perf(ROCm): replace software Dp4a with hardware v_dot4_i32_iu8 - #2363

Closed

ghazni101 wants to merge 5 commits into

mudler:mainfrom

ghazni101:row/ROCM-HW-DP4A

Commits on Sep 2, 2026

spec(BACKEND-ROCM): hardware Dp4a for KQuantGemmK (v_dot4_i32_iu8)
ghazni101
committed
9d18372
perf(ROCm): replace software Dp4a with hardware v_dot4_i32_iu8
ghazni101
committed
1c438d3
record(BACKEND-ROCM): re-measure A/B on the final PR head
ghazni101
committed
f337f4d
fix(ROCM-HW-DP4A): register the new checker's creation mutation, which pr-size demands
mudler
committed
e0454b3
fix(ROCm): correct instruction name v_dot4_i32_iu8 → v_dot4_i32_i8
ghazni101
committed
ad67e79