hooks/pyramid_attention_broadcast: fix redundant recompute at iteration 0 and free stale cache when outside timestep range#13467
Open
GitGlimpse895 wants to merge 1 commit intohuggingface:mainfrom
Conversation
…on 0 and free stale cache when outside timestep range
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Fixes two bugs in
PyramidAttentionBroadcastHook.new_forward:Redundant
iteration == 0condition —self.state.cache is Nonealreadycovers the first-call case after every
reset_state, making the extra guarddead code that creates a misleading impression of two independent invariants.
Stale cache leaking GPU VRAM — when outside the active timestep range,
the hook was still writing
self.state.cache = output, holding a fullhidden-state activation tensor on GPU until the next generation's
reset_statecall. For video transformers with dozens of PAB-hooked layersthis accumulates hundreds of MBs of unreleased VRAM. The fix sets
self.state.cache = Noneimmediately when outside the range.Fixes # (issue)
Before submitting
Who can review?
@yiyixuxu @sayakpaul @DN6