GQA (grouped-query attention)
Multiple query heads share a smaller number of key/value heads, shrinking the cache.
Almost every current model uses GQA. The number that matters for memory is num_key_value_heads, not num_attention_heads — a model with 32 query heads and 8 KV heads caches a quarter of what full multi-head attention would. When a model omits the KV head count, it defaults to the query head count, which is plain multi-head attention.