Paper: E8-vs-KVarN iso-bit-width head-to-head + frontier-leadership positioning#43
Merged
Conversation
…ntier-leadership positioning claim
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Adds a faithful iso-bit-width comparison against KVarN (the nearest calibration-free rotation-based KV-quant rival; Sinkhorn port verified exact against the authors' reference, 0.00 fp32 deviation). At matched K4V2 on Mistral-Inst-v0.3: NQ reaches +0.267% PPL at 3.125 bpe vs KVarN +0.421% at 3.375 bpe (lower delta at lower bit-budget), with KVarN edging the single 4K NIAH cell (5/5 vs 4/5, within FP16 4/5 noise). Framed honestly as comparable, not a win. Also adds a source-verified positioning claim: NQ K2V2 (~1.1 entropy-bpe) is, to our knowledge, the lowest-bit calibration-free operating point with needle-retrieval validation (every calibration-free NIAH/RULER method sits at >=2 bits; the only sub-1.1-bit needle-validated methods are calibrated). Every cell independently verified against the source JSON; compiles clean (0 errors, 88pp). PDF + KVarN result sidecar.