Module model 

Llama-family transformer model for inference.

Composes trueno primitives (rms_norm, Q4K matmul, fused attention) into a complete transformer that loads GGUF weights and generates text.

Structs

ForwardArena

Pre-allocated scratch buffers for the forward pass. Eliminates all per-token heap allocations (FALSIFY-ARENA-001). Contract: contracts/cgp/cgp-inference-arena-v1.yaml

KvCache

KV cache for incremental decoding.

LayerWeights

Weights for a single transformer layer.

LlamaModel

Complete transformer model ready for inference.

ModelConfig

Model hyperparameters extracted from GGUF metadata.

ModelWeights

Full model weights.

Enums

WeightMatrix

A weight matrix that may be Q4K (bytes) or any-other-quant dequantized to F32.