Compile bug: Entry function flash_attn_tile (mangled) uses too much shared data (0xd100 bytes, 0xc000 max)

[simlay]

Git commit

7b8443a and fc2b005

Operating systems

Linux

GGML backends

CUDA

Problem description & steps to reproduce

I'm trying to build on archlinux with cuda 12.9 (for an nvidia 1080Ti). Possibly I'm not using the right C++ compiler but this has worked up until #22286.

Admittedly I've been getting a lot of warnings saying: nvcc warning : Support for offline compilation for architectures prior to '<compute/sm/lto>_75' will be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).

Apologies if this is redundant or duplicated (I looked but found nothing related).

First Bad Commit

7b8443a

Compile command

cmake -B build -DGGML_CCACHE=OFF -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="61" -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j8

Relevant log output

[  8%] Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o
nvcc warning : Support for offline compilation for architectures prior to '<compute/sm/lto>_75' will be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
ptxas error   : Entry function '_Z15flash_attn_tileILi320ELi256ELi1ELi32ELb1EEvPKcS1_S1_S1_S1_PKiPfP6float2ffffjfi5uint3iiiiiiiiiiiliiliiiiil' uses too much shared data (0xd100 bytes, 0xc000 max)
ptxas error   : Entry function '_Z15flash_attn_tileILi320ELi256ELi1ELi32ELb0EEvPKcS1_S1_S1_S1_PKiPfP6float2ffffjfi5uint3iiiiiiiiiiiliiliiiiil' uses too much shared data (0xd100 bytes, 0xc000 max)
make[2]: *** [ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make:1040: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-tile-instance-dkq320-dv256.cu.o] Error 255
make[2]: *** Waiting for unfinished jobs....
make[1]: *** [CMakeFiles/Makefile2:2412: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/all] Error 2
make: *** [Makefile:146: all] Error 2

[wbulot]

I have a script that pulls the latest version, compiles it, and starts it.
I'm getting the same error too, starting today because of one of the latest commits.

[voidcentury]

Facing the same issue since today (GTX 1050 Ti). Perhaps only affects older gpus?

[jm-pro]

I also have a script that pulls latest version and builds it.
Same error.
NVIDIA-SMI 580.142 Driver Version: 580.142 CUDA Version: 13.0
Quadro P2000

[jason6822]

tesla p4 same Compile command same error

[Yeetster-cmd]

Adding more to this as well, compute61 just seems to not be working. Might be intentional?

[am17an]

Probably #22286, cc @JohannesGaessler @lnigam

[JohannesGaessler]

Should be fixed by #22541 .