What changed

PyTorch 2.6 changed torch.load to default to weights-only deserialization, a security improvement that can break older checkpoints. The project also stopped publishing official Conda packages and introduced binary and ABI changes for selected Linux builds.

Why it matters

The release changed both a sensitive model-loading boundary and supported installation paths. Silent assumptions about pickle-compatible checkpoints, Conda provenance, native extensions, or Linux ABI compatibility could break training and serving workflows.

What to do

Inventory every torch.load call and checkpoint source. Keep weights-only loading where possible, explicitly allowlist required globals, and regression-test native extensions and packaging against a currently supported PyTorch release before migration.

Caveat

Version 2.6 is now historical, so production users should select a currently supported patched release rather than install 2.6. Loading untrusted checkpoints with weights_only disabled can permit arbitrary code execution.