QuARX (GPU-efficient 4-bit-weight / 8-bit-activation quantization) - UNVERIFIED PLACEHOLDER

GROUP SIZE LEFT NULL 2026-10-04 and the record itself flags the uncertainty: the bits field reads 'w4a8 (believed; not confirmed against a primary source)'. QuARX has only a publications-listing source in this repo, no kernel and no published config. Filling a group size here would be inventing a second unverified number on top of an unverified first one. DO NOT TRUST THIS RECORD AS WRITTEN. It exists because the quantization schema has a 'quarx' enum value and this slice was asked to cover it, but I could not locate a primary source for QuARX and therefore could not fill any field with a sourced fact. Searched on 2026-10-03, all negative: the arXiv API (all fields, and title), arXiv's web search and its full-text search; GitHub repo, code and issue search; HuggingFace model and paper search; Semantic Scholar; OpenReview's search API; OpenAlex; Crossref; the NeurIPS 2024 proceedings index (grepped locally, no match); and two general web-search backends. Separately, Albert Tseng's own publication list - read directly, and cited here as the one relevant negative result - does not mention QuARX, though it does list the neighbouring work (QuIP#, QTIP, YAQA) from the same Cornell-RelaxML lineage that QuARX is associated with in community usage. bits='w4a8' is my recollection of the scheme's positioning inferred from its name and lineage, NOT a sourced figure, which is exactly why confidence is 0.15. Empty native_support/emulated_support means 'unknown', not 'none' - do not read it as a claim that no hardware supports it. ACTION FOR A HUMAN: locate the QuARX paper or repo, register it as a source, and either fill or delete this record. Everything in the other records in this directory that discusses rotation-based W4A4 (SpinQuant, QuIP#, QuaRot lineage) is independently sourced; QuARX is the only gap. RE-VERIFICATION ATTEMPT (2026-10-03, later the same day, independent of the search log already recorded above): searched the arXiv API across all-fields and title-field queries built around the tokens 'RoRot', 'ZipQuant', 'Macaron Hessian', 'affine 4-bit quantization', 'FlatRot' and 'QuARX' itself, plus arXiv title queries for each, and searched OpenAlex by title and by full-text, and Crossref by title and by bibliographic query. Zero hits for QuARX, FlatRot, ZipQuant, or a Macaron LLM mixed-precision paper on any of them. This is a third independent negative result on the same repo, arrived at through different backends, which strengthens the case that QuARX has no locatable primary source rather than that previous searches were too narrow. The nearest real neighbours that DID resolve - Atom (mixed-precision block-outliers, W4A4), LQER (weight + low-rank correction), LLM-FP4 (W4A4 FP4) - are now recorded in this directory, so if QuARX turns out to be real it is most likely one of those under a different name. RECORD STILL UNVERIFIED; keep confidence at 0.15 and status draft until a human locates a primary source. NATIVE_SUPPORT CLASSIFIED 2026-10-04: this field is still [] and that is CORRECT, but the empty list used to be ambiguous between three different states and now is not. Reason class on this record: W4A4/W8A8 RESEARCH FORMAT - this empty list is a REAL FINDING about the absence of a silicon path, not a tautology. GENUINELY UNCHECKED, and the only record in this repo with BOTH native_support and emulated_support empty. Its own bits field already reads 'w4a8 (believed; not confirmed against a primary source)', so the format itself is unverified, let alone its hardware support. Consequence for a reader: this is the class where a vendor could in principle have shipped something later, so it is the only class worth re-checking against a new hardware release. The full statement of the three-way distinction, and the worked list of all 72 empty records by class, is in gotcha quant-native-support-empty-means-three-different-things. Rule of thumb for this field everywhere in the repo: if the bits field carries an 'a16', or the scheme is KV-cache-only or a llama.cpp integer storage type, an empty native_support is a DEFINITION and needs no verification; only a W4A4/W8A8/sub-8-bit format with research kernels can have an empty list that is a real finding.