diff -rub ollama-0.1.27/llm/generate/gen_linux.sh ollama-0.1.27-patched/llm/generate/gen_linux.sh
--- ollama-0.1.27/llm/generate/gen_linux.sh	2024-02-22 23:41:43.000000000 +0100
+++ ollama-0.1.27-patched/llm/generate/gen_linux.sh	2024-02-25 03:16:43.566940450 +0100
@@ -48,7 +48,7 @@
         export CUDACXX=$(command -v nvcc)
     fi
 fi
-COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
+COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_LTO=on -DCMAKE_BUILD_TYPE=Release -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
 source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
@@ -59,7 +59,7 @@
     # llama.cpp, and we'll build only 1 CPU variant in that case as the default.
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_LTO=on -DCMAKE_BUILD_TYPE=Release ${CMAKE_DEFS}"
         BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu"
         echo "Building custom CPU"
         build
@@ -75,7 +75,7 @@
         # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
         # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
 
-        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off"
+        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_LTO=on -DCMAKE_BUILD_TYPE=Release -DLLAMA_NATIVE=off"
         if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
             #
             # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)