soft3/glia.md

glia

universal .model runtime. graph-agnostic.

glia does not know about cybergraph, particles, or cyberlinks. it receives a .model and runs it. the same runtime that executes a graph-compiled model executes any other model.

tru  →  .model artifact  →  glia  →  inference outputs (tensors, features, neural activations)
                                          ↓
                                         mir  (neural features for rendering)

two crates

crate binary what
import/ mi source format → canonical .model. boundary where format variance ends and .model invariance begins
run/ mr universal .model runtime. backends: honeycrisp (ANE+AMX), wgpu (cross-platform), cpu

import depends on run. run knows nothing about import.

import — source format → .model

mi import <DIR> converts external model formats to the canonical .model file that run/ operates on.

format source status
GGUF (single or sharded) llama.cpp, ollama wired
safetensors (single or sharded) HuggingFace loader present, not wired
ONNX generic loader present, not wired
MLX Apple MLX not implemented

after import, the runtime never sees BF16, GGUF block ordering, fused KV tensors, or nested VL configs. format variance ends at the import boundary. workarounds in runtime are bugs.

see import/specs/ for the full import contract.

run — universal .model runtime

three execution paths:

path what when
curated known-arch fast path (LlamaStyle, BertStyle, DiTDiffusion, ...) production inference
graph IR generic graph walk over .model IR nodes unknown/novel architectures
nox trident-compiled programs running on nox VM proved computation

three backends:

backend hardware status
honeycrisp ANE (rane) + AMX (acpu) + Metal GPU (aruminium) primary
wgpu Vulkan / Metal / DX12 / WebGPU cross-platform fallback
cpu single-threaded, deterministic reference

hardware dispatch on macOS: large matrix ops → acpu (AMX); NRF head inference → rane (ANE).

see run/specs/ for the full runtime spec (architecture, ops, IR, quantization, tokenizer, execution contract).

what glia is not

glia does not compile .graph.model — that is tru. glia does not render — that is mir. glia receives a .model and produces outputs.

in the stack

tru (compile model)   →  .model + graph field (φ*, eigenvectors)
glia (run model)      →  inference outputs (features, activations)
mir (render)          →  R-1.0 world

see stack

Folder

Graph