Single-dispatch Metal flash-attention forward. q,k,v are fp16 ([B,Hq,Tq,D] / [B,Hkv,Tkv,D]). Returns [O fp16, L fp32 (log-sum-exp)].
Single-dispatch Metal flash-attention forward. q,k,v are fp16 ([B,Hq,Tq,D] / [B,Hkv,Tkv,D]). Returns [O fp16, L fp32 (log-sum-exp)].