Cowardly refusing to serialize non-leaf tensor

Something still attached to the autograd graph was handed to the checkpoint writer. Saving it would serialise the graph as well as the values, so torch refuses rather than writing a checkpoint that cannot be loaded cleanly. This entry explains how to confirm the cause, apply the fix, and separate it from adjacent pytorch-fsdp-ddp failures.

The literal signature is "Cowardly refusing to serialize non-leaf tensor". It is a distributed training failure associated with PyTorch Distributed, FSDP, DDP, and TorchElastic. The line identifies the failing operation or subsystem, while the surrounding evidence decides whether it is the initiating fault or a downstream symptom.

Something still attached to the autograd graph was handed to the checkpoint writer. Saving it would serialise the graph as well as the values, so torch refuses rather than writing a checkpoint that cannot be loaded cleanly. 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 "Cowardly refusing to serialize non-leaf tensor".
• 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.

• PyTorch Distributed, FSDP, DDP, and TorchElastic
• production-shaped multi-accelerator workloads
• containerized and bare-metal deployments of the same stack

Apply the following checklist to a small reproduction: each box below is a positive signal that you are looking at this exact failure rather than a sibling in the same taxonomy.

• ✓Find the first occurrence of "Cowardly refusing to serialize non-leaf tensor" and preserve at least 100 lines before it.
• ✓Identify which rank, node, device, or process emitted the first related warning.
• ✓this is almost always a metric or a buffer accumulated inside the loop without detaching (running loss, a cached activation) that ended up in the checkpoint dict. Print which keys carry requires_grad: [k for k,v in sd.items() if torch.is_tensor(v) and v.requires_grad].
• ✓Repeat the same input after the targeted change and require the signature to disappear.
• ✓Resume from the previous good checkpoint only after the control passes.

Cowardly refusing to serialize non-leaf tensor

Timestamps and exact values vary across runs, but the pattern. An info-level start, an early WARN, an ERROR carrying the symptom. Is the actual fingerprint you should alert on. The Denpex platform flags this combination automatically.

detach before saving.detach() on the offending tensor, or save model.state_dict() rather than tensors captured from the training loop. This changes the condition in the causal diagnosis instead of hiding the outer exception. The repeated control proves ownership before recovery from the previous good checkpoint.

# Preserve evidence before restarting
rg -n -i 'error|exception|timeout|failed' <log-file>
nvidia-smi
python -m torch.utils.collect_env

# Find the exact signature in the complete log
rg -n -F -- "Cowardly refusing to serialize non-leaf tensor" <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.

• Something still attached to the autograd graph was handed to the checkpoint writer. Saving it would serialise the graph as well as the values, so torch refuses rather than writing a checkpoint that cannot be loaded cleanly.
• The decisive evidence is the first log line that precedes "Cowardly refusing to serialize non-leaf tensor" and differs from a healthy run.
• A nearby failure remains a competing hypothesis until a control separates configuration, capacity, transport, and hardware causes.