transformer_lens.model_bridge.supported_architectures.qwen3_vl module¶
Qwen3-VL architecture adapter.
Alibaba’s Qwen3-VL (Qwen3VLForConditionalGeneration): a ViT tower at
model.visual matching the Qwen3.5 vision layout (learned pos_embed +
2D rotary, qkv/proj attention, fc1/fc2 MLP) plus DeepStack — extra patch
mergers on early vision blocks whose features the text model injects
into the residual stream at visual token positions during the first
decoder layers. The injection itself is a tensor add inside the HF text
loop (not a module call); the per-level DeepStack mergers are wrapped so
their features are hookable at the source. Because the injection happens
BETWEEN block calls, blocks.{i}.hook_out omits it while blocks.{i+1}’s
input contains it on image runs — resid_post[i] != resid_pre[i+1] at
visual positions for the first DeepStack layers, and patching resid_post
there drops the injection (prefer blocks.{i+1}.hook_in). Text attention
is Qwen3-style (per-head QK RMS-norm) with interleaved mRoPE, HF-native.
- class transformer_lens.model_bridge.supported_architectures.qwen3_vl.Qwen3VLArchitectureAdapter(cfg: Any)¶
Bases:
ArchitectureAdapterArchitecture adapter for Qwen3VLForConditionalGeneration models.
- __init__(cfg: Any) None¶
Initialize the Qwen3-VL architecture adapter.
- component_mapping: ComponentMapping | None¶
- required_libraries: list[str] = ['torchvision']¶
- required_libraries_group: str = 'multimodal'¶
- uses_split_attention: bool¶
- weight_processing_conversions: Dict[str, ParamProcessingConversion | str] | None¶