NVIDIA/cudnn-frontend

View on GitHub

cuDNN Frontend is NVIDIA's modern, open-source entry point to the cuDNN library and a growing collection of high-performance open-source kernels.

Is NVIDIA/cudnn-frontend beginner-friendly?

In the last 30 days, NVIDIA/cudnn-frontend merged 232 of 288 pull requests from outside contributors, and its maintainers usually reply within 1 day. 1 beginner-friendly issue is open now.

Stars: 989
Forks: 326
Open beginner issues: 1
Indexed issues: 201
Avg merge: 19h 5m
Merged PRs (30d): 666

Dominant language: Python
License: Apache-2.0
Last GitHub push: Oct 29, 2026
Latest indexed: Sep 19, 2026
Contributing guide: Contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

How to contribute to NVIDIA/cudnn-frontend

Read the contributing guide first: it says how the maintainers want changes proposed, tested and reviewed.
Your contributions will be published under the project's Apache-2.0 license.
Pick the open beginner-friendly issue below and comment that you want to work on it before you start.

Beginner friendly Everything

Open All

Issues somebody may already be working on are listed last. List everything by date

[Bug]: v1.32.1 tag has version 1.33.0
Open
- Is something wrong?
cat-bug

Difficulty 1/5 Under an hour Newbie friendliness 62/100

NVIDIA/cudnn-frontend#2205 · 1 comment · 2 reactions · Opened Nov 2, 2026

Maintainers usually reply within 1 day
[Bug]: SDPA backward (backend engines 5 / 7) gives wrong dQ / dK, NaN dQ or an illegal memory access when O, dO or dQ spans more than 4 GiB (cuDNN 9.27.0.42; GH200 and B300)
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 8/100

NVIDIA/cudnn-frontend#2174 · Opened Nov 1, 2026

Maintainers usually reply within 1 day
SM107 paged D192/V128 prefill: default dispatch lands on backend engine 8, which faults (illegal address) past 2^31-byte Q
Open
- Is something wrong?
area:global_attention cat-bug mod-backend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 8/100

NVIDIA/cudnn-frontend#2150 · Opened Nov 1, 2026

Maintainers usually reply within 1 day
MXFP8 SDPA backward: dK_amax came back NaN on a deterministic rerun (once on GB200 CI)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 25/100

NVIDIA/cudnn-frontend#2122 · 1 comment · Opened Oct 31, 2026

Maintainers usually reply within 1 day
python/*/cutedsl:be-9.27.0-cutedsl-4.8.1:sm100 7 jobs failed at test_bsa_fp8_blk128_dispatch_preserves_v_layout_and_metadata
Open
- Is something wrong?
cat-bug mod-frontend orig-qa

Difficulty 4/5 3-5 days Newbie friendliness 45/100

NVIDIA/cudnn-frontend#2098 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
python/gemm/frost:be-9.27.0-cutedsl-4.8.0:sm100 3 jobs failed at test_sm120_quant_scale_k_padding_is_zeroed
Open
- Is something wrong?
cat-bug mod-frontend orig-qa

Difficulty 3/5 1-2 days Newbie friendliness 45/100

NVIDIA/cudnn-frontend#2097 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
[Bug]: MOE_GROUPED_MATMUL with MXFP8 block scales reads the wrong token-scale rows for every group after the first (B200, cuDNN 9.26.0.51, CUDA 13.4 libraries); same cuDNN under CUDA 13.0 is correct
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 20/100

NVIDIA/cudnn-frontend#2081 · 1 comment · Opened Oct 30, 2026

Maintainers usually reply within 1 day
SDPA bwd THD on SM80: wrong gradients when dO has its own token stride / ragged offset (backend eng7)
Open
- Is something wrong?
area:global_attention

Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA/cudnn-frontend#2051 · Opened Oct 29, 2026

Maintainers usually reply within 1 day
SM80 FROST SDPA rows stage gather/scatter copies on layouts they now serve by default (Rule 2)
Open
- Is something wrong?
area:frost

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/cudnn-frontend#1986 · 1 comment · Opened Oct 28, 2026

Maintainers usually reply within 1 day
[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)
Open
- Is something wrong?
cat-perf-bug mod-backend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/cudnn-frontend#1788 · Opened Oct 22, 2026

Maintainers usually reply within 1 day
SM100 FP8 SDPA forward under-reports Amax_S at short s_q (max over the first 128 KV columns only)
Open
- Is something wrong?
area:global_attention cat-bug mod-backend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 8/100

NVIDIA/cudnn-frontend#1613 · Opened Oct 13, 2026

Maintainers usually reply within 1 day
SM100 SDPA backward d=256: the heuristic's knob set fails NVRTC compilation (hidden by the frontend override)
Open
- Is something wrong?
area:global_attention cat-bug mod-backend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 15/100

NVIDIA/cudnn-frontend#1612 · Opened Oct 13, 2026

Maintainers usually reply within 1 day
SDPA rng_dump at s_q == 1 leaves 75% of the dropout mask unwritten on SM80/SM89/SM90
Open
- Is something wrong?
area:global_attention cat-bug mod-backend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 15/100

NVIDIA/cudnn-frontend#1611 · 1 comment · Opened Oct 13, 2026

Maintainers usually reply within 1 day
SDPA dropout on SM100 cannot go below 1/256 and is biased high for small p
Open
- Is something wrong?
area:global_attention cat-bug mod-backend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 12/100

NVIDIA/cudnn-frontend#1610 · Opened Oct 13, 2026

Maintainers usually reply within 1 day
SDPA: b or h > 65535 passes check_support/build_plans on every plan, then fails at kernel launch (native and FROST)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 46/100

NVIDIA/cudnn-frontend#1573 · 1 comment · Opened Oct 11, 2026

Maintainers usually reply within 1 day
sdpa fwd: execute-time override_strides are accepted but ignored (kernel uses the declared strides; silently wrong, cuDNN 9.27)
Open
- Is something wrong?
area:global_attention cat-bug

Difficulty 4/5 3-5 days Newbie friendliness 45/100

NVIDIA/cudnn-frontend#1570 · Opened Oct 11, 2026

Maintainers usually reply within 1 day
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)
Open
- Is something wrong?
area:global_attention cat-bug

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/cudnn-frontend#1567 · 1 comment · Opened Oct 11, 2026

Maintainers usually reply within 1 day
SDPA: SM107 forward test387 watchdog after graph execution (padded D144/V104)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 12/100

NVIDIA/cudnn-frontend#1538 · 3 comments · Opened Oct 9, 2026

Maintainers usually reply within 1 day
test_mhas_v2 on cc 10.7 outgrew the SM107-test_mhas job's 2400 s timeout (13252 -> 17007 items)
Open
- Is something wrong?
area:frost cat-ci

Difficulty 4/5 1-2 days Newbie friendliness 18/100

NVIDIA/cudnn-frontend#1526 · Opened Oct 9, 2026

Maintainers usually reply within 1 day
[FROST][SM107] BF16 paged sink decode remains 2.6–9x slower than vLLM FlashInfer GEN on identical page16 inputs
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 5/100

NVIDIA/cudnn-frontend#1517 · Opened Oct 9, 2026

Maintainers usually reply within 1 day
[FROST][SM107] Selected BF16 sink SDPA plan is 1.5x slower than native backend for high-batch verification
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA/cudnn-frontend#1472 · Opened Oct 8, 2026

Maintainers usually reply within 1 day
sdpa fwd (frost): nonpaged THD split rule overcounts CTAs (no GQA packing, ragged max length, one-wave cap); 1.4-3x left unselected
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 38/100

NVIDIA/cudnn-frontend#1439 · 6 comments · Opened Oct 7, 2026

Maintainers usually reply within 1 day
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)
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/cudnn-frontend#1435 · 1 comment · Opened Oct 7, 2026

Maintainers usually reply within 1 day
[Bug]: causal SDPA with attn_scale=0.0 returns NaN on the first-ranked backend engines (B200 eng10/eng5, H100 eng9; cuDNN 9.26)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA/cudnn-frontend#1428 · Opened Oct 7, 2026

Maintainers usually reply within 1 day
frost(sdpa): synthesized-padding path (mask-free, S_kv off the KV tile) costs ~70 us host per eager execute and disables split-KV
Open
- Is something wrong?
area:frost cat-perf-bug mod-frost orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 38/100

NVIDIA/cudnn-frontend#1425 · 1 comment · Opened Oct 7, 2026

Maintainers usually reply within 1 day
sdpa(paged): backend FP8 forward reads page-table slots past the live pages (tile-rounded); NaN pages poison O
Open
- Is something wrong?
area:global_attention cat-bug mod-backend orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 42/100

NVIDIA/cudnn-frontend#1416 · Opened Oct 6, 2026

Maintainers usually reply within 1 day
[Bug]: cuDNN 9.27 BF16 SDPA forward nondeterminism on GB10 (COMPOSITE and UNIFIED; fixed-input repro)
Open
- Is something wrong?
cat-bug mod-backend orig-external

Difficulty 5/5 Over a week Newbie friendliness 20/100

NVIDIA/cudnn-frontend#1412 · 3 comments · Opened Oct 6, 2026

Maintainers usually reply within 1 day
frost(sdpa): SM89 / Ada (L20) d64 dense forward row
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA/cudnn-frontend#1137 · Opened Sep 18, 2026

Maintainers usually reply within 1 day
[Feature]: SM100 sparse indexer: stable log-probabilities and unclipped KL backward
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/cudnn-frontend#1136 · Opened Sep 18, 2026

Maintainers usually reply within 1 day
[Bug]: FROST SDPA fwd engine ignores execute-time shape overrides and rejects the graph by its declared shape
Open
- Is something wrong?
cat-bug mod-frost orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 52/100

NVIDIA/cudnn-frontend#1098 · 2 comments · Opened Sep 16, 2026

Maintainers usually reply within 1 day
[Doc]: SDPA execute-time shape override cannot cross s_q == 1; heuristic engine flips between declared max_len 128 and 256
May be free again @0z5a claimed this 45 days ago, and no pull request is open. Open
- Is something wrong?
cat-doc mod-frontend orig-nv-eng

Difficulty 3/5 1-2 days Newbie friendliness 68/100

NVIDIA/cudnn-frontend#1087 · 3 comments · Opened Sep 16, 2026

Maintainers usually reply within 1 day
[Perf]: pygraph.execute host overhead ~30 us above cudnnBackendExecute per call
May be free again @0z5a claimed this 41 days ago, and no pull request is open. Open
- Is something wrong?
cat-perf-bug mod-frontend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 45/100

NVIDIA/cudnn-frontend#1086 · 4 comments · Opened Sep 16, 2026

Maintainers usually reply within 1 day
[Perf]: cuDNN 9.26+ ragged/THD SDPA fwd d192/128 regression (2-CTA configs gated off), +22-33% at s >= 4k on B200
Open
- Is something wrong?
cat-perf-bug mod-backend orig-nv-eng

Difficulty 5/5 Over a week Newbie friendliness 20/100

NVIDIA/cudnn-frontend#1085 · 1 comment · Opened Sep 16, 2026

Maintainers usually reply within 1 day
[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
Open
- Is something wrong?
cat-bug mod-backend orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 45/100

NVIDIA/cudnn-frontend#1084 · Opened Sep 16, 2026

Maintainers usually reply within 1 day
[Bug]: SDPA fwd ragged Stats (LSE) written only for the first head of each kv group when s_q == 1 with GQA (cuDNN 9.26/9.27, all archs)
Open
- Is something wrong?
cat-bug mod-backend orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 45/100

NVIDIA/cudnn-frontend#1083 · Opened Sep 16, 2026

Maintainers usually reply within 1 day
[Bug]: cuDNN 9.26 SDPA backward returns NaN dK/dV for Sq=1 with negative LSE on H200
Open
- Is something wrong?
cat-bug

Difficulty 4/5 3-5 days Newbie friendliness 35/100

NVIDIA/cudnn-frontend#1072 · Opened Sep 15, 2026

Maintainers usually reply within 1 day
[Bug]: cuDNN 9.26 SDPA backward plan creation segfaults when B=Hq=Sq=1
Open
- Is something wrong?
cat-bug

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA/cudnn-frontend#1071 · Opened Sep 15, 2026

Maintainers usually reply within 1 day
[Bug]: cuDNN 9.10 regression: EXECUTION_PLAN runtime compilation adds ~1.5s cold-start for depthwise convolutions
Open
- Is something wrong?
cat-bug

Difficulty 5/5 Over a week Newbie friendliness 30/100

NVIDIA/cudnn-frontend#1066 · Opened Sep 15, 2026

Maintainers usually reply within 1 day
FROST fp8 prefill claims d=256 SDPA graphs the native backend declines, and returns wrong results (qwen3.5 preset)
Open
- Is something wrong?
area:frost area:global_attention cat-bug mod-frost orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 68/100

NVIDIA/cudnn-frontend#978 · 1 comment · Opened Sep 10, 2026

Maintainers usually reply within 1 day
FROST mxfp8 THD prefill: returned amax off by ~5% on fuzz configs (sm100/sm103, opt-in lane)
Open
- Is something wrong?
area:frost area:global_attention cat-bug mod-frost orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 42/100

NVIDIA/cudnn-frontend#977 · Opened Sep 10, 2026

Maintainers usually reply within 1 day
sdpa/suites: paged + ragged (THD Q/O over paged KV) coverage for f16 and fp8 harnesses
Open
- Is something wrong?
area:global_attention cat-enhancements mod-infra orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA/cudnn-frontend#966 · Opened Sep 9, 2026

Maintainers usually reply within 1 day
sdpa/suites: raise bprop sequence caps to 16k (needs KV-blocked fp8 backward reference)
Open
- Is something wrong?
area:global_attention cat-enhancements mod-infra orig-nv-eng

Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA/cudnn-frontend#965 · Opened Sep 9, 2026

Maintainers usually reply within 1 day
[FROST][SDPA] Paged-KV decode entry with FlashInfer/TRTLLM contract (block_tables + device seq_lens, split-KV, graph-safe)
Open
- Is something wrong?
area:global_attention

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA/cudnn-frontend#920 · Opened Sep 4, 2026

Maintainers usually reply within 1 day
frost(sdpa): no SM100 backward engine — sdpa_bwd_sm120 / sdpa_bwd_sm80 only
Open
- Is something wrong?
area:global_attention cat-enhancements

Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA/cudnn-frontend#836 · 2 comments · Opened Sep 1, 2026

Maintainers usually reply within 1 day
test: general indexer + top-k coverage + specialized DSv4/QSA/GLM shape suites
Open
- Is something wrong?
area:sparse_attention

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA/cudnn-frontend#832 · Opened Aug 31, 2026

Maintainers usually reply within 1 day
benchmark: GLM-5.x indexer + top-k (5.2 IndexShare, 5.3-Flash pooled indexer)
Open
- Is something wrong?
area:sparse_attention

Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA/cudnn-frontend#831 · Opened Aug 31, 2026

Maintainers usually reply within 1 day
benchmark: Qwen3.8-Flash-Next QSA indexer + top-k
Open
- Is something wrong?
area:sparse_attention

Difficulty 4/5 3-5 days Newbie friendliness 52/100

NVIDIA/cudnn-frontend#830 · Opened Aug 31, 2026

Maintainers usually reply within 1 day
benchmark: DeepSeek-V4 CSA indexer + top-k
Open
- Is something wrong?
area:sparse_attention

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/cudnn-frontend#829 · Opened Aug 31, 2026

Maintainers usually reply within 1 day
test: general sparse_attention forward coverage + specialized DSv4/QSA/GLM shape suites
Open
- Is something wrong?
area:sparse_attention

Difficulty 4/5 3-5 days Newbie friendliness 38/100

NVIDIA/cudnn-frontend#828 · Opened Aug 31, 2026

Maintainers usually reply within 1 day
benchmark: GLM-5.x (5.2 + 5.3-Flash) coverage for sparse_attention forward
Open
- Is something wrong?
area:sparse_attention

Difficulty 4/5 3-5 days Newbie friendliness 52/100

NVIDIA/cudnn-frontend#827 · Opened Aug 31, 2026

Maintainers usually reply within 1 day

Showing the newest 100

This page lists what was indexed most recently. Advanced filter has the whole inventory, narrowed by language, difficulty and how long a task takes.

Open advanced filter

NVIDIA/cudnn-frontend

Is NVIDIA/cudnn-frontend beginner-friendly?

How to contribute to NVIDIA/cudnn-frontend

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Showing the newest 100

Get new issues in your inbox