]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 72ff528)
metal : faster q4_0 (#1775)
authorKawrakow <redacted>
Fri, 9 Jun 2023 07:39:59 +0000 (10:39 +0300)
committerGitHub <redacted>
Fri, 9 Jun 2023 07:39:59 +0000 (10:39 +0300)
commit245fc3c37da5ac5963f9f11a9f4f2ac08d96afc6
treeb2312b5b19a6887526d9e25d41b29eb4fdbcd49etree
parent72ff5282bf0388c60821f504c4c8cc2b1f491aa6commit | diff
metal : faster q4_0 (#1775)

* metal : 8% faster q4_0

Avoid copying into local uchar4 anf float4.

* metal : 17% faster Q4_0

Use 64 threads in a thread group.

---------

Co-authored-by: Iwan Kawrakow <redacted>
ggml-metal.m diff | blob | history
ggml-metal.metal diff | blob | history
Packaging of ggml-org/llama.cpp