]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
CUDA: better coalesce data-access for contiguous concat (llama/22330)
authorOliver Simons <redacted>
Sun, 26 Apr 2026 07:21:45 +0000 (09:21 +0200)
committerGeorgi Gerganov <redacted>
Thu, 30 Apr 2026 08:29:20 +0000 (11:29 +0300)
commit9bf6c3c8602b976b79139d908fd63ffe048749ba
treedf584233f843f90f143509d805ee24968be97511
parent2f3df42cddca762047c2884342b683549420be71
CUDA: better coalesce data-access for contiguous concat (llama/22330)

Also, distribute all elements across CTAs evenly instead of launching
one CTA per dim
ggml/src/ggml-cuda/concat.cu