PyTorch DeepSeek schema update deprecates v1 compatibility

The PyTorch project has announced changes to its DeepSeek indexing and contraction validation logic, as detailed in its GitHub release notes. This update addresses issues with scale storage assumptions and expands compatibility by honoring actual strides and logical axes. The v1 schema for DeepSeek native choices has been deprecated, with the v2 schema now being the default. Additionally, unsupported contraction axes will now be explicitly rejected rather than defaulting to a computed product. Tensorwise and rowwise paths remain unchanged, while native and fake metadata enforce stricter validation rules.

Operators should carefully review the schema changes and test their workloads against the updated validation logic before upgrading. Any reliance on the v1 schema or silent handling of unsupported contraction axes could lead to failures or unexpected behavior. Regression tests and compatibility checks are recommended, especially for deployments involving Hopper, XPU, or ROCm execution environments, which remain unverified locally according to the release notes.