]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 95f84d5)
Allow multiple copy function pointers for CUDA graph kernel param updates (#7565)
authoragray3 <redacted>
Mon, 27 May 2024 17:33:42 +0000 (18:33 +0100)
committerGitHub <redacted>
Mon, 27 May 2024 17:33:42 +0000 (19:33 +0200)
commit197c00681b80f9dea17d11a4436b6b8ef1be0ce8
tree300aaa9d0d9c3741f62a3e21bc0c041aef197410tree
parent95f84d5ce8b449a9b16009434aca800df504a02ecommit | diff
Allow multiple copy function pointers for CUDA graph kernel param updates (#7565)

CUDA graphs require parameter updates to kernels associated with
GGML_OP_CPY nodes. Previously the implementation only checked for a
single CUDA kernel in such nodes, but this caused a bug in cases where
2 such kernels exist. This fixes the issue by using a vector to allow
multiple function pointers to be stored and checked against.

Fixes #7942
ggml-cuda.cu diff | blob | history
Packaging of ggml-org/llama.cpp