Jev_CeNN — CUDA-optimized hybrid cellular attention

The model weights, tokenizer, config and all custom Python runtime code are included in this repository. The original research package is NOT required.

Install: pip install torch transformers==4.57.6 safetensors huggingface_hub

Download the load_model.py file or the entire repository, then run:

from load_model import load_model, predict
model, tokenizer = load_model("USERNAME/MODEL", device="cpu")
output = predict("USERNAME/MODEL", state={"speed": 20},
                 question={"type": "choice", "instructions": "Select action",
                           "criteria": {"accelerate": None, "brake": None}})
print(output)

The bundled loader executes bundled runtime.py. Review/pin trusted code before use. The model is a typed-decision encoder, NOT a text-generation model.

W1024 at 512-token input is a quadratic full-context control and does NOT demonstrate a faster or attention-free replacement. Consult training_report.json for actual held-out metrics.

Optional GPU acceleration uses Triton and falls back to PyTorch on CPU. Install triton>=3.0 for NVIDIA GPU inference. Speed and numerical parity must be measured on the target GPU.