Repository navigation

perf(cuda): widen the attention query tile on sm_89 and sm_120 - #202

Open

sysuTobo wants to merge 6 commits into

InferMatrixLab:mainfrom

sysuTobo:cuda/attention-wide-tile

Open

perf(cuda): widen the attention query tile on sm_89 and sm_120#202
sysuTobo wants to merge 6 commits into
InferMatrixLab:mainfrom
sysuTobo:cuda/attention-wide-tile

Commits on Nov 1, 2026
- cuda: widen the attention query tile on sm_89 and sm_120
  sysuTobo
  committed
  a3bb8c4
- cuda: keep only the attention record, not the measurement programs
  sysuTobo
  committed
  7e5b124
- cuda: keep the attention record to the frozen protocol and its results
  sysuTobo
  committed
  18049cb
- cuda: trim the attention change and its record
  sysuTobo
  committed
  4e69f6c
- cuda: fix the attention record after review
  sysuTobo
  committed
  e9aeb07
- cuda: key the wide-tile decision to the current device
  sysuTobo
  committed
  2bbf255