Transformers adds small model fixtures to catch regressions

Hugging Face describes how Transformers runs integration tests on CPU with small fixture models—the UMT5 fixture uses 66 MB versus 1,227 MB for the source checkpoint. The approach follows a real incident: after SDPA was enabled in July 2026, UMT5's decoder could attend to future tokens, and the bug stayed in main for nearly two months.

• Hugging Face adds native TorchTPU PyTorch backend2026-10-20
• Hugging Face launches Trainium model support lookup tool2026-10-23
• Developer proposes speculative decoding using 2-bit draft plus RAM residual2026-10-22
• PyTorch 2.15 brings FP8 and faster decoding to Mac MPS2026-10-20
• Hugging Face adds native Trainium support via TorchNeuron2026-10-19