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....