Is this a hardware or a software issue?

Software failures can look like either layer depending on the symptom. The discriminator is the resolution of the symptom to the recommended fix: if applying the steps under "How to fix it" eliminates the symptom, this is a software/configuration class issue regardless of the hardware it surfaced on. Repeat the fix on a fresh node to confirm.

How long does the fix typically take?

Most resolutions from this entry take 10 to 30 minutes of engineer time once the diagnosis is confirmed. Long fixes come from skipping the diagnosis step and chasing apparent symptoms (validation drift, memory growth, slow throughput) instead of the actual root cause listed above. Apply the fix in a small experiment first, then roll it forward.

Does this affect inference or only training?

Primarily this is a software failure observed during training, but the same misconfiguration can leak into downstream evaluation and serving. Treat the prevention checklist as a guardrail: any pipeline that touches a trained checkpoint should re-validate the failure mode before promoting the model.

What's the cheapest way to confirm this is the problem?

Cheapest signal: reproduce the symptom on a 1-GPU toy run with the same configuration. If the toy run exhibits the same behavior, the cause is the configuration rather than the cluster. If only the multi-GPU run is affected, the failure is a distributed interaction, so DeepSpeed is the most likely layer to inspect.

Can the Denpex platform detect this automatically?

Yes. Denpex correlates the symptoms in this entry against your training logs in real time and surfaces the failure taxonomy, root cause, and exact remediation command on the console. Three full diagnoses are included free during signup, and the encyclopedia entry you are reading is generated from the same signal library that powers the platform.

What related failures should I check next?

Closely related failures cluster around #Gradient Explosion. After applying this fix, validate against the related entries below to confirm no overlapping cause is masking as this one. Engineers routinely discover that the dominant failure mode is actually a second, sibling failure compounded on top.

Will this recur after the fix?

Applied once with the prevention checklist, this failure does not recur under normal operation. The recurring variant most engineers encounter is the re-emergence of the same symptom on a different model or stack after Gradient Explosion is upgraded. Pin compatible versions and rerun the prevention smoke test on every dependency change.

Is there a fingerprint I can monitor?

Yes. Training halts abruptly with a deepspeed internal error when calculating the gradient norm for clipping. is the highest-signal fingerprint. Wire a metric or alert on that signal (or its absence after the fix) so you can detect the recurrence automatically. The Denpex platform encodes this exact fingerprint and surfaces it directly in the dashboard.

What is the official documentation that covers this?

Reference documentation is listed in the References section at the bottom of this entry. The most authoritative URLs are the upstream framework docs and the issue trackers cited; the Denpex entry synthesizes both into the executive, engineer, and on-call perspectives so each audience can act without reading every linked source.

Does this entry include working code?

Yes. The "Code Examples" section contains copy-paste-ready snippets that target the most common stacks (PyTorch, Hugging Face Trainer, DeepSpeed, Megatron-LM, vLLM/Transformers). Each snippet is annotated with why it works so you can adapt the pattern to your own framework without losing fidelity.

DeepSpeed ZeRO-3 Stale Sync during Grad Clipping

Q: What does this failure look like in production?

In production this failure typically appears during the first tens-to-hundreds of training steps: loss curves become non-monotonic, validation metrics plateau above the baseline, and downstream checkpoints either diverge or fail to improve. Engineers frequently misread it as data drift when the underlying mechanism is in zero stage 3, parameters and gradients are partitioned. when computing the global gradient norm for clipping, all ranks must synchronize their local norms. if there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing nans. the global all-reduce aggregates these nans, causing the entire step to fail.. Validating against the symptom list in this entry is the fastest way to confirm the diagnosis.

Q: How does this scale from 1 GPU to 1000 GPUs?

At single-GPU scale the symptom usually looks like slow convergence or a single anomalous loss value. At 64+ GPU scale, distributed interactions amplify in zero stage 3, parameters and gradients are partitioned. when computing the global gradient norm for clipping, all ranks must synchronize their local norms. if there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing nans. the global all-reduce aggregates these nans, causing the entire step to fail. through collectives, and the failure can manifest as a full job abort. Always reproduce the fix at the smallest cluster configuration that still exhibits the symptom before promoting to production scale.

In ZeRO Stage 3, parameters and gradients are partitioned. When computing the global gradient norm for clipping, all ranks must synchronize their local norms. If there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing NaNs. The global all-reduce aggregates these NaNs, causing the entire step to fail.

Quick answer

In ZeRO Stage 3, parameters and gradients are partitioned.

Symptom: AssertionError: Gradient clipping returned NaN
Root cause: In ZeRO Stage 3, parameters and gradients are partitioned. When computing the global gradient norm for clipping, all ranks must synchronize their local norms. If there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing NaNs.
Recommended fix: Adjust DeepSpeed Bucket Sizes and Prefetching { "zero_optimization": { "stage": 3, "stage3_prefetch_bucket_size": 0, "use_multi_rank_bucket_allreduce": false } } Disabling prefetch and multi-rank bucket allreduce simplifies the communication graph, preventing the sync issues that lead to uninitialized buffers being aggregated.
How Denpex helps: Denpex matches DeepSpeed ZeRO-3 Stale Sync during Grad Clipping across every rank in a distributed run and reports which rank failed first, so you act on the initiating node instead of the loudest one.

Software#Gradient Explosion

What this failure is

DeepSpeed ZeRO-3 Stale Sync during Grad Clipping is a Software failure seen during ML training runs. In ZeRO Stage 3, parameters and gradients are partitioned. When computing the global gradient norm for clipping, all ranks must synchronize their local norms. If there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing NaNs. The global all-reduce aggregates these NaNs, causing the entire step to fail. Common tags: Gradient Explosion.

Live diagnosis, no signup

Is this what broke your run? Paste your log.

You're reading about DeepSpeed ZeRO-3 Stale Sync during Grad Clipping. Paste your own crash log or traceback below and get the real root cause for YOUR run, not this generic entry. No account, no card. Logs are masked at ingress and never saved to account history.

training_logs.txt

No log to hand? Try one:

3 free diagnoses/day

Want 14 days on the Scale plan?

Request an evaluation code. A verified workplace organization activates up to 50 diagnoses a day, alerts, history, and follow-up questions. No credit card or automatic subscription.

Evaluate one incident

Why it happens (the mechanism)

It looks like the model itself exploded due to hyperparameter choices, but the NaNs actually originated from a distributed communication bug where empty or un-synced buffers were reduced.

What you'll observe

AssertionError: Gradient clipping returned NaN
DeepSpeed gradient norm is NaN

Common symptoms and what they mean

Symptom	Why it happens
Training halts abruptly with a deepspeed internal error when calculating the gradient norm for clipping.	In ZeRO Stage 3, parameters and gradients are partitioned. When computing the global gradient norm for clipping, all ranks must synchronize their local norms. If there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing NaNs. The global all-reduce aggregates these NaNs, causing the entire step to fail.
Only occurs on multi-node setups running ZeRO-3, whereas ZeRO-2 might train perfectly fine.	In ZeRO Stage 3, parameters and gradients are partitioned. When computing the global gradient norm for clipping, all ranks must synchronize their local norms. If there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing NaNs. The global all-reduce aggregates these NaNs, causing the entire step to fail.

Symptom

Why it happens

Training halts abruptly with a deepspeed internal error when calculating the gradient norm for clipping.

Only occurs on multi-node setups running ZeRO-3, whereas ZeRO-2 might train perfectly fine.

Which systems are affected

DeepSpeed
ZeRO-3
PyTorch

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.

✓Disable gradient clipping temporarily; if training continues without NaNs (though unstable), the issue lies in the clipping synchronization.
✓Check NCCL logs for asynchronous timeouts or errors on specific ranks.

Searchable error signature

search key

AssertionError: Gradient clipping returned NaN

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.

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

DeepSpeed errors in context

DeepSpeed changes when parameters, gradients and optimizer state are created, partitioned, gathered and offloaded. The hub separates ZeRO, memory, checkpoint and pipeline failures by lifecycle phase.

Compare every deepspeed error side by side

Root cause

In ZeRO Stage 3, parameters and gradients are partitioned. When computing the global gradient norm for clipping, all ranks must synchronize their local norms. If there is a communication timeout, or if some ranks process empty partitions in a specific mixed-precision context, a rank might send uninitialized or stale gradient buffers containing NaNs. The global all-reduce aggregates these NaNs, causing the entire step to fail.

The fix and how to prevent it

Evaluate Denpex on your own logs

Request a Scale evaluation code. A verified workplace organization activates 14 days with up to 50 diagnoses a day. Every account keeps its current diagnosis allowance and gets a verification path. No card or automatic subscription.

References

↗DeepSpeed ZeRO-3 Grad Norm NaN

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 Browse all failures

Related Software errors

Watchdog Timeout from Inconsistent Tensor Shapes

Software · medium

Tensor Parallel RNG State Context Divergence Deadlock

Software · high

Xid 13: Graphics Engine Exception due to Kernel Out-of-Bounds Access

Software · high

Adam Epsilon Underflow in FP16 Mixed Precision

Software · critical