Skip to content

Expose attribution forward progress for runtime estimation (#1933) - #1933

Open
Rollingkeyboard wants to merge 1 commit into
meta-pytorch:masterfrom
Rollingkeyboard:export-D120368627
Open

Rollingkeyboard wants to merge 1 commit into
meta-pytorch:masterfrom
Rollingkeyboard:export-D120368627

Conversation

@Rollingkeyboard

@Rollingkeyboard Rollingkeyboard commented Sep 16, 2026 •

Copy link
Copy Markdown

Summary:

Replace the thread-local progress-bar interception with an explicit model-forward progress contract. Attribution algorithms publish an exact batch-scoped forward plan, including grouped and individual-feature work, and FeatureAttributor reports only successfully completed model forwards. Failed forwards invalidate the estimate.

This keeps the existing FeatureAttributor.set_progress_callback(total, increment) consumer API while supporting first-batch ETA after the first completed Captum forward. Unsupported algorithms disable reporting instead of publishing an unsafe estimate.

Differential Revision: D120368627

@meta-cla meta-cla Bot added the cla signed label Sep 16, 2026
@meta-codesync

meta-codesync Bot commented Sep 16, 2026

Copy link
Copy Markdown
Contributor

@Rollingkeyboard has exported this pull request. If you are a Meta employee, you can view the originating Diff in D120368627.

@meta-codesync meta-codesync Bot changed the title Expose attribution progress for runtime estimation Expose attribution forward progress for runtime estimation Sep 18, 2026
@meta-codesync meta-codesync Bot changed the title Expose attribution forward progress for runtime estimation Expose attribution forward progress for runtime estimation (#1933) Sep 21, 2026
Rollingkeyboard pushed a commit to Rollingkeyboard/captum that referenced this pull request Sep 21, 2026
…rch#1933)

Summary:

Replace the thread-local progress-bar interception with an explicit model-forward progress contract. Attribution algorithms publish an exact batch-scoped forward plan, including grouped and individual-feature work, and FeatureAttributor reports only successfully completed model forwards. Failed forwards invalidate the estimate.

This keeps the existing FeatureAttributor.set_progress_callback(total, increment) consumer API while supporting first-batch ETA after the first completed Captum forward. Unsupported algorithms disable reporting instead of publishing an unsafe estimate.

Differential Revision: D120368627
…rch#1933)

Summary:

Replace the thread-local progress-bar interception with an explicit model-forward progress contract. Attribution algorithms publish an exact batch-scoped forward plan, including grouped and individual-feature work, and FeatureAttributor reports only successfully completed model forwards. Failed forwards invalidate the estimate.

This keeps the existing FeatureAttributor.set_progress_callback(total, increment) consumer API while supporting first-batch ETA after the first completed Captum forward. Unsupported algorithms disable reporting instead of publishing an unsafe estimate.

Differential Revision: D120368627

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant