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

varlen_selective_state_update_gpu

varlen_selective_state_update_gpu[kernel_dtype: DType, DSTATE: Int, state_LT: TensorLayout, x_LT: TensorLayout, dt_LT: TensorLayout, A_LT: TensorLayout, B_LT: TensorLayout, C_LT: TensorLayout, D_LT: TensorLayout, z_LT: TensorLayout, output_LT: TensorLayout, dt_bias_LT: TensorLayout, state_batch_indices_LT: TensorLayout](total_threads: Int, batch: Int, nheads: Int, dim: Int, nheads_ngroups_ratio: Int, pad_slot_id: Int32, dt_softplus: Int8, has_state_batch_indices: Int8, state: TileTensor[kernel_dtype, state_LT, MutExternalOrigin], x: TileTensor[kernel_dtype, x_LT, MutExternalOrigin], dt: TileTensor[kernel_dtype, dt_LT, MutExternalOrigin], A: TileTensor[kernel_dtype, A_LT, MutExternalOrigin], B: TileTensor[kernel_dtype, B_LT, MutExternalOrigin], C: TileTensor[kernel_dtype, C_LT, MutExternalOrigin], D: TileTensor[kernel_dtype, D_LT, MutExternalOrigin], z: TileTensor[kernel_dtype, z_LT, MutExternalOrigin], output: TileTensor[kernel_dtype, output_LT, MutExternalOrigin], dt_bias: TileTensor[kernel_dtype, dt_bias_LT, MutExternalOrigin], state_batch_indices: TileTensor[DType.int32, state_batch_indices_LT, MutExternalOrigin], state_strides: IndexList[4], x_strides: IndexList[3], dt_strides: IndexList[3], dt_bias_strides: IndexList[2], A_strides: IndexList[3], B_strides: IndexList[3], C_strides: IndexList[3], D_strides: IndexList[2], z_strides: IndexList[3], out_strides: IndexList[3])

GPU kernel for selective state update with multi-head support.