FAIE: an experimental inference engine
FAIE is research code for local inference of large Mixture-of-Experts models on consumer hardware. It covers layer and expert paging, ternary and low-bit kernels (ternary_kernel.cu), dequantization, an Arabic-root tokenizer probe, and proof scripts.
This is an inference engine, not a model. The weights converted by FAIE (about 53 GB, per-layer and per-expert files) are not redistributed here, pending confirmation of the source model's license.
What's here
- Engine and kernels:
faie.py,faie_kernel.py,faie_runner.py,faie_generate.py,ternary_kernel.cu,ternary.py,quant_repack.py,fast_dequant.py - Conversion:
faie_convert.py,run_conversion.py,reconstruct_layer.py,gguf_analyze.py - Research modules:
neocortex.py,hive.py,memory.py,kan.py,liquid.py,poincare.py,wave.py,ttt.py - Proof scripts and small result JSONs:
prove_*.py,faie_*_proof.json(raw local results, not independently verified) - Docs:
THE_BRAIN_SPEC.md,ROOT_CALCULUS.md,FAIE_HANDOFF_TO_FABLE.md,NEXT_AGENT_INSTRUCTIONS.md cloud/: QLoRA distillation scripts.data/: Quranic-root corpus parsers.
Run logs, caches and converted weights are left out on purpose. Treat everything here as experimental.
About the lab
ABUZ8 is a one-person lab. Ahmad builds it alone, with no team and no funding, and has spent the last year living inside AI work. The long-term goal is the SUPERBRAIN fusion: Arabic-root reasoning (SUPERBRAIN), tool use and memory (Qadir), fast local inference (FAIE) and new architecture experiments (MUTHALLATH), all joined in one open system.