Issues

• [Feature Request]: Allow Inspector JSON detail output alongside OTLP metrics

  NVIDIA/nccl#2591 · Filter by author lyquid617 opened on Nov 12, 2026

• [Question]: question about team LSA crossing multiple nvl domain

  question
  NVIDIA/nccl#2590 · Filter by author fishautumn opened on Nov 12, 2026

• [RFC]: [Help Wanted] Test for NCCL Determinism Guarantees

  Community Discussionenhancementhelp wanted
  NVIDIA/nccl#2588 · Filter by author xiaofanl-nvidia opened on Nov 12, 2026

• [Issue]: ncclCommAbort on a parent writes to freed memory after a failed child init

  NVIDIA/nccl#2586 · Filter by author kodlan opened on Nov 11, 2026
  1

• [Issue]: ncclCommAbort and ncclCommDestroy free nothing after a failed or incomplete init

  NVIDIA/nccl#2585 · Filter by author kodlan opened on Nov 11, 2026
  1

• [RFE]: Reduce enqueue time with many registered buffer ranges

  NVIDIA/nccl#2583 · Filter by author ZenAlexa opened on Nov 11, 2026
  

• [Issue]: In ncclTuningCostModelInit strLength is not enough after new algorithms were added

  NVIDIA/nccl#2579 · Filter by author PatriosTheGreat opened on Nov 11, 2026
  

• [Issue]: PAT ReduceScatter with ncclAvg / PreMulSum returns the sum (no 1/n scaling), 2.31.2 and 2.32.6

  NVIDIA/nccl#2577 · Filter by author DanielSun11 opened on Nov 11, 2026
  2

• [Issue]: Ring/LL over NET: the send proxy's scan of the LL lines makes 2-8 MiB all-reduce 14-57x slower than Simple (AWS g7.48xlarge, EFA)

  NVIDIA/nccl#2574 · Filter by author KeitaW opened on Nov 10, 2026
  

• [RFE]: Connect only the collective channels that are actually used (lazy channel prefix, follow-up to #2527 direction B)

  NVIDIA/nccl#2569 · Filter by author VikingDeng opened on Nov 7, 2026
  1

• [RFE]: Log when ncclCommFinalize is waiting for CUDA graphs that still reference the communicator

  NVIDIA/nccl#2568 · Filter by author MogicianWu opened on Nov 7, 2026
  1

  

• [Issue]: Hang after port failover with NCCL_IB_RESILIENCY_PORT_RECOVERY=1 when a request is retransmitted on both QPs (v2.32.3, roughly 2-7% of runs)

  NVIDIA/nccl#2567 · Filter by author daehyun-kim-94 opened on Nov 6, 2026
  

• [RFE]: Each registered buffer allocates a 2 MiB device block for its peer address array

  enhancement
  NVIDIA/nccl#2565 · Filter by author kodlan opened on Nov 4, 2026
  3

• [Question]: Can NCCL support 0sm broadcast?

  question
  NVIDIA/nccl#2564 · Filter by author MARD1NO opened on Nov 4, 2026
  5

• [Issue]: ncclCommSplit child of a splitShare parent reports hostRmaSupport=0 (ncclRmaInitFromParent does not inherit the RMA plugin)

  NVIDIA/nccl#2560 · Filter by author ngoyal2707 opened on Nov 3, 2026
  1

• [Issue]: GPU flush requests can hang after successful IB port recovery

  NVIDIA/nccl#2555 · Filter by author BOY332 opened on Nov 3, 2026
  3

  

• [Issue]: Repeated port flaps cause recovery initialization to fail with ENOMEM and permanently disable recovery

  NVIDIA/nccl#2554 · Filter by author BOY332 opened on Nov 3, 2026
  6

• [RFE]: Reduce initialization memory by packing channel metadata

  enhancement
  NVIDIA/nccl#2553 · Filter by author arkadip-maitra opened on Nov 3, 2026
  3

  

• [Issue]: NET/Socket loses consumed header state on EAGAIN before inline payload

  NVIDIA/nccl#2549 · Filter by author KazukiKomon opened on Nov 2, 2026
  2

  

• [RFE]: Collective connections keep an LL128 buffer on communicators that can never select LL128 (#2527, direction C)

  enhancement
  NVIDIA/nccl#2548 · Filter by author kodlan opened on Nov 1, 2026
  2

• [RFE]: Send/recv P2P connections allocate an unused LL128 buffer (#2527, direction C)

  enhancement
  NVIDIA/nccl#2540 · Filter by author kodlan opened on Oct 31, 2026
  5

• [RFE]: Add memory overhead tracker tooling

  Community Discussionenhancement
  NVIDIA/nccl#2538 · Filter by author xiaofanl-nvidia opened on Oct 30, 2026
  4

  

• [Issue]: CUDA Graph replay: later all-to-all Send/Recv work slows earlier Ring AllGather/ReduceScatter

  NVIDIA/nccl#2532 · Filter by author minsii opened on Oct 29, 2026

• [Question]: Does NCCL automatically recover if an NVLink fails during training?

  question
  NVIDIA/nccl#2531 · Filter by author slevchuk opened on Oct 29, 2026
  1

• [Issue]: NCCL_ALGO with a function prefix leaves every other function without an algorithm when NCCL_PROTO is set (since 2.31)

  NVIDIA/nccl#2528 · Filter by author abhiMishra98 opened on Oct 28, 2026
  1