vllm.models.step5.common.ops

Modules:

• indexer_norm_rope –

  Indexer normalization, rotary embedding and FP8 query preparation.

• prefill_indexer_metadata –

  Step-local indexer metadata shared across sparse attention layers.

• sparse_decode_attention –

  DSA decode attention read directly from paged KV by selected regions.

• sparse_indexer –

  Packing and stable region selection for the sparse indexer.

• sparse_indexer_scores –

  FP8 indexer scores directly from paged summary keys.

• summary_update –

  CSA region summaries, compressed once when a region's eighth row arrives.