🛠️[v0.3.46] Release Note: Extended Model APIs, MTMD Binding Updates, and Improved Runtime Compatibility

[JamePeng]

0.3.46

Extended Model APIs, MTMD Binding Updates, and Improved Runtime Compatibility

Version 0.3.46 is a small but important iteration focused on keeping the Python bindings aligned with the latest llama.cpp development, improving model introspection capabilities, and preparing the foundation for upcoming multimodal and generation features.

This release mainly focuses on API synchronization and binding improvements. Some of the newly exposed MTMD interfaces are experimental API adaptations at this stage; the corresponding higher-level features have not yet been integrated.

Expanded Model Inspection APIs

I added new model-level APIs to expose more information from the native llama.cpp model:

• target_layer_ids()

  • Retrieve target model layer indices from the loaded model.

• get_tok_embd()

  • Extract the token embedding matrix as a NumPy array.

These APIs include additional validation for native return values, covering:

• null pointer handling
• unexpected embedding sizes
• incomplete data copies

This helps provide clearer runtime errors when native model data is unavailable or inconsistent.

NextN and Layer Embedding API Support

The internal context layer now exposes additional NextN embedding capabilities:

• access NextN and layer input embeddings
• configure NextN layer offsets
• access auxiliary context handles

The Python layer also adds validation for:

• invalid layer IDs
• invalid offsets
• unavailable embedding outputs

These APIs provide the foundation for future model architectures and advanced inference workflows.

MTMD API Binding Synchronization

This release updates the MTMD bindings with the latest experimental APIs from llama.cpp.

The changes include:

• input chunk serialization and restoration APIs
• experimental generated-audio API structures and processing interfaces
• HunyuanVL decoder position support
• synchronized enum values and corrected ctypes signatures

The MTMD audio generation interfaces are currently exposed at the binding level only. The higher-level Python workflow and user-facing features will be introduced in future updates as the upstream API becomes more stable.

Improved Windows Runtime Compatibility

Improved Windows dynamic library handling for complex environments such as ComfyUI and third-party plugin ecosystems.

The update addresses potential conflicts caused by:

• multiple OpenMP runtimes loaded in the same process
• conflicting ggml dynamic libraries from external environments

Additionally, deprecated /bin directory searching has been removed from ggml dynamic library loading to make backend resolution behavior more predictable.

llama.cpp Synchronization

Updated llama.cpp to:

ggml-org/llama.cpp
commit 69bf6437914596fbbc4caf09a7ac16f2acdd1a94

The llama, MTMD, and GGML API bindings were synchronized on:

2026-08-08

Summary

This release continues the effort of keeping llama-cpp-python closely synchronized with upstream llama.cpp while improving the stability and extensibility of the Python binding layer.

The focus of this version is API readiness: exposing new native capabilities safely, improving compatibility across environments, and preparing the foundation for future multimodal and advanced generation features.

Full comparison:

d9d27a7...3397ecb

JamePeng