Skip to content

v0.5.9 — reliability and validated CUDA wheels

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 15 Sep 06:48
v0.5.9
662a120
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Reliability fixes

  • Fix a GPU coordinate-publication race that could create duplicate, zero-feature rows in generated maps, and repair empty-map lifetime/launch handling.
  • Preserve residual, explicit-output-target, and TensorField relationships in deterministic convolution; cache canonical maps and use a per-call non-fused backend.
  • Honor explicit targets even for expanding/generative convolutions.
  • Embed native torch/CUDA/ABI build facts, reject incompatible runtimes early, and make standalone diagnostics tolerate older installations.
  • Repair public gradcheck and independent analytic tests; make long real-data benchmarks opt-in.
  • Add a reproducible cold/warm U-Net benchmark and remove interpolation's host scalar synchronization, with a fast path for full neighborhoods.

Determinism covers convolution forward values on the same device/build/math settings, not bitwise deterministic training or all sparse operations. This remains a best-effort fork.

Wheel compatibility

Four wheels: CPython 3.10, 3.11, 3.12, 3.13, Linux x86-64, RTX 50-series / sm_120, CUDA 12.8, torch 2.9.x cu128. All were built with torch 2.9.1+cu128 and the C++11 ABI enabled.

These are not universal manylinux wheels. Built on Ubuntu 24.04; they require glibc 2.38 or newer and the OpenBLAS runtime (libopenblas.so.0, e.g. Ubuntu's libopenblas0 package). Torch supplies the matching CUDA libraries; a compatible NVIDIA driver is required for GPU execution. Use a source build on older distributions or other GPU architectures.

Install the intended torch runtime first, then the wheel matching your Python:

python -m pip install "torch~=2.9.0" --index-url https://download.pytorch.org/whl/cu128
python -m pip install ./minkowskiengine-0.5.9-cp312-cp312-linux_x86_64.whl  # Python 3.12 example

Validation

  • All eight CPU CI combinations passed: Python 3.10–3.13 × torch 2.7/2.9. Post-merge CI and documentation builds also passed.
  • All four release wheels passed isolated install/import/CPU smoke checks before being attached to the draft. Downloaded asset SHA-256 digests and build metadata were checked.
  • The actual downloaded cp312 release wheel, installed outside the checkout, passed CPU/CUDA smoke and the regression suite on Python 3.12.3, torch 2.9.1+cu128, RTX 5090: 142 passed, 15 skipped, 102 subtests passed. The other three wheels were CPU-smoke-tested in CI, not GPU-tested.
  • Ten independent local determinism/generated-map runs passed, each including 40 generated-map trials. Independent whole-diff review plus native and packaging scouts completed; both findings were repaired and independently confirmed.

The 15 regression skips are 12 opt-in real-data/stress jobs and three pre-existing legacy API/raw-gradcheck tests; they are not passes.

Full changelog · Validation/release procedure · Merged PR #19

Assets 6
Loading