Fix PAB cond/uncond cache cross-contamination - #25
Draft
cursor[bot] wants to merge 1 commit into
Draft
Conversation
PyramidAttentionBroadcastHook stored iteration and cache on a single hook-level state object that ignored cache_context. Pipelines running separate cond and uncond forwards (CogView4, Flux true CFG, etc.) would reuse conditional attention outputs during the uncond pass when PAB skipped computation, silently corrupting guidance. Migrate PAB to StateManager so cond/uncond contexts maintain isolated state, matching MagCache and FirstBlockCache. Add a regression test and update pipeline PAB tests to use state_manager. Co-authored-by: Simon Lynch <srlynch1@users.noreply.github.com>
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Fixes a critical correctness bug in Pyramid Attention Broadcast (PAB) where hook state was shared across
cache_contextboundaries.Bug and impact
When PAB is enabled on pipelines that run separate conditional and unconditional transformer forwards (e.g. CogView4 with CFG, Flux with true CFG, QwenImage, etc.), the uncond pass could reuse cached cond attention outputs on skip steps. This silently corrupts classifier-free guidance and produces wrong generations.
Root cause
PyramidAttentionBroadcastHookstorediterationandcacheon a single hook-levelself.stateobject.ModelHook._set_context()only propagates toStateManagerattributes, so enteringcache_context("cond")vscache_context("uncond")did not partition PAB state.Fix
StateManager(same pattern as MagCache / FirstBlockCache)"inference"context when none is set (batch-concatenated pipelines)test_pab_isolates_state_between_cond_and_uncond_contextsstate_managerValidation
pytest tests/hooks/test_pyramid_attention_broadcast.pypasses