improve CUDA cpy memory bandwidth when copying transposed tensor #16841

Open

bssrdf wants to merge 11 commits intoggml-org:masterfrom

bssrdf:cuda-transpose-cpy

Commits on Oct 29, 2025

Commits on Oct 30, 2025

reformulated to handle more complicated transpose cases
bssrdf
committed
merged ok
bssrdf
committed
bring back 2D transpose for higher performance
bssrdf
committed
allow build on windows
bssrdf
committed
tranpose copy more shapes
bssrdf
committed

Commits on Oct 31, 2025

minor tweak
bssrdf
committed
final clean up
bssrdf
committed