docs: the tagline named the codebook the flagship path replaced - #63
Merged
Conversation
"E8-lattice post-training quantization" described one of four codebooks, and not the
recommended one. `--codebook` accepts e8_shell, e8_relaxed, e8p and trellis; the trellis
path is QTIP TCQ, which reaches effective dimension 256 with a lookup-free decode and is
explicitly not an 8-dimensional lattice.
The contradiction was visible on one screen: the first line said E8-lattice, and a few lines
later "NEW in v0.7 — trellis (TCQ) codebook" plus a checkpoint table marking the trellis
build as the fastest GLQ decode. It also undersold the work — 2 bpw trellis beats 2 bpw e8p
on SmolLM3-3B at PPL 11.94 vs 13.79 (bf16 9.12), precisely because it is not a lattice.
The mechanism paragraph had the same problem: it described E8 index encoding as what "GLQ"
does. It now leads with trellis as the recommended path and keeps E8 in its actual role —
which is more than backward compatibility. `--codebook` still defaults to e8_shell, and
trellis hard-errors on anything but a uniform integer bpw ("requires a uniform integer bpw
(2-8)"), so fractional and per-layer mixed precision are E8-only. Claiming E8 was kept
merely to load old checkpoints would have been the opposite error to the one being fixed.
The GitHub repo description is updated to match (takes effect immediately). The pyproject
description reaches PyPI only on the next release, so 0.8.7's page keeps the old wording.
cnygaard
force-pushed
the
docs/tagline-lattice-and-trellis
branch
from
August 17, 2026 21:48
ccb80a0 to
ee07ec5
Compare
cnygaard
enabled auto-merge (squash)
August 17, 2026 21:48
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.
README.mdopened with "E8-lattice post-training quantization", which describes one of four codebooks — and not the one the project recommends.--codebookacceptse8_shell,e8_relaxed,e8pandtrellis. The trellis path is QTIP TCQ; from its own module docstring it "reaches effective quantization dimension 256 with a lookup-free / small-LUT decode, beating GLQ's 8-D E8 lattice at low bit-rates".The contradiction was visible on a single screen: line 3 said E8-lattice, line 17 said "NEW in v0.7 — trellis (TCQ) codebook", and the checkpoint table marks
SmolLM3-3B-trellis-3inst-4bpw-kernelas "fastest GLQ decode".It also undersold the result: 2 bpw trellis beats 2 bpw e8p at PPL 11.94 vs 13.79 (bf16 9.12) because it is not an 8-dimensional lattice.
Changed
pyproject.tomldescription — same correction. This reaches PyPI only on the next release, so 0.8.7's page keeps the old wording.gh repo edit, already live.No code change; no claim about speed, footprint or quality is altered.