Issues

• [Issue]: ncclCommFinalize/ncclCommDestroy hang when a host-local peer has died (intranode barrier added in 2.31)

  NVIDIA/nccl#2506 · Filter by author d4l3k opened on Oct 22, 2026
  4

  

• [Question]: gin device poll cq

  question
  NVIDIA/nccl#2505 · Filter by author shenhongqian opened on Oct 22, 2026
  

• [Issue]: NET/Socket small-message latency and long stalls across NCCL 2.29.3, 2.30.7 and 2.32.3 on two H20 hosts

  NVIDIA/nccl#2503 · Filter by author zeyuan-lcl opened on Oct 22, 2026
  1

  

• [Issue]: NCCL init fails under Nsight Systems on GB300 aarch64/MNNVL (2.30.7 stack; CUDA error 400)

  NVIDIA/nccl#2502 · Filter by author Phoenix3334 opened on Oct 22, 2026
  

• [RFE]: Native per-operation timeouts for collectives and P2P

  NVIDIA/nccl#2501 · Filter by author d4l3k opened on Oct 21, 2026
  5

• [Issue]: Socket transport returns wrong results or hangs with NCCL_SOCKET_NTHREADS > 0 (helper thread progresses a socket's tasks out of order after the queue wraps)

  NVIDIA/nccl#2499 · Filter by author mraad opened on Oct 21, 2026
  

• [Issue]: Two ncclCommSplit calls on the same parent in one group hang or silently create 1-rank children

  NVIDIA/nccl#2488 · Filter by author kodlan opened on Oct 19, 2026
  1

  

• NCCL Roadmap: Nov 2026 - Jan 2027

  roadmap
  NVIDIA/nccl#2486 · Filter by author gab9talavera opened on Oct 16, 2026

• [Issue]: A failed thread creation aborts the process (std::terminate) instead of returning ncclSystemError

  NVIDIA/nccl#2484 · Filter by author kodlan opened on Oct 15, 2026
  

• [Issue]: GIN GDAKI: a failed put is invisible to device code (flush returns, get data never arrives), and the host learns of it once, at the 10 s query

  NVIDIA/nccl#2483 · Filter by author unionxic opened on Oct 15, 2026
  

• [Issue]: GIN GDAKI: NCCL_GIN_GDAKI_NIC_HANDLER=6 is accepted at init, but GIN operations never complete

  resolved
  NVIDIA/nccl#2480 · Filter by author unionxic opened on Oct 15, 2026
  2

  

• [Issue]: BF16/FP16 NVLS AllReduce and ReduceScatter are not repeatable when count % 4 != 0

  NVIDIA/nccl#2479 · Filter by author qcliao opened on Oct 14, 2026
  4

  

• [Question]: How much benefit can NcclCommRegister provide？

  questionresolved
  NVIDIA/nccl#2476 · Filter by author ZhouMinhao98 opened on Oct 13, 2026
  3

  

• [Question]: Will receiver-side PXN be implemented in future NCCL

  question
  NVIDIA/nccl#2475 · Filter by author caixj311 opened on Oct 13, 2026
  2

• [Issue]: Early error returns inside an internal group leave the thread stuck one group level deep (later work never launches)

  NVIDIA/nccl#2471 · Filter by author kodlan opened on Oct 11, 2026
  

• [Issue]: bootstrapCreateRoot leaks the socket descriptor when ncclSocketListen fails, so each ncclGetUniqueId that fails in bind() leaks one fd

  NVIDIA/nccl#2469 · Filter by author dmvevents opened on Oct 10, 2026
  4

  

• [Issue]: More inline data than it reserved space for, when receives get grouped

  NVIDIA/nccl#2467 · Filter by author jeseszhang1010 opened on Oct 9, 2026
  2

  

• [RFE]: AMD SYS-path restriction may limit ring diversity and AllReduce performance on multi-socket PCIe GPU systems

  enhancement
  NVIDIA/nccl#2466 · Filter by author Y7cc opened on Oct 9, 2026
  

• [Issue]: P2P operations fail at first call with NCCL_NET_GDR_READ=0 when PXN is selected

  NVIDIA/nccl#2463 · Filter by author lsavers opened on Oct 7, 2026
  2

  

• [Issue]: Collectives and send/recv on a suspended communicator (ncclCommSuspend) are accepted and crash instead of returning an error

  NVIDIA/nccl#2461 · Filter by author kodlan opened on Oct 6, 2026
  1

  

• [Question]: Fluctuating performance of GIN puts from a hybrid GIN+LSA device kernel

  question
  NVIDIA/nccl#2454 · Filter by author parinayc20 opened on Sep 29, 2026
  5

  

• [Issue]: ncclCommInitRankConfig / ncclCommInitRankScalable / ncclCommInitAll crash on invalid nranks, nId or ndev instead of returning ncclInvalidArgument

  NVIDIA/nccl#2449 · Filter by author kodlan opened on Sep 26, 2026
  1

  

• [Issue]: ncclAllReduce with ncclAvg on symmetric-memory buffers fails with "unhandled cuda error" (no avg AllReduce symmetric kernels, but ncclSymkImplemented says there are)

  NVIDIA/nccl#2447 · Filter by author kodlan opened on Sep 25, 2026
  1

  

• [Issue]: p8 sum AllReduce over symmetric memory crashes on sm_120 (RTX 5070 Ti) with CUDA 12.8 - 13.0: ptxas drops the upper 32 bits of the LL buffer address

  NVIDIA/nccl#2445 · Filter by author kodlan opened on Sep 24, 2026
  2

  

• [RFE]: is_all_reduce_fp32_accumulate_enabled (NVL OR symmetric memory)

  enhancement
  NVIDIA/nccl#2442 · Filter by author stas00 opened on Sep 23, 2026
  1