]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: d01b3c4)
convert : use f32 outtype for bf16 tensors (#6106)
authorRomain D <redacted>
Mon, 18 Mar 2024 08:04:41 +0000 (09:04 +0100)
committerGitHub <redacted>
Mon, 18 Mar 2024 08:04:41 +0000 (10:04 +0200)
commit3a6efdd03c46c5ba08e43880d34260c02dd9999b
treeb7de4172662ad4b5fb78e2578230fbb40adefd3btree
parentd01b3c4c32357567f3531d4e6ceffc5d23e87583commit | diff
convert : use f32 outtype for bf16 tensors (#6106)

The old behaviour is to use f16, but bf16 to f16 is not a lossless conversion.
Change the outtype to f32 to default to a lossless conversion.
convert.py diff | blob | history
Packaging of ggml-org/llama.cpp