tqchen opened a new pull request, #610:
URL: https://github.com/apache/tvm-ffi/pull/610
[FIX] is_contiguous returns True for empty tensors (numel() == 0)
An empty tensor — one where any shape dimension is 0 — has no elements
and its stride values carry no meaning. The previous implementation
iterated over strides and returned False when it encountered a stride
that did not match the expected row-major stride (e.g. strides all-zero
for a shape like (4, 0, 4)), which is incorrect.
This commit adds an early return when numel == 0 to both:
- C++: `IsContiguous(const DLTensor&)` in
`include/tvm/ffi/container/tensor.h` (checks any shape dim == 0
before the stride loop)
- Python/Cython: `is_contiguous()` in
`python/tvm_ffi/cython/tensor.pxi` (same shape-zero scan before the
stride loop)
Semantics now match NumPy and PyTorch: an empty tensor is trivially
contiguous.
Tests added:
- C++: `TEST(Tensor, EmptyTensorIsContiguous)` in
`tests/cpp/test_tensor.cc` — constructs a strided empty tensor with
strides (0,0,0) and asserts `IsContiguous() == true`.
- Python: `test_empty_tensor_attributes` in
`tests/python/test_tensor.py` — existing test updated to use the
correct numpy-reported strides (0,0,0) for shape (4,0,4) and now
asserts `is_contiguous() == True`.
Closes https://github.com/apache/tvm-ffi/pull/607#issuecomment-4634436950
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
