]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
CUDA: Fix non-contig rope (llama/19338)
authorOliver Simons <redacted>
Sun, 8 Feb 2026 13:12:51 +0000 (14:12 +0100)
committerGeorgi Gerganov <redacted>
Sat, 14 Feb 2026 22:20:18 +0000 (00:20 +0200)
commitdfde635249cd390496eaea524529da41038c3c4a
treeffa607881674e4e56d7b1d60b1b62c62af091fa6
parent5cecdad692d868e28dbd2f7c468504770108f30c
CUDA: Fix non-contig rope (llama/19338)

* Rename variables + fix rope_neox

Seems memory layout is shared with Vulkan so we can port fix from
https://github.com/ggml-org/llama.cpp/pull/19299

* Fix rope_multi

* Fix rope_vision

* Fix rope_norm

* Rename ne* to ne0* for consistent variable naming

* cont : consistent stride names

---------

Co-authored-by: Georgi Gerganov <redacted>
src/ggml-cuda/rope.cu