bridge.models.qwen_audio.data.collate_fn#
Qwen audio collator implementations.
Module Contents#
Functions#
Collate function for Qwen2-Audio model. |
API#
- bridge.models.qwen_audio.data.collate_fn.qwen2_audio_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,
Collate function for Qwen2-Audio model.
Uses HF-compatible label construction:
Backward search for assistant text spans (matching HF Trainer convention)
No skipped_tokens masking on labels (model learns to predict EOS/im_end)
Loss mask derived directly from active label positions
Qwen2-Audio keeps packing in
audio_lm_stepbecause audio features must remain batch-aligned until model input preparation.