minja.hpp
A minimalistic C++ Jinja templating engine specifically designed for LLM chat templates, used in llama.cpp and other projects.
What is minja.hpp?
Minja is a minimalistic C++ implementation of the Jinja templating engine specifically designed for formatting chat templates used by large language models (LLMs). It solves the problem of integrating Jinja-like templating into C++ LLM projects without the bloat of a full general-purpose engine, providing just enough functionality to support templates from models on HuggingFace.
C++ developers working on LLM inference engines, such as those building or integrating with projects like llama.cpp, gemma.cpp, or Jan, who need to process HuggingFace-style chat templates efficiently.
Developers choose Minja for its focused feature set, ease of integration as a header-only library, and lightweight codebase that avoids unnecessary complexity while supporting the exact Jinja subset required for LLM chat templates.
Overview
A minimalistic C++ Jinja templating engine for LLM chat templates
Use Cases
Best For
- Integrating Jinja templating into C++ LLM inference projects like llama.cpp
- Processing HuggingFace chat templates in C++ applications
- Adding chat template support to custom LLM inference engines
- Reducing dependencies in C++ projects that need minimal templating
- Ensuring compatibility with major LLM model templates
- Avoiding the overhead of full Jinja implementations like Jinja2Cpp
Not Ideal For
- Web applications requiring full Jinja2 compliance for HTML templating with proper XSS escaping
- Projects needing extensive Jinja features like template inheritance, includes, or all built-in filters
- Teams that cannot implement additional message sanitization to guard against prompt injection risks
- Use cases involving general-purpose templating beyond LLM chat formats, such as configuration files or report generation
Pros & Cons
Pros
Implements only the Jinja subset needed for LLM chat templates, keeping the codebase at 2.5k lines and avoiding feature creep, as stated in the design goals.
Header-only with a single dependency on nlohmann::json and no Boost required, making it straightforward to add to projects like llama.cpp with simple CMake integration.
Supports templates from major HuggingFace models and is integrated in projects like Jan and GPT4All, ensuring reliability for its intended use case.
Includes a chat-template wrapper that handles quirks in message history formatting, providing a consistent interface across different model templates.
Cons
Lacks many Jinja features such as full filter set, tuples, and template inheritance, making it unsuitable for general templating needs beyond LLM chat templates.
No protection against prompt injection and unsafe for HTML due to lack of escaping, requiring users to implement their own security measures, as warned in the README.
Has known line difference problems on Windows that need fixing, as mentioned in the roadmap, which could affect cross-platform consistency.
Frequently Asked Questions
Related Projects
stbstb single-file public domain libraries for C/C++
{fmt}A modern formatting library
xxHashExtremely fast non-cryptographic hash algorithm
single_file_libsList of single-file C/C++ libraries, with emphasis on clause-less licenses.
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.