NVIDIA/nccl

Optimized primitives for collective multi-GPU communication

Stars: 5.1k
Forks: 1.4k
Open beginner issues: 0
Indexed issues: 276
Avg merge: 2h 3m
Merged PRs (30d): 2

Dominant language: C++
License: No license data
Last GitHub push: Sep 17, 2026
Latest indexed: Sep 19, 2026
Contributing guide: Contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Beginner friendly Everything

Open All

276 open issues indexed

ncclAlltoAllConfig does not honor per-call maxCTAs Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

NVIDIA/nccl#2421 · 1 comment · Opened Sep 18, 2026
NVCC_GENCODE_LDMC_FP8 device-code gencode ignores the caller's NVCC_GENCODE target-arch list, breaking single-arch builds Open

Difficulty 3/5 1-2 days Newbie friendliness 76/100

NVIDIA/nccl#2420 · 1 comment · Opened Sep 18, 2026
[Issue]: Dual RTX PRO 6000 Blackwell: NCCL 2.26.2 illegal memory access for >=512 KiB collectives, not reproducible with 2.31.2 Open

Difficulty 4/5 3-5 days Newbie friendliness 42/100

NVIDIA/nccl#2418 · 1 comment · Opened Sep 17, 2026
[RFE]: Add optional receive queues to the GDAKI backend Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/nccl#2417 · 3 comments · Opened Sep 17, 2026
[RFE]:Pluggable external device-memory allocator: hand all non-registration buffer classes to a framework-supplied allocator (e.g. PyTorch's caching allocator) Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/nccl#2416 · 2 comments · Opened Sep 17, 2026
NCCL_CHUNK_SIZE / NCCL_P2P_NET_CHUNKSIZE >= 1MiB gives incorrect results and Xid 31 GPU MMU faults (H100, 8x400G RoCE) Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

NVIDIA/nccl#2408 · 3 comments · Opened Sep 14, 2026
[Bug]: Deadlock in bootstrap when ranks are unevenly distributed across nodes (7+8), contiguousRanksPerHost falls back to INT_MAX Open

NVIDIA/nccl#2385 · 8 comments · 1 assignee · Opened Sep 3, 2026
[GIN/GDAKI] A `TRANSPORT_RETRY_EXC_ERR` completion on a GDAKI QP hangs the device forever: the device-side wait loops treat an error CQE as "not yet" Open

NVIDIA/nccl#2383 · 3 comments · 1 assignee · Opened Sep 2, 2026
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 Open

NVIDIA/nccl#2379 · 3 comments · 2 assignees · Opened Aug 29, 2026
[Question]: potential memory consistency issue with rmda ce Open

question

NVIDIA/nccl#2375 · 3 comments · 1 assignee · Opened Aug 26, 2026
[Question]: NCCL GIN GDAKI (type 3) hangs during Pure GIN AlltoAll, PROXY (type 2) works Open

question

NVIDIA/nccl#2372 · 1 comment · 1 assignee · Opened Aug 26, 2026
[Issue]: Single-rank communicator initializes a large NET shared P2P buffer Open

NVIDIA/nccl#2363 · 2 comments · 1 assignee · Opened Aug 21, 2026
[Bug] Hopper NVLS BF16 results depend on multicast allocation history and differ across H800 hosts with identical software Open

NVIDIA/nccl#2360 · 7 comments · 1 assignee · Opened Aug 19, 2026
[Issue]: SIGSEGV in bootstrapRoot getenv() via ncclSocketDefaultMagic, racy under glibc <2.41 Open

NVIDIA/nccl#2354 · 1 comment · 1 assignee · Opened Aug 18, 2026
[Question]: Question about LLA2A endEpoch, why could an error occur when epoch = 0xFFFFFFFE and the next epoch becomes 2? Open

question

Difficulty 4/5 3-5 days Newbie friendliness 45/100

NVIDIA/nccl#2352 · Opened Aug 18, 2026
[Issue]: NCCL Checkpoint - Restore deadlocks forever in SHM's cuMem host-NUMA path; NCCL_CUMEM_HOST_ENABLE=0 fixes it Open

NVIDIA/nccl#2350 · 1 assignee · Opened Aug 17, 2026
[Issue]: NCCL Checkpoint - ncclCheckpointRestore() can block indefinitely with no timeout and no diagnostic Open

NVIDIA/nccl#2349 · 4 comments · 1 assignee · Opened Aug 17, 2026
[Issue]: NCCL Checkpoint - getHostHash() is cached across checkpoint/restore; restored communicators use a stale host identity Open

NVIDIA/nccl#2348 · 3 comments · 1 assignee · Opened Aug 17, 2026
[Issue]: NCCL RAS null communicator status leading to segfault on 2.29.7 Open

bug ongoing triaged

NVIDIA/nccl#2346 · 3 comments · 1 assignee · Opened Aug 17, 2026
[Issue]: Ring and tree both create cross-rail inter-node links, violating rail isolation and despite `NCCL_CROSS_NIC=0` Open

NVIDIA/nccl#2345 · 6 comments · 1 assignee · Opened Aug 16, 2026
[RFE]: Add NVLS suspend / resume to ncclCommSuspend / ncclCommResume Open

enhancement

NVIDIA/nccl#2337 · 7 comments · 1 reaction · 2 assignees · Opened Aug 10, 2026
RoCEv2 receiver deadlocks after successful RDMA_WRITE completions land on 'unused' request slots (net_ib/p2p.cc resiliency path) Open

NVIDIA/nccl#2334 · 4 comments · 1 assignee · Opened Aug 9, 2026
[Issue]: `ncclCommRegister` fails with `CUDA_ERROR_NOT_FOUND` (500) on memory allocated with CUDA VMM. Open

NVIDIA/nccl#2332 · 3 comments · 1 assignee · Opened Aug 7, 2026
[Question]: Why NCCL_ALGO_NVLS only support NCCL_PROTO_SIMPLE, no NCCL_PROTO_LL/NCCL_PROTO_LL128? Open

question

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

NVIDIA/nccl#2327 · 4 comments · Opened Aug 7, 2026
[Issue]: Memory leak risk of function allocateSpilled Open

NVIDIA/nccl#2324 · 2 comments · 1 assignee · Opened Aug 6, 2026
[RFE]: Add persistent backoff recovery mechanism for IB resiliency port recovery Open

enhancement

NVIDIA/nccl#2323 · 7 comments · 1 assignee · Opened Aug 6, 2026
[RFE]: Add active-standby paths for fused IB NICs Open

NVIDIA/nccl#2310 · 1 comment · 2 assignees · Opened Jul 27, 2026
[Announcement] NCCL EP and NCCL M2N are moving to NVIDIA/nccl-extensions Open

Community Discussion

NVIDIA/nccl#2308 · 2 comments · 4 reactions · 1 assignee · Opened Jul 27, 2026
[Issue]: Blackwell symmetric ReduceScatter selects a slower launch Open

NVIDIA/nccl#2305 · 1 comment · 3 assignees · Opened Jul 23, 2026
[RFE]: Add GPU Hardware Identifier to the NCCL Profiler Plugin API Open

enhancement

NVIDIA/nccl#2297 · 1 comment · 2 assignees · Opened Jul 17, 2026
[RFE]: Discuss per-peer P2P buffer merge to reduce NCCL communicator init time Open

enhancement

NVIDIA/nccl#2294 · 5 comments · 1 assignee · Opened Jul 16, 2026
[Question]: Inconsistent busMultiplier for MC kernels: why 0.55*nRanks only for AllGather and ReduceScatter? Open

question

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/nccl#2293 · 3 comments · 1 reaction · Opened Jul 16, 2026
[Question]: What are the hardware and software version requirements for running GPI (including DOCA and DPA)？ Open

question

NVIDIA/nccl#2287 · 2 comments · 1 assignee · Opened Jul 14, 2026
[Issue]: NET/IB out-of-bounds write / crash when a node exposes more than MAX_IB_DEVS active IB ports Open

NVIDIA/nccl#2282 · 4 comments · 1 assignee · Opened Jul 13, 2026
[Issue]: Tuning model appears to conservatively penalize Ring/Simple NET latency for flush on Hopper/Blackwell paths where runtime needFlush is none Open

NVIDIA/nccl#2274 · 1 comment · 1 assignee · Opened Jul 8, 2026
NCCL Roadmap: Aug - Oct 2026 Open

roadmap

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

NVIDIA/nccl#2272 · 13 comments · 8 reactions · Opened Jul 6, 2026
[Question]: ProxyOp and ProxyStep in NCCL Inspector Open

question

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

NVIDIA/nccl#2239 · 5 comments · Opened Jun 17, 2026
[Issue]: Performance and functional issues when using ncclPutSignal/ncclWaitSignal in Pipeline Parallelism Open

NVIDIA/nccl#2238 · 13 comments · 1 assignee · Opened Jun 16, 2026
[RFE]: Include qp_num in detailed RoCE completion error log Open

enhancement

Difficulty 3/5 1-2 days Newbie friendliness 68/100

NVIDIA/nccl#2219 · Opened Jun 7, 2026
[RFE]: Add opt-in auto-selection for IB NIC merge mode on two-node multi-HCA systems Open

NVIDIA/nccl#2216 · 1 comment · 1 assignee · Opened Jun 5, 2026
[Bug] SIGSEGV in ncclLocalOpAppend after ~1.1M collectives on CPU-path — dual DGX Spark (aarch64, GB10, NCCL 2.28.9+cuda13.0) Open

NVIDIA/nccl#2213 · 4 comments · 1 assignee · Opened Jun 4, 2026
[Question]: Potential `ncclIntruQueueTransfer` bug + memory leak? Open

question

NVIDIA/nccl#2204 · 2 comments · 1 assignee · Opened Jun 3, 2026
[RFE]: MXFP8 / MXFP4 / NVFP4 AllReduce support and implementation considerations Open

enhancement

NVIDIA/nccl#2199 · 3 comments · 1 assignee · Opened Jun 2, 2026
[RFE]: Misleading RING_LL suffix in AllGather/ReduceScatter kernel names Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA/nccl#2196 · 7 comments · Opened May 26, 2026
[Question]: Disable NCCL CROSS NIC Open

question

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

NVIDIA/nccl#2193 · 10 comments · Opened May 24, 2026
[Question]: NIC resiliancy/failover and VNICs Open

question

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

NVIDIA/nccl#2192 · 2 comments · Opened May 23, 2026
[Issue]: NCCL WARN Error while attaching to shared memory segment Open

NVIDIA/nccl#2190 · 7 comments · 1 assignee · Opened May 22, 2026
[Issue]: Potential memory leak in ncclIbInitDevices error paths after wrap_ibv_get_device_list() Open

NVIDIA/nccl#2189 · 2 comments · 1 assignee · Opened May 21, 2026
[Question]: Does NCCL have any plan to provide QP counter metrics of GDAKI 's QPs? Open

question

NVIDIA/nccl#2181 · 4 comments · 2 assignees · Opened May 18, 2026
[Question]: cudaMemcpyBatchAsync with cudaMemcpyFlagPreferOverlapWithCompute: D2D (local copy) bandwidth is unexpectedly low compared to SM-based copy Open

question

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

NVIDIA/nccl#2177 · 3 comments · Opened May 14, 2026

A short digest of beginner-friendly GitHub issues.