inference/research
verifiedhigh · 0.88updated 2026-10-03

QuaRot reference implementation (spcl/QuaRot GitHub repository README)

repo

Fields

id
sources/qlab-quarot-github
Publisher
spcl (Stanford / Sparsity in ML) on GitHub
Kind
repo
Published
2024-03-30
Accessed
2026-10-03
notes
Code repository referenced by qlab-quarot-paper. Confirms the repo ships a fake_quant (simulation) directory and custom kernels, i.e. the W4A4 numbers are research code, not a silicon int4 path. CAUTION: the README abstract states 'losses of at most 0.29 WikiText perplexity' for LLaMa2-70B while the paper's v2 abstract on arXiv states 'at most 0.47 WikiText perplexity'. Same claim, two values, different revision of the text. Recorded rather than resolved; treat both as unconfirmed against the experiments section. | SOURCE AUDIT 2026-10-03: url was https://arxiv.org/abs/github-spcl-quarot, an unexpanded template that arXiv returns 404 for ('Article identifier not recognized'); kind was also 'paper' although the document is a code repository. Repointed to the real repo https://github.com/spcl/QuaRot (confirmed live 2026-10-03 via gh api: "Code for Neurips24 paper: QuaRot, an end-to-end 4-bit inference of large language models", 534 stars) and kind set to repo. Claimed content (fake_quant dir, README PPL 0.29) re-verified against that repo.

Notes

Code repository referenced by qlab-quarot-paper. Confirms the repo ships a fake_quant (simulation) directory and custom kernels, i.e. the W4A4 numbers are research code, not a silicon int4 path. CAUTION: the README abstract states 'losses of at most 0.29 WikiText perplexity' for LLaMa2-70B while the paper's v2 abstract on arXiv states 'at most 0.47 WikiText perplexity'. Same claim, two values, different revision of the text. Recorded rather than resolved; treat both as unconfirmed against the experiments section. | SOURCE AUDIT 2026-10-03: url was https://arxiv.org/abs/github-spcl-quarot, an unexpanded template that arXiv returns 404 for ('Article identifier not recognized'); kind was also 'paper' although the document is a code repository. Repointed to the real repo https://github.com/spcl/QuaRot (confirmed live 2026-10-03 via gh api: "Code for Neurips24 paper: QuaRot, an end-to-end 4-bit inference of large language models", 534 stars) and kind set to repo. Claimed content (fake_quant dir, README PPL 0.29) re-verified against that repo.

Cross-references

This record references no other records.