For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo struct
Struct_latent_sparse_attention_ragged_paged
struct Struct_latent_sparse_attention_ragged_paged
Registers the mo.latent_sparse_attention.ragged.paged graph op with the graph compiler.
Sparse attention over a shared K=V latent read from two paged leaves: the
sliding-window leaf by position range and the compressed leaf by an
explicit per-query entry list, with a per-head attention sink in the
denominator. See nn.attention.latent_sparse_attention.
Implemented traits
Methods
execute
static def execute[q_type: DType, swa_type: DType, comp_type: DType, //, window: Int, target: StringSpan[ImmStaticOrigin]](output: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output.static_spec], q: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=q.static_spec], input_row_offsets: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input_row_offsets.static_spec], comp_indices: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=comp_indices.static_spec], attn_sink: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=attn_sink.static_spec], swa_kv_blocks: ManagedTensorSlice[IOSpec[_, _].MutableInput, static_spec=swa_kv_blocks.static_spec], swa_page_stride: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=swa_page_stride.static_spec], swa_cache_lengths: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=swa_cache_lengths.static_spec], swa_kv_lookup_table: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=swa_kv_lookup_table.static_spec], swa_max_prompt_length: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=swa_max_prompt_length.static_spec], swa_max_cache_length: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=swa_max_cache_length.static_spec], comp_kv_blocks: ManagedTensorSlice[IOSpec[_, _].MutableInput, static_spec=comp_kv_blocks.static_spec], comp_page_stride: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=comp_page_stride.static_spec], comp_cache_lengths: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=comp_cache_lengths.static_spec], comp_kv_lookup_table: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=comp_kv_lookup_table.static_spec], comp_max_prompt_length: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=comp_max_prompt_length.static_spec], comp_max_cache_length: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=comp_max_cache_length.static_spec], layer_swa: UInt32, layer_comp: UInt32, scale: Float32, context: DeviceContext)