Coverage for transformer_lens/model_bridge/supported_architectures/qwen2_audio.py: 100%

1"""Qwen2-Audio architecture adapter. 

2 

3``Qwen2AudioForConditionalGeneration``: a Whisper-style audio encoder at 

4``model.audio_tower``, a linear projector at ``model.multi_modal_projector``, 

5and a Qwen2 text decoder at ``model.language_model`` with a top-level 

6``lm_head`` (the transformers >= 5.13 layout). Text-only forwards work without 

7audio; audio features enter via ``input_features``. 

8""" 

9 

10from typing import Any 

11 

12from transformer_lens.model_bridge.generalized_components.base import ( 

13 GeneralizedComponent, 

14) 

15from transformer_lens.model_bridge.supported_architectures.qwen2 import ( 

16 Qwen2ArchitectureAdapter, 

17) 

18 

19 

20class Qwen2AudioArchitectureAdapter(Qwen2ArchitectureAdapter): 

21 """Architecture adapter for Qwen2AudioForConditionalGeneration models.""" 

22 

23 _testing_lm_attr = "model.language_model" 

24 _testing_eager = "config" 

25 

26 def __init__(self, cfg: Any) -> None: 

27 """Initialize the Qwen2-Audio architecture adapter.""" 

28 super().__init__(cfg) 

29 

30 self.cfg.is_multimodal = True 

31 self._reprefix_components("model.", "model.language_model.") 

32 

33 # The Whisper-style encoder and projector are wrapped opaquely 

34 # (hook_in/hook_out only) — audio features are injected by the HF 

35 # forward when input_features is passed. 

36 self.components["audio_encoder"] = GeneralizedComponent(name="model.audio_tower") 

37 self.components["audio_projector"] = GeneralizedComponent( 

38 name="model.multi_modal_projector" 

39 )