From 3b2b27ddc7cf47fcb7b92189b757b61208a4bd88 Mon Sep 17 00:00:00 2001 From: peterjunpark Date: Thu, 12 Feb 2026 14:34:54 -0500 Subject: [PATCH] [docs/7.11.0] Add llama.cpp known issue (#5962) --- RELEASE.md | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/RELEASE.md b/RELEASE.md index 742094a18..fa2b9a559 100644 --- a/RELEASE.md +++ b/RELEASE.md @@ -1729,6 +1729,19 @@ known to be affected. As a workaround, use `-Og` optimization level instead of llama.cpp builds successfully but might fail at runtime with the error "HIP kernel mul_mat_q has no device code". This issue is under investigation. +### llama.cpp prompt processing performance regression + +llama.cpp experiences reduced prompt processing performance across multiple AMD +GPU architectures. As a workaround, pass the compiler flag `-mllvm +--amdgpu-unroll-threshold-local=600` to `hipcc` or `amdclang` when compiling +llama.cpp: + +```bash +hipcc -mllvm --amdgpu-unroll-threshold-local=600 ... +``` + +This issue will be fixed in a future release. + ### PyTorch model training validation issues The following models failed validation on PyTorch for ROCm 7.11.0 due to