os.add_dll_directory() raises FileNotFoundError [WinError 3] when the
directory does not exist, so a stale HIP_PATH or VULKAN_SDK left behind by
an uninstalled SDK makes "import llama_cpp" fail outright on Windows.

The CUDA_PATH branch above already guards each candidate directory with
os.path.exists(); this applies the same pattern to the HIP and Vulkan
branches. Valid directories are still added individually, so a partially
removed SDK contributes whichever of bin/lib remain instead of raising.

…-missing-path

fix: guard HIP_PATH and VULKAN_SDK dirs with os.path.exists before add_dll_directory

Signed-off-by: JamePeng <jame_peng@sina.com>

… APIs

- Add missing `_Z` Itanium C++ ABI symbol variants to ctypes function
lookup lists. This improves compatibility with Linux and macOS builds
where C++ symbols are exported using GCC/Clang name mangling.

Signed-off-by: JamePeng <jame_peng@sina.com>

Inspect PE, ELF, and Mach-O exports and normalize platform-specific symbol names. Validate optional llama_ext ctypes aliases across Windows, Linux, and macOS builds. Keep artifacts and timestamped privacy-safe reports local to the repository.

Signed-off-by: JamePeng <jame_peng@sina.com>

Signed-off-by: JamePeng <jame_peng@sina.com>

- Update model loading configuration to use the new `load_mode` field from
llama_model_params and align with the latest llama.cpp API changes.

- Remove deprecated internal handling of legacy loading flags and keep
backward compatibility by warning users when `use_mmap`, `use_direct_io`,
or `use_mlock` are still used.

- This prepares the Python bindings for the updated llama.cpp model loading
interface while providing a smoother migration path for existing users.

Signed-off-by: JamePeng <jame_peng@sina.com>

- Replace references to the legacy model loading flags with load_mode, document all supported loading modes for the Python API and server, and update the performance tuning example.

Signed-off-by: JamePeng <jame_peng@sina.com>

Signed-off-by: JamePeng <jame_peng@sina.com>

Signed-off-by: JamePeng <jame_peng@sina.com>

- add `no_alloc` and `load_mtp` parameters
- enable `extra buffer types` by default

Signed-off-by: JamePeng <jame_peng@sina.com>

- use uint32_t for layer IDs
- fix void return type for embedding extraction control
- return target layer count as uint32_t

Signed-off-by: JamePeng <jame_peng@sina.com>

- This release focuses on reactivating and modernizing Llama’s built-in embedding capabilities, aligning the Python bindings with the latest llama.cpp APIs, and improving reliability across platforms.

Signed-off-by: JamePeng <jame_peng@sina.com>

Signed-off-by: JamePeng <jame_peng@sina.com>

Signed-off-by: JamePeng <jame_peng@sina.com>

- Allow duplicate OpenMP runtimes in complex environments such as ComfyUI
- Stop searching the deprecated /bin directory for ggml dynamic libraries

Signed-off-by: JamePeng <jame_peng@sina.com>

- Add accessors for NextN and layer input embeddings
- Support selecting the NextN layer offset
- Expose the auxiliary context handle
- Validate layer IDs, offsets, and unavailable outputs

Signed-off-by: JamePeng <jame_peng@sina.com>

Signed-off-by: JamePeng <jame_peng@sina.com>