i tried ternary decomposition instead of quantization. it works as good at q4km but takes slightly more vram. while being completly ternary. and completly PTQ (no QAT)
i tried ternary decomposition instead of quantization. it works as good at q4km but takes slightly more vram. while being completly ternary. and completly PTQ (no QAT) — reported by reddit.com, aggregated and ranked by ClawDigest.