perf: add low-latency nz quantized decode matmul. - #42

Open

maojunx99 wants to merge 3 commits into

xLLM-AI:mainfrom

maojunx99:perf/qwen3vl-decode-nofusion

Commits on Aug 17, 2026

perf: optimize non-fused nz decode matmul.
maojunx99
committed
a1a6de0

Commits on Aug 20, 2026

perf: reduce decode matmul vector overhead.
maojunx99
committed
aafec86
perf: specialize decode matmul shapes
maojunx99
committed
74686ce