Network-level configuration for Qwen 3.8 transformer networks.
Carries the published Qwen/Qwen3.8-27B geometry (Specifications/Qwen3.8.md section 1). Two things separate it from GemmaConfig, and both are the reason a distinct family exists:
- The interleave is over MIXERS, not geometries. Gemma alternates two attention layers that differ in head width and window; Qwen alternates full attention with a Gated DeltaNet recurrence, which shares no projection shape with it at all. So the DeltaNet geometry is its own set of fields rather than a "global" variant of the attention ones.
- The query projection is DOUBLE WIDTH. attn_output_gate makes it emit [query | gate], which changes the packed QKV width and the split the block performs.
What is deliberately NOT here:
- The gate function. It selects a type, so it rides the AttentionOutputGate template parameter, per the discriminating principle in Qwen3.8.md's overview: template axes are for types and layouts, runtime config is for arithmetic. Note the published output_gate_type: "swish" is a DEAD key – the reference implementation reads it nowhere and applies sigmoid – so it would be the wrong thing to carry here anyway.
- The mrope section split [11, 11, 10]. It is a multimodal positional layout, and the vision tower is out of scope for this chassis; text-only mrope degenerates to standard RoPE because all three sections then carry the same position ids.
- The MTP draft head and the vision tower (Qwen3.8.md section 1, "out of scope").
1.17.0