Overflow detected during fp16 gradient unscaling. Skipping step, loss scale halved

FP16 gradient overflow forced the loss scaler to halve repeatedly. Occasional skipped steps early in training are normal; a scale that keeps collapsing means gradients are genuinely exploding or the model is numerically unstable in FP16. This entry explains how to confirm the cause, apply the fix, and separate it from adjacent deepspeed failures.

The literal signature is "Overflow detected during fp16 gradient unscaling. Skipping step, loss scale halved". It is a distributed training failure associated with DeepSpeed ZeRO and optimizer sharding. The line identifies the failing operation or subsystem, while the surrounding evidence decides whether it is the initiating fault or a downstream symptom.

FP16 gradient overflow forced the loss scaler to halve repeatedly. Occasional skipped steps early in training are normal; a scale that keeps collapsing means gradients are genuinely exploding or the model is numerically unstable in FP16. 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 "Overflow detected during fp16 gradient unscaling. Skipping step, loss scale halved".
• 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.

• DeepSpeed ZeRO and optimizer sharding
• 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 "Overflow detected during fp16 gradient unscaling. Skipping step, loss scale halved" and preserve at least 100 lines before it.
• ✓Identify which rank, node, device, or process emitted the first related warning.
• ✓log the gradient norm per step. A norm that grows without bound is an optimization problem (learning rate, warm-up, missing gradient clipping), not a precision problem. A norm that is fine while the scale still collapses points at a single layer overflowing. LayerNorm and attention logits are the usual sites.
• ✓Repeat the same input after the targeted change and require the signature to disappear.
• ✓Resume from last checkpoint before the scale began collapsing only after the control passes.

Overflow detected during fp16 gradient unscaling. Skipping step, loss scale halved

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.

distinguish transient from terminal, if the scale recovers and training continues, this is normal warm-up behaviour. If it keeps halving toward the minimum, stop the run. This changes the condition in the causal diagnosis instead of hiding the outer exception. The repeated control proves ownership before recovery from last checkpoint before the scale began collapsing.

# 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 -- "Overflow detected during fp16 gradient unscaling. Skipping step, loss scale halved" <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.

• FP16 gradient overflow forced the loss scaler to halve repeatedly. Occasional skipped steps early in training are normal; a scale that keeps collapsing means gradients are genuinely exploding or the model is numerically unstable in FP16.
• The decisive evidence is the first log line that precedes "Overflow detected during fp16 gradient unscaling. Skipping step, loss scale halved" and differs from a healthy run.
• A nearby failure remains a competing hypothesis until a control separates configuration, capacity, transport, and hardware causes.