NVIDIA/nccl-extensions

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

• nccl-ep requires GIN VA signals but only uses indexed signals

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 2/5 1-3 hours Newbie friendliness 62/100

  NVIDIA/nccl-extensions#26 · Opened Oct 15, 2026

• Add an AWS EFA guide for nccl_ep

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 2/5 1-3 hours Newbie friendliness 42/100

  NVIDIA/nccl-extensions#24 · Opened Oct 10, 2026

• comm.abort() does not return when the reshard peer dies or stalls after in-flight ncclReshard calls

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 6/100

  NVIDIA/nccl-extensions#23 · 4 comments · Opened Oct 9, 2026

• nightly builds of nccl extensions & nccl ep

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

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

  NVIDIA/nccl-extensions#21 · Opened Oct 2, 2026

• NCCL EP Roadmap: Q4

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  roadmap

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

  NVIDIA/nccl-extensions#20 · Opened Oct 1, 2026

• [BUG] ncclReshard hangs silently when per-rank staging area reaches 2^32 bytes (16 GiB global tensor, cross-dim 4:4)

  May be free again @0z5a claimed this 27 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

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

  NVIDIA/nccl-extensions#15 · 1 comment · Opened Sep 18, 2026

• [Bug]: nccl_ep HT dispatch underutilizes CTAs because chunk size scales with max_num_sms

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

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

  NVIDIA/nccl-extensions#12 · 2 comments · Opened Aug 21, 2026

• RFC: ncclCpCombine — a context-parallel attention combine collective

  May be free again @0z5a claimed this 27 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

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

  NVIDIA/nccl-extensions#10 · 1 comment · Opened Aug 14, 2026

• nccl_m2n: full replication over a 1-D mesh requires a 2-D [Replicate, Shard(0)] spelling

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

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

  NVIDIA/nccl-extensions#9 · 2 comments · Opened Aug 12, 2026

• RFC: nccl_m2n — quantized tensor support (coupled scale forwarding + on-the-fly quantization)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement

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

  NVIDIA/nccl-extensions#7 · Opened Aug 8, 2026

• nccl-m2n: support for HPE Slingshot / CXI

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

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

  NVIDIA/nccl-extensions#5 · 4 comments · Opened Jul 29, 2026

• ncclReshard on a non-blocking communicator costs ~0.4 ms of host time per call, independent of size

  Possibly taken @kaushik-ks claimed this 2 days ago. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 22/100

  NVIDIA/nccl-extensions#22 · 1 comment · 1 assignee · Opened Oct 9, 2026