docs(research): GPTQ max-exactness fix -- partial, closes ~35% of the RTN gap - #554
Open
gHashTag wants to merge 1 commit into
Open
docs(research): GPTQ max-exactness fix -- partial, closes ~35% of the RTN gap#554gHashTag wants to merge 1 commit into
gHashTag wants to merge 1 commit into
Conversation
… RTN gap, does not clear the gate Runs the fix proposed but not executed in THEOREM_2026-08-09.md: protect each group's per-row maximal element from GPTQ's sequential compensation (equivalent to processing it first, since an exactly-quantised column propagates zero error). - max-exactness restored: 95.3% -> 0.4% blocks with an inexact row-max - perplexity: GPTQ-original 17.7047 -> GPTQ-fixed 17.5860 (-0.1186, same run) - still worse than RTN 4-bit (17.3667) by +0.2194 -- gptq_gate.py's pass condition remains unmet - reproduction note: this environment's GPTQ-original (17.7047) differs from THEOREM's (17.7846) by 0.08; all comparisons here are within-run to control for that Conclusion: max-exactness destruction explains part (~35%) of GPTQ's gap to RTN under block-max scaling, not all of it. Residual mechanism unidentified -- open question. Adds: research/block/gptq_maxexact_fix.py, research/block/GPTQ_MAXEXACT_FIX_2026-08-11.md
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Runs the fix proposed but not executed in THEOREM_2026-08-09.md: protect each group's per-row maximal element from GPTQ's sequential compensation (equivalent to processing it first, since an exactly-quantised column propagates zero error).
gptq_gate.py's pass condition remains unmetConclusion: max-exactness destruction explains part (~35%) of GPTQ's gap to RTN under block-max scaling, not all of it. Residual mechanism unidentified -- open question, not run further (5-bit / promote-only / cross-model out of scope for this PR).
Adds
research/block/gptq_maxexact_fix.py,research/block/GPTQ_MAXEXACT_FIX_2026-08-11.md.