nx_f32_gqa_attention.nx

nx_f32_gqa_attention.nx -- software-f32 Grouped-Query multi-head attention (Ainslie 2023 GQA). sd-server -> Nishi migration: Qwen3-4B (the Z-Image text encoder + companion chat model) uses GQA with 32 query heads sharing 8 kv heads (4:1, confirmed from the real GGUF: q_out 4096, kv_out 1024). This wraps the gated single-head `nx_f32_attention`: for each query head h it gathers Q_h and the SHARED K/V of kv head (h / group), runs scaled-dot-product attention, and scatters the head output back. This is the attention core of a real Qwen (and DiT) transformer layer. Q,out: flat *i64 f32 bits [n_tokens, n_q_heads*head_dim]. K,V: [n_tokens, n_kv_heads*head_dim]. license_tier: ORIGINAL