bridge.models.stepfun.modelling_step37.transformer_block#
Step3.7 transformer-block spec builder for the text decoder.
Mirrors qwen_vl/modelling_qwen3_vl/transformer_block.py: returns the
per-layer TransformerLayerSubmodules spec consumed by Megatron’s
GPTModel. Step3.7 reuses Step-3.5’s hybrid full/sliding decoder layer
(_build_step35_layer_spec) verbatim — the function lives in
step35_bridge to keep all Step-3.5 spec-construction logic in one place.
Module Contents#
Functions#
Return the Step-3.5 hybrid layer spec used as Step3.7’s text decoder. |
Data#
API#
- bridge.models.stepfun.modelling_step37.transformer_block.get_step37_text_layer_spec(*args, **kwargs)#
Return the Step-3.5 hybrid layer spec used as Step3.7’s text decoder.
- bridge.models.stepfun.modelling_step37.transformer_block.__all__#
[‘get_step37_text_layer_spec’]