Skip to content

v0.3.0

Latest

Choose a tag to compare

@icavan icavan released this 17 Nov 04:03
a09d3f7

What's Changed

  • feat(kda): add bitwise-aligned SM100 CuTeDSL forward by @icavan in #123
  • feat(kda): enable FlashKDA on SM100 by @icavan in #124
  • [KDA] Add SM90 fused WY-DqKG backward by @zheyang0825 in #109
  • docs: remove Contact section from README by @icavan in #125
  • fix(kda): preserve SM90 CP transitions in FP32 by @icavan in #127
  • chore: configure CodeRabbit reviews by @icavan in #129
  • tests: bound default suite runtime to five minutes by @icavan in #128
  • Fix SM90 WY input buffer reuse before WGMMA completion by @zheyang0825 in #130
  • refactor(kda): replace SM100 native forward with CuTeDSL by @icavan in #131
  • fix(cutedsl): close synchronization races by @icavan in #132
  • refactor(kda): remove native CUDA extension by @icavan in #133
  • refactor(kernels)!: make CuTeDSL runtime standalone by @icavan in #135
  • feat(kda): add cuDNN concatenated-summary Kimi CP by @icavan in #136
  • feat(distributed): add packed hybrid runtime and THD planner by @icavan in #137
  • fix(kda): improve SM90 inverse precision by @zheyang0825 in #134
  • fix(kda): guard recompute beta tail loads by @icavan in #139
  • feat(distributed): add calibrated THD selection routing by @icavan in #140

Full Changelog: v0.2.0...v0.3.0

What's Changed

  • feat(kda): add bitwise-aligned SM100 CuTeDSL forward by @icavan in #123
  • feat(kda): enable FlashKDA on SM100 by @icavan in #124
  • [KDA] Add SM90 fused WY-DqKG backward by @zheyang0825 in #109
  • docs: remove Contact section from README by @icavan in #125
  • fix(kda): preserve SM90 CP transitions in FP32 by @icavan in #127
  • chore: configure CodeRabbit reviews by @icavan in #129
  • tests: bound default suite runtime to five minutes by @icavan in #128
  • Fix SM90 WY input buffer reuse before WGMMA completion by @zheyang0825 in #130
  • refactor(kda): replace SM100 native forward with CuTeDSL by @icavan in #131
  • fix(cutedsl): close synchronization races by @icavan in #132
  • refactor(kda): remove native CUDA extension by @icavan in #133
  • refactor(kernels)!: make CuTeDSL runtime standalone by @icavan in #135
  • feat(kda): add cuDNN concatenated-summary Kimi CP by @icavan in #136
  • feat(distributed): add packed hybrid runtime and THD planner by @icavan in #137
  • fix(kda): improve SM90 inverse precision by @zheyang0825 in #134
  • fix(kda): guard recompute beta tail loads by @icavan in #139
  • feat(distributed): add calibrated THD selection routing by @icavan in #140

Full Changelog: v0.2.0...v0.3.0