Skip to content

NaN Loss from TransformerEngine FP8 Scaling Saturation

FP8 training with delayed scaling produces NaNs when an activation outlier saturates the scaling factor (amax history too short / margin too small). Loss and gradients are healthy until the exact step the scale saturates. Distinguishing it from LR- or data-driven NaN.

Quick answer

FP8 training with delayed scaling produces NaNs when an activation outlier saturates the scaling factor (amax history too short / margin too small).

Training Stability#nan-loss#fp8#transformer-engine#amax#delayed-scaling#h100

What this failure is

NaN Loss from TransformerEngine FP8 Scaling Saturation is a Training Stability failure seen during ML training runs. FP8 training with delayed scaling produces NaNs when an activation outlier saturates the scaling factor (amax history too short / margin too small). Loss and gradients are healthy until the exact step the scale saturates. Distinguishing it from LR- or data-driven NaN. Common tags: Nan Loss, Fp8, Transformer Engine, Amax.

Live diagnosis, no signup

Is this what broke your run? Paste your log.

You're reading about NaN Loss from TransformerEngine FP8 Scaling Saturation. 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)

DelayedScaling margin=0 with short amax_history on outlier-prone layers. Per-tensor scaling on layers with heavy-tailed activations. Skipped amax sync across TP ranks after topology change. Taken together, these mechanisms explain why the failure is reproducible, why it tends to surface on specific workloads or scales, and why generic mitigation attempts often fall short without addressing the underlying cause.

What you'll observe

  • loss=nan appears with stable prior grad_norm history
  • TE logs an amax/scaling warning for a specific layer shortly before
  • bf16 rerun of the same steps is clean

Common symptoms and what they mean

SymptomWhy it happens
amax_history overflow / scaling factor saturated messagesDelayedScaling margin=0 with short amax_history on outlier-prone layers
NaN first appears in one layer's activations then propagatesper-tensor scaling on layers with heavy-tailed activations
per-step numerics fine with fp8 disabledskipped amax sync across TP ranks after topology change

Which systems are affected

  • H100/H200 FP8 recipes (TransformerEngine)
  • Megatron/NeMo FP8 configs
  • custom DelayedScaling recipes

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.

  • Reproduce the failure from a clean checkpoint/seed: the symptom must appear without warm-up state from a previous run.
  • Verified signal present: amax_history overflow / scaling factor saturated messages
  • Verified signal present: NaN first appears in one layer's activations then propagates
  • Verified signal present: per-step numerics fine with fp8 disabled
  • A targeted fix from the "How to fix it" section eliminates or substantially reduces the symptom within one validation pass.

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.

Sign up free. Unlock the full analysis

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

Root cause

  • DelayedScaling margin=0 with short amax_history on outlier-prone layers
  • per-tensor scaling on layers with heavy-tailed activations
  • skipped amax sync across TP ranks after topology change

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.

We send a single-use code tied to that address. Static provider and TLD rules do not reject valid addresses. Account trust determines the benefit after signup.

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.