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

mamba2_ssd_chunk_scan_varlen_fwd_gpu

def mamba2_ssd_chunk_scan_varlen_fwd_gpu[kernel_dtype: DType, DSTATE: Int, x_LT: TensorLayout, dt_LT: TensorLayout, A_LT: TensorLayout, B_LT: TensorLayout, C_LT: TensorLayout, D_LT: TensorLayout, dt_bias_LT: TensorLayout, initial_states_LT: TensorLayout, y_LT: TensorLayout, final_states_LT: TensorLayout, query_start_loc_LT: TensorLayout, has_initial_state_LT: TensorLayout, Storage: TensorStorage = PointerStorage](nheads: Int, head_dim: Int, ngroups: Int, nheads_ngroups_ratio: Int, batch: Int, dt_softplus: Int8, x: TileTensor[kernel_dtype, x_LT, MutUntrackedOrigin, Storage=Storage], dt: TileTensor[kernel_dtype, dt_LT, MutUntrackedOrigin, Storage=Storage], A: TileTensor[kernel_dtype, A_LT, MutUntrackedOrigin, Storage=Storage], B: TileTensor[kernel_dtype, B_LT, MutUntrackedOrigin, Storage=Storage], C: TileTensor[kernel_dtype, C_LT, MutUntrackedOrigin, Storage=Storage], D: TileTensor[kernel_dtype, D_LT, MutUntrackedOrigin, Storage=Storage], dt_bias: TileTensor[kernel_dtype, dt_bias_LT, MutUntrackedOrigin, Storage=Storage], initial_states: TileTensor[DType.float32, initial_states_LT, MutUntrackedOrigin, Storage=Storage], y: TileTensor[kernel_dtype, y_LT, MutUntrackedOrigin, Storage=Storage], final_states: TileTensor[DType.float32, final_states_LT, MutUntrackedOrigin, Storage=Storage], query_start_loc: TileTensor[DType.int32, query_start_loc_LT, MutUntrackedOrigin, Storage=Storage], has_initial_state: TileTensor[DType.bool, has_initial_state_LT, MutUntrackedOrigin, Storage=Storage], x_strides: IndexList[Int(3)], dt_strides: IndexList[Int(2)], A_strides: IndexList[Int(1)], B_strides: IndexList[Int(3)], C_strides: IndexList[Int(3)], D_strides: IndexList[Int(1)], dt_bias_strides: IndexList[Int(1)], initial_states_strides: IndexList[Int(4)], y_strides: IndexList[Int(3)], final_states_strides: IndexList[Int(4)])

GPU kernel: Mamba-2 SSD varlen prefill scan, one thread per (head, channel).

Grid: (ceildiv(head_dim, BLOCK), nheads, batch). Each thread owns one (b, h, p) channel, carries the dstate-vector state in registers, and walks its sequence [seq_start, seq_end) sequentially.