mlx-lattice-cuda13 0.1.9

CUDA 13 backend for mlx-lattice

MLX Lattice

Sparse point cloud convolution library for Apple MLX.

[!CAUTION] We're working on the proofing of math correctness between those operators in different backends. Currently, only the Metal operators have the identical math property, while the CUDA operators still need to be verified.

Operations

Currently, the following operations are supported:

• Tensor: SparseTensor, sparse_collate, cat, prune, topk_rows
• Coordinates: downsample, build_kernel_map, build_generative_map, build_transposed_kernel_map
• Features: linear, relu, sigmoid
• Sparse convolution: conv3d with stride, padding, and dilation; conv_transpose3d; generative_conv_transpose3d
• Sparse pooling: pool3d, max_pool3d, avg_pool3d, global_pool, global_sum_pool, global_avg_pool, global_max_pool
• Modules: Linear, Conv3d, ConvTranspose3d, GenerativeConvTranspose3d, SumPool3d, MaxPool3d, AvgPool3d, GlobalPool, GlobalSumPool, GlobalAvgPool, GlobalMaxPool, BatchNorm, ReLU, Sigmoid

Usage

import mlx.core as mx
import mlx_lattice as ml
import mlx_lattice.nn as mln

coords = mx.array(
    [[0, 0, 0, 0], [0, 1, 0, 0], [0, 2, 0, 0]],
    dtype=mx.int32,
)
feats = mx.array([[1.0, 0.0], [0.5, 1.0], [0.0, 2.0]], dtype=mx.float32)
x = ml.SparseTensor(coords, feats)

conv = mln.Conv3d(2, 8, kernel_size=3, bias=True)
pool = mln.SumPool3d(kernel_size=2, stride=2)

y = pool(conv(x))
mx.eval(y.feats)

Coordinates follow the sparse point convention (batch, x, y, z). The module weight layout follows MLX convolution modules: (out_channels, kx, ky, kz, in_channels).

Development

uv sync
uv run ruff check .
uv run ty check
uv build --wheel

The native extension is built with CMake, scikit-build-core, nanobind, and the MLX C++ backend toolchain. Metal builds are enabled on macOS; CUDA kernels are enabled on non-Apple hosts when CMake finds a CUDA compiler and toolkit.

For native editor indexing:

uv run cmake --preset clangd

Install and run hooks with:

prek install
prek run --all-files

License

Copyright © 2026 Yu

Open sourced under MIT license