Repository navigation

refactor(att_resources): unify attention state and cache transfers - #1664

Merged

hiworldwzj merged 56 commits into

mainfrom

wzj_att

Nov 2, 2026

Merged

refactor(att_resources): unify attention state and cache transfers#1664
hiworldwzj merged 56 commits into
mainfrom
wzj_att

Commits on Oct 31, 2026
Commits on Nov 1, 2026
- fix: address attention-state regressions and synchronize NCCL receives
  hiworldwzj
  committed
  43ffe86
- test: align PD and V4 fixtures with runtime initialization
  hiworldwzj
  committed
  472a689
- perf: remove redundant CPU cache offload stream wait
  hiworldwzj
  committed
  26b1f67
- fix: unify NCCL send and receive status polling
  hiworldwzj
  committed
  7c4d82f
- refactor MLA logical configuration and factories
  hiworldwzj
  committed
  30650a3
- refactor: organize attention states by type
  hiworldwzj
  committed
  038c4db
- refactor: group quantized states by attention type
  hiworldwzj
  committed
  161c3b1
- docs: simplify attention resource design
  hiworldwzj
  committed
  22cae3c
- fix packed MLA test state helpers
  hiworldwzj
  committed
  597080c
- avoid dynamic SWA kernel specializations
  hiworldwzj
  committed
  52e0d6b
- avoid dynamic DSpark SWA specializations
  hiworldwzj
  committed
  4bf0441
- bind startup barriers to current CUDA device
  hiworldwzj
  committed
  f806782
- Revert "bind startup barriers to current CUDA device"
  hiworldwzj
  committed
  5c3d7a9
- Revert "Revert "bind startup barriers to current CUDA device""
  hiworldwzj
  committed
  98ad420
- Revert "Revert "Revert "bind startup barriers to current CUDA device"""
  hiworldwzj
  committed
  f25a8ec
- fix flashinfer allreduce subgroup rendezvous
  hiworldwzj
  committed
  5d323c6
Commits on Nov 2, 2026
- refactor attention transfer operators
  hiworldwzj
  committed
  d998618
- refactor: extract attention state interfaces
  hiworldwzj
  committed
  227d8ac
- fix: pass pre-commit checks
  sufubao
  committed
  3debd74
- fix: handle absent draft attention states
  sufubao
  committed
  a3ecc4b
- fix: allow separate INT4 KV storage in prefill
  sufubao
  committed
  9aabee3
- fix: preserve draft role when committing DFlash KV
  sufubao
  committed
  2c4ba35
- fix: evict radix cache before allocating DP state pages
  sufubao
  committed
  0878642
- fix: pass state lengths in CPU cache benchmark
  sufubao
  committed
  245fc55
- fix: rebind HOLD page after resuming KV memory
  sufubao
  committed
  aa3bef2
- docs: move attention resource design into Chinese docs
  sufubao
  committed
  4aa8075

Commits

Commits on Oct 31, 2026

Commits on Nov 1, 2026

Commits on Nov 2, 2026