Module recurrent_gemma

Module recurrent_gemma 

Source
Expand description

Recurrent Gemma model implementation

Recurrent Gemma is a version of the Gemma language model that incorporates recurrent memory. This allows the model to maintain state between predictions and have longer-range memory.

Key characteristics:

  • Real-gated linear recurrent units (RGLRU)
  • 1D convolution for local context
  • RMSNorm for layer normalization
  • Rotary positional embeddings (RoPE)
  • Grouped query attention

References:

This implementation is based on the python version from huggingface/transformers. https://github.com/huggingface/transformers/blob/b109257f4fb8b1166e7c53cc5418632014ed53a5/src/transformers/models/recurrent_gemma/modeling_recurrent_gemma.py#L2

Structs§

Config
Model

Enums§

TemporalBlockType