Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Restore the facts the frontend comment trim reworded away (#10292)

Daniel Han55418be5eb23
diff --git a/studio/frontend/src/features/audio/audio-page.tsx b/studio/frontend/src/features/audio/audio-page.tsx
index 3ada13607d2..503c82dc823 100644
--- a/studio/frontend/src/features/audio/audio-page.tsx
+++ b/studio/frontend/src/features/audio/audio-page.tsx
@@ -2,8 +2,8 @@
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
// The Audio page: Generate (TTS via the main inference slot) and Transcribe (STT via the
-// dictation sidecar). The page stays mounted across tab switches, so `active` gates polling,
-// popovers and the recorder rather than lifecycle.
+// dictation sidecar). The page stays mounted across tab switches (see __root.tsx), so `active`
+// gates polling, popovers and the recorder rather than lifecycle.
import { TestTubeOutlineIcon } from "@/lib/hugeicons-derived";
import {
@@ -880,7 +880,8 @@ export function AudioPage({
ggufFilename?: string | null,
// Where the weights actually are: a row cached in a NON-ACTIVE HF cache is loadable only by its
// snapshot path, which the picker supplies as meta.loadId, and sending the display repo id
- // instead failed offline or re-downloaded into the active cache.
+ // instead failed offline or re-downloaded into the active cache. Chat threads the same field
+ // (chat-page.tsx).
loadId?: string | null,
audioType?: string | null,
remoteCodeApproval?: RemoteCodeApproval,
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 10bfa0813e2..22b0ca470b3 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -407,7 +407,8 @@ type OpenAIStreamAdapterOptions = {
/** Tracks which user messages were sent with an audio file (messageId → filename). */
export const sentAudioNames = new Map<string, string>();
-// Synthetic provider-side tool names; mirror of backend _SERVER_SIDE_BUILTIN_TOOL_NAMES.
+// Synthetic provider-side tool names; the backend stamps args._server_tool so user functions with
+// the same name aren't dropped. Mirror of backend _SERVER_SIDE_BUILTIN_TOOL_NAMES.
const SERVER_SIDE_BUILTIN_TOOL_NAMES = new Set<string>([
"web_search",
"web_fetch",
@@ -1129,8 +1130,8 @@ function serializeToolResultPart(
isSearchImagesToolResult(result) ||
isSandboxWrapper(result, tc.toolName ?? "")
) {
- // Replay the stdout the model saw, not the wrapper; image tokens go with it, since a token
- // resolves only against the message whose search produced it.
+ // Replay the stdout the model saw, not the card's sessionId/images/files; image tokens go with
+ // it, since a token resolves only against the message whose search produced it.
const replayText = isSearchImagesToolResult(result)
? stripSearchImageTokens(result.text)
: result.text;
@@ -5408,7 +5409,8 @@ export function createOpenAIStreamAdapter(
});
// Backend tool ids ("call_0", ...) restart every response, so a bare id as store key lets a
// later turn's stream overwrite a finished card's output. Mint one run-unique part id per
- // backend id, resolved through this map and dropped at tool_end.
+ // backend id; every tool_start/output/args/end resolves the same id through this map, which
+ // is dropped at tool_end.
const toolPartIdByBackendId = new Map<string, string>();
const resolveToolPartId = (backendToolCallId: string): string =>
resolveToolCallPartId(
@@ -6715,7 +6717,8 @@ export function createOpenAIStreamAdapter(
parsedResult = mcpImages;
} else if (imgIdx !== -1) {
const text = rawResult.slice(0, imgIdx);
- // Fall back to "_default", the backend sandbox dir used when there is no session_id.
+ // Fall back to "_default", the backend sandbox dir used when there is no
+ // session_id (see tools.py _get_workdir).
const sessionId = sandboxSessionId || "_default";
try {
const images = JSON.parse(
diff --git a/studio/frontend/src/features/chat/attachment-content.ts b/studio/frontend/src/features/chat/attachment-content.ts
index 2c43ff1c054..25892db3e8a 100644
--- a/studio/frontend/src/features/chat/attachment-content.ts
+++ b/studio/frontend/src/features/chat/attachment-content.ts
@@ -48,7 +48,7 @@ const DOCX_MIME =
// may be called anything and still be parsed as XML, while an .xml part nothing points at
// is never opened. The bound therefore follows docx-reader.js: the two package parts it
// always reads plus the five resolved out of those relationships, each with mammoth's own
-// "word/<name>.xml" fallback. Image targets stay lazy.
+// "word/<name>.xml" fallback. Image targets stay lazy and are never read by extractRawText.
const DOCX_CONTENT_TYPES_PART = "[Content_Types].xml";
const DOCX_PACKAGE_RELATIONSHIPS = "_rels/.rels";
const DOCX_RELATIONSHIP_NAMESPACE =
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index a0bf66fd400..1fc3eb2e607 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -624,7 +624,8 @@ export function ChatSettingsPanel({
const [seedDraft, setSeedDraft] = useState<string | null>(null);
// What blur would commit, available before it runs. Clicking Save blurs the box during
// mousedown, but React has not re-rendered by the time onClick fires, so a handler reading
- // `params` still sees the seed from before the entry. NumericValueInput bridges the same gap.
+ // `params` still sees the seed from before the entry. NumericValueInput bridges the same gap
+ // with its imperative commit().
const committedSeed = useMemo<number | null>(() => {
if (seedDraft === null) return params.seed ?? null;
// Measured after the padding: a zero-padded seed is short enough to keep, and truncating
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index 184ff32df28..c80e90e61f3 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -974,7 +974,8 @@ export function useChatModelRuntime() {
) &&
// The id names the weights, not how the server was invoked: a remembered context length,
// drafter, placement or extra arg the resident load does not run is a real reload. Not
- // `pendingConfig`: with no saved record the caller has already reset the store to defaults,
+ // `pendingConfig`: with no saved record the caller has already run
+ // applyModelLoadConfigToRuntime(null), which resets the store to DEFAULT_PER_MODEL_CONFIG,
// and performLoad reads the store for every field the config does not carry.
residentRuntimeMatchesConfig(status, comparedConfig, {
// What the applier fills an unset field with, so the comparison is against what /load would send
@@ -1866,10 +1867,9 @@ export function useChatModelRuntime() {
maxTokensCap: loadedContextCap,
},
);
- // Qwen3.5/3.6 small models disable thinking by default. Anchored regex: the first "Xb" after
- // start-of-string or [-_/.], so "Qwen3.5-35B-A3B" gives 35 rather than 3.
- // Qwen3.5/3.6 small models are 0.8B, 2B, 4B and 9B. The regex matches the first "Xb" or "X.Xb"
- // after start or [-_/.], so "Qwen3.5-35B-A3B" yields 35 (total), not 3 (MoE active).
+ // Qwen3.5/3.6 small models (0.8B, 2B, 4B, 9B) disable thinking by default. Anchored regex:
+ // first "Xb" / "X.Xb" after start-of-string or [-_/.] so the version literal in "qwen3.5" /
+ // "qwen3.6" does not match first, and "Qwen3.5-35B-A3B" yields 35 (total), not 3 (MoE active).
let reasoningDefault = loadResponse.supports_reasoning ?? false;
if (reasoningDefault) {
const mid = modelId.toLowerCase();
diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
index 441812d4184..02a85cb84c1 100644
--- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
+++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
@@ -628,7 +628,8 @@ export function applyActiveModelStatusToStore(
storedReasoningEnabled === null
) {
// Anchored regex: first "Xb"/"X.Xb" after start or [-_/.] so the version literal in
- // "qwen3.5" does not match first, and "Qwen3.5-35B-A3B" yields 35 (total), not 3 (active).
+ // "qwen3.5" / "qwen3.6" does not match first, and "Qwen3.5-35B-A3B" yields 35 (total), not 3
+ // (active).
// Mirrors use-chat-model-runtime.ts and the inline regex in llama_cpp.py.
let reasoningDefault = true;
const mid = checkpointId.toLowerCase();
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index 3efdf24612e..c20a7f5fdb3 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -505,8 +505,10 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
},
mistral: OPENAI_COMPAT_BASE,
// Gemini's generationConfig accepts temperature, topP, topK, presencePenalty; minP and
- // repetitionPenalty are not in the contract. Shaping lives in _stream_gemini.
- // Gemini also accepts a frequencyPenalty this panel does not surface.
+ // repetitionPenalty are not in the contract, see
+ // https://ai.google.dev/api/rest/v1beta/GenerationConfig.
+ // Gemini also accepts a frequencyPenalty this panel does not surface. Shaping lives in
+ // _stream_gemini in studio/backend/core/inference/external_provider.py.
gemini: {
temperature: true,
topP: true,
@@ -765,8 +767,12 @@ function resolveKimiReasoningCapabilities(modelId: string): ExternalReasoningCap
return withEnableThinkingStyle();
}
-// Gemini's thinking ladder: 3.x uses the string thinkingLevel (Pro rejects MINIMAL);
-// 2.5 Flash/Pro use integer thinkingBudget (0=off on Flash, -1=dynamic, Pro rejects 0);
+// Gemini's thinking ladder: 3.x (Pro + Flash + Flash-Lite) and the gemini-pro-latest /
+// gemini-flash-latest aliases use the string `thinkingConfig.thinkingLevel`. Flash takes
+// MINIMAL/LOW/MEDIUM/HIGH, 3.1+ Pro takes LOW/MEDIUM/HIGH, and undotted legacy `gemini-3-pro`
+// takes LOW/HIGH only, so the backend coerces MINIMAL to LOW on any Pro and MEDIUM to HIGH on
+// the legacy one. 2.5 Flash/Pro use the integer
+// `thinkingConfig.thinkingBudget` (0=off on Flash, -1=dynamic, N>0=cap, Pro rejects 0);
// 2.5 Flash-Lite and image ids get none. 3.x minors match by pattern so a new
// `gemini-3.6-flash` cannot fall into the 2.5 branch. Mirrors _GEMINI3_FAMILY/_GEMINI3_PRO.
const GEMINI3_PRO_PATTERN = /^gemini-3(\.\d+)?-pro/;
@@ -797,6 +803,7 @@ function resolveGeminiReasoningCapabilities(
// Gemini 2.5 Flash-Lite: thinkingBudget 0 = off, positive from 512. Must be checked
// BEFORE the broader `gemini-2.5-flash` prefix.
// The backend maps "minimal" to that 512 floor in _stream_gemini.
+ // https://ai.google.dev/gemini-api/docs/thinking
if (m.startsWith("gemini-2.5-flash-lite")) {
return withReasoningEffortStyle({
supportsReasoning: true,
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index 896c06cf1f2..ece53085b88 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -1117,7 +1117,8 @@ function scheduleGenerationRecovery(
}
if (followStalled || generationNeedsRecovery(currentMetadata)) {
// Fence the run before presenting the reply as resumable: settling only in this tab leaves the
- // row active, and create_run refuses a thread that already has one, so Continue would 409.
+ // row queued/running/cancelling, and create_run refuses a thread that already has an active
+ // generation, so Continue and the next message would both 409.
// Best effort: a producer wedged inside the engine stays in `cancelling` for the sweeper.
serverCancel();
// The follow returned with the run still non-terminal, so its no-progress deadline expired.
diff --git a/studio/frontend/src/features/chat/search-images/search-images.ts b/studio/frontend/src/features/chat/search-images/search-images.ts
index 40a10bfb303..e6ae982a8b4 100644
--- a/studio/frontend/src/features/chat/search-images/search-images.ts
+++ b/studio/frontend/src/features/chat/search-images/search-images.ts
@@ -293,8 +293,9 @@ export function placeSubjectImages(
}
// Split marker from subject, and never let two whitespace quantifiers span the same run: the
-// single-regex form backtracked at ~O(n^3.5), so 250 spaces in one bullet froze the thread
-// for 13s. The body is anchored on non-space edges so the surrounding runs cannot overlap.
+// single-regex form nested `\s+`, `\s*` and a body class that also matched spaces, which
+// backtracked at ~O(n^3.5), so 250 spaces in one bullet froze the thread for 13s. The body is
+// anchored on non-space edges so the surrounding runs cannot overlap.
const LIST_ITEM_MARKER_RE = /^[ \t]*(?:\d{1,2}[.)]|[-*+•])[ \t]+/;
const LIST_ITEM_RE =
/^(?:\*\*|__)?[ \t\r]*([^\s*_\n][^*_\n]{0,58}?[^\s*_\n])(?:[ \t\r]*(?:\*\*|__))?[ \t\r]*(?::|[-–—][ \t\r]|\(|$)/;
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index 6f515ac2728..98647085fd6 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -751,7 +751,8 @@ export function SharedComposer({
const reasoningLockedOn =
effectiveSupportsReasoning &&
(effectiveReasoningAlwaysOn || !effectiveSupportsReasoningOff);
- // Kimi's $web_search builtin mandates thinking=disabled. Both pills stay clickable, but turning
+ // Kimi's $web_search builtin mandates thinking=disabled
+ // (https://platform.kimi.ai/docs/guide/use-web-search). Both pills stay clickable, but turning
// one on flips the other off, so the visible state matches what the backend sends.
const isKimiExternal = selectedExternalProvider?.providerType === "kimi";
const effectiveReasoningEnabled = reasoningLockedOn ? true : reasoningEnabled;
@@ -1410,7 +1411,8 @@ export function SharedComposer({
}
// A pane's OWN saved split is sent instead of being forced to Auto (#7574); the shared Send-time
// snapshot is not, since its layer count is bounded by another GGUF. Knobs the runner has no
- // equivalent for stay hard-forced, and an UNCLASSIFIED GGUF is pinned too.
+ // equivalent for (MoE offload, tensor parallel) stay hard-forced, and an UNCLASSIFIED GGUF
+ // is pinned too: see lib/gpu-placement.ts.
const {
gpuMemoryMode: effectiveGpuMemoryMode,
gpuLayers: effectiveGpuLayers,
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 6a739452a6f..6928489d9a2 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -158,7 +158,8 @@ export const CHAT_RAG_OCR_KEY = "unsloth_chat_rag_ocr_scanned";
export const CHAT_RAG_CAPTION_KEY = "unsloth_chat_rag_caption_figures";
// Only the model-agnostic intents (auto/ngram/off) persist: a saved drafter mode no-ops on a
// model with no MTP head or DSpark sidecar.
-// The model-specific drafter modes and spec_draft_n_max stay session-only.
+// The model-specific drafter modes (mtp/mtp+ngram/dspark/dflash) and spec_draft_n_max stay
+// session-only. Unknown -> auto.
const PERSISTED_SPEC_MODES = new Set(["auto", "ngram", "off"]);
export type RagSource = { type: "thread" } | { type: "kb"; kbId: string };
@@ -313,7 +314,8 @@ function loadRagNumber(
}
// External picks ride in `params.checkpoint` as `external::<providerId>::<modelId>`.
-// PersistedChatSettings omits `checkpoint`: only the local side is mirrored by /status.
+// PersistedChatSettings omits `checkpoint`: only the local side is mirrored by the backend's
+// /api/inference/status.active_model.
const LAST_EXTERNAL_CHECKPOINT_KEY = "unsloth_chat_last_external_checkpoint";
function loadLastExternalCheckpoint(): string | null {
@@ -4358,7 +4360,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
options?.persist !== false && state.settingsHydrated;
// A sampling key moved with a chat open belongs to that chat, so it reaches neither the
// defaults nor the model's memory. NOT gated on hydration, unlike the global write: the
- // composer is live while the initial request is out, and an uncaptured edit leaked.
+ // composer is live while the initial /api/chat/settings request is out, and an uncaptured
+ // edit leaked.
const sharedParams =
options?.persist !== false
? withoutCapturedThreadEdits(changedParams, fromModelDefaults)
diff --git a/studio/frontend/src/features/chat/utils/fork-count-store.ts b/studio/frontend/src/features/chat/utils/fork-count-store.ts
index 38dcccc1606..3e2d6df13dc 100644
--- a/studio/frontend/src/features/chat/utils/fork-count-store.ts
+++ b/studio/frontend/src/features/chat/utils/fork-count-store.ts
@@ -47,7 +47,8 @@ async function refresh(threadId: string): Promise<void> {
// returns without a row. The row decides that, not the id, since a `__LOCALID_` prefix belongs
// to every chat the app creates. Skipped because the answer is already the empty map the entry
// holds, not to dodge a failure: fork_counts_for_thread GROUPs without looking the source up,
- // so an unknown thread gets 200 and an empty map.
+ // so an unknown thread gets 200 and an empty map here, unlike the per-thread routes, which
+ // reject a missing thread.
if (isThreadIncognito(threadId)) return;
const seq = ++entry.seq;
let counts: Counts;
@@ -82,8 +83,8 @@ function onHistoryUpdated(): void {
// Clear and reschedule, as the sidebar refresh does. Returning while a timer exists would make
// this a leading-edge throttle: streaming fires this event per chunk, so the timer would
// expire mid-stream and the next chunk would start another window, costing one whole-thread
- // fetch per debounce window for as long as the reply runs. Fork counts cannot change during
- // generation, so every one of those is wasted.
+ // fetch every FORK_COUNT_REFRESH_DEBOUNCE_MS for as long as the reply runs. Fork counts cannot
+ // change during generation, so every one of those is wasted.
if (pendingRefresh) clearTimeout(pendingRefresh);
pendingRefresh = setTimeout(runRefresh, FORK_COUNT_REFRESH_DEBOUNCE_MS);
// The bound. Deliberately not restarted while it is already running, or a per-chunk event stream
diff --git a/studio/frontend/src/features/images/images-page.tsx b/studio/frontend/src/features/images/images-page.tsx
index de60327fe1d..60cbe0b5703 100644
--- a/studio/frontend/src/features/images/images-page.tsx
+++ b/studio/frontend/src/features/images/images-page.tsx
@@ -180,9 +180,10 @@ function useImageModels(host: HostClass): ModelOption[] {
return useMemo(() => catalogToModelOptions(IMAGE_CATALOG, host), [host]);
}
-// Workflow tabs. `requires` is the backend workflow id the model must support; null = always.
-// Each conditioned workflow names the images it consumed for the restore toast, since a
-// recipe keeps the scalar settings but not the uploads. txt2img restores completely.
+// Workflow tabs. `requires` is the backend workflow id (status.workflows) the model must
+// support; null = always. Each conditioned workflow names the images it consumed for the restore
+// toast, since a recipe keeps the scalar settings but not the uploads. Keys are the backend's own
+// workflow strings; txt2img is absent because it restores completely.
const CONDITIONED_WORKFLOW_INPUTS: Record<string, string> = {
img2img: "the source image",
inpaint: "the source image and mask",
@@ -2324,7 +2325,7 @@ export function ImagesPage({
// Reseed the Advanced selects from the LOADED build, so a declined request snaps to what
// engaged and Precision never advertises a scheme the model is not running. Keyed on the
- // LOAD-TIME half of the record: the backend rewrites the speed/attention entries at
+ // LOAD-TIME half of the record: the backend rewrites the speed/attention/cache entries at
// GENERATION time, and the whole record threw away a Precision picked but not yet loaded.
const resolvedKey = status?.loaded ? resolvedSeedKey(status.resolved) : null;
useEffect(() => {
@@ -3821,7 +3822,7 @@ export function ImagesPage({
onClick={() => setExtendSides((s) => ({ ...s, [key]: !s[key] }))}
className={cn(
// No border in either mode: the fill alone marks the state, and a ring would not survive
- // mouse focus anyway.
+ // mouse focus anyway (index.css blanks it).
"rounded-lg px-2 py-1.5 text-xs font-medium transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
on
? "bg-primary/15 text-foreground hover:bg-primary/20 dark:bg-primary/25 dark:hover:bg-primary/30"
diff --git a/studio/frontend/src/features/model-picker/api/model-overrides.ts b/studio/frontend/src/features/model-picker/api/model-overrides.ts
index d92acbcba7f..1153caea684 100644
--- a/studio/frontend/src/features/model-picker/api/model-overrides.ts
+++ b/studio/frontend/src/features/model-picker/api/model-overrides.ts
@@ -112,7 +112,8 @@ export function foldOverrideKey(key: string): string {
}
// splitQuantSuffix, not the last colon: a colon is legal in a POSIX filename, so
// "/models/foo:Bar.gguf" is a whole path and reading "Bar.gguf" as a quant would fold it
- // onto a different real file. Mirrors the backend's split_quant_suffix.
+ // onto the real, different file "/models/foo:bar.gguf". Mirrors the backend's
+ // split_quant_suffix.
const split = splitQuantSuffix(key);
const id = split ? split[0] : key;
const quant = split ? `:${split[1].toLowerCase()}` : "";
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-capabilities.ts b/studio/frontend/src/features/model-picker/components/model-selector/model-capabilities.ts
index 419db742942..3ac1d71560a 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/model-capabilities.ts
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-capabilities.ts
@@ -69,7 +69,7 @@ const AUDIO_NAME_RE = new RegExp(
// read as an image generator from its name alone. Video is matched FIRST below, since
// "HunyuanVideo" and "hunyuanimage" share a stem and the video families are narrower. These
// end on a letter boundary rather than END, since a family stem runs into its version
-// ("flux1", "sd3"); without it "fluxion-7b" and "SVDQuant" read as generators.
+// ("flux1", "sd3", "Wan2.2"); without it "fluxion-7b" and "SVDQuant" read as generators.
const FAMILY_END = "(?![a-z])";
const IMAGE_GEN_NAME_RE = new RegExp(
`${SEP}(?:flux|sdxl|sd3|stable[-_]?diffusion|z[-_]?image|qwen[-_]?image|hidream|ideogram|lumina|hunyuanimage|krea|kolors|playground|pixart)${FAMILY_END}`,
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts b/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts
index 3e1a1bb4a10..0dedd80f456 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts
@@ -276,11 +276,13 @@ export const IMAGE_CATALOG: CatalogGroup[] = [
artifacts: [bf16Pipeline("Alpha-VLLM/Lumina-Image-2.0", 11, { totalParams: 2609769152 })],
},
{
- // 17B dual-stream 2K-native DiT with a Qwen2.5-VL encoder. bf16 only: the QuantStack GGUF
- // consumer route was unpublished, and an entry the Hub cannot serve renders as a one-click
- // download that fails partway through. No vetted replacement: unsloth has an FP8 mirror but
- // no GGUF, and the third-party GGUF repos are a different lineage. At 24 GB the bf16 still
- // fits the 61.6 GB budget, so the group stays visible; the quant ladder is what is gone.
+ // 17B dual-stream 2K-native DiT with a Qwen2.5-VL encoder. bf16 only: this used to carry
+ // gguf("QuantStack/HunyuanImage-2.1-GGUF") as the consumer route, and that repo was
+ // unpublished. An entry the Hub cannot serve renders as a one-click download that fails
+ // partway through, the exact case model-catalog-network-check exists to find. No vetted
+ // replacement: unsloth has an FP8 mirror but no GGUF, and the third-party GGUF repos are a
+ // different lineage. At 24 GB the bf16 still fits the 61.6 GB budget, so the group stays
+ // visible; the quant ladder is what is gone.
// The mirror guider components load natively on diffusers 0.39. The retired GGUF repo 404s to an
// authed request and 401s anonymously; QuantStack itself is alive and still ships its other GGUF
// repos, so this is one model withdrawn rather than a publisher going away.
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
index b8fc160e081..760d1c7e86b 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
@@ -1769,8 +1769,9 @@ function GgufVariantExpander({
}
return recommended;
}, [variantGroups, preferredByGroup, anyBudgetGb, budgetKnown, getGgufFit]);
- // `effectiveRecommendedByGroup` is keyed by PRESENTATION group while the footprint pass
- // buckets by the backend's dependency_key, so that pass asks through the variant itself.
+ // `effectiveRecommendedByGroup` is keyed by PRESENTATION group ("quantizations", "text-frames",
+ // "reference-media") while the footprint pass buckets by the backend's dependency_key
+ // ("flux.2-klein:<digest>"), so that pass asks through the variant itself.
const recommendedQuantForVariant = useMemo(() => {
const byVariant = new Map<GgufVariantDetail, string>();
for (const group of variantGroups) {
@@ -3729,7 +3730,9 @@ export function HubModelPicker({
catalogSeedRows,
]);
- // Ordered by the On Device dropdown. The gate keeps diffusion GGUFs in the Images/Video picker and out of chat.
+ // Ordered by the On Device dropdown (Recent / Size / Name / Downloaded). The gate keeps a
+ // supported diffusion GGUF listed here so picking one routes to Images or Video; only the
+ // never-loadable tag is dropped.
const sortedCachedGguf = useMemo(
() =>
sortCachedRepos(
diff --git a/studio/frontend/src/features/model-picker/model-config/llama-extra-args.ts b/studio/frontend/src/features/model-picker/model-config/llama-extra-args.ts
index 27994e58bf2..eae0ed79ecc 100644
--- a/studio/frontend/src/features/model-picker/model-config/llama-extra-args.ts
+++ b/studio/frontend/src/features/model-picker/model-config/llama-extra-args.ts
@@ -866,7 +866,8 @@ export function diagnoseExtraArgs(
continue;
}
// Before the de-duplication below, because llama.cpp reads the LAST occurrence: in `-ngl 20 -ngl
- // many` it is the second one the backend refuses.
+ // many` it is the second one the backend parses and refuses, so checking only the first would
+ // leave Load enabled for a request that 400s.
if (INTEGER_VALUE_FLAGS.has(flag) || VALUE_REQUIRED_FLAGS.has(flag)) {
const attached = valueIsAttached(token, flag);
const value = attached ? token.split("=")[1] : tokens[index + 1];
diff --git a/studio/frontend/src/features/video/video-page.tsx b/studio/frontend/src/features/video/video-page.tsx
index 242fef9264c..73a00409b35 100644
--- a/studio/frontend/src/features/video/video-page.tsx
+++ b/studio/frontend/src/features/video/video-page.tsx
@@ -1909,8 +1909,10 @@ function VideoGenerator({
}, [setStatusIfNewest]);
// A generation can be refused because the runtime went away under the page: an idle auto-unload
- // frees it server-side and the browser hears nothing, so the refusal is the news. Also
- // clears the state that only means anything while one is resident.
+ // frees it server-side and the browser hears nothing. Without a re-read here Generate stays
+ // enabled off the stale flag and every retry 409s again, so the refusal is the news. Also
+ // clears the state that only means anything while one is resident (the Reapply target, a
+ // replacement load's tracking).
const resyncAfterGenerateRefusal = useCallback(async () => {
// A model picked while this read is in flight makes the answer stale rather than wrong:
// /video/status reports committed state, so it says loaded: false for the load that has just