cyb/pipeline.md

the full cycle of intelligence — from cyberlink to inference and back

link → graph → .cyb → compile → inference
 ↑                                    │
 └────── reverse (model → graph) ─────┘

stages

link

neurons create and edit cyberlinks — the atomic units of knowledge. every link is a signed, staked assertion binding two particles

tools: cyb browser (react/, bevy/), scripting (nu/)

render

the cybergraph visualized — pages, navigation, search, graph minimap

tools: optica publisher, cyb browser

format

.cyb — single-file container for graph + weights + config. the universal exchange format between all stages

sections: config (TOML), graph IR (nodes + edges), tensor index, tensor data (Q4/Q8/F16/F32)

canonical tensor naming: HuggingFace style (model.layers.0.self_attn.q_proj.weight)

reverse

extract computation graph from trained models. ONNX protobuf → IR nodes. safetensors → weight tensors. GGUF → quantized weights. all formats converge into one Graph

loaders: safetensors, GGUF, ONNX, .cyb

compile

graph optimization and weight quantization at import time. F16/F32 weights → Q4_0 (4-bit, 3.5x compression). tensor name normalization. config extraction

pipeline: cyb-llm import → canonicalize + quantize + pack → .cyb

inference

.cyb → GPU compute → tokens. three backend paths reading the same .cyb:

decode loop: embed → (norm → QKV → RoPE → KV cache → attention → O proj → residual → norm → FFN → residual) × layers → LM head → sample

the cycle

inference produces new knowledge. neurons observe the output, create new cyberlinks, and the graph grows. the pipeline is a loop — not a line

invariants

one format (.cyb) for all stages. one set of tensor names (HF canonical). one quantization (Q4_0) for all backends. no format conversion at runtime — all conversion happens at import

discover all concepts