Repository Issue Activity (beta)

#2181 [Question]: Does NCCL have any plan to provide QP counter metrics of GDAKI 's QPs?

#2332 [Issue]: `ncclCommRegister` fails with `CUDA_ERROR_NOT_FOUND` (500) on memory allocated with CUDA VMM.

#2417 [RFE]: Add optional receive queues to the GDAKI backend

#2416 [RFE]:Pluggable external device-memory allocator: hand all non-registration buffer classes to a framework-supplied allocator (e.g. PyTorch's caching allocator)

#2420 NVCC_GENCODE_LDMC_FP8 device-code gencode ignores the caller's NVCC_GENCODE target-arch list, breaking single-arch builds

#2421 ncclAlltoAllConfig does not honor per-call maxCTAs

#2033 [Issue]: CPU affinity is not restored at the end of `initTransportsRank()` (introduced in 2.29.2-1)

#1270 Provenance of NVTX headers in NCCL

#2111 [RFC] nccl-ep: support Attention–MoE disaggregation

#1026 half precision reduction accumulation in fp32?

#2308 [Announcement] NCCL EP and NCCL M2N are moving to NVIDIA/nccl-extensions

#2418 [Issue]: Dual RTX PRO 6000 Blackwell: NCCL 2.26.2 illegal memory access for >=512 KiB collectives, not reproducible with 2.31.2

#2372 [Question]: NCCL GIN GDAKI (type 3) hangs during Pure GIN AlltoAll, PROXY (type 2) works

#2397 [Issue]: ncclCommShrink leads to IB transport errors and hang when excluded ranks exit without calling shrink, and also affects the allreduce operations after the shrink.

#2360 [Bug] Hopper NVLS BF16 results depend on multicast allocation history and differ across H800 hosts with identical software

#2325 [Issue]: After the rollback of `PciInfo` the `nvmlDeviceGetPciInfo_v3` is used with `v2` struct causing stack buffer overflow.

#2226 [Question]: [GIN] GPI vs GDAKI: who selects the QP, and what NIC support does GPI need?

#2408 NCCL_CHUNK_SIZE / NCCL_P2P_NET_CHUNKSIZE >= 1MiB gives incorrect results and Xid 31 GPU MMU faults (H100, 8x400G RoCE)

#2409 Low AllReduce bandwidth on 2x8x H100 with 8x 400G RoCE: per GPU<->GPU connection capped at ~13 GB/s while perftest reaches 392 Gb/s line rate

#2160 NCCL_TOPO_XML_MAX_NODES=256 limit hit during intra-node XML fusion on 32-NIC hosts (AWS p5.48xlarge)

#2367 [Issue]: dlopen() of a truncated plugin .so causes SIGBUS crash instead of graceful fallback (NCCL_PROFILER_PLUGIN / NET / TUNER / ENV)

#2345 [Issue]: Ring and tree both create cross-rail inter-node links, violating rail isolation and despite `NCCL_CROSS_NIC=0`

#2379 Probabilistic multi-communicator init hang in the cuMem UDS fd exchange on single-node PCIe (no NVLink); NCCL_CUMEM_ENABLE=0 isolates it, 20/20 vs 5/9 A/B

#2355 [Issue]: Profiler plugin overhead regression after upgrading to NCCL 2.31.2

#2238 [Issue]: Performance and functional issues when using ncclPutSignal/ncclWaitSignal in Pipeline Parallelism