bridge.models.stepfun.modelling_step37.rope#
Step3.7 RoPE — per-layer partial rotary applied by Step37DecoderLayer.
The text decoder uses Step-3.5’s per-layer rotary_percents /
rotary_base_per_layer mechanism (see
stepfun/step35_provider.Step35DecoderLayer.__init__). The vision tower’s
own 2D RoPE lives next to the attention block, in
modelling_step37/utils.EncoderRope2D. This module exists to mirror
qwen_vl/modelling_qwen3_vl/rope.py; it re-exports those entry points so
downstream code can import them from a single, Step3.7-namespaced location.
Module Contents#
Data#
API#
- bridge.models.stepfun.modelling_step37.rope.__all__#
[‘EncoderRope2D’, ‘apply_rotary_emb’, ‘rotate_half’]