]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: d86c7d6)
ggml-hip : remove -funsafe-math-optimizations (#26696)
authorJim Wu <redacted>
Thu, 13 Aug 2026 06:38:02 +0000 (23:38 -0700)
committerGitHub <redacted>
Thu, 13 Aug 2026 06:38:02 +0000 (08:38 +0200)
commite79e4bf660e19f2ad851e06c6913f7a8c5852621
tree364577d98371c27103a0ebe13b76ce4515c0606atree
parentd86c7d62dfa7fb7227f0b699e3af58c11b295549commit | diff
ggml-hip : remove -funsafe-math-optimizations (#26696)

It enables -fassociative-math, which reassociates FP reductions and can flip
greedy argmax on RDNA3.5 (e.g. MTP speculative decode diverging from the
non-speculative baseline). Drop it so HIP builds are IEEE-conformant.

Co-authored-by: Jim Wu <redacted>
ggml/src/ggml-hip/CMakeLists.txt diff | blob | history
Packaging of ggml-org/llama.cpp