ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf

Quick answer

ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf means The AMP gradient scaler reached a state where every gradient was zero or non-finite, so the scale update divided by zero. The scaler is reporting a numerical collapse upstream, not causing one. Preserve the first preceding error, then run the targeted control below.

Symptom: ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf
Root cause: The AMP gradient scaler reached a state where every gradient was zero or non-finite, so the scale update divided by zero. The scaler is reporting a numerical collapse upstream, not causing one. The decisive evidence is the first log line that precedes "ZeroDivisionError float division by zero GradScaler.
Recommended fix: do not patch the scaler. Find the step where gradients first became non-finite, log torch.isfinite on the gradient norm each step and locate the first failure.
How Denpex helps: Denpex investigates ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf 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.

Training Stability#numerics#adamw#grad#scaler#zero#division

What this failure is

The literal signature is "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf". It is a training stability failure associated with mixed precision, gradient scaling, and numerical kernels. The line identifies the failing operation or subsystem, while the surrounding evidence decides whether it is the initiating fault or a downstream symptom.

Live diagnosis, no signup

Is this what broke your run? Paste your log.

You're reading about ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf. Paste your own traceback and relevant evidence for an investigation of your workload, with a next action or a specific missing fact. A reference entry does not establish your cause. No account or card for the free diagnosis. Review data handling before submitting sensitive logs.

Before uploading, review cloud data handling and local options.

training_logs.txt

No log to hand? Try one:

3 free diagnoses/day

Why it happens (the mechanism)

The AMP gradient scaler reached a state where every gradient was zero or non-finite, so the scale update divided by zero. The scaler is reporting a numerical collapse upstream, not causing one. 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.

What you'll observe

The workload stops or loses forward progress after emitting "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf".
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.

Common symptoms and what they mean

Symptom	Why it happens
ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf	The AMP gradient scaler reached a state where every gradient was zero or non-finite, so the scale update divided by zero. The scaler is reporting a numerical collapse upstream, not causing one.
The same operation fails at a consistent stage of mixed precision, gradient scaling, and numerical kernels.	The decisive evidence is the first log line that precedes "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" and differs from a healthy run.
The first related warning appears before the final exception and names the causal subsystem.	A nearby failure remains a competing hypothesis until a control separates configuration, capacity, transport, and hardware causes.
A known-good control changes one variable and either reproduces or clears the failure.	The AMP gradient scaler reached a state where every gradient was zero or non-finite, so the scale update divided by zero. The scaler is reporting a numerical collapse upstream, not causing one.

Which systems are affected

mixed precision, gradient scaling, and numerical kernels
production-shaped multi-accelerator workloads
containerized and bare-metal deployments of the same stack

How to confirm this is the problem

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 "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" and preserve at least 100 lines before it.
✓Identify which rank, node, device, or process emitted the first related warning.
✓all-zero gradients usually mean the loss detached from the graph (an in-place op, a .detach(), or a branch that returns a constant). All-Inf usually means an exploding forward pass. Check which by inspecting a single gradient tensor at the failing step.
✓Repeat the same input after the targeted change and require the signature to disappear.
✓Resume from last checkpoint with finite gradients only after the control passes.

Root cause

The AMP gradient scaler reached a state where every gradient was zero or non-finite, so the scale update divided by zero. The scaler is reporting a numerical collapse upstream, not causing one.
The decisive evidence is the first log line that precedes "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" and differs from a healthy run.
A nearby failure remains a competing hypothesis until a control separates configuration, capacity, transport, and hardware causes.

The fix and how to prevent it

Searchable error signature

search key

ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf

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.

The fix and the prevention pattern

The root cause is on this page and stays free. A free account adds the exact remediation steps, saved history, and the fix on every entry in the encyclopedia.

No credit card. Daily allowance follows verified trust tier. Instant access.

Why the recommended fix works

do not patch the scaler. Find the step where gradients first became non-finite, log torch.isfinite on the gradient norm each step and locate the first failure. This changes the condition in the causal diagnosis instead of hiding the outer exception. The repeated control proves ownership before recovery from last checkpoint with finite gradients.

Code examples

snippet

# 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 -- "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" <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.

Best practices by model family

Model / Stack	Recommendation	Notes
First response	Preserve the first failure	Keep the context before "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" so aggregation does not erase causality.
Confirmation	Change one variable	Use a known-good node, rank, input, or configuration as the control.
Recovery	Resume from last checkpoint with finite gradients	Resume only after the literal signature no longer appears in the same control.

With the fix vs without the fix

Dimension	With the fix	Without the fix
Evidence	First preceding error and one controlled comparison	Only the final aggregated exception
Fix	do not patch the scaler. Find the step where gradients first became non-finite, log torch.isfinite on the gradient norm each step and locate the first failure.	Retrying the unchanged workload
Exit criterion	"ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" is absent in the repeated control	The job happened to run once

Diagnostic note

“Treat "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" as a search key and an investigation checkpoint, not as proof of every cause associated with the phrase. The high-value evidence is what changed immediately before it and whether the failure follows the workload, node, or configuration.”

Visual fingerprint

Decision path for ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf

literal error captured
        |
        v
find first preceding failure
        |
        v
run one known-good control
        |
        +-- follows workload --> inspect input or configuration
        +-- follows node ------> inspect hardware or platform
        +-- disappears --------> validate the targeted fix

The control separates workload, configuration, and node ownership before recovery from last checkpoint with finite gradients.

Diagnose this failure in VS Code

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.

Install the free VS Code extension

Related failures to investigate next

Frequently asked questions

Questions engineers and on-call staff commonly ask about this failure.

What does "ZeroDivisionError float division by zero GradScaler.step all gradients unscaled to zero or Inf" mean?

The AMP gradient scaler reached a state where every gradient was zero or non-finite, so the scale update divided by zero. The scaler is reporting a numerical collapse upstream, not causing one.

Is this line always the root cause?

No. It can be the direct failure or the point where an earlier failure becomes visible. The first preceding error and a controlled comparison decide which.

What should I collect before restarting?

Collect complete log context, the emitting rank or node, component versions, resolved configuration, and the diagnostic output shown above.

What is the fastest confirmation?

all-zero gradients usually mean the loss detached from the graph (an in-place op, a .detach(), or a branch that returns a constant). All-Inf usually means an exploding forward pass. Check which by inspecting a single gradient tensor at the failing step.

How do I prevent it from recurring?

prefer BF16, which needs no loss scaling at all on Ampere and later. If FP16 is required, add gradient clipping and verify the loss is finite before backward().

References

Don't just read the fix, diagnose your run

The encyclopedia tells you what went wrong. Denpex tells you what went wrong in YOUR training run. With your logs, your config, and your stack.

Diagnose your logs free Compare a solved incident Browse all failures

Related Training Stability errors

NaN Detection and Skip

Training Stability · high

LR Without Warmup or Decay

Training Stability · high

Learning Rate Finder Result Misuse

Training Stability · medium

Gradient Accumulation BatchNorm Issue

Training Stability · high