Post Training Quantization of Large Language Models with Microscaling Formats

SmoothQuant, AWQ and GPTQ were each designed for fixed-point targets; none of them can emit the microscaling formats that new hardware natively supports.

Extends SmoothQuant, AWQ and GPTQ so each can quantize into MX formats (shared per-block power-of-two scale plus a narrow element type), and then studies how the three compose when stacked: activation smoothing (SmoothQuant) before activation-aware weight saliency (AWQ) before error-compensating rounding (GPTQ). The contribution is largely the compatibility work plus the interaction analysis, not a new algorithm.

Paper's own claim (abstract): combining the PTQ methods allows quantizing to 4-bit weights and 8-bit activations in the MXINT format 'with negligible accuracy loss compared to the uncompressed baseline'. COMPRESSION RATIO: 4-bit weights + 8-bit activations in MXINT (~4x on weights vs FP16, plus the shared-scale overhead). QUALITY COST: 'negligible' - no paired delta in the abstract. Venue recorded as 'other': ENLSP 2024 is not in this schema's venue enum.

arXiv ID 2405.07135 verified via /abs/. Venue ENLSP 2024 from Semantic Scholar; the schema's venue enum has no ENLSP member, so venue='other' and the actual venue is named in notes. Year = venue year 2024. The repo already has source quant-mx-ptq-paper for this paper. arXiv ID and title verified by fetching https://arxiv.org/abs/2405.07135 and reading the citation_title meta tag on 2026-10-03. Venue and year from the Semantic Scholar graph API batch endpoint (paper/arXiv:2405.07135) unless the record's notes say otherwise. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. Peer-reviewed venue confirmed as ENLSP 2024 (workshop-style venue, enum has no member). This is an additive track annotation; `venue` is unchanged at 'other' and no status change is implied by the track alone. VENUE RECLASSIFIED 2026-10-10 (crawler/w12-venue-rest): venue 'other' -> 'enlsp', venue_kind 'conference'. Confirmed by PMLR v262 proceedings HTML at https://proceedings.mlr.press/v262/sharify24a.html ('Post Training Quantization of Large Language Models with Microscaling Formats', PMLR 262:232-240) and agreeing source record notes.