SSRQ: Structured Stochastic Rounding for Quantization Aware Post Training on FP3
17 Jul 2026
Most low-bit work stops at FP4 / INT4. SSRQ is the first to close the FP3–FP16 gap: packed FP3 (~81% smaller) within ~1.5 pp of FP16 float on mlp9 (94.2% vs 95.7%; PTQ collapses to 71.3%) — plain RTNE deploy, no extra params.