[Important for Distribution] Distribute this on nixpkgs since it is essential

[izelnakri]

Summary

Hi there,

There is an absolute requirement to have this library distributed on nixpkgs. I'll try to assist best of my abilities but maintainers of this project should maintain this library on nixpkgs as well. I might create a PR to this project soon to make it possible for now(before nixpkgs) this:

nix build github:uxlfoundation/oneMath

Problem statement

I managed to run whisper.cpp project on my NixOS system, Lunar Lake laptop can utilize the GPU by installing a custom proprietary OneAPI Toolkit distribution of Intel. However all this process has many setup, installion and build pitfalls and I suspect even 2025.2 distribution is Intels OneMKL, instead of oneMath.

I suspect ggml-sycl project also currently set up to target vendor specific oneMKL currently for this reason:
https://github.com/ggml-org/whisper.cpp/blob/master/ggml/src/ggml-sycl/CMakeLists.txt#L112

Current mkl package on nixpkgs is quite old and lacks plenty of features:
https://search.nixos.org/packages?channel=25.05&from=0&size=50&sort=relevance&type=packages&query=mkl

Preferred solution

nix profile install nixpkgs#onemath

# or declared in flake.nix and $ nix develop picks it up.

When we have this package (along with other missing uclfoundation packages on nixpkgs. One of the pain points of edge computing, LLM inferencing now will be cleared up and made much easier to utilize the latest good tools and features. At very least, this will allow for much more minimal and straightforward inference docker containers. Nix has awesome tool to help with this, if we can package this for nixpkgs.

I'll try to help to best of my abilities but as I mentioned, some maintainers hopefully will take on the initiative(s) for all uxlfoundation projects, not only this one ;)

[Rbiessy]

Hello @izelnakri, having worked both on oneMath and ggml-sycl (as part of llama.cpp) I can answer to some of your statements.

I managed to run whisper.cpp project on my NixOS system, Lunar Lake laptop can utilize the GPU by installing a custom proprietary OneAPI Toolkit distribution of Intel. However all this process has many setup, installion and build pitfalls and I suspect even 2025.2 distribution is Intels OneMKL, instead of oneMath.

1. Not sure if that would have helped you but there is some documentation in llama.cpp to build that project with SYCL: https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/SYCL.md
2. ggml-sycl will try to use oneDNN by default if found on the system to dispatch gemm operations. oneMKL (for Intel devices) or oneMath (for Nvidia or AMD) is still a dependency for historical reasons but it could be removed in favor of oneDNN only. This is not a priority anymore for us but a contribution would be welcomed. It would be a cleaner solution to your problem in my opinion. This would simplify ggml-sycl and reduce the number of dependencies.
3. In case you have missed it, oneMath is setup to be automatically downloaded and built as part of ggml-sycl if needed, see https://github.com/ggml-org/llama.cpp/blob/1c872f71fb8a25589efa3ee9b6bf8b517cb8caa4/ggml/src/ggml-sycl/CMakeLists.txt#L142. I would think this makes the setup easier but I understand that this is not be desirable in all scenarios.
4. Just to confirm oneAPI packages indeed ship oneMKL for Intel devices only and not oneMath.

I suspect ggml-sycl project also currently set up to target vendor specific oneMKL currently for this reason:
https://github.com/ggml-org/whisper.cpp/blob/master/ggml/src/ggml-sycl/CMakeLists.txt#L112

ggml-sycl still uses oneMKL for Intel devices due to some limitations of oneMKL and oneMath on Windows, see #654.

On a final note I am rather in favor of making oneMath easier to use but I won't be able to help much more with that. Note there is a somewhat related issue to provide pre-built packages of oneMath: #467.

[izelnakri]

Useful points @Rbiessy , thank you for your response. Given the amount of contributions you've made here, I suspect you are one of the maintainers?

1. I've been able to already set up with the custom/proprietary OneAPI Toolkit.
2. ggml-sycl will try to use oneDNN by default good to know! I see that nixpkgs oneDNN is v3.7.3 but more oddly it points to another repo than the uxlfoundation repo, should we revert it to the uxlfoundation repo?
3. I suspect this line, prevents fetching on my Lunar Lake laptop: https://github.com/ggml-org/llama.cpp/blob/1c872f71fb8a25589efa3ee9b6bf8b517cb8caa4/ggml/src/ggml-sycl/CMakeLists.txt#L111
4. 👍

On a final note I am rather in favor of making oneMath easier to use

a) Thank you! I've always found it nice that all Intel AI stack developers are willing to help, however all these packages need to be nix -ified. I suspect the reluctance comes from unfamiliarity with nix. Nix is amazing, solves many build and distribution problems and wrong assumptions inside builds, so all these projects should use it more in the CI and flake.nix in my view. Given one in practice can even package a nix package in any format(be it a docker container or .deb etc), in a very minimal and customizable way, I think flake.nix should be a part of every software project of oneAPI and its CI/CD eventually.

I've even manage to build my entire apps(and their environment/dependencies) as flatpak. So I really think nix should be as prevalent as git is today in software development for DX: https://github.com/izelnakri/project-templates/blob/main/c/Makefile#L23

b) Lastly, I suspect nixpkgs oneDNN misses .pc file so cmake inside whisper.cpp can't find it even if I have it in my dev environment 🤦 😅 Thats why would be great if main maintainers also maintain/review the nixpkgs packages, despite it being an open community effort always ;) :

19:51:00 izelnakri ~/Github/whisper.cpp ❄ -> master rm -rf build && cmake -B build -DWHISPER_SDL2=ON -DWHISPER_OPENVINO=1 -DGGML_SYCL=ON -DCMAKE_C_COMPILER=$CC -DCMAKE_CXX_COMPILER=$CXX

CMake Deprecation Warning at CMakeLists.txt:1 (cmake_minimum_required):
  Compatibility with CMake < 3.10 will be removed from a future version of
  CMake.

  Update the VERSION argument <min> value.  Or, use the <min>...<max> syntax
  to tell CMake that the project requires at least <min> but has been updated
  to work with policies introduced by <max> or earlier.


-- The C compiler identification is Clang 21.0.0
-- The CXX compiler identification is Clang 21.0.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /home/izelnakri/Github/intel-nix/result/bin/clang - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /home/izelnakri/Github/intel-nix/result/bin/clang++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found Git: /home/izelnakri/.nix-profile/bin/git (found version "2.49.0")
-- Setting GGML_NATIVE_DEFAULT to OFF
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
-- Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF
-- CMAKE_SYSTEM_PROCESSOR: x86_64
-- GGML_SYSTEM_ARCH: x86
-- Including CPU backend
-- Could NOT find OpenMP_C (missing: OpenMP_C_FLAGS OpenMP_C_LIB_NAMES)
-- Could NOT find OpenMP_CXX (missing: OpenMP_CXX_FLAGS OpenMP_CXX_LIB_NAMES)
-- Could NOT find OpenMP (missing: OpenMP_C_FOUND OpenMP_CXX_FOUND)
CMake Warning at ggml/src/ggml-cpu/CMakeLists.txt:77 (message):
  OpenMP not found
Call Stack (most recent call first):
  ggml/src/CMakeLists.txt:361 (ggml_add_cpu_backend_variant_impl)


-- x86 detected
-- Adding CPU backend variant ggml-cpu:
-- GGML_SYCL_TARGET=INTEL
-- Performing Test SUPPORTS_SYCL
-- Performing Test SUPPORTS_SYCL - Success
CMake Warning at ggml/src/ggml-sycl/CMakeLists.txt:12 (message):
  Using open-source SYCL compiler (clang++).  Didn't detect ENV
  {ONEAPI_ROOT}.

          If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
          source /opt/intel/oneapi/setvars.sh


-- SYCL found
CMake Warning at ggml/src/ggml-sycl/CMakeLists.txt:39 (find_package):
  By not providing "FindIntelSYCL.cmake" in CMAKE_MODULE_PATH this project
  has asked CMake to find a package configuration file provided by
  "IntelSYCL", but CMake did not find one.

  Could not find a package configuration file provided by "IntelSYCL" with
  any of the following names:

    IntelSYCLConfig.cmake
    intelsycl-config.cmake

  Add the installation prefix of "IntelSYCL" to CMAKE_PREFIX_PATH or set
  "IntelSYCL_DIR" to a directory containing one of the above files.  If
  "IntelSYCL" provides a separate development package or SDK, be sure it has
  been installed.


CMake Warning at ggml/src/ggml-sycl/CMakeLists.txt:54 (find_package):
  By not providing "FindDNNL.cmake" in CMAKE_MODULE_PATH this project has
  asked CMake to find a package configuration file provided by "DNNL", but
  CMake did not find one.

  Could not find a package configuration file provided by "DNNL" with any of
  the following names:

    DNNLConfig.cmake
    dnnl-config.cmake

  Add the installation prefix of "DNNL" to CMAKE_PREFIX_PATH or set
  "DNNL_DIR" to a directory containing one of the above files.  If "DNNL"
  provides a separate development package or SDK, be sure it has been
  installed.


-- oneDNN not found, disabling oneDNN support
CMake Error at ggml/src/ggml-sycl/CMakeLists.txt:117 (find_package):
  By not providing "FindMKL.cmake" in CMAKE_MODULE_PATH this project has
  asked CMake to find a package configuration file provided by "MKL", but
  CMake did not find one.

  Could not find a package configuration file provided by "MKL" with any of
  the following names:

    MKLConfig.cmake
    mkl-config.cmake

  Add the installation prefix of "MKL" to CMAKE_PREFIX_PATH or set "MKL_DIR"
  to a directory containing one of the above files.  If "MKL" provides a
  separate development package or SDK, be sure it has been installed.


-- Configuring incomplete, errors occurred!
19:51:00 izelnakri ~/Github/whisper.cpp ❄ -> master pkg-config --list-all | rg pango
19:51:00 izelnakri ~/Github/whisper.cpp ❄ -> master pkg-config --list-all
openvino     OpenVINO - OpenVINO™ Toolkit
libze_loader Level Zero Loader - Runtime Library Loader for Level Zero
level-zero   Level Zero - Level Zero
19:51:00 izelnakri ~/Github/whisper.cpp ❄ -> master
19:51:00 izelnakri ~/Github/whisper.cpp ❄ -> master echo $buildInputs
/nix/store/x4cz3spvw0bwwz5sjsdn2qm4f89rcryn-glibc-2.40-66-dev /nix/store/sa7j7cddyblhcb3ch3ds10w7nw75yjj1-gcc-14.3.0 /nix/store/99mambdfxnpm6rkdjbf1hhgdg87caxa3-zlib-1.3.1-dev /nix/store/f91yrpz9kjvsdv62yxkmi9apghx0vikk-level-zero-1.21.9 /nix/store/vifmg9yvzqjkhy5yq520qsw2swn5279b-sdl2-compat-2.32.56-dev /nix/store/pmqp1c9i7vffk1zvjsknlc336qy820jd-openvino-2025.2.0 /nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev /nix/store/swpsnjk7ff610dqd07l7ybkz02p39ksw-whisper-cpp-dev
19:51:00 izelnakri ~/Github/whisper.cpp ❄ -> master cd /nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev
19:51:00 izelnakri /nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev ❄ -> master ls
 include   lib   nix-support
19:51:00 izelnakri /nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev ❄ -> master cd ..
19:51:00 izelnakri /nix/store ❄ -> master cd /nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev
19:51:00 izelnakri /nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev ❄ -> master nvim .
19:51:00 izelnakri /nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev ❄ -> master cd ..
19:51:00 izelnakri /nix/store ❄ -> master tree z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev
z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev
├── include
│   ├── dnnl.h
│   ├── dnnl.hpp
│   ├── dnnl_config.h
│   ├── dnnl_debug.h
│   ├── dnnl_ocl.h
│   ├── dnnl_ocl.hpp
│   ├── dnnl_sycl.h
│   ├── dnnl_sycl.hpp
│   ├── dnnl_sycl_types.h
│   ├── dnnl_threadpool.h
│   ├── dnnl_threadpool.hpp
│   ├── dnnl_threadpool_iface.hpp
│   ├── dnnl_types.h
│   ├── dnnl_version.h
│   └── oneapi
│       └── dnnl
│           ├── dnnl.h
│           ├── dnnl.hpp
│           ├── dnnl_common.h
│           ├── dnnl_common.hpp
│           ├── dnnl_common_types.h
│           ├── dnnl_config.h
│           ├── dnnl_debug.h
│           ├── dnnl_graph.h
│           ├── dnnl_graph.hpp
│           ├── dnnl_graph_ocl.h
│           ├── dnnl_graph_ocl.hpp
│           ├── dnnl_graph_sycl.h
│           ├── dnnl_graph_sycl.hpp
│           ├── dnnl_graph_types.h
│           ├── dnnl_ocl.h
│           ├── dnnl_ocl.hpp
│           ├── dnnl_ocl_types.h
│           ├── dnnl_sycl.h
│           ├── dnnl_sycl.hpp
│           ├── dnnl_sycl_types.h
│           ├── dnnl_threadpool.h
│           ├── dnnl_threadpool.hpp
│           ├── dnnl_threadpool_iface.hpp
│           ├── dnnl_types.h
│           ├── dnnl_ukernel.h
│           ├── dnnl_ukernel.hpp
│           ├── dnnl_ukernel_types.h
│           ├── dnnl_version.h
│           └── dnnl_version_hash.h
├── lib
│   └── cmake
│       └── dnnl
│           ├── dnnl-config-version.cmake
│           ├── dnnl-config.cmake
│           ├── dnnl-targets-release.cmake
│           └── dnnl-targets.cmake
└── nix-support
    └── propagated-build-inputs

8 directories, 48 files

c) I've also just found that oneDNN package inside nixpkgs is compiled with wrong flags and possibly needs refresh and SYCL argument. Things like these is the reason we need core maintainers to have a look, update and ideally even review(put themselves as reviewers) to nixpkgs packages:

20:08:52 izelnakri ~/Github/whisper.cpp ❄ -> master cmake -B build \
  -DWHISPER_SDL2=ON \
  -DWHISPER_OPENVINO=1 \
  -DGGML_SYCL=ON \
  -DCMAKE_C_COMPILER=$CC \
  -DCMAKE_CXX_COMPILER=$CXX \
  -DDNNL_DIR="/nix/store/z70svjsr04zmy9gsvpsnkppsch57gmsc-oneDNN-3.8.1-dev/lib/cmake/dnnl"

CMake Deprecation Warning at CMakeLists.txt:1 (cmake_minimum_required):
  Compatibility with CMake < 3.10 will be removed from a future version of
  CMake.

  Update the VERSION argument <min> value.  Or, use the <min>...<max> syntax
  to tell CMake that the project requires at least <min> but has been updated
  to work with policies introduced by <max> or earlier.


-- Setting GGML_NATIVE_DEFAULT to OFF
-- Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF
-- CMAKE_SYSTEM_PROCESSOR: x86_64
-- GGML_SYSTEM_ARCH: x86
-- Including CPU backend
-- Could NOT find OpenMP_C (missing: OpenMP_C_FLAGS OpenMP_C_LIB_NAMES)
-- Could NOT find OpenMP_CXX (missing: OpenMP_CXX_FLAGS OpenMP_CXX_LIB_NAMES)
-- Could NOT find OpenMP (missing: OpenMP_C_FOUND OpenMP_CXX_FOUND)
CMake Warning at ggml/src/ggml-cpu/CMakeLists.txt:77 (message):
  OpenMP not found
Call Stack (most recent call first):
  ggml/src/CMakeLists.txt:361 (ggml_add_cpu_backend_variant_impl)


-- x86 detected
-- Adding CPU backend variant ggml-cpu:
-- GGML_SYCL_TARGET=INTEL
CMake Warning at ggml/src/ggml-sycl/CMakeLists.txt:12 (message):
  Using open-source SYCL compiler (clang++).  Didn't detect ENV
  {ONEAPI_ROOT}.

          If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
          source /opt/intel/oneapi/setvars.sh


-- SYCL found
CMake Warning at ggml/src/ggml-sycl/CMakeLists.txt:39 (find_package):
  By not providing "FindIntelSYCL.cmake" in CMAKE_MODULE_PATH this project
  has asked CMake to find a package configuration file provided by
  "IntelSYCL", but CMake did not find one.

  Could not find a package configuration file provided by "IntelSYCL" with
  any of the following names:

    IntelSYCLConfig.cmake
    intelsycl-config.cmake

  Add the installation prefix of "IntelSYCL" to CMAKE_PREFIX_PATH or set
  "IntelSYCL_DIR" to a directory containing one of the above files.  If
  "IntelSYCL" provides a separate development package or SDK, be sure it has
  been installed.


CMake Warning at ggml/src/ggml-sycl/CMakeLists.txt:74 (message):
  oneDNN must be compiled for the same target as llama.cpp.

                   llama.cpp: INTEL, oneDNN: NONE.
                   Disabling oneDNN support.

[Rbiessy]

Yes I am indeed one of the maintainers of oneMath and of the ggml-sycl backend of llama.cpp.

2. oneDNN was moved from oneapi-src organization to uxlfoundation but they are the same repository. GitHub will automatically redirect the old links to the new one but it wouldn't hurt to update the links to uxlfoundation.
3. The CMake line you linked controls whether ggml-sycl will link against oneMKL or oneMath. GGML_SYCL_TARGET is set to INTEL by default and we can see it in your CMake output so ggml-sycl tries to link to oneMKL as expected for your device. Your issue is that oneAPI doesn't seem to be installed with oneMKL or CMake is not able to find it. The SYCL backend expects that oneAPI is installed with oneMKL and the related environment variables are set, typically using /opt/intel/oneapi/2025.2.0/setvars.sh will add the paths to CMAKE_PREFIX_PATH. If you want to build all the library dependencies from source your only solution is to add the option to disable oneMKL and oneMath in ggml-sycl. Using oneMath will also need to dispatch to oneMKL so this won't help you.

I'm confident now that supporting nixpkgs won't help you solve your particular issue. That doesn't mean it can't be done of course. Pinging @uxlfoundation/onemath-arch-write for visibility so they can consider this suggestion.

[izelnakri]

The CMake line you linked controls whether ggml-sycl will link against oneMKL or oneMath. GGML_SYCL_TARGET is set to INTEL by default and we can see it in your CMake output so ggml-sycl tries to link to oneMKL as expected for your device. Your issue is that oneAPI doesn't seem to be installed with oneMKL or CMake is not able to find it. The SYCL backend expects that oneAPI is installed with oneMKL and the related environment variables are set, typically using /opt/intel/oneapi/2025.2.0/setvars.sh will add the paths to CMAKE_PREFIX_PATH. If you want to build all the library dependencies from source your only solution is to add the option to disable oneMKL and oneMath in ggml-sycl.

This is what we discovered as well, hopefully someone(maybe us) will PR to whisper.cpp and adjust the cmake build to fix this.

Using oneMath will also need to dispatch to oneMKL so this won't help you.

Could you elaborate? @Rbiessy Do I need to compile oneMKL in addition to oneMath as 2 separate repositories to make it work for whisper.cpp (I hope not). Isn't oneMath includes perhaps the oneMKL portion inside the repository?

This is the issue: Intel currently has many build assumptions in its AI software stack and nixpkgs can help you be more explicit and better to distribute these packages in repositories and system configurations. So I think nixpkgs support is absolutely essential. This will allow you to package better/optimized builds with less assumptions and size for all targets as well. You can create a minimal docker containers, .deb packages etc all sourced from nix and can eliminate wrong/outdated assumptions on the custom cmake build configurations. In other words, nix is just a far superior build system for C and C++ and reduces the need for custom build assumptions & configurations.

More developers are familiar with nix than cmake and its far superior/modern approach(nix internally uses cmake as well but better defaults obviously since everything is reproducable). So I think intel should make this the top priority if they want wide adoption compared to competition with CUDA, RoCM, Mojo/Max etc. I appreciate cross-industry initiative uxlfoundation but reallly useful cross-vendor solution would be a nix based build system.

[Rbiessy]

Do I need to compile oneMKL in addition to oneMath as 2 separate repositories to make it work for whisper.cpp (I hope not).

No my point is that oneMath is not used in ggml-sycl for Intel devices. The errors in your CMake output show that your issue is with CMake failing to find oneMKL. You need to make sure you install oneAPI with oneMKL and set the environment variables I described.

Isn't oneMath includes perhaps the oneMKL portion inside the repository?

I don't get what you mean here. oneMath dispatches to other math libraries and relies on oneMKL for Intel devices.

[vmalia]

@izelnakri Please check the first FAQ for the difference between oneMath and oneMKL: https://github.com/uxlfoundation/oneMath?tab=readme-ov-file#onemath