Skip to content

YarnRotaryConfig

Module: fast_llm.layers.attention.rotary.config

Variant of: RotaryConfig — select with type: yarn

Inherits from: DefaultRotaryConfig, RotaryConfig, ModuleConfig

Fields

attention_factor — architecture
Type: float or None    Default: None
beta_fast — architecture
Type: float    Default: 32.0
beta_slow — architecture
Type: float    Default: 1.0
original_context_length — architecture
Type: int    Default: 8192
scale_factor — architecture
Type: float    Default: 8.0
theta — architecture

Type: float    Default: 10000

Scale for the rotary positional embeddings