Realization.Nvidia.SM86.StreamingAttentionSM86

dQ from dS^T: parameters K 1, dQ 7, s 10 (the query pass's) and dS^T 11 (the key pass's argument 14); grid as the query pass's. Three shared stages of K's tile and dS^T's (64 rows of 128 bytes each, unpadded: a row's chunk c at c ^ (row % 8), so the eight rows an LDSM reads fall in distinct banks), the copies two tiles ahead. K's and dS^T's offsets of the thread's copy rows r + 16 m (LDGSTS's global immediate is 11 bits: each row group keeps its own offset)

1991def sqdKeyOffset = (lambda unrestricted m : Nat . (naturalAdd 180 m))