Skip to content

Init v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@icavan icavan released this 12 Aug 09:48
v0.2.0
4cc51c5
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's Changed

  • Adjust README by @icavan in #6
  • Add Slack link and modify readme contact by @KevinZeng08 in #7
  • Prepare for pypi installation by @icavan in #22
  • [KDA] Optimize recompute_wu kernel and support disable_recompute=True by @KevinZeng08 in #27
  • [KDA] Change to transposed VK state layout for Hopper fused forward by @KevinZeng08 in #33
  • Add PR template by @KevinZeng08 in #39
  • doc: update README by @icavan in #40
  • support bf16 beta for KDA (SM90 & SM10X) by @cherhh in #34
  • [KDA] Support CuTeDSL KDA Decode by @yuan-luo in #37
  • [KDA] Refactor csrc kernel selection, add tests and sanity check by @KevinZeng08 in #42
  • [Doc] Update Kimi CP link by @KevinZeng08 in #45
  • docs: Update method for joining WeChat group by @zheyang0825 in #46
  • doc: update README with bigger logo by @icavan in #53
  • [KDA] Adapt recompute_wu and delta_h for backward by @KevinZeng08 in #54
  • perf(kda_decode): optimize GMEM coalescing, V-tile parallelism by @icavan in #52
  • fix(linter): fix linter & updates kda decode perf by @icavan in #60
  • feat: BHVK (K-last) state layout for Lightning Attention prefill & decode by @icavan in #56
  • perf(la_decode): boost small-batch (B<=32) by 33% via 8-warp CTA by @icavan in #59
  • [KDA] Optimize recompute_wu with better register allocation by @KevinZeng08 in #61
  • fix output_final_state wrapper issue by @yechenzhi in #63
  • [KDA] fix internal output_final_state wrapper issue in SM90 by @yechenzhi in #66
  • [Fix] change to umma pipelines for SM100 KDA by @KevinZeng08 in #68
  • [Fix] delta_h race risk by @KevinZeng08 in #69
  • [KDA] sm90 GVA enhance by @sjmshsh in #64
  • [Feat] upgrade FLA to v0.5.0 by @KevinZeng08 in #72
  • [Fix] add cross-proxy fence for recomp_wu kernel by @KevinZeng08 in #77
  • [KDA] sm100 GVA enhance by @sjmshsh in #65
  • fix: wire h0_indices into Lightning Attention decode for state-pool indexing by @Emre-Dinc in #75
  • Unify the naming convention by @icavan in #79
  • Consolidate duplicated benchmark helper logic into benchmarks/utils.py, including shared timing, reporting, and accuracy utilities. by @icavanyu in #80
  • [KDA] support GVA for SM100 end-to-end by @KevinZeng08 in #73
  • [KDA] add backward chunk_wy_dqkg kernel for SM10X by @KevinZeng08 in #74
  • [KDA] split unit tests into fast/slow modes by @Longxmas in #84
  • [KDA] Add intra-card CP for chunk_delta_h forward in SM100 by @cherhh in #70
  • Relax dependency minimum versions by @icavan in #87
  • [CI] Add GitHub workflow for building and releasing fat wheels by @tongke6 in #91
  • [ci] add prek lint workflow by @tongke6 in #98
  • ci: align pytorch version requirements in wheel build by @tongke6 in #99
  • refactor(kda): reorganize KDA backends into arch-first layout and add lazy imports by @cherhh in #100
  • feat: intracard cp for sm90 by @Hyaloid in #86
  • [KDA] KDA MTP decode: recurrent + KVBuffer chunkwise verify + flush by @Longxmas in #96
  • [LA] Lightning Attention MTP decode + KVBuffer parallel verify / commit by @fkuner in #97
  • feat: add packed kda decode by @zhouaihui in #102
  • [Feature] Add GVA support for Lightning by @sjmshsh in #85
  • [Fix] Wire cudac bindings by @cherhh in #105
  • feat(kda): SM90 FlashKDA prefill, intracard CP, and backend dispatch by @cherhh in #106
  • fix(kda): add the missing cross-proxy fence before TMA stores by @cherhh in #114
  • [Lightning Attention] Add Hopper SM90 CuTe DSL prefill by @Aharrypotter in #111
  • [KDA] Fuse causal conv1d into MTP verify decode by @Longxmas in #107
  • Detect tcgen05 load/store API by signature by @icavan in #119
  • Remove obsolete SM100 fully fused KDA prototype by @icavan in #116
  • Update SM90 FlashKDA benchmark results by @icavan in #120
  • feat: isolate cutlass._mlir imports behind compat gateway (#118) by @bikrammajhi in #121
  • perf(kda): preserve SM90 workspace byte layout with bulk copies by @fkuner in #115

New Contributors

Full Changelog: v0.1.0...v0.2.0

What's Changed

  • Adjust README by @icavan in #6
  • Add Slack link and modify readme contact by @KevinZeng08 in #7
  • Prepare for pypi installation by @icavan in #22
  • [KDA] Optimize recompute_wu kernel and support disable_recompute=True by @KevinZeng08 in #27
  • [KDA] Change to transposed VK state layout for Hopper fused forward by @KevinZeng08 in #33
  • Add PR template by @KevinZeng08 in #39
  • doc: update README by @icavan in #40
  • support bf16 beta for KDA (SM90 & SM10X) by @cherhh in #34
  • [KDA] Support CuTeDSL KDA Decode by @yuan-luo in #37
  • [KDA] Refactor csrc kernel selection, add tests and sanity check by @KevinZeng08 in #42
  • [Doc] Update Kimi CP link by @KevinZeng08 in #45
  • docs: Update method for joining WeChat group by @zheyang0825 in #46
  • doc: update README with bigger logo by @icavan in #53
  • [KDA] Adapt recompute_wu and delta_h for backward by @KevinZeng08 in #54
  • perf(kda_decode): optimize GMEM coalescing, V-tile parallelism by @icavan in #52
  • fix(linter): fix linter & updates kda decode perf by @icavan in #60
  • feat: BHVK (K-last) state layout for Lightning Attention prefill & decode by @icavan in #56
  • perf(la_decode): boost small-batch (B<=32) by 33% via 8-warp CTA by @icavan in #59
  • [KDA] Optimize recompute_wu with better register allocation by @KevinZeng08 in #61
  • fix output_final_state wrapper issue by @yechenzhi in #63
  • [KDA] fix internal output_final_state wrapper issue in SM90 by @yechenzhi in #66
  • [Fix] change to umma pipelines for SM100 KDA by @KevinZeng08 in #68
  • [Fix] delta_h race risk by @KevinZeng08 in #69
  • [KDA] sm90 GVA enhance by @sjmshsh in #64
  • [Feat] upgrade FLA to v0.5.0 by @KevinZeng08 in #72
  • [Fix] add cross-proxy fence for recomp_wu kernel by @KevinZeng08 in #77
  • [KDA] sm100 GVA enhance by @sjmshsh in #65
  • fix: wire h0_indices into Lightning Attention decode for state-pool indexing by @Emre-Dinc in #75
  • Unify the naming convention by @icavan in #79
  • Consolidate duplicated benchmark helper logic into benchmarks/utils.py, including shared timing, reporting, and accuracy utilities. by @icavanyu in #80
  • [KDA] support GVA for SM100 end-to-end by @KevinZeng08 in #73
  • [KDA] add backward chunk_wy_dqkg kernel for SM10X by @KevinZeng08 in #74
  • [KDA] split unit tests into fast/slow modes by @Longxmas in #84
  • [KDA] Add intra-card CP for chunk_delta_h forward in SM100 by @cherhh in #70
  • Relax dependency minimum versions by @icavan in #87
  • [CI] Add GitHub workflow for building and releasing fat wheels by @tongke6 in #91
  • [ci] add prek lint workflow by @tongke6 in #98
  • ci: align pytorch version requirements in wheel build by @tongke6 in #99
  • refactor(kda): reorganize KDA backends into arch-first layout and add lazy imports by @cherhh in #100
  • feat: intracard cp for sm90 by @Hyaloid in #86
  • [KDA] KDA MTP decode: recurrent + KVBuffer chunkwise verify + flush by @Longxmas in #96
  • [LA] Lightning Attention MTP decode + KVBuffer parallel verify / commit by @fkuner in #97
  • feat: add packed kda decode by @zhouaihui in #102
  • [Feature] Add GVA support for Lightning by @sjmshsh in #85
  • [Fix] Wire cudac bindings by @cherhh in #105
  • feat(kda): SM90 FlashKDA prefill, intracard CP, and backend dispatch by @cherhh in #106
  • fix(kda): add the missing cross-proxy fence before TMA stores by @cherhh in #114
  • [Lightning Attention] Add Hopper SM90 CuTe DSL prefill by @Aharrypotter in #111
  • [KDA] Fuse causal conv1d into MTP verify decode by @Longxmas in #107
  • Detect tcgen05 load/store API by signature by @icavan in #119
  • Remove obsolete SM100 fully fused KDA prototype by @icavan in #116
  • Update SM90 FlashKDA benchmark results by @icavan in #120
  • feat: isolate cutlass._mlir imports behind compat gateway (#118) by @bikrammajhi in #121
  • perf(kda): preserve SM90 workspace byte layout with bulk copies by @fkuner in #115

New Contributors

Full Changelog: v0.1.0...v0.2.0

Contributors

yuan-luo, yechenzhi, and 14 other contributors
Assets 6
Loading