All pull requests

Pull requests

• test: remove two unused SDPA helper modules, a test that can only skip, and unreferenced helper code

  area:frostarea:global_attentionarea:linear_attentionarea:sparse_attentioncat-cleanuporig-nv-eng
  #2545·RomanAnders90 opened Nov 13, 2026Contributor·5 tasks done·

  ·

  · Frontend 1.33.0
  9

  

• test: bring twelve host-only tests into the default L0 run

  area:frostarea:gemmcat-ciorig-nv-eng
  #2544·RomanAnders90 opened Nov 13, 2026Contributor·5 tasks done·

  ·

  · Frontend 1.33.0
  7

  

• Enable SiTU-GLU in canonical JAX grouped GEMM

  #2540·pggPL opened Nov 12, 2026·2 of 5 tasks·

  ·
  1

• la(gdn): SM90 CUDA engine for Gated DeltaNet (gdn_hopper_cuda)

  #2539·Connor-XY opened Nov 12, 2026·4 of 5 tasks·

  ·
  1

• Add native JAX bindings for GDN, GDN2, GDP, and Mamba2

  area:frostarea:linear_attentioncat-featuremod-cutedslorig-nv-eng
  #2537·mgoldfarb-nvidia opened Nov 12, 2026Collaborator·5 tasks done·

  ·
  1

• refactor(jax): remove cudnn.jax; move the JAX bridge helpers to cudnn._jax

  area:frostarea:gemmarea:global_attentionarea:linear_attentioncat-cleanuporig-nv-eng
  #2536·hwanseoc opened Nov 12, 2026Member·

  ·

  · Frontend 1.33.0

• frost(sdpa bwd sm90): deterministic key-owner kernel with an ordered dQ relay for dense d64 / d128

  area:frostarea:global_attentioncat-featuremod-cutedslmod-frostorig-nv-eng
  #2535·yanzhuo607 opened Nov 12, 2026Collaborator·

  ·

  · Frontend 1.33.0
  53

• Zhibinz/feature/moe ep sync 72796e96

  area:gemmcat-featurecat-routine-updatemod-cutedslorig-nv-eng
  #2523·zhibinz-nv opened Nov 12, 2026Contributor·5 tasks·

  ·

  · Frontend 1.33.0
  18

• perf(bsa): SM90 blk64 forward/backward optimizations, prepare_backward API and determinism suite

  area:sparse_attentioncat-featurecat-perf-bugorig-external
  #2511·shangxiaokang opened Nov 12, 2026·

  ·

  · Frontend 1.33.0
  20

• Fix SM90 BT64 KDA gate-gradient precision in local backward

  area:linear_attentioncat-bugorig-external
  #2510·icavan opened Nov 12, 2026Contributor·4 of 5 tasks·

  ·

  · Frontend 1.33.0
  5

• aot: rename to libcudnn_frontend_aot_attention; SM120 decode tile, sm_120f, quantized decode on Ampere/Ada, no abort

  area:frostarea:global_attentionarea:packagingcat-featureorig-nv-eng
  #2487·YangXu1990uiuc opened Nov 12, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.33.0
  1

• feat(sdpa): add native SM89 half backward through D256

  area:frostarea:global_attentioncat-featureorig-external
  #2485·0z5a opened Nov 11, 2026Contributor·

  ·

  · Frontend 1.33.0
  12

• Fix GEMM packing of NVFP4 grouped RHT column scales

  #2484·zhongbozhu opened Nov 11, 2026·5 tasks·

  ·
  1

• feat(sdpa bwd): serve SM100/SM103 half D256 backward for dense D_V in {192, 256} head pairs

  area:frostarea:global_attentioncat-featureorig-external
  #2483·0z5a opened Nov 11, 2026Contributor·

  ·

  · Frontend 1.33.0
  5

• Fix SM90 KDA backward precision by keeping Qg^T dO in FP32

  area:linear_attentioncat-bugfixorig-external
  #2482·icavan opened Nov 11, 2026Contributor·4 of 5 tasks·

  ·

  · Frontend 1.33.0
  17

• fix(kda): correct SM90 CP state-probe precision with FP32 summaries

  area:linear_attentioncat-featuremod-frontendorig-external
  #2473·icavan opened Nov 11, 2026Contributor·4 of 5 tasks·

  ·

  · Frontend 1.33.0
  7

• Add canonical GLU layouts to existing grouped GEMM kernels

  area:gemmcat-featureorig-nv-eng
  #2455·zhongbozhu opened Nov 11, 2026·5 tasks·

  ·

  · Frontend 1.33.0
  18

• Linear attention: eager KDA/GDN skip the torch custom-op dispatch

  area:linear_attentioncat-enhancementsorig-nv-eng
  #2450·hwanseoc opened Nov 10, 2026Member·

  ·
  1

• Add Cake GQA, sparse attention, and DSA indexer training APIs

  area:global_attentionarea:sparse_attentioncat-featureorig-nv-eng
  #2435·Anerudhan opened Nov 10, 2026Collaborator·

  ·

  · Frontend 1.33.0
  11

• feat(sdpa): SM90 (Hopper) FROST SDPA backward (sdpa_bwd_sm90)

  area:frostarea:global_attentioncat-featureorig-external
  #2427·0z5a opened Nov 10, 2026Contributor·4 of 5 tasks·

  ·

  · Frontend 1.33.0
  21

• sdpa(frost d512 role-split, sm107): join every reader before the TMEM dealloc credits

  #2405·bastianhagedorn opened Nov 10, 2026Contributor·3 of 5 tasks·

  ·
  1

• gab(qsa_topk_decode): acquire the cluster exchange barriers at cluster scope

  area:frostarea:global_attentioncat-bugfixmod-cutedslorig-nv-eng
  #2404·bastianhagedorn opened Nov 10, 2026Contributor·3 of 5 tasks·

  ·

  · Frontend 1.33.0
  4

• fix: continue graph planning after explicit backend declines

  area:frostarea:global_attentioncat-bugorig-nv-eng
  #2371·YangXu1990uiuc opened Nov 9, 2026Collaborator·5 tasks done·

  ·

  · Frontend 1.33.0
  6

• [Grouped GEMM] Hadamard-fused SReLU: round through C dtype, add tanh_clamp_scale, match consumer NVFP4 amax

  area:gemmcat-bugfixcat-featuremod-cutedslorig-nv-eng
  #2366·xrennvidia opened Nov 9, 2026·2 of 5 tasks·

  ·

  · Frontend 1.33.0
  9

• feat(frost-gemm): compose indirect row operands with grouped fusion

  area:gemmcat-featuremod-frostorig-external
  #2305·meinie0826 opened Nov 6, 2026Contributor·

  ·

  · Frontend 1.33.0
  35