Comment by dot_treo

I would probably treat the (3 GatedDeltaNet + 1 GatedAttention) Blocks as one transformer block, when generating next steps one would therefore use the kv cache for the gated attention and skip the entire delta nets.