Reference only

MPS backend out of memory on Mac: account for unified-memory pressure

On Apple Silicon, applications, browsers, model weights and tensors share memory resources. Total installed memory is not all available to one generation task.

Symptoms and scope

On Apple Silicon, applications, browsers, model weights and tensors share memory resources. Total installed memory is not all available to one generation task.

Error fragments for search; IDs, values and filenames may vary:

MPS backend out of memory
PYTORCH_MPS_HIGH_WATERMARK_RATIO

Source-supported context

PyTorch documents the MPS high-watermark limit and warns about system-level risk when disabling it. ComfyUI issue #7171 is a particular OOM report, not an endorsement of every suggested environment-variable value. Source 1 Source 2 Apple Activity Monitor

Cases to distinguish

1. Image size, batch size or video length exceeds the current memory headroom.

2. Multiple applications or models compete for unified memory.

3. A version-specific node, fallback or memory-residency problem differs from the native baseline.

Diagnostic sequence

The sequence below is an editorial procedure based on the cited context, not a diagnosis already confirmed for your environment.

Step 1. Record Mac memory capacity, Torch, failing node and task settings. Inspect memory pressure and swap use in Activity Monitor.

Step 2. Save work, stop unnecessary memory-heavy applications and concurrent generation. If the specific workflow and model support it, reduce one input factor such as batch count, dimensions or frame count for a controlled test.

Step 3. Use a documented setting to reduce model residency or decode load only when the failure stage and workflow support it. Use an MPS-compatible workflow; unsupported FP8 weights are not an automatic memory workaround.

Step 4. Adjust limits only after understanding the version's documentation and keeping a rollback route. Report failures of a minimal official workflow with logs.

Completion check

The task completes while the machine remains responsive and memory pressure does not keep worsening. Increase input gradually.

Limitations and cautions

Do not make a zero high-watermark limit the default recommendation. CPU fallback neither supplies unlimited RAM nor implements every unsupported dtype.

Original sources

Localization prepared: 2026-09-22. The source-check dates above were inherited from the existing article; this translation does not claim they were all rechecked today. No GPU run, minimum-memory measurement or runtime guarantee is supplied.

Didn't solve it?

Check the next possible cause

The same symptom can come from a different cause. Work through these related entries in order.

  1. Float8 is unsupported on MPS: distinguish stored weights from compute dtypeFloat8_e4m3fnLocate whether the failure happens during weight loading, dequantization or an operation. A file's FP8 label alone does not describe every possible quantized execution path.
Paste your full log into the error finder

Was this page helpful?

Anonymous. Only a yes/no counter is stored; no account, IP address or device details.

Sources & references

Localized from the existing Chinese research draft. Original source-check dates are retained; this translation is not new technical verification or a runtime test.

01MPS environment variablesSource checked: 2026-09-2102MPS backend out of memorySource checked: 2026-09-2103Apple Activity Monitor memory usageSource checked: 2026-09-25

Report an issue · b6b344f0-5dfd-5d61-ace6-7cc057fcd74f