bridge.models.qwen_omni.data.collate_fn

Qwen3-Omni thinker collator implementation.

Module Contents

Functions

qwen3_omni_collate_fn

Collate typed Qwen3-Omni conversations with image, video, and audio inputs.

Data

CHATML_ASSISTANT_START
CHATML_ASSISTANT_END
CHATML_OTHER_ROLE_STARTS

API

bridge.models.qwen_omni.data.collate_fn.CHATML_ASSISTANT_START

‘<|im_start|>assistant\n’

bridge.models.qwen_omni.data.collate_fn.CHATML_ASSISTANT_END

‘<|im_end|>\n’

bridge.models.qwen_omni.data.collate_fn.CHATML_OTHER_ROLE_STARTS

None

bridge.models.qwen_omni.data.collate_fn.qwen3_omni_collate_fn(

examples: list[dict[str, Any]],

processor: Any,

*,

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, Any]

Collate typed Qwen3-Omni conversations with image, video, and audio inputs.

Media resolution is delegated to the Hugging Face processor’s chat-template path so local paths and URLs follow the processor’s native conversation schema. Qwen3-Omni training currently uses dense right-padded batches; its model step rejects in-batch packing.