bridge.models.qwen3_asr.modeling_qwen3_asr.rope#

Module Contents#

Functions#

get_rope_index

Calculate the 3D rope index for Qwen3-ASR.

API#

bridge.models.qwen3_asr.modeling_qwen3_asr.rope.get_rope_index(
input_ids: torch.LongTensor,
attention_mask: torch.Tensor | None = None,
) → tuple[torch.Tensor, torch.Tensor]#

Calculate the 3D rope index for Qwen3-ASR.

Simplified version for audio-only model: just cumulative position IDs expanded to 3 MRoPE dimensions. No special vision/video handling needed.

Ported from HF Qwen3ASRPreTrainedModelForConditionalGeneration.get_rope_index.

Parameters:

  • input_ids – Input token IDs of shape (batch_size, sequence_length).

  • attention_mask – Attention mask of shape (batch_size, sequence_length).

Returns:

Position IDs of shape (3, batch_size, sequence_length). mrope_position_deltas: Position deltas of shape (batch_size, 1).

Return type:

position_ids