Issues

• sdpa fwd (frost): nonpaged THD split rule overcounts CTAs (no GQA packing, ragged max length, one-wave cap); 1.4-3x left unselected

  NVIDIA/cudnn-frontend#1439 · Filter by author YangXu1990uiuc opened on Oct 7, 2026

• frost(sdpa): forward kernels fold attn_scale into exp2 after an unscaled -inf-masked max, so attn_scale <= 0 with any mask gives NaN (SM80/SM100/SM107/SM120)

  NVIDIA/cudnn-frontend#1435 · Filter by author YangXu1990uiuc opened on Oct 7, 2026

• [Bug]: causal SDPA with attn_scale=0.0 returns NaN on the first-ranked backend engines (B200 eng10/eng5, H100 eng9; cuDNN 9.26)

  NVIDIA/cudnn-frontend#1428 · Filter by author YangXu1990uiuc opened on Oct 7, 2026

• frost(sdpa): synthesized-padding path (mask-free, S_kv off the KV tile) costs ~70 us host per eager execute and disables split-KV

  area:frostcat-perf-bugmod-frostorig-nv-eng
  NVIDIA/cudnn-frontend#1425 · Filter by author YangXu1990uiuc opened on Oct 7, 2026

• sdpa(paged): backend FP8 forward reads page-table slots past the live pages (tile-rounded); NaN pages poison O

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1416 · Filter by author YangXu1990uiuc opened on Oct 6, 2026

• frost(sdpa): FP8 forward rows decline graphs that declare Amax_S (parity with backend)

  area:frostcat-enhancementsmod-frostorig-nv-eng
  NVIDIA/cudnn-frontend#1415 · Filter by author YangXu1990uiuc opened on Oct 6, 2026

• [Bug]: cuDNN 9.27 BF16 SDPA forward nondeterminism on GB10 (COMPOSITE and UNIFIED; fixed-input repro)

  cat-bugmod-backendorig-external
  NVIDIA/cudnn-frontend#1412 · Filter by author Jaeyeong-CHOI opened on Oct 6, 2026 · Frontend 1.31.0
  1

• [CI] python/sdpa/frost:cutlass-4.8:sm100: test/python/sdpa/frost/test_sdpa_bwd_config_sm107.py::test_rejects_every_shared_cfg_predicate[SEQ_KV_LENS_PRESENT-f16]

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1408 · Filter by author liujane-dev opened on Oct 6, 2026
  1

• [CI] python/*/cutedsl:cutlass-4.8:sm100 and python:dev:sm100 FE regression at 85b669dc affects cutedsl BSA/GAB and FP4 matmul jobs

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1404 · Filter by author liujane-dev opened on Oct 6, 2026
  1

• [CI] python:dev:sm80: test_mha_gat.py::test_gnn_mha_gat_uses_current_stream Mismatched

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1297 · Filter by author liujane-dev opened on Sep 30, 2026

• [Feature]: Integrate Mamba2 CuTeDSL implementation to cudnnFE

  area:linear_attentioncat-featuremod-frostorig-nv-eng
  Enhancement

  NVIDIA/cudnn-frontend#1292 · Filter by author ghadiaravi13 opened on Sep 29, 2026 · Frontend 1.31.0
  

• [CI] cutedsl:cutlass-4.8:sm107 test_gemm_swiglu_wrapper numerical mismatch in cluster_shape_mn1-mma_tiler_mn1-c_dtype0-acc_dtype0-ab12_dtype1-ab_dtype2-n-n-m

  cat-bugmod-cutedslorig-qa
  Bug

  NVIDIA/cudnn-frontend#1281 · Filter by author liujane-dev opened on Sep 29, 2026
  2

• [Feature]: JAX API for Rubin-specific grouped GEMM fusion

  cat-feature
  Task

  NVIDIA/cudnn-frontend#1275 · Filter by author jberchtold-nvidia opened on Sep 28, 2026 · Frontend 1.31.0

• [CI] oss:cutlass-4.8:sm107 gated-attention FROST GEMM plans missing after #1177

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1206 · Filter by author liujane-dev opened on Sep 23, 2026

• [CI] qa_cpp_samples:9.24.0.43_13.3.0:sm100: fp16_fwd_with_block_mask.cpp:189 FAILED

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1175 · Filter by author liujane-dev opened on Sep 22, 2026
  

• [bug][sdpa][sm120] FROST SM120 per-tensor FP8 forward returns NaN rows on dense (skv_tile=0) shapes

  NVIDIA/cudnn-frontend#1156 · Filter by author 0z5a opened on Sep 20, 2026
  2

• frost(sdpa): SM89 / Ada (L20) d64 dense forward row

  NVIDIA/cudnn-frontend#1137 · Filter by author 0z5a opened on Sep 18, 2026

• [Feature]: SM100 sparse indexer: stable log-probabilities and unclipped KL backward

  NVIDIA/cudnn-frontend#1136 · Filter by author C-TC opened on Sep 18, 2026

• [Bug]: qa_py_test:9.24.0.43_13.3.0:sm90 6 jobs failed with Mismatched elements

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1130 · Filter by author liujane-dev opened on Sep 18, 2026
  1

  

• [Bug]: FROST SDPA fwd engine ignores execute-time shape overrides and rejects the graph by its declared shape

  cat-bugmod-frostorig-nv-eng
  NVIDIA/cudnn-frontend#1098 · Filter by author YangXu1990uiuc opened on Sep 16, 2026
  2

• Optimize SM100 blk64 BSA backward single-group dK/dV epilogue

  NVIDIA/cudnn-frontend#1089 · Filter by author wplf opened on Sep 16, 2026

• [Doc]: SDPA execute-time shape override cannot cross s_q == 1; heuristic engine flips between declared max_len 128 and 256

  cat-docmod-frontendorig-nv-eng
  NVIDIA/cudnn-frontend#1087 · Filter by author YangXu1990uiuc opened on Sep 16, 2026
  2

• [Perf]: pygraph.execute host overhead ~30 us above cudnnBackendExecute per call

  cat-perf-bugmod-frontendorig-nv-eng
  NVIDIA/cudnn-frontend#1086 · Filter by author YangXu1990uiuc opened on Sep 16, 2026
  4

• [Perf]: cuDNN 9.26+ ragged/THD SDPA fwd d192/128 regression (2-CTA configs gated off), +22-33% at s >= 4k on B200

  cat-perf-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1085 · Filter by author YangXu1990uiuc opened on Sep 16, 2026

• [Bug]: SDPA fwd decode class (s_q == 1): first engine config fails with CUDNN_STATUS_INTERNAL_ERROR_COMPILATION_FAILED when s_kv is not a tile multiple

  cat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1084 · Filter by author YangXu1990uiuc opened on Sep 16, 2026