mstar.engine.resources.attn#
Modules
The attention resource's shared machinery: the spec-time factory, the custom op every backend attends through, and the workspace pool. |
|
What a model declares about attention: its backend, its spec, its step. |
|
Cross-attention over a context written once and never extended. |
|
Attention as one dense FlashAttention-3 varlen pass. |
|
Paged attention through FlashInfer's prefill/decode wrappers. |
|
FlashInfer utility wrappers for batched paged attention. |