Mojo function

flash_attention

flash_attention[dtype: DType, q_layout: Layout, //, use_score_mod: Bool = False, config: MHAConfig = MHAConfig(dtype, UInt(Int.__init__[IntTuple[ComptimeOrigin]](q_layout.shape[2])), UInt(Int.__init__[IntTuple[ComptimeOrigin]](q_layout.shape[3])), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), 4, 1, FlashAttentionAlgorithm(-1), OptionalReg[UInt](None), TensorMapSwizzle(3)), decoding_warp_split_k: Bool = False, naive_kernel: Bool = False, sink: Bool = False](output: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], q: LayoutTensor[dtype, q_layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], k: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], v: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], mask: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], scale: Float32, context: DeviceContextPtr = DeviceContextPtr(), num_partitions: OptionalReg[Int] = None, sink_weights: OptionalReg[LayoutTensor[dtype, Layout.row_major(-1), MutableAnyOrigin]] = None)

flash_attention[cache_t: KVCacheT, mask_t: MHAMask, score_mod_t: ScoreModTrait, dtype: DType, q_layout: Layout, //, use_score_mod: Bool = False, config: MHAConfig = MHAConfig(dtype, UInt(Int.__init__[IntTuple[ComptimeOrigin]](q_layout.shape[(q_layout.rank() - 2)])), UInt(Int.__init__[IntTuple[ComptimeOrigin]](q_layout.shape[(q_layout.rank() - 1)])), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), 4, 1, FlashAttentionAlgorithm(-1), OptionalReg[UInt](None), TensorMapSwizzle(3)), ragged: Bool = False, sink: Bool = False, decoding_warp_split_k: Bool = False, naive_kernel: Bool = False](output: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], q: LayoutTensor[dtype, q_layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], k: cache_t, v: cache_t, mask_functor: mask_t, score_mod_functor: score_mod_t, valid_length: ManagedTensorSlice[io_spec, static_spec=static_spec], scale: Float32, ctx: DeviceContext, q_max_seq_len: OptionalReg[Int] = None, kv_input_row_offsets: OptionalReg[LayoutTensor[DType.uint32, Layout.row_major(-1), MutableAnyOrigin]] = None, num_partitions: OptionalReg[Int] = None, sink_weights: OptionalReg[LayoutTensor[dtype, Layout.row_major(-1), MutableAnyOrigin]] = None)

Flash attention 2 algorithm. Compute: (1) Transpose (Q) BSHD -> BHSD; (2) Transpose (K) BSHD -> BHSD; (3) Transpose (V) BSHD -> BHSD; (4) P = Bmm(Q, K), P is also called "score"; (5) P = P * scale + mask; (6) P = softmax(P); (7) O = Bmm(P, V) (8) Output = Transpose(O).

B, S, H, D denote batch size, sequence length, head count and depth, respectively. (1), (2), (3) happens while loading the data into shared memory. (8) happens when writing output to global memory.

All inputs (query, key, and value) must have BSHD layout. The mask can be BSS or BHSS.

This kernel also handles grouped attention optimization. In this case the shape of K and V are BShD where h = H / num_groups.

This kernels handles batches with different valid lengths (i.e., before the padding). Such lengths are passed in valid_length argument.

flash_attention[mask_t: MHAMask, score_mod_t: ScoreModTrait, dtype: DType, q_layout: Layout, //, use_score_mod: Bool = False, config: MHAConfig = MHAConfig(dtype, UInt(Int.__init__[IntTuple[ComptimeOrigin]](q_layout.shape[2])), UInt(Int.__init__[IntTuple[ComptimeOrigin]](q_layout.shape[3])), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), OptionalReg[UInt](None), 4, 1, FlashAttentionAlgorithm(-1), OptionalReg[UInt](None), TensorMapSwizzle(3)), decoding_warp_split_k: Bool = False, _use_valid_length: Bool = False, _padded_ndbuffer: Bool = False, naive_kernel: Bool = False, sink: Bool = False](output: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], q: LayoutTensor[dtype, q_layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], k: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], v: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], mask_functor: mask_t, score_mod_functor: score_mod_t, scale: Float32, ctx: DeviceContext, num_partitions: OptionalReg[Int] = None, valid_length: OptionalReg[ManagedTensorSlice[IOSpec[True, IO(-1)](), static_spec=StaticTensorSpec.create_unknown[DType.uint32, 1]()]] = None, sink_weights: OptionalReg[LayoutTensor[dtype, Layout.row_major(-1), MutableAnyOrigin]] = None)