bit-jev

bit-jev scores explicit options over a 1.58-bit BitNet backbone. Its I2_S GGUF inference path loads a resident model and returns structured answers, logits, and probabilities without generating answer tokens.

GitHub documentation · GGUF model and model card · 中文说明

Install

pip install bit-jev

The Windows x64 wheel includes a precompiled CPU runner for AVX2 processors. CPU inference on those machines needs no Git, CMake, or C++ compiler. The 1.19 GB GGUF downloads from Hugging Face on first use; installation itself does not download model weights. Other platforms and GPU backends build the native runner on demand and require Git, CMake 3.28+, and a C++17 compiler (Windows C++ Build Tools). GPU builds also require a Vulkan SDK or CUDA Toolkit. The precompiled program uses pinned BitNet and llama.cpp source with the ReLU² runtime patch and carries their MIT license notices.

Resident inference

from bit_jev.gguf import BitJev

request = {
    "state": "A customer reports a duplicate charge.",
    "questions": {
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this?",
            "criteria": {"billing": "Payment and refund issues", "shipping": "Delivery issues"},
        }
    },
}

with BitJev.from_pretrained(device="cpu", threads=8) as model:
    result = model.infer(request)
    print(result["answers"], result["latency_ms"])

Use device="gpu" for Vulkan or device="cuda" for an NVIDIA CUDA build. GPU requests fail clearly if a backend or visible GPU is unavailable. A local model directory can replace the default Hugging Face repo, and binary="/path/to/bit-jev-cpu" can select an existing native runner. The model stays loaded for repeated infer() calls; latency_ms reports native compute only, excluding download, build, loading, encoding, and IPC.

The CLI accepts UTF-8 JSONL input:

bit-jev --device cpu --input requests.jsonl --output results.jsonl

Training and distillation dependencies are optional: pip install 'bit-jev[train]'. The native runner evaluates one causal row per question; multiple questions repeat the shared state. The repository's Apache-2.0 license covers source code, while the checkpoint has no standalone open-weights license. The model card documents Yelp training-data provenance and its unresolved permission request.