Accuracy cost of quantization: measured deltas on named benchmarks

PROXY ENUM VALUE - DECLARED 2026-10-04 (minor): this is a cross-method SUMMARY of measured accuracy deltas, filed under 'gptq'. Its three sources are Marlin, SpQR and SignRound, of which only SpQR and SignRound are GPTQ-adjacent and neither IS GPTQ. Use it for the accuracy comparison, not as a GPTQ record. GROUP SIZE LEFT NULL 2026-10-04: this is a SUMMARY record that aggregates measured accuracy deltas across methods whose group sizes differ (the underlying records are gptq-128, spqr-128, signround-128). Recording one value would attribute a single configuration to a cross-method comparison. This record addresses the repo's priority: 'most of this literature reports speed and omits accuracy - prioritize sources that report measured accuracy DELTA on a named benchmark.' The summary above is compiled from the records in this repo. The key finding: 4-bit weight-only quantization is nearly lossless for most models (0.1-0.3 ppl), but 3-bit and below show significant degradation. KV cache quantization at 4-bit is nearly lossless, but at 2-3-bit the accuracy cost is model-dependent. The most dramatic accuracy delta is SignRound's 6.91-33.22% improvement at 2-bit, which shows that optimized rounding matters enormously at very low bit widths. The least documented area is FP8/FP4 KV cache quantization, where no controlled benchmark with named accuracy delta was found. This is a significant gap in the literature. 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: NOT A STORAGE FORMAT, SO THE FIELD IS DEFINITIONALLY EMPTY. This record describes a calibration file, a search orchestrator, removed enum slots or a cross-method summary rather than a wire format, and no accelerator can natively support any of those. Consequence for a reader: nothing further to verify - the question does not apply to this record. 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.