]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: d5a409e)
Q6_K AVX improvements (#10118)
authorEve <redacted>
Mon, 4 Nov 2024 22:06:31 +0000 (22:06 +0000)
committerGitHub <redacted>
Mon, 4 Nov 2024 22:06:31 +0000 (23:06 +0100)
commit340736477651095a98a3b10e19b038ec62593a1d
tree61bf8d4d5e70793f76848de87e4a7aeef47d64b5tree
parentd5a409e57fe8bd24fef597ab8a31110d390a6392commit | diff
Q6_K AVX improvements (#10118)

* q6_k instruction reordering attempt

* better subtract method

* should be theoretically faster

small improvement with shuffle lut, likely because all loads are already done at that stage

* optimize bit fiddling

* handle -32 offset separately. bsums exists for a reason!

* use shift

* Update ggml-quants.c

* have to update ci macos version to 13 as 12 doesnt work now. 13 is still x86
.github/workflows/build.yml diff | blob | history
ggml/src/ggml-quants.c diff | blob | history
Packaging of ggml-org/llama.cpp