Compile bug: gcc-11: error: unrecognized command-line option '-compress-mode=size'

[googx]

Git commit

2c9f833

Operating systems

Linux

GGML backends

CUDA

Problem description & steps to reproduce

NVIDIA GeForce RTX 4090
Driver Version: 570.124.04

lsb_release -a

No LSB modules are available.
Distributor ID: Ubuntu
Description: Ubuntu 22.04.4 LTS
Release: 22.04
Codename: jammy

First Bad Commit

No response

Compile command

cmake -B build -DGGML_CUDA=ON 
cmake --build build --config Release

Relevant log output

root # >     cmake -B build -DGGML_CUDA=ON                                                        141 ⨯
-- Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF
    -- CMAKE_SYSTEM_PROCESSOR: x86_64
-- Including CPU backend
-- x86 detected
-- Adding CPU backend variant ggml-cpu: -march=native
-- CUDA Toolkit found
-- Using CUDA architectures: 50;61;70;75;80
-- CUDA host compiler is GNU 11.4.0

-- Including CUDA backend
-- Configuring done
-- Generating done

root # >   cmake --build build --config Release
Consolidate compiler generated dependencies of target ggml-base
    [  2%] Built target ggml-base
    [  2%] Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o
nvcc warning : The 'compute_35', 'compute_37', 'compute_50', 'sm_35', 'sm_37' and 'sm_50' architectures are deprecated, and may be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
gcc-11: error: unrecognized command-line option '-compress-mode=size'
gmake[2]: *** [ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make:76: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o] Error 1
gmake[1]: *** [CMakeFiles/Makefile2:1788: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/all] Error 2
gmake: *** [Makefile:146: all] Error 2

[googx]

[googx]

have any one can help me ?? !!

[googx]

I also have another machine running Ubuntu 22.04 with a 3090 GPU. The NVIDIA-SMI version is 550.78, the Driver Version is 550.78, and the CUDA Version is 12.4. This setup was installed earlier and does not have this issue. I'm not sure if there's any tool that can run a one-click environment check script to output information that might be helpful to you. I've looked up this error: the -compress-mode=size parameter should not appear in the GCC arguments. It should be a parameter for nvcc. I'm not sure why this incorrect option is showing up in GCC. I simply cloned the code and followed the instructions in the CUDA section of docs/build.md. No matter how many times I try, I keep encountering this error. I can ensure that the build directory is deleted before each build. Can anyone help us? Thank you very much.

[MilanBojic1999]

Just happend to me, i changed in llama.cpp/Makefile, on line 174; from cc to gcc.
And it worked. I have a6000 with 12.4 cuda version on Ubuntu 22.04

[Green-Sky]

I've looked up this error: the -compress-mode=size parameter should not appear in the GCC arguments. It should be a parameter for nvcc. I'm not sure why this incorrect option is showing up in GCC.

Pretty much, not sure why this gets passed down to gcc.

[Green-Sky]

CUDA Version is 12.4

This is very funny, since there is a version check for >= 12.8.

llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt

Lines 105 to 112 in ba76543

	if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
	# Options are:
	# - none (not recommended)
	# - speed (nvcc's default)
	# - balance
	# - size
	list(APPEND CUDA_FLAGS -compress-mode=${GGML_CUDA_COMPRESSION_MODE})
	endif()

@googx According to your screenshots you have the 12.8.93 toolkit installed, but the 11.5.119 compiler.

[googx]

I'm not quite clear about the following log output. What exactly does this version number refer to? Is it the CUDA toolkit version? The nvcc version? Or GCC? . How should I investigate this issue and get it to work properly?

-- CUDA host compiler is GNU 11.4.0

and 11.5.119 , what is it ?

[googx]

I think I know the reason now. Although I've installed CUDA Toolkit version 12.8, there are still other versions of dynamic libraries on my machine. It's likely that these old dynamic libraries are still being loaded, causing the compiler to use an older CUDA version and attempt to compile code with new features from CUDA 12.8. Haha, another piece of useless knowledge added to my collection :)

[googx]

update !! new problem !

After uninstalling the old version of CUDA, I successfully made progress in the compilation. However, a new error occurred. This time, I confirmed that the ptxas library is correctly using the CUDA 12.8 version.

[Green-Sky]

The warning support for offline compilation for arches prior to *_75 will be removed... is expected (they start popping up with 12.8)

I see a ^C before the nvcc error..., which means you hit ctrl+c and canceled the process.

[mohammad-matini]

I did a git bisect, and found that the first commit where the error:

gcc-12: error: unrecognized command-line option ‘-compress-mode=size’

starts appearing, is commit: 80c41dd, so commits before it build okay.

Additionally, don't make my mistake; having installed CUDA from NVIDIA's repos, I did not set up the environment correctly, which means I wasn't using nvcc from the latest CUDA package, but from the nvidia-cuda-toolkit Ubuntu package, which has an older nvcc.

[Green-Sky]

Can someone test replacing CUDAToolkit_VERSION here
https://github.com/ggml-org/llama.cpp/blob/84d547554123a62e9ac77107cb20e4f6cc503af4/ggml/src/ggml-cuda/CMakeLists.txt#L105C9-L105C28
with CMAKE_CUDA_COMPILER_VERSION ?

And add a message(STATUS "cuda compiler version: ${CMAKE_CUDA_COMPILER_VERSION}") right before the if.

[Green-Sky]

Did someone try it?