returns true as long as the ordinals appear in the probe. _resolve_gguf_gpu_ids_for_request then unconditionally rejects GPU selection for Vulkan diffusion models, so the auto-switch load still fails with 400 instead of dropping the incompatible saved field as intended for cross-host settings.

Useful? React with 👍 / 👎.

Retry detail revisions skipped by the in-flight guard

When a selected streaming entry already has cached details, a new updated_at revision can arrive while the preceding detail fetch is still running. requestDetail then returns false, but completing that fetch changes neither selectedIsMissing nor any other effect dependency, so if this is the terminal revision it is never fetched and the panel remains on the truncated row preview. Trigger a retry when the in-flight request finishes or track the skipped revision explicitly.

Useful? React with 👍 / 👎.

Run the desktop backfill after authentication is ready

In the Tauri app, this RootLayout effect runs before AppProvider has finished starting the backend and performing tauriAutoAuth; if the backend remains unavailable beyond authFetch's startup retries, backfillModelOverrides returns with its done flag unset. Since isAuthFlowRoute never changes when desktop authentication later succeeds, the effect does not retry, so upgraded desktop users' existing per-model settings are not mirrored during that session and API loads continue using defaults.

Useful? React with 👍 / 👎.

Serialize writes for each model override

Each save launches an independent full-replacement PUT without sequencing it against earlier saves for the same model. If a user applies two edits quickly and the first request completes after the second, its older payload becomes the final server value even though localStorage contains the newer config, causing subsequent API-triggered loads to use stale settings until another save succeeds. Queue or supersede writes per override key so completion order cannot reverse the user's changes.

Useful? React with 👍 / 👎.

Preserve case sensitivity for concrete local paths

When a concrete POSIX path has no exact override, this fallback can return the settings for a different file whose path differs only by case. For example, an override for /models/foo.gguf is applied when the resolver loads /models/Foo.gguf if only the former entry exists, potentially replaying an incompatible context, chat template, or GPU configuration. The fresh evidence beyond the earlier migration-casing issue is that the fallback is also used for concrete filesystem targets, where casing identifies distinct models; restrict case folding to repository/quant identifiers rather than local paths.

Useful? React with 👍 / 👎.

Do not parse arbitrary POSIX colons as quant separators

On POSIX, a valid unqualified local model path can contain a colon, such as /models/foo:bar.gguf. This parser treats the suffix as a quant because it is short and contains no separator, so the first settings save can look up /models/foo and silently graft that unrelated model's llama_extra_args onto the local model. Only perform the bare-key carry-over when the caller actually supplied a GGUF variant or validate the suffix as a known quant.

Useful? React with 👍 / 👎.

Detect API traffic present in the initial snapshot

If the first external API request starts after the Studio page mounts but before this initial monitor fetch returns, the first snapshot is already nonempty and its request ID is seeded as history. Subsequent polls see the same ID, so the advertised floating monitor never opens for that request, even if it remains active or completes; it opens only when another request arrives. Seed existing history separately from requests that are live during initialization, or establish the baseline before starting the polling request.

Useful? React with 👍 / 👎.

Done. The first snapshot seeds finished entries only:

seenIdsRef.current = new Set(
  data.entries.filter((entry) => entry.status !== "running").map((entry) => entry.id),
);

A request still running when that snapshot lands started while Studio was loading, so it is unseen live traffic rather than history, and the seeding branch falls through to the open check when any such entry came in via an API key.

Restrict API-backed settings to auto-switchable models

This exposes the settings action for every downloaded non-dataset row, including Transformers/safetensors models, and the resulting page says and persists that configuration as settings used by API loads. However, _maybe_auto_switch_model delegates to resolve_local_gguf, whose index explicitly excludes Transformers/safetensors models, so when another model is loaded an API request naming such a row does not load it or apply these settings and can instead be served by the current model. Gate the API-load claim/server mirror to GGUF rows or extend the resolver to support the other formats.

Useful? React with 👍 / 👎.

The mirror is already GGUF gated. model-config-page.tsx writes the local config for any model but only calls syncModelOverride when target.isGguf, for the reason given: the auto-switch resolver indexes GGUFs, so a safetensors entry would advertise settings on the applied-on-API-load list that no request can apply.

Keeping the row action ungated is intended. A safetensors model is still configurable by the picker, and its settings are honoured on a normal load. Only the server mirror is meaningless for it, and that is where the gate sits. There is no API-load claim in the page copy to gate.

Added test_only_gguf_configs_are_mirrored_to_the_server in 6ad8cb6 so the gate cannot be removed silently.

Drop saved Vulkan diffusion GPU pins before loading

When auto-switching a DiffusionGemma GGUF on a Vulkan build whose saved override contains gpu_ids, this helper returns true whenever the ordinals exist. _load_model_impl then reaches _resolve_gguf_gpu_ids_for_request, whose _classify_diffusion_gguf(config) is True branch rejects every Vulkan GPU selection with HTTP 400, so the saved setting prevents the API load instead of being discarded as intended. Include model classification in this reconciliation and remove the pin for Vulkan diffusion models.

Useful? React with 👍 / 👎.

Not reproducible: this path already succeeds. The retry directly below the pre-flight catches a 400 whose detail mentions gpu, drops the pin and loads again, and the Vulkan diffusion rejection is exactly the case it was written for. test_load_retries_without_gpu_ids_when_the_loader_rejects_the_pin in test_openai_auto_switch.py asserts it, with the diffusion refusal as the injected error.

The pre-flight deliberately does not classify the model: _classify_diffusion_gguf needs a ModelConfig that is not built at that point, and duplicating loader rules there means they drift. What was wrong is the docstring, which claimed the pre-flight mirrors every check the loader makes. Corrected in 6ad8cb6 to say it covers the device-availability rules (the ones that go stale) and that the retry covers the rest, including rules added later.

Avoid saving required-variant settings under a bare key

When variant discovery fails for a downloaded GGUF row that requires a variant, this fallback still opens the editor with ggufVariant = null. Saving then writes the browser config under the bare model ID, but resolveInitialConfig matches variants exactly, so a later picker load for Q4_K_M, for example, will not find these settings; meanwhile the backend auto-switch deliberately falls back to the bare override and does apply them. This recreates a UI/API configuration mismatch, so either resolve a concrete variant before opening or prevent saving until one is available.

Useful? React with 👍 / 👎.

Done, and taken further than suggested. When a required variant cannot be resolved the editor does not open at all: it shows an error toast and returns, rather than opening against a bare key. Saving under a key the picker matches exactly but the API falls back to is the mismatch worth avoiding, so refusing to open is the safer half of the either/or.

Invalidate pending row lookups when opening detail settings

If a row's listGgufVariants request is still pending, opening settings from another model's detail card does not increment settingsOpenSeq. The older row request therefore still passes the guard at line 1287 when it finishes and replaces settingsTarget with the wrong model. Increment the shared sequence in this entry point before setting the target so every new settings open invalidates earlier asynchronous lookups.

Useful? React with 👍 / 👎.

Done. settingsOpenSeq.current += 1 is in the detail-card handler as well as in openModelSettings, so an open from either entry point invalidates a lookup still pending from the other.

Make the covered detail view inert

When settings are opened from a model detail view, this higher-z-index overlay visually covers the detail pane, but only the catalog sibling is made inert. The still-mounted HubDetailView remains in keyboard focus order and the accessibility tree, so tabbing past the settings controls can operate or announce controls hidden behind the overlay. Mark the detail pane inert while settingsTarget is set, or unmount it while the settings page is displayed.

Useful? React with 👍 / 👎.

Done. Both detail wrappers carry inert={settingsTarget !== null || undefined}, alongside the existing inert={catalogCovered || undefined} on the catalog sibling.

Retry a detail revision rejected by the in-flight guard

When a selected request already has cached details and a newer list revision arrives while its previous detail fetch is still in flight, requestDetail returns false and this effect records no retry. The eventual fetch completion changes details but leaves selectedIsMissing false, so none of the effect dependencies change; if that newer revision is the terminal one, no further fetch starts and the full final reply remains stale or truncated indefinitely. Track the rejected revision and retry after the in-flight request settles.

Useful? React with 👍 / 👎.

Done. The effect tracks the last revision a fetch actually started for, and when requestDetail refuses it schedules a 250 ms nudge through retryTick rather than recording the revision. Nothing else in the dep list moves when the in-flight fetch settles, which is the gap described.

Make explicit removal override supplied config fields

For a direct API request that sends remove: true together with any config field, is_removal only disables launch-flag carry-over; all supplied fields are still passed to set_model_override, so the entry is updated rather than removed. Since remove was added as the explicit operation discriminator, callers reasonably expect it to clear the entry regardless of stale form fields in the payload. Short-circuit to an empty override when payload.remove is true.

Useful? React with 👍 / 👎.

Done, and it landed before this review. payload.remove is True short-circuits to set_model_override(target_id, llama_extra_args = [], max_seq_length = None) and never reaches the branch that passes the config fields. Only the explicit flag short-circuits; the legacy inferred path still just gates launch-flag carry-over, so the existing carry-over test stays green.

It also removes under resolve_model_override_key(payload.model_id) or payload.model_id, since the browser normalizes casing before storing and a literal-key removal would leave the stored entry behind.

Remove server overrides evicted from local storage

When saving a new configuration pushes the local map over its 500-entry or 1 MiB budget, savePerModelConfig silently evicts older model entries and still returns success, but this call mirrors only the model currently being saved. The evicted models therefore retain server-side overrides that the UI no longer shows and cannot forget, so later API auto-switches continue applying stale settings. Track the keys evicted by enforceStorageBudget and remove their corresponding server overrides as part of the synchronization.

Useful? React with 👍 / 👎.

Already handled. savePerModelConfig takes an evicted array and model-config-page.tsx clears each dropped model server side right after the mirror:

for (const dropped of evicted) {
  syncModelOverride(dropped.modelId, dropped.ggufVariant, null);
}

There was no test holding it in place, which is a fair catch. Added test_evicted_local_configs_drop_their_server_overrides in 6ad8cb6, which also pins that the eviction path decodes the storage key back into a model id and variant rather than reporting the normalized key.

Seed known traffic after resuming overlay polling

When the user leaves the full monitor while an API request is still running, this resets seededRef, but the next snapshot seeds only terminal entries and consequently treats that already-observed running request as new, reopening the floating panel immediately. A similar stale-baseline problem occurs after auto-open is disabled and later re-enabled because polling stops without resetting the baseline, so completed traffic accumulated while disabled can trigger the panel. Track whether polling is resuming from a deliberate hiatus and seed all IDs from its first snapshot rather than applying the fresh-page running-request rule.

Useful? React with 👍 / 👎.