flashinfer.mla.DSV4HCAMetadata

class flashinfer.mla.DSV4HCAMetadata(hca_swa_indices: Tensor, hca_compressed_block_tables: Tensor, hca_seq_lens: Tensor, swa_topk_lens: Tensor, hca_is_causal: bool = True)

Reusable metadata for the CuTe DSL causal HCA backend.

hca_swa_indices contains absolute rows into the flattened SWA cache; hca_compressed_block_tables contains physical compressed-cache page IDs. Both are ordered by flattened query row (B * Q). hca_seq_lens describes the backing HCA footprint scheduled by TMA and may exceed the effective per-row sparse_topk_lens.

__init__(hca_swa_indices: Tensor, hca_compressed_block_tables: Tensor, hca_seq_lens: Tensor, swa_topk_lens: Tensor, hca_is_causal: bool = True) → None

Methods

__init__(hca_swa_indices, ...[, hca_is_causal])

Attributes

hca_is_causal
hca_swa_indices
hca_compressed_block_tables
hca_seq_lens
swa_topk_lens