🛠️[v0.3.44] Release Note: Improved Windows DLL(OpenMP) Loading Reliability for GGML Backends

[JamePeng]

Release Note: 0.3.44

Improved Windows DLL(OpenMP) Loading Reliability for GGML Backends

This release is a small but important maintenance update focused on improving Windows dynamic library loading reliability.

The change comes from feedback and debugging reports from the ComfyUI community. During integration with ComfyUI and some third-party plugins, I found that external applications may modify the process DLL search behavior before llama-cpp-python is initialized.

Although llama-cpp-python already prioritizes the packaged library directory:

site-packages\llama_cpp\lib

in its dynamic loading logic, the Windows DLL loader behavior can still be affected by the surrounding application environment. In some cases, paths such as:

C:\Windows\System32

may be searched earlier than expected.

If an incompatible libomp140.x86_64.dll already exists in the system directory (for example, a version built against a different Visual Studio runtime such as a non-VC143 build), Windows may load that DLL instead of the bundled runtime shipped with llama-cpp-python.

This can cause downstream dependency failures when loading CPU backend libraries:

ggml-cpu-*.dll
        |
        v
libomp140.x86_64.dll
        |
        X incompatible runtime

and eventually result in ggml loading failures.

---

Fix: Preload Bundled OpenMP Runtime on Windows

To make the loading behavior more deterministic, I added an explicit preload step for:

llama_cpp/lib/libomp140.x86_64.dll

before loading:

ggml-base.dll

The bundled OpenMP runtime is now loaded first on Windows for:

• llama-cpp-python >= 0.3.39
• packaged builds containing the bundled runtime DLL

This ensures that CPU backend libraries resolve against the runtime shipped with the package instead of an incompatible system-wide DLL.

Users no longer need to:

• modify the global PATH environment variable
• manually install a matching OpenMP runtime
• remove conflicting system DLLs

---

llama.cpp Synchronization

Updated llama.cpp to:

• ggml-org/llama.cpp
• commit: 846e991ec3c7ccec49112ff2c5b00b710e5f551d

---

Thanks to the ComfyUI community for the detailed feedback and testing that helped identify this Windows DLL loading edge case.

I will continue improving cross-platform compatibility and making the dynamic backend loading process more predictable across different application environments.

JamePeng