DeepSpeed Out of Memory
DeepSpeed OOM errors crash ZeRO-optimized training when memory optimization settings don't match model architecture.
DeepSpeed OOM errors crash ZeRO-optimized training when memory optimization settings don't match model architecture.
What this failure is
DeepSpeed Out of Memory is a Distributed Training failure seen during ML training runs. DeepSpeed OOM errors crash ZeRO-optimized training when memory optimization settings don't match model architecture. Common tags: Deepspeed, Oom, Zero, Memory.
Is this what broke your run? Paste your log.
You're reading about DeepSpeed Out of Memory. 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.
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.
Why it happens (the mechanism)
ZeRO stage 2 still keeps model parameters on every GPU; only optimizer states and gradients are partitioned. For very large models, peak activation memory can still exceed GPU capacity. ZeRO stage 3 partitions all states but requires collectives for every parameter access, creating temporary memory spikes that can cause OOM. DeepSpeed memory fragmentation: repeated offload/onload cycles fragment GPU memory, causing OOM after many steps even with sufficient aggregate free memory. Mixed-precision training with DeepSpeed AMP (FP16) doubles optimizer state memory (Adam holds fp32 copies). 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
- DeepSpeed training crashes with CUDA OOM despite using ZeRO stage 2 or 3
- The OOM appears at different points across different GPU counts
- DeepSpeed's memory optimizations don't seem to reduce memory consumption as expected
Common symptoms and what they mean
| Symptom | Why it happens |
|---|---|
| CUDA out of memory during DeepSpeed engine initialization | ZeRO stage 2 still keeps model parameters on every GPU; only optimizer states and gradients are partitioned. For very large models, peak activation memory can still exceed GPU capacity |
| RuntimeError: CUDA out of memory when ZeRO partitions optimizer states | ZeRO stage 3 partitions all states but requires collectives for every parameter access, creating temporary memory spikes that can cause OOM |
| DeepSpeed forward pass OOM on the first batch | DeepSpeed memory fragmentation: repeated offload/onload cycles fragment GPU memory, causing OOM after many steps even with sufficient aggregate free memory |
| Memory usage in nvidia-smi shows near-total utilization despite ZeRO being enabled | Mixed-precision training with DeepSpeed AMP (FP16) doubles optimizer state memory (Adam holds fp32 copies) |
Which systems are affected
- Large language models (7B-70B) trained with DeepSpeed ZeRO
- Multi-node DeepSpeed training with model parallelism
- DeepSpeed with CPU offload and NVMe offload stages
- DeepSpeed with activation checkpointing enabled
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: CUDA out of memory during DeepSpeed engine initialization
- ✓Verified signal present: RuntimeError: CUDA out of memory when ZeRO partitions optimizer states
- ✓Verified signal present: DeepSpeed forward pass OOM on the first batch
- ✓Verified signal present: Memory usage in nvidia-smi shows near-total utilization despite ZeRO being enabled
- ✓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 analysisNo 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 extensionDeepSpeed 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 sideRelated failures to investigate next
Root cause
- ZeRO stage 2 still keeps model parameters on every GPU; only optimizer states and gradients are partitioned. For very large models, peak activation memory can still exceed GPU capacity
- ZeRO stage 3 partitions all states but requires collectives for every parameter access, creating temporary memory spikes that can cause OOM
- DeepSpeed memory fragmentation: repeated offload/onload cycles fragment GPU memory, causing OOM after many steps even with sufficient aggregate free memory
- Mixed-precision training with DeepSpeed AMP (FP16) doubles optimizer state memory (Adam holds fp32 copies)
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.
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.