Repository navigation

refactor(attention): separate backends, runtime and persistent state - #1700

Closed

sufubao wants to merge 23 commits into

ModelTC:mainfrom

sufubao:refactor/attention-backend-plans

Closed

refactor(attention): separate backends, runtime and persistent state#1700
sufubao wants to merge 23 commits into
ModelTC:mainfrom
sufubao:refactor/attention-backend-plans

Commits on Nov 12, 2026
- refactor(attention): separate backends, runtime and persistent state
  sufubao
  committed
  24fb7d3
- refactor(attention): drop unsupported hardware adapters
  sufubao
  committed
  1eab574
- fix(attention): keep sparse query metadata on plans
  sufubao
  committed
  880592c
- style(attention): apply repository formatting
  sufubao
  committed
  1eb2ad7
- fix(attention): enforce backend ownership and selector fallback
  sufubao
  committed
  c4ab271
- fix(attention): enforce rank-consistent backend selection
  sufubao
  committed
  72b438e
- test(attention): cover remote rank selection failures
  sufubao
  committed
  45bc3f9
- fix(attention): update DeepSeek4 cache test fixture
  sufubao
  committed
  4ff64d7
- fix(attention): share sparse MLA capability checks
  sufubao
  committed
  4068feb
- style(attention): use pinned Black formatting
  sufubao
  committed
  e1a0bc9
- fix(attention): honor sliding windows in MHA decode
  sufubao
  committed
  400f1fc
- fix(attention): migrate offline model initialization
  sufubao
  committed
  f4eadc3
- docs(attention): update model initialization guide
  sufubao
  committed
  b9618da
- style(attention): normalize imports
  sufubao
  committed
  51b06e3
- refactor(attention): raise unsupported capability errors
  sufubao
  committed
  5f077f4
- refactor(attention): make layer forward calls explicit
  sufubao
  committed
  1808721
- refactor(attention): rename LayerAttentionBinding
  sufubao
  committed
  0954237
- refactor(attention): name execution contexts explicitly
  sufubao
  committed
  419a8d2
- refactor(attention): separate varlen operators from backends
  sufubao
  committed
  4c79bac
- Merge upstream/main into attention backend refactor
  sufubao
  committed
  8e27692
- Merge upstream/main KDA prefill fusion
  sufubao
  committed
  ad6abb9
- refactor(attention): clarify context and layer state ownership
  sufubao
  committed
  0e22d05
- Merge upstream/main KDA diagonal solve optimization
  sufubao
  committed
  d649b22

Commits

Commits on Nov 12, 2026