PyTorch's MetaConverter Bug Was Silently Corrupting AI Model Memory — The Patch Just Landed

A critical bug in PyTorch’s `MetaConverter` has been fixed that could silently corrupt model weights during export. The issue, reported as #168908, affected `nn.Parameter`s created from tensor views. When a parameter shares backing storage with another parameter—for example, a view of a larger tensor that is later wrapped into a `Parameter`—`MetaConverter` incorrectly memoized the smaller, offset‑zero tensor’s storage as representative of the entire source storage. During `draft_export` with real tensor propagation, subsequent parameters pointing to the same backing storage but with nonzero offsets would then reuse this undersized storage, triggering an out‑of‑bounds error in `real_tensor.set_`.

The problem was compounded by symbolic storage sizes: Python’s attempt to bool‑cast symbolic comparisons made the representative‑storage check unsafe. The patch, authored by jansel, introduces a conservative memoization rule: only when the freshly allocated storage is statically known to cover the full source storage will it be cached. For all other cases—particularly when storage sizes are symbolic or non‑trivially shared—the code falls back to the existing `meta_storage(s) + set_` path, which always allocates storage large enough for the entire source. This preserves the fast path for simple tensors while preventing the corruption for views and parameters with offsets. The fix includes new tests (`test_parameter_views_keep_full_storage` and `test_shared_storage_parameters_with_offset`) to lock in correct behavior. As deep‑learning models grow and quantization/export pipelines become more aggressive, such low‑level memory correctness patches are essential for production reliability.