feat(#1189): packed-tensor mapped staging — Q4_K/Q6_K served straight from the mmap on Android - #1190

Merged

michalharakal merged 6 commits into

feature/1189-mapped-packed-staging

Aug 27, 2026

Commits on Aug 27, 2026

feat(#1189): packed-tensor mapped staging — serve Q4_K/Q6_K blocks straight from the mmap

michalharakal
and
claude
committed
5fa5858
fix(#1189): plan charges mapped weights against device RAM, not the heap budget

michalharakal
and
claude
committed
ac28fb5
docs(#1189): kernel-support matrix grows a mapped-serving section

michalharakal
and
claude
committed
576af98
feat(#1195): thread the Q4_K/Q6_K matmul kernels — 153 → 61 ms/step on the Pixel 8a

michalharakal
and
claude
committed
3c53ea6
harness(#1193): residency knob for the M2-A5 harness + the mixed-quant fallback trap, measured

michalharakal
and
claude
committed
f281e2f
Merge pull request #1196 from SKaiNET-developers/feature/1195-threaded-packed-kernels
michalharakal
authored
5d1395f