Skip to content

Latest commit

 

History

History
47 lines (45 loc) · 10 KB

File metadata and controls

47 lines (45 loc) · 10 KB

NCCL Error Pattern Reference

Quick-lookup table of NCCL log patterns → code → root cause → fix. Used by the diagnostic script to map log lines to a remediation section in debugging-guide.md.

Log pattern Code Root cause Fix
Rendezvous / connection
Timeout waiting for TIMEOUT_RENDEZVOUS Peers not joining init SG self-ref, NetworkPolicy, MASTER_ADDR
Connection refused CONN_REFUSED Rank-0 not listening Fix MASTER_ADDR + headless service
Address already in use PORT_CONFLICT Port 29500 bound Change MASTER_PORT to 29501
NCCL WARN Connect to CONNECT_FAIL NCCL peer blocked SG self-ref + NetworkPolicy
network is unreachable NET_UNREACHABLE No route to MASTER_ADDR DNS + VPC routing + SG
Error in Store / DistStoreError STORE_ERR c10d rendezvous timeout Fix network first
RendezvousConnectionError RDZV_CONN_ERR Elastic rendezvous failed MASTER_ADDR DNS + SG
RendezvousTimeout RDZV_TIMEOUT Elastic rendezvous timed out Peers not reachable
Name or service not known DNS_FAIL DNS resolution failed Create headless service
getaddrinfo failed DNS_FAIL DNS resolution failed CoreDNS + headless service
Runtime / AllReduce
Watchdog timeout WATCHDOG_TIMEOUT AllReduce timed out Extend PyTorch init_process_group(timeout=...); find straggler
unhandled system error SYSTEM_ERROR GPU/EFA hardware SSM: dmesg XID errors; reboot node
unhandled cuda error CUDA_ERROR CUDA runtime error GPU driver crash or hardware fault
peer access is not supported P2P_FAIL GPU P2P blocked by ACS/IOMMU Disable ACS; check IOMMU
NCCL WARN Cuda failure CUDA_ERROR CUDA failure inside NCCL GPU hardware or driver
Call to ncclCommAbort NCCL_COMM_ABORT Communicator aborted Check for straggler or hardware fault
EFA / libfabric
fi_getinfo failed EFA_INIT_FAIL EFA not available Fix EFA; use gloo on non-EFA
NCCL_OFI_RDMA OFI_ERROR aws-ofi-nccl broken Check plugin + EFA version
Call to ibv_reg_mr failed RDMA_REG_FAIL memlock=0 blocks EFA RDMA ulimit -l 8388608
NET/OFI Using TCP EFA_TCP_FALLBACK Fell back to TCP Fix EFA device plugin + env
Failed to load NCCL NCCL_LOAD_FAIL libnccl.so missing Check LD_LIBRARY_PATH
libnccl-net.so OFI_LOAD_FAIL OFI plugin missing Install aws-ofi-nccl
OOM / resource limits
OOMKilled OOM_KILL Pod out of memory Reduce batch size; increase limits
CUDA out of memory / cudaMalloc failed CUDA_OOM GPU VRAM exhausted Reduce batch size, enable ZeRO
failed to extend /dev/shm / Bus error SHM_FULL /dev/shm too small emptyDir medium:Memory 10Gi
ENOMEM ENOMEM Memory alloc/registration failure Check memlock + GPU memory
Version / config
NCCL function not found NCCL_VERSION_MISMATCH Mixed NCCL versions Use identical container images
Incompatible NCCL version NCCL_VERSION_MISMATCH Mixed NCCL versions Use identical container images
Could not find interface IFACE_NOT_FOUND Bad NCCL_SOCKET_IFNAME Set ^lo,docker,efa,veth,virbr
world_size mismatch WORLD_SIZE_MISMATCH WORLD_SIZE ≠ ranks WORLD_SIZE = pods × GPUs/pod
doesn't have NCCL built in NCCL_NOT_BUILT PyTorch without NCCL Use AWS DLC image
CUDA_VISIBLE_DEVICES CUDA_VIS_DEV GPUs hidden from training Remove CUDA_VISIBLE_DEVICES
invalid alignment CUDA_ALIGN_ERR CUDA alignment error Check driver/NCCL version compat
Stale state / topology
unlink shared memory SHM_STALE Stale /dev/shm/nccl-* files Set RemoveIPC=no; clean up
MNNVL topology MNNVL_TOPO_FAIL NCCL topology search failure Try fixed memlock (e.g. ulimit -l 8388608) — field-observed workaround, not NCCL-documented; see debugging-guide.md § 17