mirror of
https://github.com/luxfi/precompile.git
synced 2026-09-24 00:26:21 +00:00
Captures the validated path from the toy fixture to a real on-chain LLM (zen-nano / Qwen3-0.6B = Beluga's brain 'blue'): - ZEN_NANO_PORT.md: architecture (GQA 16/8, head_dim 128, q/k-norm, RoPE θ1e6, SwiGLU, tied embed, vocab 151936), the validated quantization recipe (per-channel int8 weights; dynamic-integer or static activation scales; gemmlowp requant; integer norm/softmax/silu/ rope LUTs), coherence results (float/dynamic/static int8 all coherent on real weights -> deterministic on-chain inference is VIABLE), and the Go-port blueprint + provisioning via the model-registry precompile. - cmd/blue: CLI that invokes the inference precompile (token-id prompt -> generated tokens), the same code path an on-chain tx hits. Talks to the toy fixture today; to real zen-nano once the port lands. Research validated this session against ~/work/zen/models/zen-nano-0.6b; remaining work is the Go engine port (generalize tiny_llm -> Qwen3) + a quant exporter, no unknowns. Co-authored-by: Hanzo Dev <[email protected]>