ComfyUI WorldBETA
Sign in
Reference only

Float8 is unsupported on MPS: distinguish stored weights from compute dtype

Locate 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.

Updated 2026-09-24Published revision: 4Source baseline: not recorded

Symptoms and scope

Locate 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.

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

Float8_e4m3fn
does not have support for that dtype

Source-supported context

ComfyUI issue #8988 records a Float8_e4m3fn-to-MPS failure. PyTorch's conditional MPS fallback does not automatically turn arbitrary CUDA/Float8 algorithms into equivalent CPU implementations. Source 1 Source 2 Source 3

Cases to distinguish

1. The loader tries to compute directly in a dtype the backend does not support.

2. A custom node hard-codes device or dtype choices outside the core's adaptation path.

3. The quantization format needs its own loader; FP8, GGUF and other formats are not interchangeable because they are all lower precision.

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 exact filename, model family, loader class and Torch version. Capture the first dtype error rather than the final queue message.

Step 2. Check Mac support for that specific model and loader, including the intended storage/compute precision pair.

Step 3. In a copy, try a documented non-FP8 weight or the correct dequantizing loader. Check memory requirements before downloading larger weights.

Step 4. If the minimal example still fails, report file identity and versions rather than adding .float() to every node or blindly forcing fallback.

Completion check

The loader no longer sends an unsupported dtype to MPS and a small task completes. Loading alone does not test all operators.

Limitations and cautions

The cited report is time-specific, not a permanent claim about future PyTorch support. Installing NVIDIA-only CUDA wheels is not a Mac repair.

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. Float, Half and BFloat16 mismatch: trace the failing operationmat1 and mat2 must have the same dtypeA dtype failure differs from a shape or device failure. Start with the traceback and incompatible operand types; a KSampler error alone does not identify the component that caused them.
  2. A GGUF file will not load: match the architecture and the specialized loaderUnetLoaderGGUFGGUF is a container format, not a guarantee that an arbitrary chat-model file can serve as a diffusion model or text encoder in a particular ComfyUI node.
  3. MPS backend out of memory on Mac: account for unified-memory pressureMPS backend out of memoryOn Apple Silicon, applications, browsers, model weights and tensors share memory resources. Total installed memory is not all available to one generation task.
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.

01Float8_e4m3fn unsupported on MPSSource checked: 2026-09-2102MPS environment variablesSource checked: 2026-09-2103PyTorch local installationSource checked: 2026-09-21

Report an issue · 8191a873-08ab-547c-b2ba-102a4c681867