Compile bug: CUDA build uses too much shared data

[steampunque]

Git commit

b6498

Operating systems

Linux

GGML backends

CUDA

Problem description & steps to reproduce

export CUDA_ARCH="60;61;70;75"

build

First Bad Commit

b6498

Compile command

cmake
make

Relevant log output

[ 27%] Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o
ptxas error   : Entry function '_Z15flash_attn_tileILi128ELi32ELb1EEvPKcS1_S1_S1_S1_PKiPfP6float2ffffjfiiiiiiiiiiiiiliiliiiiil' uses too much shared data (0xc400 bytes, 0xc000 max)
ptxas error   : Entry function '_Z15flash_attn_tileILi128ELi32ELb0EEvPKcS1_S1_S1_S1_PKiPfP6float2ffffjfiiiiiiiiiiiiiliiliiiiil' uses too much shared data (0xc400 bytes, 0xc000 max)
make[3]: *** [ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make:314: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile.cu.o] Error 255
make[3]: *** Waiting for unfinished jobs....
make[3]: Leaving directory '/dev/shm/tmp/build'
make[2]: *** [CMakeFiles/Makefile2:1024: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/all] Error 2
make[2]: Leaving directory '/dev/shm/tmp/build'
make[1]: *** [Makefile:136: all] Error 2

[JohannesGaessler]

Please confirm whether #16091 fixes the issue.

[steampunque]

Please confirm whether #16091 fixes the issue.

Confirmed. Thank for fast response!