Compile bug: broken build for CUDA NVIDIA card

[mx237]

Git commit

bef3727

Operating systems

Linux

GGML backends

CUDA

Problem description & steps to reproduce

Got build error on CUDA (for NVIDIA card), likely from commit bef3727. After I revert this commit build has no issue. Seems flash attention 512 is not supported if V_DOT2_F32_F16_AVAILABLE is undefined.

Build environment:
Ubuntu 26.04.1 LTS x86_64
Card: 4090 (CMAKE_CUDA_ARCHITECTURES=89-real detected by cmake)
CMake 4.2.3
make 4.4.1
gcc 15.2.0
nvcc 13.1
CUDA 13.3

First Bad Commit

bef3727

Compile command

cmake -B build -DBUILD_SHARED_LIBS=OFF -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=ON -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BUILD_APP=OFF -DLLAMA_BUILD_UI=OFF -DLLAMA_USE_PREBUILT_UI=OFF -DGGML_CUDA=ON -DGGML_CCACHE=OFF && cmake --build build --target llama-server

Relevant log output

[ 37%] Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-instance-q8_0-turbo3_0.cu.o
ptxas error   : Entry function '_Z18flash_attn_ext_vecILi512ELi1EL9ggml_type8ELS0_44ELb1EEvPKcS2_S2_S2_S2_PKiPfP6float2ffffjfi5uint3iiiiiiiiiiiliiliiiiil' uses too much shared data (0x10100 bytes, 0xc000 max)
ptxas error   : Entry function '_Z18flash_attn_ext_vecILi512ELi1EL9ggml_type8ELS0_44ELb0EEvPKcS2_S2_S2_S2_PKiPfP6float2ffffjfi5uint3iiiiiiiiiiiliiliiiiil' uses too much shared data (0x10100 bytes, 0xc000 max)
gmake[3]: *** [ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make:2645: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-instance-q8_0-turbo3_0.cu.o] Error 255
gmake[2]: *** [CMakeFiles/Makefile2:833: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/all] Error 2
gmake[1]: *** [CMakeFiles/Makefile2:1758: tools/server/CMakeFiles/llama-server.dir/rule] Error 2
gmake: *** [Makefile:520: llama-server] Error 2

[Tbotsche]

Thank You.
git log --oneline
git reset --hard 5ed81c8
Saves my day ;-)