]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
overview / pkg / ggml / sources / ggml / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 2a79110)
Allow multiple copy function pointers for CUDA graph kernel param updates (llama...
authoragray3 <redacted>
Mon, 27 May 2024 17:33:42 +0000 (18:33 +0100)
committerGeorgi Gerganov <redacted>
Wed, 29 May 2024 10:16:38 +0000 (13:16 +0300)
commite31cd3a07647ecf75a8ffb8cb78b17c51ce37d2a
tree0b3b83d196cf41d2372916d6bfbb27c87c52223ftree
parent2a791109462855f1c793219b4301706d6039c552commit | diff
Allow multiple copy function pointers for CUDA graph kernel param updates (llama/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
src/ggml-cuda.cu diff | blob | history
Packaging of ggml-org/ggml