Skip to content

Pull requests: NVIDIA/cudnn-frontend

New pull request New
74 Open 753 Closed
74 Open 753 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

Add native Mamba-2 SSD forward and backward graph engine area:frost area:linear_attention cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1133 opened Sep 18, 2026 by YangXu1990uiuc Collaborator Loading…
3
pygraph: native one-crossing overrides; FROST SDPA f16 SM100/SM107 dense through the prepared positional entry (#1086, #1098) area:frost area:global_attention cat-perf-bug Performance regressions or cases where behavior is correct but too slow. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1132 opened Sep 18, 2026 by YangXu1990uiuc Collaborator Loading…
4 tasks done
Frontend 1.30.0
10
frost(sdpa): Rubin fp8/mxfp8 prefill +5..17 % -- Amax_O fold via max.f32 (FMNMX3) and a CTA-scope scheduler credit arrive area:frost cat-perf-bug Performance regressions or cases where behavior is correct but too slow. mod-cutedsl CuTeDSL kernels, generated kernels, examples, or related integration work. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1129 opened Sep 18, 2026 by RomanAnders90 Contributor • Draft
1
Norm torch custom ops orig-nv-eng Reported or requested by NVIDIA engineering.
#1128 opened Sep 17, 2026 by rmhaskarnvidia Collaborator Loading…
4 tasks done
Frontend 1.30.0
3
fix: acquire SM100 grouped-wgrad tensor maps before TMA use cat-bugfix mod-cutedsl CuTeDSL kernels, generated kernels, examples, or related integration work. orig-nv-eng Reported or requested by NVIDIA engineering.
#1126 opened Sep 17, 2026 by JF-D Loading…
3 of 4 tasks
Frontend 1.30.0
@JF-D
3
CI heartbeat cat-ci CI failures, test flakiness, workflow breakage, or automation issues. mod-infra Infrastructure, CI/CD, build systems, packaging, releases, or repo maintenance. orig-nv-eng Reported or requested by NVIDIA engineering.
#1125 opened Sep 17, 2026 by hwanseoc Member • Draft
2
[Draft] frost(sdpa): prevent causal leakage in D256 forward and backward area:global_attention
#1123 opened Sep 17, 2026 by elfiegg Contributor Loading…
3 of 4 tasks
Frontend 1.30.0
@elfiegg
6
[DSv4.1] Fuse native-layout vision RoPE backward and QKV gradient packing area:frost cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1122 opened Sep 17, 2026 by YangXu1990uiuc Collaborator • Draft
1
[DSv4.1] Add FROST prepared BF16 tail RoPE for prefill area:frost cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1116 opened Sep 17, 2026 by YangXu1990uiuc Collaborator • Draft
1
[DSv4.1] Extend FROST RoPE QDQ with group16 E4M3-scale FP4 area:frost cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1115 opened Sep 17, 2026 by YangXu1990uiuc Collaborator • Draft
1
[DSv4.1] Add FROST tail RoPE and group32 FP4/FP8 QDQ area:frost cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1114 opened Sep 17, 2026 by YangXu1990uiuc Collaborator • Draft
1
[DSv4.1] Add FROST saved-state Engram gate forward and backward area:frost cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1113 opened Sep 17, 2026 by YangXu1990uiuc Collaborator • Draft
1
frost(sdpa_bwd_sm80): serve THD ports at their own head stride area:frost area:global_attention cat-enhancements orig-nv-eng Reported or requested by NVIDIA engineering.
#1108 opened Sep 16, 2026 by egilliam-nv Collaborator • Draft Frontend 1.30.0
@egilliam-nv
4
frost(sdpa): paged KV for the SM100 per-tensor FP8 d128 engine area:frost area:global_attention cat-feature Requests for new functionality, APIs, examples, or behavior improvements. mod-cutedsl CuTeDSL kernels, generated kernels, examples, or related integration work. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1106 opened Sep 16, 2026 by vedaanta Collaborator Loading…
4 tasks done
20
frost(sdpa): d128 decode-shaped heuristics -- one-cluster scheduler rule, split-KV combine costing area:frost area:global_attention cat-perf-bug Performance regressions or cases where behavior is correct but too slow. mod-frontend cuDNN frontend APIs, operation graph construction, plans, and user-facing wrappers. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1100 opened Sep 16, 2026 by vedaanta Collaborator Loading…
4 tasks done
18
fix(bsa): lazily import torch for cached device capability
#1097 opened Sep 16, 2026 by OutWhite Loading…
4 tasks done
5
frost(sdpa): wire paged KV into the SM100 d512 f16/bf16 flavor (DSv4-class decode) area:frost area:global_attention cat-feature Requests for new functionality, APIs, examples, or behavior improvements. mod-cutedsl CuTeDSL kernels, generated kernels, examples, or related integration work. mod-frontend cuDNN frontend APIs, operation graph construction, plans, and user-facing wrappers. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1093 opened Sep 16, 2026 by vedaanta Collaborator Loading…
4 tasks done
27
BSA: fuse single-group SM100 blk64 dK/dV epilogue
#1091 opened Sep 16, 2026 by wplf Loading…
3 of 4 tasks
1
6
FROST SDPA fp8: block-scaled O epilogues (NVFP4 / MXFP8) with an sf_o output area:frost area:global_attention cat-feature Requests for new functionality, APIs, examples, or behavior improvements. mod-cutedsl CuTeDSL kernels, generated kernels, examples, or related integration work. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1088 opened Sep 16, 2026 by vedaanta Collaborator Loading…
3 tasks
52
[Draft] Integrate Frost grouped GEMM and FP8 MoE support for FlashInfer area:frost cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1080 opened Sep 15, 2026 by YangXu1990uiuc Collaborator • Draft
2 of 4 tasks
1
sdpa(fp8): decline padding mask + left band bound on Hopper (#1009) cat-bugfix mod-frontend cuDNN frontend APIs, operation graph construction, plans, and user-facing wrappers. orig-nv-eng Reported or requested by NVIDIA engineering.
#1069 opened Sep 15, 2026 by vedaanta Collaborator Loading…
2 of 4 tasks
1
9
frost(sdpa_bwd_sm80): read the bound ragged offsets on device (issue #737) area:frost area:global_attention cat-feature Requests for new functionality, APIs, examples, or behavior improvements. orig-nv-eng Reported or requested by NVIDIA engineering.
#1058 opened Sep 14, 2026 by egilliam-nv Collaborator • Draft Frontend 1.30.0
8
sdpa(paged): harden the paged KV decline paths and bound the page walk (#964 review) cat-bugfix mod-frontend cuDNN frontend APIs, operation graph construction, plans, and user-facing wrappers. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1042 opened Sep 13, 2026 by vedaanta Collaborator Loading…
12
Add cuDNN GAT and GATv2 Python APIs
#1021 opened Sep 11, 2026 by tingyu66 Member • Draft
4 tasks
9
Prefer FROST for NVFP4 QAT backward with automatic Triton fallback area:frost area:global_attention cat-enhancements mod-cutedsl CuTeDSL kernels, generated kernels, examples, or related integration work. mod-frost orig-nv-eng Reported or requested by NVIDIA engineering.
#1003 opened Sep 10, 2026 by YangXu1990uiuc Collaborator Loading…
10
ProTip! no:milestone will show everything without a milestone.