Module gemma4_engine 

Gemma-4 dedicated text engine (E2B / E4B-style GGUF).

Distinct from generic Gemma-2/3 crate::decoder::Decoder:

• per-layer token embeddings (per_layer_token_embd + proj)
• shared KV layers (attention.shared_kv_layers)
• SWA vs full layers with different head dims (key_length / key_length_swa) and a bool SWA pattern array

Graph mirrors .scratch/llama.cpp/src/models/gemma4.cpp.

Structs

Gemma4AttnWeights

Gemma4DecodeState

Gemma4Engine

Gemma4Hparams

Hyperparameters from {arch}.* GGUF metadata.

Gemma4LayerWeights

Gemma4Weights

Constants

GEMMA4_ARCHES

Architectures served by this engine.