bridge.models.glm_vl.data.collate_fn

GLM VL collator implementations.

Module Contents

Functions

_normalize_glm4v_assistant_content	Flatten text-only structured assistant content for GLM chat templates.
_glm4v_assistant_mask_boundary_config	Build GLM-4.5V role boundaries and exclude its empty thinking prefix.
_build_glm4v_assistant_loss_mask	Build GLM-4.5V loss spans using a virtual final turn terminator.
glm4v_collate_fn	Collate function for GLM-4.5V model.

Data

GLM4V_ASSISTANT_START
GLM4V_ASSISTANT_END
GLM4V_EMPTY_THINK
GLM4V_NEXT_ROLE_MARKERS

API

bridge.models.glm_vl.data.collate_fn.GLM4V_ASSISTANT_START

‘<|assistant|>\n’

bridge.models.glm_vl.data.collate_fn.GLM4V_ASSISTANT_END

‘<|endoftext|>’

bridge.models.glm_vl.data.collate_fn.GLM4V_EMPTY_THINK

‘\n’

bridge.models.glm_vl.data.collate_fn.GLM4V_NEXT_ROLE_MARKERS

(‘<|system|>\n’, ‘<|user|>\n’, ‘<|observation|>\n’)

bridge.models.glm_vl.data.collate_fn._normalize_glm4v_assistant_content(

example: dict[str, Any],

) → dict[str, Any]

Flatten text-only structured assistant content for GLM chat templates.

bridge.models.glm_vl.data.collate_fn._glm4v_assistant_mask_boundary_config(

processor: Any,

) → megatron.bridge.data.conversation_processing.AssistantMaskBoundaryConfig

Build GLM-4.5V role boundaries and exclude its empty thinking prefix.

bridge.models.glm_vl.data.collate_fn._build_glm4v_assistant_loss_mask(

example: dict,

input_ids: torch.Tensor,

processor: Any,

skipped_tokens: torch.Tensor,

boundary_config: megatron.bridge.data.conversation_processing.AssistantMaskBoundaryConfig,

) → torch.Tensor

Build GLM-4.5V loss spans using a virtual final turn terminator.

bridge.models.glm_vl.data.collate_fn.glm4v_collate_fn(

examples: list,

processor,

*,

visual_keys: object = None,

min_pixels: int | None = None,

max_pixels: int | None = None,

sequence_length: int | None = None,

pad_to_max_length: bool = False,

pad_to_multiple_of: int = 128,

enable_in_batch_packing: bool = False,

in_batch_packing_pad_to_multiple_of: int = 1,

) → dict[str, torch.Tensor]

Collate function for GLM-4.5V model.

GLM-4.5V requires mm_token_type_ids to distinguish image (1) and video (2) tokens from text (0) when computing 3D MRoPE positions. The processor returns this field by default (return_mm_token_type_ids=True in Glm4vProcessor defaults). We wrap all visual tensors — including mm_token_type_ids — in

Class:

GenericVisualInputs so they flow through vlm_step.py to the model.