Batch decode tuning - #7

Closed

SimonTeixidor wants to merge 6 commits into

Nathanw1014:strix-halo-vulkanfrom

SimonTeixidor:strix-halo-vulkan

Commits on Aug 30, 2026

vulkan: use 4 rows per workgroup for mat-vec on RDNA3 above 3 columns
SimonTeixidor
committed
f9b8073
vulkan: use 4 rows per workgroup for mul_mat_vec_id on RDNA3
SimonTeixidor
committed
d10419c
vulkan: add integer-dot mat-vec pipeline for IQ4_XS
SimonTeixidor
committed
dd76dfb
vulkan: enable MMVQ for Q6_K on RDNA3 above one column
SimonTeixidor
committed
dbf549b
vulkan: hoist the Q2_K activation sum out of the row loop
SimonTeixidor
committed
04c9576
vulkan: gate the RDNA3 mat-vec tuning on UMA, and Q6_K MMVQ from four columns
SimonTeixidor
committed
88a6ff0