For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

selective_scan_fwd_gpu_minimal

def selective_scan_fwd_gpu_minimal[kernel_dtype: DType, DSTATE: Int, output_LT: TensorLayout, x_LT: TensorLayout, u_LT: TensorLayout, delta_LT: TensorLayout, A_LT: TensorLayout, B_LT: TensorLayout, C_LT: TensorLayout](total_batch_dim: Int, batch: Int, dim: Int, seqlen: Int, group_size: Int, delta_softplus: Int8, output: TileTensor[kernel_dtype, output_LT, MutUntrackedOrigin], x: TileTensor[kernel_dtype, x_LT, MutUntrackedOrigin], u: TileTensor[kernel_dtype, u_LT, MutUntrackedOrigin], delta: TileTensor[kernel_dtype, delta_LT, MutUntrackedOrigin], A: TileTensor[kernel_dtype, A_LT, MutUntrackedOrigin], B: TileTensor[kernel_dtype, B_LT, MutUntrackedOrigin], C: TileTensor[kernel_dtype, C_LT, MutUntrackedOrigin], output_strides: IndexList[Int(3)], x_strides: IndexList[Int(4)], u_strides: IndexList[Int(3)], delta_strides: IndexList[Int(3)], A_strides: IndexList[Int(2)], B_strides: IndexList[Int(4)], C_strides: IndexList[Int(4)])

Minimal GPU kernel for selective scan forward - no D, z, or delta_bias.

Each thread processes one (batch, dim) pair and iterates through the sequence. Omits the D skip connection, z gating, and delta_bias supported by selective_scan_fwd_gpu.

Parameters:

kernel_dtype (DType): Element type of the input and output tensors.
DSTATE (Int): Number of SSM state elements per (batch, dim) pair.
output_LT (TensorLayout): Memory layout of the output tensor.
x_LT (TensorLayout): Memory layout of the x checkpoint tensor.
u_LT (TensorLayout): Memory layout of the u input tensor.
delta_LT (TensorLayout): Memory layout of the delta time-step tensor.
A_LT (TensorLayout): Memory layout of the A recurrence matrix.
B_LT (TensorLayout): Memory layout of the B input projection tensor.
C_LT (TensorLayout): Memory layout of the C output projection tensor.

Args:

total_batch_dim (Int): Total number of (batch, dim) pairs launched, equal to batch * dim, used for thread bounds checking.
batch (Int): Number of sequences processed in parallel.
dim (Int): Hidden dimension, equal to the number of channels per sequence position.
seqlen (Int): Number of timesteps in each sequence.
group_size (Int): Number of dims per group; dims in the same group share B and C inputs.
delta_softplus (Int8): Nonzero applies softplus to delta before the scan recurrence.
output (TileTensor[kernel_dtype, output_LT, MutUntrackedOrigin]): Output tensor of shape (batch, dim, seqlen), written.
x (TileTensor[kernel_dtype, x_LT, MutUntrackedOrigin]): Checkpoint tensor of shape (batch, dim, n_chunks, 2*DSTATE) storing per-chunk cumulative A and B values, written.
u (TileTensor[kernel_dtype, u_LT, MutUntrackedOrigin]): Selective scan input tensor of shape (batch, dim, seqlen), read.
delta (TileTensor[kernel_dtype, delta_LT, MutUntrackedOrigin]): Time-step tensor of shape (batch, dim, seqlen), read.
A (TileTensor[kernel_dtype, A_LT, MutUntrackedOrigin]): SSM recurrence matrix of shape (dim, DSTATE), read.
B (TileTensor[kernel_dtype, B_LT, MutUntrackedOrigin]): SSM input projection of shape (batch, n_groups, DSTATE, seqlen), read.
C (TileTensor[kernel_dtype, C_LT, MutUntrackedOrigin]): SSM output projection of shape (batch, n_groups, DSTATE, seqlen), read.
output_strides (IndexList[Int(3)]): 3D strides (batch, dim, seqlen) for indexing output.
x_strides (IndexList[Int(4)]): 4D strides (batch, dim, n_chunks, 2*DSTATE) for indexing x.
u_strides (IndexList[Int(3)]): 3D strides (batch, dim, seqlen) for indexing u.
delta_strides (IndexList[Int(3)]): 3D strides (batch, dim, seqlen) for indexing delta.
A_strides (IndexList[Int(2)]): 2D strides (dim, DSTATE) for indexing A.
B_strides (IndexList[Int(4)]): 4D strides (batch, n_groups, DSTATE, seqlen) for indexing B.
C_strides (IndexList[Int(4)]): 4D strides (batch, n_groups, DSTATE, seqlen) for indexing C.