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

- Refactor MTMD chat template handling to resolve and analyze the chat template only
once per handler instance instead of on every request.

- Add template initialization state, cache parsed media placeholder tags, and support
explicit chat template overrides through a dedicated field. Improve lifecycle cleanup
by resetting cached template state and MTMD resources during handler close.

- This keeps runtime processing focused on message rendering and media tokenization
while avoiding repeated chat template resolution overhead.

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

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

Rename the 'llama_cpp.llama_cpp' import alias to 'llama_cpp_lib' to avoid potential namespace conflicts with the local '.llama_cpp' imports. Update all affected call sites in llama_embedding.py.

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

- Allow ctypes_function_for_shared_library to accept either a single
symbol name or an ordered iterable of ABI-compatible aliases.

- Resolve aliases in order and bind the first exported symbol found while
preserving the selected symbol name for runtime diagnostics. Also improve
error reporting for empty alias lists and missing symbols.

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

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

- Ensure MTMDChatHandler initialization remains compatible with specialized chat
handlers that define their own chat_format before calling super().__init__().

- Initialize chat_format only when it is not already provided by the subclass,
then apply chat_format_override or fallback to the built-in MTMD template.
This prevents AttributeError during inherited handler initialization while
keeping template override behavior unchanged.

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

- Allow ctypes bindings to mark symbols as optional through the `required`
flag.

- Missing symbols caused by ABI naming differences, API changes, or experimental
extensions will no longer break library loading. Optional APIs emit diagnostic
warnings and provide runtime unavailable stubs instead.

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

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

…nctions

- Add explicit validation for ctypes function argument declarations before
assigning them to the loaded shared library function.

- This provides clearer error messages when invalid Python types are passed
to `argtypes`, instead of exposing the internal ctypes error about missing
`from_param()` methods.

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

- Add Python ctypes bindings for the experimental APIs exposed by
llama-ext.h, including NextN/MTP embeddings, and model metadata extraction.

- Extension symbols are loaded optionally to handle ABI changes, renamed
symbols, and builds that do not export experimental APIs without breaking
the main Python bindings.

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

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

…template

- fix: chat template — null handling, reasoning preservation, turn-tag balance, input validation
- https://huggingface.co/google/gemma-4-31B-it/commit/68abe48010cbe15293462fa11e901a60639a44e5

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

- This release focuses on improving compatibility with the rapidly evolving llama.cpp ecosystem, especially around experimental APIs, ABI changes, MTMD processing, and latest model integrations.

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

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

- Preload the packaged libomp140.x86_64.dll on Windows before initializing
ggml-base to ensure CPU backend DLLs can resolve their OpenMP runtime
dependency.

- This only applies to Windows builds with llama-cpp-python >= 0.3.39 and
uses the bundled runtime from the package lib directory, avoiding the need
for users to configure system PATH or install additional OpenMP runtimes.

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

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

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

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

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