[Bug] CUDA sparse-conv backend fails to compile with CUDA 12.8 / torch 2.8 / GCC 11+

[WhaleK1123]

[Bug] CUDA sparse-conv backend fails to compile with CUDA 12.8 / torch 2.8 / GCC 11+

Summary

MinkowskiEngine master (0.5.4, commit 02fc608) does not compile its CUDA
backend on a modern toolchain: CUDA 12.8 (thrust/CCCL), torch 2.8, GCC 11 or 13.
A set of small compatibility fixes is available (see below) that lets it build and
run the sparse-conv backend on an RTX 5070 Ti (sm_120).

Environment

• OS: Ubuntu 22.04
• Host compiler: GCC 11.4 (also reproduced on GCC 13)
• CUDA: 12.8 (nvcc 12.8.93, supports compute_120)
• PyTorch: 2.8.0+cu128
• Build: SETUPTOOLS_USE_DISTUTILS=stdlib, TORCH_CUDA_ARCH_LIST="12.0",
  CC=gcc-13 CXX=g++-13

Steps to reproduce

git clone https://github.com/NVIDIA/MinkowskiEngine.git
cd MinkowskiEngine
pip install torch==2.8.0+cu128 numpy<2 setuptools wheel
export CUDA_HOME=/usr/local/cuda-12.8 TORCH_CUDA_ARCH_LIST="12.0" SETUPTOOLS_USE_DISTUTILS=stdlib
export CC=gcc-13 CXX=g++-13
python setup.py build_ext --inplace

Errors encountered (in order after each previous fix)

1. namespace "thrust" has no member "sort_by_key" (also unique, sequence,
   reduce, etc.) in src/spmm.cu, src/coordinate_map_gpu.cu, pooling kernels.
   New thrust in CUDA 12.8 requires explicit includes; MinkowskiEngine relied
   on the old transitive includes.

2. identifier "CUDF_FUNC_RANGE" is undefined in
   src/3rdparty/concurrent_unordered_map.cuh. The bundled cudf nvtx macro was
   disabled when its headers are removed.

3. std::__to_address overload ambiguity in <bits/shared_ptr_base.h>
   (both GCC 11 and GCC 13), triggered by
   concurrent_unordered_map::create() returning a std::unique_ptr that is
   converted to std::shared_ptr. This is the hard blocker.

Root causes

• MinkowskiEngine (2021) predates the CUDA 12.x thrust/CCCL change that made
  algorithm headers non-transitive.
• The bundled nvtx3.hpp/ranges.hpp collide with torch's nvtx3.
• default_allocator / c10_allocator were only minimally declared (no
  pointer/rebind/operator== etc.), which modern libstdc++ shared_ptr
  fails on.
• concurrent_unordered_map::create() returns unique_ptr; the
  unique_ptr -> shared_ptr conversion hits a libstdc++ std::__to_address
  overload-ambiguity bug on both GCC 11 and 13.

Proposed fix (verified, builds on RTX 5070 Ti)

• Add explicit thrust includes (<thrust/sort.h>, <thrust/unique.h>,
  <thrust/sequence.h>, <thrust/reduce.h>, <thrust/remove.h>,
  <thrust/count.h>, <thrust/copy.h>, <thrust/fill.h>,
  <thrust/device_vector.h>, <thrust/iterator/zip_iterator.h>,
  <thrust/tuple.h>, <thrust/execution_policy.h>).
• No-op the bundled nvtx3 headers and add #define CUDF_FUNC_RANGE() ((void)0).
• Make both allocators fully conformant (pointer, const_pointer, size_type,
  difference_type, rebind, operator==/!=).
• Change concurrent_unordered_map::create() to return std::shared_ptr<Self>
  (pointer + deleter) instead of std::unique_ptr.

After the fix, a real sparse convolution runs on GPU:

>>> MinkowskiEngine 0.5.4 GPU sparse conv LIVE on RTX 5070 Ti (CUDA 12.8, sm_120)
>>> is_cuda_available: True, cuda_version: 12080

A ready-made patch is in the fix/build-cuda12.8-gcc13 branch (commit 6fd7c64)
of my fork. I verified it compiles and runs on CUDA 12.8 / torch 2.8 / sm_120.

Notes

• This affects downstream users: FUSER (Jiang-HB/FUSER) pins
  MinkowskiEngine==0.5.4 and, on a Blackwell GPU, needs torch>=2.7 — which
  then requires this compatibility patch to compile. See the related
  FUSER requirements.txt issue #1.