All issues

Search results

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

Status: Open.

#12 In NVIDIA/nccl-extensions;

· EvanZhang118 opened on Aug 21, 2026

RFC: ncclCpCombine — a context-parallel attention combine collective

Status: Open.

#10 In NVIDIA/nccl-extensions;

· kwen2501 opened on Aug 14, 2026

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

Status: Open.

#9 In NVIDIA/nccl-extensions;

· kwen2501 opened on Aug 12, 2026

nccl_ep low latency combine: missing async proxy fence before mbarrier_arrive causes stochastic output corruption

Status: Open.

Bug

#8 In NVIDIA/nccl-extensions;

· Oseltamivir opened on Aug 8, 2026

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

enhancementNew feature or requestNew feature or request

Status: Open.

#7 In NVIDIA/nccl-extensions;

· kwen2501 opened on Aug 8, 2026

nccl-m2n: support for HPE Slingshot / CXI

Status: Open.

#5 In NVIDIA/nccl-extensions;

· parinayc20 opened on Jul 29, 2026