René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"Bump tensor bridge version cap for PyTorch 2.12","articleBody":"## Summary\n\nThe tensor bridge's `_torch_version_check()` in `_memoryview.pyx` caps the supported PyTorch version range at `(2, 3) \u003c= (major, minor) \u003c= (2, 11)`. PyTorch 2.12 was released on ~May 14, 2026, causing the version check to fail. When this happens, `_is_torch_tensor()` returns `False` and torch tensors fall through to the generic CAI/DLPack paths instead of the fast AOTI tensor bridge.\n\nFor `from_cuda_array_interface`, this causes a hard failure because PyTorch reports `__cuda_array_interface__` version 2, while cuda-python requires version 3+:\n\n```\nBufferError: only CUDA Array Interface v3 or above is supported\n```\n\n## Root cause\n\n```python\n# _memoryview.pyx\ncdef inline bint _torch_version_check():\n    ...\n    _torch_version_ok = (2, 3) \u003c= (major, minor) \u003c= (2, 11)  # ← needs bumping\n```\n\nThe upper bound exists because the tensor bridge relies on undocumented PyTorch internals (`THPVariable` struct layout, `AtenTensorHandle == at::Tensor*`). The docstring says: \"Bump the upper bound after verifying a new PyTorch release.\"\n\n## Action needed\n\n1. Verify PyTorch 2.12's `THPVariable` struct layout is unchanged (check `torch/csrc/autograd/python_variable.h`)\n2. Bump the upper bound to `(2, 12)`\n3. Re-add `latest` to the nightly CI pytorch test matrix (currently pinned to `2.11.0` as a workaround in #1987)\n\n## Discovered by\n\nNightly optional-dependency CI (#1987) — the `latest` torch entry started failing when PyTorch 2.12 was released.\n\n-- Leo's bot","author":{"url":"https://github.com/leofang","@type":"Person","name":"leofang"},"datePublished":"2026-05-15T02:36:58Z","url":"https://github.com/NVIDIA/cuda-python/issues/2089"}