- {statusFresh ? (
+ {status ? (
- {statusFresh && status.connected === false && (() => {
+ {status && status.connected === false && (() => {
const missingCount = status.missingPackages?.length || 0;
const hasPath = !!status.pythonPath;
return (
@@ -1392,7 +1401,7 @@ export default function VideoGen() {
backend={backend}
backendDisclosures={status?.backendDisclosures}
model={isGrok ? null : currentModel}
- systemMemoryGb={status?.systemMemoryGb}
+ systemMemoryGb={modelContext?.systemMemoryGb}
/>
{!isGrok && (
@@ -1563,10 +1572,11 @@ export default function VideoGen() {
onClose={() => setInstallModalOpen(false)}
onComplete={() => {
refreshByovStatus();
- // The capability probe is part of /video-gen/status's model
- // decoration. Refresh it after install/repair so H3's LoRA picker
- // and warning react without a manual page reload.
+ // The capability probe decorates the model list, and the install
+ // also moves python health. Refresh both after install/repair so
+ // H3's LoRA picker and warning react without a manual page reload.
refreshStatus();
+ refreshModelContext();
}}
/>
diff --git a/client/src/pages/VideoGen.modelLoading.test.jsx b/client/src/pages/VideoGen.modelLoading.test.jsx
index 9840dece63..a6636bbc46 100644
--- a/client/src/pages/VideoGen.modelLoading.test.jsx
+++ b/client/src/pages/VideoGen.modelLoading.test.jsx
@@ -7,90 +7,82 @@ import {
resetVideoGenMockState,
state,
videoGenModel,
+ videoGenModelContext,
videoGenStatus,
} from '../test/videoGenPageMocks.jsx';
/**
- * The Model picker paints before /status lands.
+ * The Model picker does not wait on /status.
*
- * /status shells out to python and rebuilds the hardware-aware model list on
- * every call, so the field used to be absent for a second or two and then pop
- * into the middle of the form. It now holds its place with a loading
- * placeholder, and a session-cached payload paints the real list immediately —
- * while every connectivity claim keeps waiting for the live probe.
+ * /status shells out to python on every call, so a cold load used to leave the
+ * field absent for a second or two and then pop it into the middle of the form.
+ * The list now comes from the probe-free /video-gen/model-context, which lands
+ * on its own — while every connectivity claim keeps waiting for the live probe.
+ * The loading placeholder remains for the window before either answers.
*/
const MODEL_ONE = videoGenModel('example-one');
const MODEL_TWO = videoGenModel('example-two');
const statusPayload = (overrides = {}) => videoGenStatus([MODEL_ONE, MODEL_TWO], overrides);
+const modelContextPayload = (overrides = {}) => videoGenModelContext([MODEL_ONE, MODEL_TWO], overrides);
await loadVideoGenPage();
-const { VIDEO_GEN_STATUS_CACHE_KEY } = await import('../lib/videoGenStatusCache.js');
-// A /status call the test settles by hand, so the page can be asserted mid-probe.
-const deferredStatus = () => {
+// A call the test settles by hand, so the page can be asserted mid-flight.
+const deferred = (mock) => {
let settle;
- state.getVideoGenStatus.mockReturnValue(new Promise((resolve) => { settle = resolve; }));
+ mock.mockReturnValue(new Promise((resolve) => { settle = resolve; }));
return async (payload) => { await act(async () => { settle(payload); }); };
};
+const deferredStatus = () => deferred(state.getVideoGenStatus);
+const deferredModelContext = () => deferred(state.getVideoGenModelContext);
-describe('VideoGen model picker while /status is in flight', () => {
+describe('VideoGen model picker vs the /status python probe', () => {
beforeEach(() => {
localStorage.clear();
sessionStorage.clear();
resetVideoGenMockState();
state.getVideoGenStatus.mockResolvedValue(statusPayload());
+ state.getVideoGenModelContext.mockResolvedValue(modelContextPayload());
state.attach.mockResolvedValue({ filename: 'example.mp4' });
});
it('keeps the Model field with a loading placeholder until the model list lands', async () => {
- const resolveStatus = deferredStatus();
+ const resolveModelContext = deferredModelContext();
await renderVideoGenPage();
const field = screen.getByLabelText('Model');
expect(field).toBeDisabled();
expect(field).toHaveTextContent('Loading models…');
- await resolveStatus(statusPayload());
+ await resolveModelContext(modelContextPayload());
await waitFor(() => expect(screen.getByLabelText('Model')).toHaveValue(MODEL_ONE.id));
expect(screen.getByLabelText('Model')).toBeEnabled();
});
- it('paints the cached model list on the next load instead of waiting for the probe', async () => {
- const first = await renderVideoGenPage();
- await waitFor(() => expect(screen.getByLabelText('Model')).toHaveValue(MODEL_ONE.id));
- // Only the model-shaping slice is persisted — python health never is.
- expect(Object.keys(JSON.parse(sessionStorage.getItem(VIDEO_GEN_STATUS_CACHE_KEY))).sort())
- .toEqual(['defaultModel', 'models', 'systemMemoryGb']);
- first.unmount();
-
+ it('paints the model list on a cold load while every python claim waits', async () => {
+ // No prior visit and no priming — the regression this pins is the Model
+ // field sitting on its placeholder for the whole /status round trip, and
+ // the converse: nothing may report the interpreter before it answers.
const resolveStatus = deferredStatus();
await renderVideoGenPage();
const field = screen.getByLabelText('Model');
expect(field).toBeEnabled();
expect(field).toHaveValue(MODEL_ONE.id);
- await resolveStatus(statusPayload());
- });
-
- it('never reports python health from a cached entry', async () => {
- // A hand-written entry carrying a FAILED probe — the belt to the
- // projection's braces. The model list may come from storage; the diagnosis
- // may not, because the interpreter can have been fixed since.
- sessionStorage.setItem(VIDEO_GEN_STATUS_CACHE_KEY, JSON.stringify(statusPayload({
- connected: false,
- reason: 'Python probe failed',
- missingPackages: ['torch'],
- })));
- const resolveStatus = deferredStatus();
- await renderVideoGenPage();
-
- expect(screen.getByLabelText('Model')).toHaveValue(MODEL_ONE.id);
expect(screen.getByText('Checking…')).toBeInTheDocument();
expect(screen.queryByText(/Install missing Python packages/)).toBeNull();
- expect(screen.queryByText(/Python probe failed/)).toBeNull();
await resolveStatus(statusPayload({ connected: true, pythonVersion: '3.12.1' }));
await waitFor(() => expect(screen.getByText('Python 3.12.1')).toBeInTheDocument());
});
+
+ it('takes the Model field away when the context fetch names no model at all', async () => {
+ // A failed /model-context leaves nothing to offer. The field must not hold
+ // its placeholder forever — the rest of the form closes over the gap.
+ state.getVideoGenModelContext.mockRejectedValue(new Error('offline'));
+ await renderVideoGenPage();
+
+ await waitFor(() => expect(screen.queryByLabelText('Model')).toBeNull());
+ });
});
diff --git a/client/src/pages/VideoGen.terms.test.jsx b/client/src/pages/VideoGen.terms.test.jsx
index 76c5d4ece8..e3fc1f78d6 100644
--- a/client/src/pages/VideoGen.terms.test.jsx
+++ b/client/src/pages/VideoGen.terms.test.jsx
@@ -7,6 +7,7 @@ import {
resetVideoGenMockState,
state,
videoGenModel,
+ videoGenModelContext,
videoGenStatus,
videoGenTermsGate,
} from '../test/videoGenPageMocks.jsx';
@@ -33,6 +34,7 @@ describe('VideoGen MiniMax H3 orchestration', () => {
state.generateVideo.mockResolvedValue({ jobId: 'job-1' });
state.repair.mockResolvedValue({ ok: true });
state.getVideoGenStatus.mockResolvedValue(videoGenStatus([H3_ONE, H3_TWO]));
+ state.getVideoGenModelContext.mockResolvedValue(videoGenModelContext([H3_ONE, H3_TWO]));
state.attach.mockImplementation(async (_jobId, handlers) => {
handlers.onComplete({ result: { filename: 'example.mp4' } });
return { filename: 'example.mp4' };
@@ -98,10 +100,14 @@ describe('VideoGen MiniMax H3 orchestration', () => {
it('refreshes the model capability payload after runtime setup completes', async () => {
await renderVideoGenPage();
await waitFor(() => expect(screen.getByLabelText('Model')).toHaveValue(H3_ONE.id));
- const before = state.getVideoGenStatus.mock.calls.length;
+ const beforeStatus = state.getVideoGenStatus.mock.calls.length;
+ const beforeContext = state.getVideoGenModelContext.mock.calls.length;
await act(async () => { await state.runtimeInstallComplete(); });
- await waitFor(() => expect(state.getVideoGenStatus).toHaveBeenCalledTimes(before + 1));
+ // The install moves BOTH halves: the hardware decoration on the model list
+ // and the python health the connectivity banner reads.
+ await waitFor(() => expect(state.getVideoGenModelContext).toHaveBeenCalledTimes(beforeContext + 1));
+ expect(state.getVideoGenStatus).toHaveBeenCalledTimes(beforeStatus + 1);
});
});
diff --git a/client/src/pages/VideoGen.textEncoderAutoDownload.test.jsx b/client/src/pages/VideoGen.textEncoderAutoDownload.test.jsx
index b22a213f35..78ab72171d 100644
--- a/client/src/pages/VideoGen.textEncoderAutoDownload.test.jsx
+++ b/client/src/pages/VideoGen.textEncoderAutoDownload.test.jsx
@@ -17,6 +17,7 @@ import {
resetVideoGenMockState,
state,
videoGenModel,
+ videoGenModelContext,
videoGenStatus,
videoGenTermsGate,
} from '../test/videoGenPageMocks.jsx';
@@ -44,6 +45,7 @@ describe('VideoGen substitute text-encoder auto-download', () => {
beforeEach(() => {
resetVideoGenMockState();
state.getVideoGenStatus.mockResolvedValue(videoGenStatus([MODEL]));
+ state.getVideoGenModelContext.mockResolvedValue(videoGenModelContext([MODEL]));
// The substitute is never resident: what these cases pin down is the
// request, and a cached encoder would short-circuit it.
state.getModelStatus = (id) => (String(id).startsWith('__text_encoder_option__:')
diff --git a/client/src/services/apiImageVideo.js b/client/src/services/apiImageVideo.js
index 6dd2bc59ba..6690b48d45 100644
--- a/client/src/services/apiImageVideo.js
+++ b/client/src/services/apiImageVideo.js
@@ -165,6 +165,13 @@ export const getVideoGenStatus = (options = {}) => request('/video-gen/status',
export const listVideoModels = ({ includeUnavailable = false, ...options } = {}) =>
request('/video-gen/models', options)
.then((models) => filterHardwareCompatibleModels(models, { includeUnavailable }));
+// `{ models, defaultModel, systemMemoryGb, fflfLtx2PixelBudget }` — the model
+// list plus the numbers its auto-select reads, with no python probe behind it.
+// getVideoGenStatus() returns the same fields, but only after shelling out to
+// the interpreter; fetch this alongside it so the Model picker paints first.
+export const getVideoGenModelContext = (options = {}) =>
+ request('/video-gen/model-context', options)
+ .then((ctx) => ({ ...ctx, models: filterHardwareCompatibleModels(ctx?.models) }));
// `{ models: [...], textEncoder: { repo, cached, sizeBytes } }`. Same shape
// contract as the image variant + a text-encoder block since the active
// encoder is a separate multi-GB pull.
diff --git a/client/src/test/videoGenPageMocks.jsx b/client/src/test/videoGenPageMocks.jsx
index 937fe98776..cf09048e3d 100644
--- a/client/src/test/videoGenPageMocks.jsx
+++ b/client/src/test/videoGenPageMocks.jsx
@@ -42,6 +42,11 @@ export const state = {
peers: [],
/** `getVideoGenStatus`; a spy so a suite can defer it, count calls or vary the payload. */
getVideoGenStatus: vi.fn(),
+ /**
+ * `getVideoGenModelContext`; the probe-free half the Model picker reads (#5835).
+ * A spy for the same reasons — a suite defers it to assert the page mid-flight.
+ */
+ getVideoGenModelContext: vi.fn(),
generateVideo: vi.fn(),
attach: vi.fn(),
eventSourceRef: { current: null },
@@ -61,7 +66,7 @@ export const state = {
universeStyle: DEFAULT_UNIVERSE_STYLE,
};
-const SPIES = ['getVideoGenStatus', 'generateVideo', 'attach', 'start', 'startWhenIdle', 'repair', 'cancel', 'refresh'];
+const SPIES = ['getVideoGenStatus', 'getVideoGenModelContext', 'generateVideo', 'attach', 'start', 'startWhenIdle', 'repair', 'cancel', 'refresh'];
/** Restore every documented default, including fresh spies. Call it first in `beforeEach`. */
export function resetVideoGenMockState() {
@@ -105,6 +110,19 @@ export const videoGenTermsGate = (termsId) => ({
licenseUrl: 'https://example.com/license',
});
+/**
+ * A `/model-context` payload over `models` — the model list plus the three
+ * shaping numbers the picker's auto-select reads, with no python probe behind
+ * them. The first model is the default unless overridden.
+ */
+export const videoGenModelContext = (models, overrides = {}) => ({
+ models,
+ defaultModel: models[0]?.id ?? null,
+ systemMemoryGb: 128,
+ fflfLtx2PixelBudget: 8_000_000,
+ ...overrides,
+});
+
/** A `/status` payload over `models`; the first model is the default unless overridden. */
export const videoGenStatus = (models, overrides = {}) => ({
connected: true,
@@ -122,6 +140,7 @@ vi.mock('../services/api', () => ({
// a media provider the picker renders nothing and every local path is unchanged.
getInstances: vi.fn(async () => ({ peers: state.peers })),
getVideoGenStatus: (...args) => state.getVideoGenStatus(...args),
+ getVideoGenModelContext: (...args) => state.getVideoGenModelContext(...args),
generateVideo: (...args) => state.generateVideo(...args),
cancelVideoGen: vi.fn(async () => ({})),
listVideoHistory: vi.fn(async () => []),
diff --git a/server/lib/apiRouteCatalog.generated.json b/server/lib/apiRouteCatalog.generated.json
index 9227c19c57..02ca6e6af8 100644
--- a/server/lib/apiRouteCatalog.generated.json
+++ b/server/lib/apiRouteCatalog.generated.json
@@ -16454,6 +16454,14 @@
"server/routes/videoGen.js"
]
},
+ {
+ "method": "GET",
+ "path": "/api/video-gen/model-context",
+ "mountPath": "/api/video-gen",
+ "sources": [
+ "server/routes/videoGen.js"
+ ]
+ },
{
"method": "GET",
"path": "/api/video-gen/model-terms",
@@ -17497,8 +17505,8 @@
],
"stats": {
"mounts": 147,
- "operations": 2167,
- "declarations": 2175,
+ "operations": 2168,
+ "declarations": 2176,
"sourceFiles": 230
}
}
diff --git a/server/routes/videoGen.js b/server/routes/videoGen.js
index 7e4e5cfc88..659dea1dac 100644
--- a/server/routes/videoGen.js
+++ b/server/routes/videoGen.js
@@ -88,6 +88,33 @@ const hardwareAwareVideoModels = async () => {
};
};
+// The model list plus the three numbers that decide which entry the picker
+// auto-selects. Deliberately free of any python probe: /status shells out to
+// the interpreter on every call (~1-2s) and the Model field used to wait on it,
+// so `/model-context` serves the same fields off the registry and the cached
+// hardware probe alone. /status keeps returning them for its other readers —
+// this is the single builder both routes share, so the two can't drift.
+const videoModelContext = async () => {
+ const { capabilities, models } = await hardwareAwareVideoModels();
+ return {
+ // Each entry carries its optional `disclosure` block (provenance, weights/
+ // runtime licenses, pinned-snapshot download size) straight off the
+ // registry — absent for custom models, which the UI renders as Unknown.
+ models,
+ defaultModel: defaultVideoModelId(capabilities),
+ // Total system memory in GB — the client uses this to auto-select the
+ // highest-memory mode-compatible model that fits on this machine.
+ // Rounded to nearest GB; sub-GB precision isn't useful for the
+ // model-size comparison and reads more cleanly in the UI.
+ systemMemoryGb: Math.round(os.totalmem() / 1024 ** 3),
+ // Effective FFLF/ltx2 stage-2 pixel-frame budget (honors
+ // FFLF_LTX2_PIXEL_BUDGET). The multi-keyframe picker mirrors the
+ // back-solve so it can reject out-of-budget keyframe indices before
+ // submit instead of letting the worker 400 mid-render.
+ fflfLtx2PixelBudget: resolveFflfLtx2PixelBudget(),
+ };
+};
+
// M4A files are stored in an MP4 container. Browsers and OS file pickers
// label them inconsistently: Safari uses `video/mp4`, Chrome/Firefox use
// `audio/mp4`, and some platforms emit `audio/x-m4a` or `audio/aac`.
@@ -400,18 +427,17 @@ router.get('/status', asyncHandler(async (_req, res) => {
const s = await getSettings();
const py = s.imageGen?.local?.pythonPath || null;
const { connected, reason, missing, pythonVersion } = await resolveLocalPythonHealth(py);
- const { capabilities, models } = await hardwareAwareVideoModels();
res.json({
connected,
pythonPath: py,
pythonVersion: pythonVersion || null,
reason,
missingPackages: missing,
- // Each entry carries its optional `disclosure` block (provenance, weights/
- // runtime licenses, pinned-snapshot download size) straight off the
- // registry — absent for custom models, which the UI renders as Unknown.
- models,
- defaultModel: defaultVideoModelId(capabilities),
+ // `models` / `defaultModel` / `systemMemoryGb` / `fflfLtx2PixelBudget` —
+ // kept here for the callers that already read them off /status. The Video
+ // Gen page takes them from GET /model-context instead, so its Model picker
+ // never waits on the python probe above.
+ ...(await videoModelContext()),
// Server-owned execution + policy scope per render backend (#3674). The
// client renders these strings verbatim so the wording can't drift between
// the two surfaces.
@@ -419,16 +445,6 @@ router.get('/status', asyncHandler(async (_req, res) => {
// Authoritative list of bring-your-own-venv runtimes — lets the client
// gate the install-banner probe without hardcoding the same Set.
byovRuntimes: Object.keys(BYOV_RUNTIME_INFO),
- // Total system memory in GB — the client uses this to auto-select the
- // highest-memory mode-compatible model that fits on this machine.
- // Rounded to nearest GB; sub-GB precision isn't useful for the
- // model-size comparison and reads more cleanly in the UI.
- systemMemoryGb: Math.round(os.totalmem() / 1024 ** 3),
- // Effective FFLF/ltx2 stage-2 pixel-frame budget (honors
- // FFLF_LTX2_PIXEL_BUDGET). The multi-keyframe picker mirrors the
- // back-solve so it can reject out-of-budget keyframe indices before
- // submit instead of letting the worker 400 mid-render.
- fflfLtx2PixelBudget: resolveFflfLtx2PixelBudget(),
// Runtime fingerprint — host chip/os + resolved ltx/mlx/torch versions per
// installed BYOV runtime — so the UI can show the exact numerical stack and
// bug reports for garbled/"mosaic" output carry the version info that makes
@@ -530,6 +546,14 @@ router.get('/models', asyncHandler(async (_req, res) => {
res.json(models);
}));
+// Everything the Model picker needs to render AND auto-select, with no python
+// probe in the way. A sibling route rather than a wrapper around /models so the
+// bare-array shape that route has always returned stays intact for its existing
+// callers (and for an older client talking to a newer server).
+router.get('/model-context', asyncHandler(async (_req, res) => {
+ res.json(await videoModelContext());
+}));
+
router.get('/models/status', asyncHandler(async (_req, res) => {
// Text encoder is shared across all video renders. A registry entry with
// `localPath` (e.g. an LM Studio install) trumps the HF cache check, so
diff --git a/server/routes/videoGen.test.js b/server/routes/videoGen.test.js
index 457c94ceb9..d4dba66796 100644
--- a/server/routes/videoGen.test.js
+++ b/server/routes/videoGen.test.js
@@ -567,6 +567,35 @@ describe('videoGen routes', () => {
});
});
+ describe('GET /model-context', () => {
+ it('serves the model list and its auto-select numbers without probing python', async () => {
+ const { checkPackages } = await import('../lib/pythonSetup.js');
+ checkPackages.mockClear();
+
+ const r = await request(app).get('/api/video-gen/model-context');
+ expect(r.status).toBe(200);
+ // The picker's whole input set, in one probe-free answer.
+ expect(r.body.models.map((m) => m.id)).toEqual(['ltx2_unified']);
+ expect(r.body.defaultModel).toBe('ltx2_unified');
+ expect(typeof r.body.systemMemoryGb).toBe('number');
+ expect(r.body.systemMemoryGb).toBeGreaterThan(0);
+ expect(typeof r.body.fflfLtx2PixelBudget).toBe('number');
+ expect(r.body.fflfLtx2PixelBudget).toBeGreaterThan(0);
+ // The point of the route: no interpreter subprocess in the request path.
+ expect(checkPackages).not.toHaveBeenCalled();
+ });
+
+ it('agrees with /status on every field the two share', async () => {
+ const [context, status] = await Promise.all([
+ request(app).get('/api/video-gen/model-context'),
+ request(app).get('/api/video-gen/status'),
+ ]);
+ for (const field of ['models', 'defaultModel', 'systemMemoryGb', 'fflfLtx2PixelBudget']) {
+ expect(status.body[field]).toEqual(context.body[field]);
+ }
+ });
+ });
+
describe('GET /models/:modelId/download — restricted terms', () => {
const h3CheckpointFiles = ['LICENSE', 'FL2VA/model_index.json', 'FL2VA/video_vae/source/model.safetensors'];
const h3 = {