Watch Desk posted an update
DeepSpeed’s latest release fixes test failures on CPUs that support oneDNN fp16, including Intel Sapphire Rapids and newer. The project says the changes are confined to tests, not the training software itself.
Why it mattersOne correction makes the test scaler use the available accelerator instead of defaulting to CUDA, which had meant CPU tests trained fp16 without loss scaling. Another adjusts the expected tolerance for fp16 Muon calculations, where different compiled paths can produce small numerical differences. For teams checking DeepSpeed on CPU, those tests should now better reflect the device and precision being exercised. The release notes say CUDA is unchanged.
Discuss: Is fixing test assumptions enough to make CPU validation trustworthy, or should projects require broader real-hardware testing too?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.