[allozaur]

Overview

Picking a model that is already loading no longer starts a second load, and a chat without a model of its own pre-selects the model that is loading. Automatic selection also stops picking models hidden from the selector.

Demo

Recorded at 3a8c7ad; click the preview for the full video.

Additional information

The demo doesn't cover hidden models.

Cause

• The selector only checked that a model wasn't loaded before loading it. status.load() skips loads this client already started, but not a model the server reports as loading (started from another tab, the manager, or before a reload), so picking it sent another load request.
• Pre-selection went from the last reply's model straight to the first loaded model, so a model being loaded lost to one already up.
• Automatic selection only skipped models flagged ui: false by the server, not the ones hidden in the manager, so it could land on a model the selector doesn't list.

What changed

• modelsStore.isModelLoading() covers both the server's loading status and a load this client started. handleSelect in use-models-selector only loads a model that is neither loaded nor loading.
• ensureFirstModelSelected() (new chat) tries a loading model right after the last reply's model, before loaded models, favorites and the first available model.
• getLoadingModelToPreselect() lets a chat without a model in its replies swap its selection for a loading model, unless the selection is loaded or loading itself. A new chat and selectModelFromLastAssistantResponse() (opening a chat) use it, and the chat form's fallback tries a loading model before a loaded one.
• getVisibleModels() also skips models hidden in the manager, and the chat form's fallback uses it instead of every model. A conversation whose replies name a hidden model still selects it.

Testing

• Type-check, lint and unit tests pass.
• Recorded the demo against a router: picking a loading model again sent no second load request, and after a reload the loading model was pre-selected over the loaded one, with the script checking the server still reported it as loading at both points.

Demo script

test('demo', async ({ demo, page }) => {
	const LOADING = 'InternScience/Sage-A1.5-Q8_0-GGUF:Q8_0';
	let loadRequests = 0;
	page.on('request', (request) => {
		if (request.method() === 'POST' && request.url().endsWith('/models/load')) loadRequests++;
	});

	const status = async () => {
		const response = await page.request.get('http://localhost:9931/models');
		const { data } = await response.json();

		return data.find((m: { id: string }) => m.id === LOADING)?.status?.value;
	};
	const trigger = (name: RegExp) => page.getByRole('button', { name }).first();
	const pickSage = async () => {
		await demo.click(trigger(/Groot|Sage/), 300);
		await demo.type(page.getByPlaceholder('Search models...'), 'Sage', 200);
		await demo.click(page.getByRole('option', { name: /Sage/ }).first(), 300);
	};

	await expect(trigger(/Groot/)).toBeVisible({ timeout: 30_000 });
	await demo.caption('A new chat starts on the model that is already loaded', 1500);

	// pick an unloaded model, then pick it again while it loads
	await pickSage();
	await pickSage();
	expect(await status()).toBe('loading');
	const requests = loadRequests;

	// a reload drops the pick: the loading model wins over the loaded one
	await page.reload();
	await expect(trigger(/Sage/)).toBeVisible({ timeout: 30_000 });
	expect(await status()).toBe('loading');
	await demo.caption('After a reload, the loading model is pre-selected over the loaded one', 2200);
	await demo.caption(`Picking it twice while it loaded sent ${requests} load request`, 2200);
});

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, Claude (Opus 5.5 Medium)

[ahoplock]

This is unrelated to this change but the "first loaded model" logic will select a model even when it's hidden from the UI.

[allozaur]

This is unrelated to this change but the "first loaded model" logic will select a model even when it's hidden from the UI.

could be addressed in this PR :) thanks for the heads up!

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[ServeurpersoCom]

One small thing: isModelLoading() also returns true while this client is unloading a model, since unload() sets the same loadingStates flag, so a model being unloaded can be pre-selected as the loading one. Checking this._status.isOperationInProgress(modelId) && !this.isModelLoaded(modelId) would keep only loads, as the status stays LOADED until UNLOADED.

[ahoplock]

could be addressed in this PR :) thanks for the heads up!

Thanks, works as expected now!