bridge.models.stepfun.modelling_step37.text_model

Step3.7 text decoder wrapper.

Mirrors qwen_vl/modelling_qwen3_vl/text_model.py: a thin subclass of megatron.core.models.gpt.GPTModel that fixes Step3.7-specific defaults without changing the GPTModel forward signature. The actual decoder layer type (Step-3.5’s hybrid full/sliding Step35DecoderLayer) is selected by the layer spec passed in by :class:Step37ModelProvider.

Module Contents

Classes

Step37GPTModel

GPTModel subclass used as Step3.7’s language tower.

Data

__all__

API

class bridge.models.stepfun.modelling_step37.text_model.Step37GPTModel

Bases: megatron.core.models.gpt.gpt_model.GPTModel

GPTModel subclass used as Step3.7’s language tower.

Currently this class only carries a Step3.7-specific name (so error / state-dict messages name the right model) and exists to mirror Qwen3VLGPTModel for structural parity. All forward behaviour is inherited from :class:megatron.core.models.gpt.GPTModel.

bridge.models.stepfun.modelling_step37.text_model.__all__

[‘Step37GPTModel’]