Skip to content

armgit / janas

Public Imported from GitHub
0

37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.

3.3 KB
Raw Blame

Contributing to Janas

Janas is an experiment in early development: formats, interfaces and the command line change without notice. Before writing code, it is worth asking in an issue whether the thing you have in mind is already being changed underneath.

What helps most

A measurement from a machine nobody here owns. Everything in the engine is decided from what it measures, and it has only ever measured one machine (a Core Ultra 9 185H with 32 GB of RAM and an NVMe disk, on Debian 13). Run:

bin/x86_64-linux/janas-bench <model.jns>

with the machine to itself — close the browser and the editor first. The engine takes the memory that is free when it starts and reads the disk while it answers, so anything else running lands in the numbers: on the development laptop, with 8 GB already taken by other programs, the same model went from 20-25 tokens per second to nearly 18.

Then open a hardware report: the form asks for the report it writes, your distribution and kernel (uname -a), your memory (free -g), the disk the model sits on, and anything that looked wrong. docs/hardware.md lists which machines are wanted most: CPUs without AVX-VNNI, AMD, discrete GPUs, machines with 16 GB or less, slower disks.

A model that will not convert, or converts and then answers nonsense. Say which file (with its link and its SHA-256) and what gguf2jns printed — there is a form for that too.

Code

  • Build and test before sending anything:

    ./build.sh release test
    ./build.sh asan test        # and tsan, if you touched anything concurrent
    

    ./build.sh --help says what each variant catches and what the actions do.

  • The code is C11 with no dependencies beyond the C library. Vulkan, when used, is loaded at run time: the engine must build and run without it.

  • Format with clang-format -i (the .clang-format in the root). Warnings are errors: -Wall -Wextra -Werror.

  • A change that touches arithmetic must keep the results it promises: kernels agree bit for bit with the scalar reference, and a pass over a block of tokens agrees bit for bit with the same tokens one at a time. The tests check this; keep them passing.

  • Numbers in a commit message are welcome, and should say on what they were measured.

Copyright

Janas is licensed under the GNU General Public License, version 3 or later (see LICENSE).

Code from contributors cannot be accepted yet. Janas's copyright is going to be held by a non-profit foundation, which does not exist yet; once it does, contributors will assign their rights in what they contribute to it, and their authorship will be recognised. A single holder is what keeps the code free for everyone to use, in the spirit of free and open source software, and lets the project defend its licence as a whole. Until the foundation is set up, pull requests with code stay open but are not merged; reports, measurements and ideas in the issues are as welcome as ever.

Models

The engine reads models; it does not carry them. Anything about redistributing converted weights is in MODELS.md — in short, the weights keep their own licence, which is not this one.