oa::ml::nn::AttentionBackend

Execution policy for scaled dot-product attention. Auto selects only a route proven by device-specific performance evidence. It currently preserves the compositional AttentionBackend::Standard provider on every device.