]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 5fa1c19)
CUDA: Fix non-contig rope (#19338)
authorOliver Simons <redacted>
Sun, 8 Feb 2026 13:12:51 +0000 (14:12 +0100)
committerGitHub <redacted>
Sun, 8 Feb 2026 13:12:51 +0000 (15:12 +0200)
commite06088da0fa86aa444409f38dff274904931c507
tree3217fe022444354ff85bd2f8a0f27b5ea0991166tree
parent5fa1c190d9fc86c02698b730a2cb933195e19d96commit | diff
CUDA: Fix non-contig rope (#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>
ggml/src/ggml-cuda/rope.cu diff | blob | history
Packaging of ggml-org/llama.cpp