Issues

• python/sdpa/frost:cutlass-4.8:sm100 - 5 SDPA FROST jobs: Capabilities append-order pin broken

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1915 · Filter by author liujane-dev opened on Oct 27, 2026

• python/*/jax:cutlass-4.8:sm100 2 JAX KDA jobs: piece-chain floor breaks 256-token case

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1865 · Filter by author liujane-dev opened on Oct 26, 2026

• python:dev:sm100 + 7 jobs: e0a12f44 stale tier IDs in test_tiers.py::test_every_listed_cell_collects[smoke_cc100.txt]

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1864 · Filter by author liujane-dev opened on Oct 26, 2026

• python/*/cutedsl:cutlass-4.8:sm100 - cutlass sm100 CuTeDSL failures after c118114f

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1845 · Filter by author liujane-dev opened on Oct 25, 2026

• python/sdpa/frost:cutlass-4.8:sm100: SM103 D192 F16 causal SASS spill pin regressed

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1835 · Filter by author liujane-dev opened on Oct 24, 2026

• deploy:build_bdist_wheels:win:amd64 6 jobs: Windows bdist wheels fail when dynamic-loading build requires CUDA 13 headers

  cat-bugorig-qa
  Bug

  NVIDIA/cudnn-frontend#1833 · Filter by author liujane-dev opened on Oct 24, 2026

• cudnn.fla KDA: route to the Hopper engine on SM90 (and its backward scale)

  NVIDIA/cudnn-frontend#1818 · Filter by author 0z5a opened on Oct 23, 2026

• FROST norm / RoPE / Engram: serve SM90 (exact (10, 0) gates refuse Hopper)

  NVIDIA/cudnn-frontend#1817 · Filter by author 0z5a opened on Oct 23, 2026

• cpp_samples:rel:sm90 cudagraph_topology G08 device memory delta assertion fails

  cat-bugmod-frontendorig-qa
  Bug

  NVIDIA/cudnn-frontend#1797 · Filter by author liujane-dev opened on Oct 23, 2026
  1

• [Perf]: backend FP8 d128 SDPA fwd on Rubin, bottom-right causal with short s_q does not scale with heads/KV split (up to 12x slower than the mask-free graph)

  cat-perf-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1788 · Filter by author YangXu1990uiuc opened on Oct 22, 2026

• [Bug]: KDA kernels return NaN when a key channel decays faster than ~5.5 nats/token (the #1644 GDN-2 failure, unfixed for KDA)

  area:linear_attentioncat-enhancementsorig-external
  Enhancement

  NVIDIA/cudnn-frontend#1773 · Filter by author andreykarailiev opened on Oct 21, 2026 · Frontend 1.33.0
  2

• SM100 FP8 SDPA forward under-reports Amax_S at short s_q (max over the first 128 KV columns only)

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1613 · Filter by author YangXu1990uiuc opened on Oct 13, 2026

• SM100 SDPA backward d=256: the heuristic's knob set fails NVRTC compilation (hidden by the frontend override)

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1612 · Filter by author YangXu1990uiuc opened on Oct 13, 2026

• SDPA rng_dump at s_q == 1 leaves 75% of the dropout mask unwritten on SM80/SM89/SM90

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1611 · Filter by author YangXu1990uiuc opened on Oct 13, 2026
  1

• SDPA dropout on SM100 cannot go below 1/256 and is biased high for small p

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1610 · Filter by author YangXu1990uiuc opened on Oct 13, 2026

• [Docs] cuDNN latest document-check found high-priority findings

  cat-docmod-frontendorig-qa
  Bug

  NVIDIA/cudnn-frontend#1576 · Filter by author liujane-dev opened on Oct 12, 2026
  1

  

• SDPA: b or h > 65535 passes check_support/build_plans on every plan, then fails at kernel launch (native and FROST)

  NVIDIA/cudnn-frontend#1573 · Filter by author YangXu1990uiuc opened on Oct 11, 2026
  1

• sdpa fwd: execute-time override_strides are accepted but ignored (kernel uses the declared strides; silently wrong, cuDNN 9.27)

  area:global_attentioncat-bug
  NVIDIA/cudnn-frontend#1570 · Filter by author YangXu1990uiuc opened on Oct 11, 2026

• sdpa fwd: execute-time s_kv override off the KV tile attends to the tile's padding rows (silently wrong; first-ranked plans on H100/B200, cuDNN 9.27)

  area:global_attentioncat-bug
  NVIDIA/cudnn-frontend#1567 · Filter by author YangXu1990uiuc opened on Oct 11, 2026
  1

• SDPA: SM107 forward test387 watchdog after graph execution (padded D144/V104)

  NVIDIA/cudnn-frontend#1538 · Filter by author YangXu1990uiuc opened on Oct 9, 2026
  3

• test_mhas_v2 on cc 10.7 outgrew the SM107-test_mhas job's 2400 s timeout (13252 -> 17007 items)

  area:frostcat-ci
  NVIDIA/cudnn-frontend#1526 · Filter by author YangXu1990uiuc opened on Oct 9, 2026 · Frontend 1.33.0

• [FROST][SM107] BF16 paged sink decode remains 2.6–9x slower than vLLM FlashInfer GEN on identical page16 inputs

  NVIDIA/cudnn-frontend#1517 · Filter by author vedaanta opened on Oct 9, 2026

• [FROST][SM107] Selected BF16 sink SDPA plan is 1.5x slower than native backend for high-batch verification

  NVIDIA/cudnn-frontend#1472 · Filter by author vedaanta opened on Oct 8, 2026

• 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
  6

• 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
  1