opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations (speculative decoding/MTP) - #26477

Draft

wanghqc wants to merge 15 commits into

ggml-org:masterfrom

qualcomm:hq/quant-gemv-opt-intel-ready-r0730

Commits on Aug 3, 2026

opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations
wanghqc
committed
opencl: guard q4_K/q6_K tiled_ns convert-kernel registration for non-Adreno build
wanghqc
committed
opencl: gate q4_K MUL_MAT+GLU fusion dispatch to Adreno
wanghqc
committed
opencl: require the noshuffle weight layout in the q4_K GLU fusion gate
wanghqc
committed
opencl: do not take the vectorized f16 mrow GEMV path on an unaligned row stride
wanghqc
committed
opencl: pass the new get_scale_min_k4 stride argument at the row-major call sites
wanghqc
committed
opencl: enable the q4_K split-K decode GEMV only where it is measured to win
wanghqc
committed
opencl: record the X1-85 split-K datapoint (neutral, exclusion confirmed)
wanghqc
committed
opencl: restrict the tiled lm_head/embed GEMV default to X2E/A8X
wanghqc
committed
opencl: fix q4_K variant kernels to read the transposed scales layout
wanghqc
committed
opencl: keep the flat-GEMV large-m escape opt-in
wanghqc
committed
opencl: guard the o4 GEMV store against the rounded-up dispatch tail
wanghqc
committed
opencl: restore the tiled q4_K/q6_K layout on tensor read-back
wanghqc
committed
opencl: split-K for the q8_0 decode GEMV at small M
wanghqc
committed
opencl: keep the q6_K noshuffle correctness escape ahead of the opt-in gate
wanghqc
committed