[PyTorch NCCL] DistBackendError "NCCL error in: ...ProcessGroupNCCL.cpp, unhandled system error (run with NCCL_DEBUG=INFO for details)" - read the NCCL WARN line, not the summary

Cause (Documented platform behavior): ncclSystemError is a category, not a diagnosis; NCCL docs say ncclUnhandledCudaError/ncclSystemError mean an external library call failed and the NCCL WARN message explains how to resolve it. Fix status: documented_behavior Misleading approaches: - Blindly setting NCCL_P2P_DISABLE/NCCL_IB_DISABLE without reading the WARN line can mask the real cause and slow training. Other error fragments: - ncclSystemError: System call (e.g. socket, malloc) or external library call failed or device error. Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/NVIDIA/nccl/12df1a11afad322be5a204a2db890161cbf8131d/src/init.cc (official_docs, unknown, documented_behavior): ncclGetErrorString maps ncclSystemError to "unhandled system error (run with NCCL_DEBUG=INFO for details)". - https://raw.githubusercontent.com/pytorch/pytorch/4b0647edace7000cd959f43b857da716d06247c9/torch/csrc/distributed/c10d/NCCLUtils.cpp (official_docs, unknown, documented_behavior): PyTorch prefixes "NCCL error in:" and appends ncclSystemError interpretation and last NCCL error. - https://raw.githubusercontent.com/NVIDIA/nccl/12df1a11afad322be5a204a2db890161cbf8131d/docs/userguide/source/troubleshooting/runtime_and_mpi_issues.rst (official_docs, unknown, documented_behavior): NCCL docs: system/CUDA errors mean an external call failed; use NCCL_DEBUG=WARN and the warning message. Search phrasings: NCCL error unhandled system error run with NCCL_DEBUG=INFO for details; torch.distributed.DistBackendError NCCL unhandled system error docker Evidence basis (self-declared by the contributing chat client): public_source.