bridge.models.glm5next.glm53flash_builder#

Model builder for GLM-5.3-Flash.

It constructs GLM53FlashModel as the legacy GLM53FlashModelProvider does: the Hugging Face vision tower first on the pre-process stage, then the Hybrid language model.

Module Contents#

Classes#

GLM53FlashModelBuilder

Combine the GLM-5.3-Flash Hybrid language model with the Hugging Face vision tower.

Data#

API#

class bridge.models.glm5next.glm53flash_builder.GLM53FlashModelBuilder#

Bases: megatron.bridge.models.hybrid.hybrid_builder.BridgeHybridModelBuilder

Combine the GLM-5.3-Flash Hybrid language model with the Hugging Face vision tower.

build_model(
pg_collection: megatron.core.process_groups_config.ProcessGroupCollection,
pre_process: bool | None = None,
post_process: bool | None = None,
vp_stage: int | None = None,
) → megatron.bridge.models.glm5next.glm53flash_bridge.GLM53FlashModel#

Build one GLM-5.3-Flash pipeline stage.

Parameters:
  • pg_collection – Process groups used for distributed construction.

  • pre_process – Whether this stage owns the embeddings and the vision tower.

  • post_process – Whether this stage owns the language-model output layer.

  • vp_stage – Virtual pipeline stage.

Returns:

A GLM-5.3-Flash model stage.

Raises:

TypeError – If the builder was given another model config.

_build_language_model(
pg_collection: megatron.core.process_groups_config.ProcessGroupCollection,
pre_process: bool,
post_process: bool,
vp_stage: int | None,
) → megatron.core.models.hybrid.hybrid_model.HybridModel#

Build the Hybrid language model with the arguments HybridModelProvider.provide passes.

GLM53FlashModel inserts vision features into full-sequence embeddings and scatters them for sequence parallelism itself, so the embedding does not scatter. Megatron Core’s Hybrid builder has no argument for that, so this builds the model directly.

bridge.models.glm5next.glm53flash_builder.__all__#

[‘GLM53FlashModelBuilder’]