FlexGEMM v1.0.0

🎉 Initial Stable Release

FlexGEMM is a cross-platform backend for high-performance sparse convolutions, providing optimized Triton kernels with flexible autotuning for sparse 3D workloads.

This is the first stable public release of FlexGEMM.

Core Features

• Efficient sparse submanifold convolution (forward & backward)
• Grid sample with nearest or trilinear interpolation for sparse 3D tensors

Platform & Environment Support

• OS

  • Linux
  • Windows

• Python

  • Python ≥ 3.8

• Dependencies

  • PyTorch ≥ 2.4.0
  • Triton ≥ 3.2.0 (Linux)
  • triton-windows ≥ 3.2.0 (Windows)

Tests & Benchmarks

• Comprehensive test coverage for:

  • Sparse convolution (forward / backward)
  • Grid sample ops
  • Hash map & neighbor cache

• Benchmark scripts included for training & inference scenarios

Known Issues & Notes

• This is the first stable release; while core APIs and functionality are stable, performance tuning and edge-case fixes will continue in upcoming 1.0.x releases.
• Autotune cache behavior may evolve as more workloads are covered.
• Users are encouraged to report platform-specific issues, especially on Windows.

Acknowledgements

We thank @PozzettiAndrea, @geekuillaume, and @DQSSSSS for their contributions to cross-platform support, autotuning infrastructure, and uint64 hashing in FlexGEMM.