Compile bug: I'm trying to compile llama.cpp with SYCL support on Arch Linux with the glibc 2.41+r2+g0a7c7a3e283a-1

[dpriedel]

Git commit

HEAD as of the time I post this

9ab42dc

Operating systems

Linux

GGML backends

SYCL

Problem description & steps to reproduce

I'm trying to compile llama.cpp with SYCL support on Arch Linux with the glibc 2.41+r2+g0a7c7a3e283a-1 and gcc 14.2.1+r730+gc061ad5a36ba-1 packages. Prior to installing these updates I had no problems with the compile.

I am also using the Intel oneAPI Base Toolkit version 2025.0.1

First Bad Commit

N/A

Compile command

I do `rm -rf build` and then

cmake -B build -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Release

followed by:

cmake --build build --config Release -j 10

Relevant log output

There are multiple instances of the following type of output

In file included from /vol_KUtil2/LLM_Models/llama.cpp/ggml/src/ggml-sycl/conv.cpp:13:
In file included from /vol_KUtil2/LLM_Models/llama.cpp/ggml/src/ggml-sycl/conv.hpp:16:
In file included from /vol_KUtil2/LLM_Models/llama.cpp/ggml/src/ggml-sycl/common.hpp:16:
In file included from /usr/lib64/gcc/x86_64-pc-linux-gnu/14.2.1/../../../../include/c++/14.2.1/fstream:40:
In file included from /usr/lib64/gcc/x86_64-pc-linux-gnu/14.2.1/../../../../include/c++/14.2.1/istream:40:
In file included from /usr/lib64/gcc/x86_64-pc-linux-gnu/14.2.1/../../../../include/c++/14.2.1/ios:40:
In file included from /usr/lib64/gcc/x86_64-pc-linux-gnu/14.2.1/../../../../include/c++/14.2.1/iosfwd:42:
In file included from /usr/lib64/gcc/x86_64-pc-linux-gnu/14.2.1/../../../../include/c++/14.2.1/bits/postypes.h:40:
In file included from /usr/lib64/gcc/x86_64-pc-linux-gnu/14.2.1/../../../../include/c++/14.2.1/cwchar:44:
In file included from /usr/include/wchar.h:30:
/usr/include/bits/floatn.h:79:52: error: unsupported machine mode '__TC__'
   79 | typedef _Complex float __cfloat128 __attribute__ ((__mode__ (__TC__)));
      |                                                    ^

[Rbiessy]

Hello @dpriedel I'm not sure the issue is coming from llama.cpp. The error seems to come from including <fstream> with your particular environment. Have you maybe tried compiling a simpler cpp application that would also include <fstream> with the same compiler and environment and similar compilation flags used to compile ggml/src/ggml-sycl/conv.cpp?

[dpriedel]

I removed the include for fstream since it's not needed but the same error recurs from the iostream header which triggers the same sequence of includes, erroring out on the complex-float typedef.

The code in the floatn.h header indicates it is for 'older' compilers. I wonder if this is then an Intel oneAPI compiler problem??

Thanks.

PS. If I build for the Vulkan back-end, llama.cpp compiles.

[qnixsynapse]

This is the SYCL compiler's fault and not llama.cpp. I too encountered that while working in this PR (comment)

I think they ship a dated version of compiler in intel-oneapi-base-toolkit. Please report on their repository. (I think it's intel/llvm)

For now, downgrade glibc to make the SYCL work again.

Also, there is a confirmed bug in the intel compute runtime.

[dpriedel]

I'm trying to make a small test case which reproduces the problem....

As it turns out, reverting the glibc 2.41 update is not an option for me.

Also, the problem is not with the fstream header since the oneapi c++ compiler will compile a simple program which writes to a file using an fstream. The problem is related to compiling SYCL and glibc 2.41 with the current version of the icpx compiler.

A way around this on a system with glibc 2.41 is to use the oneapi docker container. Compile llama.cpp in the container (sharing the locally installed code) and then run on the local machine.

02/20/2024 Update:

Arch Linux posted a patched version of glibc 2.41 which fixes the problem compiling SYCL with the ipcx compiler. They also posted the current version of the Intel oneAPI toolkts.