Repository navigation
Reduce PSRN scoring memory with ordered candidate tiles - #2
Merged
Merged
Conversation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
PSRN scoring currently materializes full candidate outputs and operand-index tables, increasing memory use for large symbolic searches. This change reconstructs selected operand indices on demand and scores ordered final-layer tiles, with automatic dense/tiled dispatch based on estimated tensor storage and sample batching.
Preserves sample accumulation order and candidate ordering, with dense fallback for autocast, nonfloating inputs, and CPU Pow. Reuses dense residual storage outside autocast and supports legacy Torch dtype/autocast APIs.
Scope: two implementation modules and three Python regression test files. No PDF, documentation, assets, datasets, workflow, or dependency changes are included.
Review: inspected the code diff and searched source/config text for Alibaba/ByteDance identifiers, private endpoints, and credential markers; no indicators were found. This is an indicator review, not proof of provenance.
Validation: tests were not executed in this session, following the user's instruction to skip extra validation. Included tests cover dense/tiled scoring, operand geometry, dispatch, dtype promotion, and autocast behavior.