OptionalcanvasDiffusionGemma: fixed denoising-canvas length (256). Undefined elsewhere.
Qwen3-MoE sparse-block gating (qwen3_moe): a layer is MoE unless it is
in mlpOnlyLayers and (idx+1) % decoderSparseStep == 0 selects it.
normTopkProb renormalizes the top-k gate weights. Defaults are inert
for non-qwen3_moe families.
26B-A4B MoE block: dense MLP + routed experts per layer.
Every fullAttentionInterval-th layer is full-attention (Qwen3.5).
e2b/e4b: per-layer input embedding width (0 = disabled).
"sliding_attention" | "full_attention" per layer.
Qwen3.5 hybrid (gated-DeltaNet) geometry; 0/false for other families.
Global (full-attention) layers can use different KV geometry.
Trailing layers that share KV with earlier same-type layers.
Rotary fraction of head_dim (Qwen3.5 full attention: 0.25 → 64 of 256).
Qwen3.5 full-attention output gate: q_proj emits 2× and gates the output.