NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect

GPUDirect RDMA was attempted between a GPU and a NIC on different NUMA nodes. The transfer has to cross the CPU interconnect (UPI/Infinity Fabric), which is far slower than a local path and on some platforms is not supported for peer-to-peer at all. This entry explains how to confirm the cause, apply the fix, and separate it from adjacent gdr-fabric failures.

Symptom

NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect

Root cause

GPUDirect RDMA was attempted between a GPU and a NIC on different NUMA nodes. The transfer has to cross the CPU interconnect (UPI/Infinity Fabric), which is far slower than a local path and on some platforms is not supported for peer-to-peer at all. The decisive evidence is the first log line that precedes "NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect" and differs from a healthy run.

Recommended fix

bind each rank to the NIC on its own NUMA node, set NCCL_IB_HCA per rank based on topology rather than letting NCCL pick. nvidia-smi topo -m shows which HCA is local to which GPU.

How Denpex helps

Denpex investigates NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect using the evidence you provide or your connected workload collects. Earlier rank, host or application evidence is needed to distinguish an initiating failure from a downstream report.

The literal signature is "NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect". It is a communication failure associated with GPUDirect RDMA, RoCE, GDRCopy, and NUMA topology. The line identifies the failing operation or subsystem, while the surrounding evidence decides whether it is the initiating fault or a downstream symptom.

GPUDirect RDMA was attempted between a GPU and a NIC on different NUMA nodes. The transfer has to cross the CPU interconnect (UPI/Infinity Fabric), which is far slower than a local path and on some platforms is not supported for peer-to-peer at all. The failure becomes visible at this call site because the operation first requires the missing resource, valid state, healthy peer, or correct result. Earlier log lines and a known-good control carry more causal value than the final wrapper exception.

• The workload stops or loses forward progress after emitting "NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect".
• A retry on the same configuration reproduces the failure because the causal state has not changed.
• The outer framework exception can hide the rank, node, allocation, or dependency that failed first.
• Increasing timeouts or reducing workload size can suppress the symptom without correcting the cause.

• GPUDirect RDMA, RoCE, GDRCopy, and NUMA topology
• production-shaped multi-accelerator workloads
• containerized and bare-metal deployments of the same stack

Use this checklist to test the hypothesis against a small reproduction. No single line proves the root cause, so preserve the preceding events and compare one variable at a time.

• ✓Find the first occurrence of "NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect" and preserve at least 100 lines before it.
• ✓Identify which rank, node, device, or process emitted the first related warning.
• ✓read the topology matrix, a PIX or PXB path is local; SYS means it crosses the CPU interconnect. Any GPU-HCA pair showing SYS will take this path. Also confirm process CPU affinity places the rank on the matching NUMA node.
• ✓Repeat the same input after the targeted change and require the signature to disappear.
• ✓Resume from job start with corrected NIC binding only after the control passes.

• GPUDirect RDMA was attempted between a GPU and a NIC on different NUMA nodes. The transfer has to cross the CPU interconnect (UPI/Infinity Fabric), which is far slower than a local path and on some platforms is not supported for peer-to-peer at all.
• The decisive evidence is the first log line that precedes "NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect" and differs from a healthy run.
• A nearby failure remains a competing hypothesis until a control separates configuration, capacity, transport, and hardware causes.

NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect

Use this text as a lookup key in logs and upstream issue trackers. It is not presented as a captured customer log. Confirm the cause from your own preceding events, versions, configuration and the cited references.

bind each rank to the NIC on its own NUMA node, set NCCL_IB_HCA per rank based on topology rather than letting NCCL pick. nvidia-smi topo -m shows which HCA is local to which GPU. This changes the condition in the causal diagnosis instead of hiding the outer exception. The repeated control proves ownership before recovery from job start with corrected NIC binding.

# Preserve evidence before restarting
NCCL_DEBUG=INFO NCCL_DEBUG_SUBSYS=INIT,NET,COLL torchrun train.py
ibstat
ethtool -S <interface>

# Find the exact signature in the complete log
rg -n -F -- "NCCL WARN Direct RDMA write between GPU NUMA and HCA NUMA failed UPI interconnect" <log-file>

Adapt the snippet to your framework. The same pattern holds for PyTorch Lightning, Hugging Face Trainer, DeepSpeed, Megatron-LM, and vLLM training wrappers. Where the wrapper exposes a config flag (for examplelr_scheduler_type in Trainer), prefer the flag over the imperative API to keep the schedule declarative and reproducible.

Select the traceback or open the failed terminal, then run Denpex locally to see the initiating rank, collateral failures, exact fix, and verification command without uploading the log.

NCCL is where a distributed job reports failure, which is not the same as where it failed. The hub lists every common NCCL error next to what it actually indicates, and the environment variables that tell them apart.