transformer_lens.model_bridge.supported_architectures.qwen3_vl_moe module¶
Qwen3-VL-MoE architecture adapter.
Alibaba’s Qwen3-VL-MoE (Qwen3VLMoeForConditionalGeneration): the
Qwen3-VL layout (DeepStack vision tower, interleaved-mRoPE text decoder)
with a sparse-MoE MLP — batched experts plus a parameter-only top-k
router returning a (logits, scores, indices) tuple, so the router stays
unwrapped and MoEBridge delegates the block. Layers listed in
mlp_only_layers hold a dense gated MLP under the same name.
- class transformer_lens.model_bridge.supported_architectures.qwen3_vl_moe.Qwen3VLMoeArchitectureAdapter(cfg: Any)¶
Bases:
Qwen3VLArchitectureAdapterArchitecture adapter for Qwen3VLMoeForConditionalGeneration models.