[Widgets] Don't call api inference if output exists

[mishig25]

Closes #1011
This PR is continuation of #1017
Follow up to #979

#979 correctly implemented behaviour for the previewInputSample function here. Which is: don't call api-inference if sample.output exists

The same should be the case for applyInputSample (which is this PR). Don't call api-inference if sample.output exists. #1011 happened because api-inference was called and there was no cached example although sample.output was in the repo

Order of merge operations

1. Merge [Widgets] Refactor examples running #1023
2. Then merge [Widgets] Don't call api inference if output exists #1063 (this PR)
3. Then merge [Widgets] Handle display logic #1019

[SBrandeis]

The logic is not implemented for most widgets. Is that intended?

[SBrandeis]

This has no effect in this widget

[mishig25]

yes, because widgets such as AudioToAudio or ImageSegmentation are not expected to have example.output in their model cards

for example here

hub-docs/js/src/lib/components/InferenceWidget/widgets/AudioToAudioWidget/AudioToAudioWidget.svelte

Line 131 in e5cdf0a

function applyInputSample(sample: WidgetExampleAssetInput, opts: ExampleRunOpts = {}) {

, if they were expected to have example.output, the typing would have been sample: WidgetExampleAssetInput<WIDGET OUTPUT TYPE>

docs here for the list of supported output types: https://huggingface.co/docs/hub/models-widgets#example-outputs from #978

@julien-c could you confirm this statement

yes, because widgets such as AudioToAudio or ImageSegmentation are not expected to have example.output in their model cards

[julien-c]

i'm not sure i understand, i'm a bit lost...

[julien-c]

from the docs (emphasis mine):

each widget example can also optionally describe the corresponding model output, directly in the output property

AudioToAudio or ImageSegmentation should support a file output, i.e. a output.url, same as text-to-image:

[mishig25]

Here is what I meant:

At the current state of widgets, uses can supply output in their modelcard for any widget. However, only subset of widgets will show/use that example output

AudioToAudio or ImageSegmentation should support a file output

But in the current implementation, AudioToAudio or ImageSegmentation do neither validation nor usage of output

Should I open a PR (extend the current PR) to make all the remaining widgets validate and use output ?

[mishig25]

As we discussed internally, lets merge this PR without the proposed changes above (remaining widgets validate and use output). And iterate in subseq PRs

[julien-c]

Should I open a PR (extend the current PR) to make all the remaining widgets validate and use output ?

Yes

[julien-c]

that was the goal to support output for all widgets