Module bitnet 

Native BitNet MoE model implementation for Hydra.

This module implements the Hydra model architecture with native Rust inference, loading weights directly from safetensors format.

Architecture (from actual model weights)

Input tokens → Embedding [32000, 192]
             ↓
4x MoE Layers:
  - Gate: Linear(192, 4) → softmax → top-k selection
  - Experts: Heterogeneous MLPs (different depths/widths)
             ↓
LayerNorm [192]
             ↓
SemanticHead: Linear(192, 192)
             ↓
CompressionHead: Linear(192, 4) → [NONE, BPE, BROTLI, ZLIB]
SecurityHead: Linear(192, 2) → [SAFE, UNSAFE]

Structs

Expert

Expert MLP with variable architecture

HydraBitNet

Complete Hydra model

HydraConfig

Model configuration derived from actual weights

LayerNorm

Layer normalization

Linear

Linear layer (dense)

MoELayer

MoE Layer with gating