All pull requests

Pull requests

• fix(bsa): compute blk64 backward O*dO products in FP32

  area:sparse_attentioncat-bugfixorig-external
  #1713·wplf opened Oct 18, 2026Contributor·3 of 4 tasks·

  ·

  · Frontend 1.32.0
  8

• perf(frost): persist common linear-attention forward artifacts

  area:frostarea:linear_attentioncat-perf-bugorig-nv-eng
  #1708·YangXu1990uiuc opened Oct 18, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.32.0
  10

• perf(frost): deduplicate shared-cache compilation across processes

  area:frostcat-perf-bugorig-nv-eng
  #1707·YangXu1990uiuc opened Oct 18, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.32.0
  3

• fix(dsa): remove runtime Torch fallback from MXFP8 scale packing

  area:sparse_attentioncat-bugfixop: DSAorig-external
  #1706·icavan opened Oct 18, 2026Contributor·3 of 4 tasks·

  ·

  · Frontend 1.32.0
  6

• Optimize causal THD forward attention for context parallel workloads

  area:frostarea:global_attentioncat-perf-bugorig-nv-eng
  #1702·YangXu1990uiuc opened Oct 17, 2026Collaborator·

  ·
  1

• perf(kda): optimize SM100 KDA backward for pretraining

  area:linear_attentioncat-enhancementsorig-external
  #1699·icavan opened Oct 17, 2026Contributor·3 of 4 tasks·

  ·

  · Frontend 1.32.0
  6

• Fix terminal dQ barrier reuse in FROST attention backward

  area:frostarea:global_attentioncat-bugfixorig-nv-eng
  #1698·YangXu1990uiuc opened Oct 17, 2026Collaborator·

  ·

  · Frontend 1.32.0
  1

• Optimize THD work lookup for context-parallel attention

  area:frostarea:global_attentioncat-perf-bugorig-nv-eng
  #1696·YangXu1990uiuc opened Oct 17, 2026Collaborator·

  ·

  · Frontend 1.32.0
  5

• Fix cached GDP summary normalization for token-major inputs

  area:linear_attentioncat-bugfixorig-nv-eng
  #1695·dingqingy-nv opened Oct 16, 2026·3 of 4 tasks·

  ·

  · Frontend 1.32.0
  5

• Add fp32 state checkpoint support for Mamba-2 prefill

  area:linear_attentioncat-enhancementsorig-nv-eng
  #1694·jhjpark opened Oct 16, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.32.0
  7

  

• Support batch-pitched D256 attention backward

  area:frostarea:global_attentioncat-enhancementsorig-nv-eng
  #1690·YangXu1990uiuc opened Oct 16, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.32.0
  5

• Relax SM restrictions in TensorIR MemBound Fusion samples

  cat-bugfixmod-frontendorig-nv-eng
  #1687·yeliu-oss opened Oct 16, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.32.0
  4

• frost(sdpa): SM120 half decode tile, whole-group PackGQA and split-KV with attention sinks

  area:frostarea:global_attentioncat-enhancementsmod-frostorig-nv-eng
  #1685·Aneureka opened Oct 16, 2026Member·4 tasks done·

  ·
  3

• fix(dsa): honor the APIBase compile contract on SM100

  area:sparse_attentioncat-bugfixop: DSAorig-external
  #1683·icavan opened Oct 16, 2026Contributor·3 of 4 tasks·

  ·

  · Frontend 1.32.0
  14

• Optimize blk64 BSA with independent Q/KV pairing and split TMA loads

  area:sparse_attentioncat-featuremod-cutedslorig-external
  #1681·wplf opened Oct 16, 2026Contributor·4 tasks done·

  ·

  · Frontend 1.32.0
  6

• Release what CUDA graphs handed back when a plan or graph is destroyed, not during execute

  cat-bugfixmod-frontendorig-nv-eng
  #1676·egilliam-nv opened Oct 15, 2026Collaborator·

  ·

  · Frontend 1.32.0
  7

• Expand deterministic FP8 BatchNorm fusion coverage.

  cat-cicat-enhancementsmod-frontendorig-nv-eng
  #1673·msalasooNV opened Oct 15, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.32.0
  1

  

• test(moe_ep): add EP8 world8 case and fix runtime teardown

  cat-bugfixmod-infraorig-nv-eng
  #1665·zhibinz-nv opened Oct 15, 2026Contributor·3 of 4 tasks·

  ·

• Index the lazy top-level cudnn names

  cat-cimod-infraorig-nv-eng
  #1663·hwanseoc opened Oct 15, 2026Member·

  ·

  · Frontend 1.32.0
  5

• benchmark/dsa: indexer + core sparse attention suite across phases, KV dtypes and paged layouts

  #1656·vedaanta opened Oct 15, 2026Collaborator·

  ·
  1

• docs: op APIs live in cudnn.ops

  cat-docmod-frontendorig-nv-eng
  #1655·hwanseoc opened Oct 14, 2026Member·

  ·

  · Frontend 1.32.0

• Expand deterministic convolution fusion coverage

  cat-enhancementsmod-frontendorig-nv-eng
  #1647·msalasooNV opened Oct 14, 2026Collaborator·3 of 4 tasks·

  ·

  · Frontend 1.32.0
  5

  

• Fix TMEM release state in SM100 MXFP8 backward kernels

  #1630·bastianhagedorn opened Oct 14, 2026·3 of 4 tasks·

  ·
  3

• Optimize context parallel attention backward workspace and vector loads

  area:frostarea:global_attentioncat-enhancementsorig-nv-eng
  #1629·YangXu1990uiuc opened Oct 14, 2026Collaborator·

  ·

  · Frontend 1.32.0
  9

• NSA selection attention: mask keys past the sequence end (non-causal THD)

  area:sparse_attentioncat-bugfixmod-cutedslorig-nv-eng
  #1624·yekerr opened Oct 13, 2026Contributor·3 of 4 tasks·

  ·

  · Frontend 1.32.0
  12