mstar.engine.resources.attn.ragged

mstar.engine.resources.attn.ragged#

Modules

base

The cacheless attention resource: the spec-time factory, and the plan machinery its backends share.

block_causal

Block-causal cacheless attention through FlashInfer's paged prefill.

config

What a model declares about cacheless (ragged) attention.

flashinfer

Cacheless varlen attention through FlashInfer's ragged prefill wrapper.

wrappers

Cacheless (ragged) prefill wrappers: queries and keys packed by segment into [total_tokens, H, D] tensors, laid out by cu_seqlens.