Docs: operations cookbook (memory, coordinate manager, DDP, torch.compile, determinism)

[alpsaur]

Summary

• Add docs/operations.md: a practical "running ME in production/training" cookbook covering the recurring pain points mined from ~250 upstream and downstream issues
• Register the page in the docs/index.rst toctree (Miscellanea section, after the performance guide)
• Add one linking line in README.md below the existing performance-guide paragraph

Sections

1. Memory management with varying point counts - torch.cuda.empty_cache() on an interval; the OOM-leak caveat from CUDA OOM error leads to GPU memory leak NVIDIA/MinkowskiEngine#359 (C++-side OOM strands GPU memory; restartable subprocess loop for unattended training)
2. Coordinate-manager lifecycle - ME.clear_global_coordinate_manager() required every iteration under SHARE_COORDINATE_MANAGER mode, or the kernel-map cache grows unbounded
3. Multi-GPU / DDP recipe - DataParallel is NOT supported (0.5 parallel.replicate cannot replicate dimension attribute for convolution NVIDIA/MinkowskiEngine#264); full DDP recipe with torchrun, NCCL, DistributedSampler(drop_last=True), find_unused_parameters=True; pickling fix from PR Fix pickling error for MinkowskiConvolutionFunction during ddp training NVIDIA/MinkowskiEngine#139. Honest note: multi-GPU is upstream functionality this fork has not re-validated on torch 2.9; single-GPU is the tested path
4. torch.compile / Dynamo - ME ops are C++ autograd Functions with opaque handles; Dynamo cannot trace. Workaround: torch._dynamo.disable() on the sparse backbone, compile only the dense head
5. Determinism - two footguns: (a) GPU vs CPU quantization produces different coordinate sets (Coordinates ordering on CPU vs GPU NVIDIA/MinkowskiEngine#441); quantize on one device consistently; (b) conv output varies in low-order bits due to atomics (Non Reproducible Outputs on GPU when using MinkowskiConvolution and stride > 1 NVIDIA/MinkowskiEngine#554); manual coord-sort workaround described, built-in ME.set_deterministic() noted as planned but not yet merged
6. ME_LAZY_SYNC single-stream safety - the flag assumes all ME work is on the default PyTorch CUDA stream; unsafe with custom side streams or CUDA-graph capture; off by default

Style matches docs/performance.md: code-snippet-first, concise, no em dashes. Upstream issue numbers cited inline as plain links.

Test plan

• Verify docs/operations.md renders correctly in the hosted Sphinx docs
• Verify docs/index.rst toctree entry is correctly placed
• Verify README link resolves to the hosted operations page
• CI green (docs build / lint checks)

Generated with Claude Code