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_mla_decode_graph_paged_fp8_sparse_cp
struct Struct_mla_decode_graph_paged_fp8_sparse_cp
Registers the attention.mla.graph.decode.paged.fp8.sparse.cp graph op.
attention.mla.graph.decode.paged.fp8.sparse for decode context parallelism.
Each query token t attends over the first sparse_lengths[t] entries of
its sparse_indices row (0 is allowed), and the op also writes lse, the
attention log-sum-exp per (token, head) in base 2:
log2(sum_k exp2(q.k * scale * log2(e))), before W_UV. A token with no
keys gets lse = -inf and output = 0. Ranks merge as
w_r = exp2(lse_r - max_r lse_r), O = sum_r w_r O_r / sum_r w_r.
split_len_bound (>0) bounds every length and sizes split-K in place of
indices_stride; it only affects performance. As in the base op with
dcp_size > 1, the fused latent store writes a new token only on the rank
that owns its position (ContextParallelLayout[dcp_size, dcp_interleave]), at that rank's local slot; dcp_rank is a host scalar
after split_len_bound.
SM100 only.
Implemented traitsโ
Methodsโ
executeโ
static def execute[m_scale_granularity: Int, n_scale_granularity: Int, k_scale_granularity: Int, mask_str: StaticString, dcp_size: Int, dcp_interleave: Int, indices_stride: Int, physical_indices: Bool](output: T, lse: T, q: T, kv: T, input_row_offsets: T, freqs_cis: T, kv_norm_gamma: T, w_uk: T, w_uv: T, kv_blocks: T, page_stride: T, cache_lengths: T, kv_lookup_table: T, max_prompt_length: T, max_cache_length: T, layer_idx: UInt32, scale: Float32, epsilon: Float32, w_uk_scale: T, w_uv_scale: T, scalar_args: T, sparse_indices: T, sparse_lengths: T, split_len_bound: Int32, dcp_rank: Int32, context: DeviceContext)