Quila

Appearance

Architecture ​

System Overview ​

Quila uses a neuron-based dynamic graph architecture with 32,768 neurons operating in parallel.

Neuron Structure ​

Each neuron maintains 8 channels of state:

  • S1: Semantic (long-term knowledge)
  • S2: Episodic (session memory)
  • S3: Working (active computation)
  • S4: Plan (goal representation)
  • S5: Tool (external interaction)
  • S6: Output (generation buffer)
  • S7: Conflict (contradiction detection)
  • S8: Meta (self-monitoring)

Computation Streams ​

4 parallel streams process each neuron:

  • Stream A: Micro-NSA (sparse attention)
  • Stream B: SSM (state space model)
  • Stream C: Linear Attention (RWKV-style)
  • Stream D: DRC (residual correction)

Inference Pipeline ​

6 phases execute sequentially:

  1. Phase 0: VQ-GAN encoding
  2. Phase 1: Replay (adaptive skip)
  3. Phase 2: LastInput (demand identification)
  4. Phase 3: Context re-read (HOT-NSA grading)
  5. Phase 4: Plan generation
  6. Phase 5: Think + Tool + Generate
  7. Phase 6: Output (Attention Merge)

Memory Hierarchy ​

  • L1 (GPU SRAM): Neuron states
  • L2 (GPU HBM): WMQ, active KFE
  • L3 (System RAM): SCT, cold KFE
  • L4 (NVRAM): Persona vector
  • L5 (NVMe): NMDB, Engram