Release v0.5.0

What's Changed

• Bump verion to 0.3.1 by @swahtz in #303
• shared memory types in IntegrateTSDF kernel by @swahtz in #307
• Viewer now supports multiple scenes by @phapalova in #308
• Update nanovdb_editor version. by @areidmeyer in #309
• Implementing morton and hilbert for Grid and GridBatch ijk by @blackencino in #311
• Google analytics in docs by @fwilliams in #312
• minor edits by @kmuseth in #305
• Fix radix sort prefetch for disjoint sorts by @matthewdcong in #315
• Rename _Cpp python extension library binary to _fvdb_cpp, include pybind11 headers first by @harrism in #313
• Morton/Hilbert fixes by @blackencino in #316
• Active grid coords cleanup by @blackencino in #318
• Use smaller runtime images for non-build CI steps by @matthewdcong in #320
• Make nightly workflows not run on forks of openvdb/fvdb-core by @harrism in #319
• Changed the _Cpp.pyi filename to _fvdb_cpp.pyi by @blackencino in #322
• Convolution default fixed, extensive tests by @blackencino in #321
• Morton/Hilbert module level standalone functions by @blackencino in #323
• conda environment files: switch to conda-forge torchvision by @swahtz in #327
• Fix documentation for setting TORCH_CUDA_ARCH_LIST by @matthewdcong in #328
• Eliminate snake_case in GaussianTileIntersection.cu by @harrism in #330
• Fix GHA error in (currently unused) nightly workflows by @harrism in #329
• Disable running tests for Draft PRs by @swahtz in #339
• build.sh debug build fixes by @swahtz in #338
• Add viz bindings for wait and add_image by @phapalova in #332
• Render all contributing gaussian IDs/weights by @swahtz in #340
• Fix Gaussian rasterization shared memory alignment by @swahtz in #342
• [Bug Fix] Improve binary search handling in JIdxForJOffsets by @iYuqinL in #325
• Debug build gtest fix by @swahtz in #344
• GaussianProjectionForward: fix camera data loading that exceeds blockDim by @swahtz in #345
• Support backgrounds in Gaussian Rasterization by @harrism in #343
• Remove tests.yml 'push' trigger by @swahtz in #347
• Temporarily disable nightly benchmark/test job by @swahtz in #346
• Fix potential oversubscription of nvcc * cmake parallel threads by @swahtz in #351
• Hot fixes for Jcat errors by @blackencino in #352
• Fix inverted logic on computing abs(gradient) in backward Gaussian rasterization by @harrism in #355
• A pedantic pull request by @swahtz in #359
• Fix JaggedTensor.from_*_and_list_ids ldim=2 issue by @swahtz in #357
• Gradients added to Convolution Ground Truth Unit Tests by @blackencino in #358
• JaggedTensor::unbind*: Reduce number of blocking GPU->CPU copies by reusing cached lsizes by @swahtz in #360
• Usescaled_dot_product_attention operator from Torch, remove our SDPA by @swahtz in #364
• Added backwards/gradient tests to the default convolution unit tests. by @blackencino in #361
• Add lineinfo option to build by @harrism in #367
• Create and use getMaxSharedMemory utility by @harrism in #368
• Remove unnecessary stream sync in GaussianTileIntersection by @harrism in #370
• plumb sparse rendering functions by @fwilliams in #348
• MCMC gaussian splatting relocation kernel and unit tests by @harrism in #374
• MCMC add noise kernel and gtests by @harrism in #377
• Bindings and Pytests for MCMC functions by @harrism in #394
• Expose min_opacity parameter in Gaussian MCMC relocate function. by @harrism in #396
• Expose k and t parameters in MCMC relocation by @fwilliams in #402
• Optimize joffsets construction via pinned memory by @matthewdcong in #403
• NVIDIA Branding in docs by @fwilliams in #405
• Fix hardcoded float dtype by @matthewdcong in #406
• Prefetch fused SSIM outputs to avoid write page faults by @matthewdcong in #407
• Fix the docstrings of fvdb.viz.Scene.camera_orbit_direction by @swahtz in #398
• Fix NaNs in rasterizeTopContributingGaussianIdsForward weights by @swahtz in #400
• Plumb Sparse Gaussian TileIntersection by @swahtz in #401
• Fix/improve radix sort synchronization by @matthewdcong in #409
• Fix the derivation of the number of cameras in the rasterization kernels that would be incorrect when in packed mode. by @swahtz in #412
• CMake: create nanovdb_editor_BINARY_DIR before using it as a working directory by @harrism in #416
• Fix crash loading GaussianPly files to CPU device by @swahtz in #417
• Switch radix sort merge from host to device side synchronization by @matthewdcong in #415
• Pin PyTorch version in CI to 2.9.1 by @matthewdcong in #424
• computeSparseInfo small optimization by @swahtz in #428
• Fix semantics of GaussianTileIntersection's torch::cumsum by @swahtz in #427
• Fix version macro and PyTorch 2.10 build error by @matthewdcong in #423
• Multi-Axis Dispatch Framework by @blackencino in #418
• SampleGridTrilinear: Vectorized float4 loads by @swahtz in #430
• Implement PrivateUse1 (multiGPU) support for MCMC kernels by @harrism in #421
• Rasterize contributing Gaussian ID kernels optimizations by @swahtz in #429
• Expose evaluate_spherical_harmonics in Python bindings by @swahtz in #431
• Fix: GaussianSphericalHarmonicsBackwards race condition for cameras/batch-size > 1 by @swahtz in #434
• Update CUDA 13 nightly test image by @matthewdcong in #438
• Disable two GCC 13.3 warnings by @matthewdcong in #439
• Improve PyTorch build configuration time by @matthewdcong in #441
• Switch to device-centric synchronization for forEach mGPU by @matthewdcong in #440
• Fix chain rule for log_scale gradient in projection backward pass by @harrism in #433
• Gaussian Projection with the Unscented Transform by @fwilliams in #420
• Build speedups, added trace and optional pip forced install by @blackencino in #443
• Fix: GaussianProjectionBackward dLossDQuat missing warpSum by @swahtz in #435
• Optimize Gaussian tile intersection for mGPU by @matthewdcong in #446
• dispatch framework for_each, views, and tag canonicalization by @blackencino in #452
• Fix: ProjectionForward initializes accessors in both the initializer list and constructor body by @swahtz in #450
• ProjectedGaussianSplats opacities uses expand/view and accessors instead of per-camera copies by @swahtz in #451
• Remove unused sparse convolution backends by @blackencino in #454
• Add developer worktree tools for parallel development workflows by @harrism in #445
• Add AGENTS.md for AI agent guidance by @harrism in #455
• Fix fvdb-issue clipboard crash in SSH sessions by @harrism in #461
• Fix CI checks showing 'waiting for status' when paths-ignore skips workflows by @swahtz in #462
• Fix JaggedTensor single-element constructor unconditionally initializing CUDA via pinned_memory by @swahtz in #468
• CI: Skip stopping runners if starting runners was skipped by @harrism in #471
• Skip stopping runners that were skipped for CUDA 12.8 and 13.0 tests. by @harrism in #472
• Add no-argument interactive mode to fvdb-issue by @harrism in #466
• Sparse Conv Default full feature support by @blackencino in #473
• Rasterization using 3d gaussians by @fwilliams in #444
• Update nanovdb to version 32.9.0 and refine grid type checks by @swahtz in #475
• Nightly build and publish action by @swahtz in #477
• Nightly publish fix for non-existent nightly packages by @swahtz in #478
• Updated nightly build install docs by @swahtz in #481
• Downgrade nano to 32.8.0 by @swahtz in #482
• Fix multibatch mGPU race condition in SH backwards op by @matthewdcong in #484
• Upgrade openvdb git tag for NanoVDB 32.9.0 by @swahtz in #483
• Fix datatype in backward projection test by @matthewdcong in #486
• Switch from inverse to linalg_inv_ex to avoid sync by @matthewdcong in #487
• Refactor CameraIntrinsics constructor and add missing include by @blackencino in #489
• SampleGridTrilinear optimization: stencil + sample by @swahtz in #474
• Refactor Gaussian rendering to use composable camera model for projection and ray generation by @fwilliams in #485
• Add masks support to all Gaussian render methods by @swahtz in #480
• Update NanoVDB to v32.9.1 by @swahtz in #493
• Feature/op consolidation by @blackencino in #492
• SaveNanoVDB: Fix voxel size/origin metadata on serialized index grids by @swahtz in #490
• viewer fix for notebook by @zlalena in #350
• Disable unreliable rasterization deadlock test by @fwilliams in #498
• SimpleUnet bug fixes by @swahtz in #496
• Added the github CLI to the conda dev environment by @blackencino in #500
• Handle duplicate pixels in sparse pixel gaussian rendering by @harrism in #488
• Add bfloat16 support to JaggedTensor reduce operators by @swahtz in #501
• Add seed initialization in TestSimpleUNet class by @swahtz in #502
• Add unit tests for fvdb.nn modules by @swahtz in #497
• Pin CI checkout refs to immutable commit SHAs to fix build/test skew by @swahtz in #503
• Docs, examples, notebooks udpates by @swahtz in #504
• Version class by @swahtz in #507
• Improve/optimize mGPU scaling via batched prefetching and sorting changes by @matthewdcong in #499
• Switch 32-bit tensor index accessors to 64-bit across all ops by @harrism in #505
• Fix CCCL version check macro by @matthewdcong in #509
• Add devtools script to report unanswered external issues by @harrism in #510
• Fix flaky test_jsum_list_of_lists bfloat16 test by @fwilliams in #517
• Update CONDA_OVERRIDE_CUDA to 13.0 by @swahtz in #519
• Add Slack output format and daily CI workflow for unanswered issues report by @harrism in #513
• Fix insider issues filtering for CI unanswered issues script by @harrism in #522
• Release Process Updates by @swahtz in #525
• Update CI workflow to include git installation in system dependencies by @swahtz in #526
• Add release branching docs and automation scripts by @harrism in #512
• Fix smoke test Python setup and open release PR as draft by @harrism in #528
• Make start-release.sh idempotent for safe re-runs by @harrism in #529
• Run unit tests only for matching test_environment.yml config by @harrism in #531
• Restore fix for dLossDQuat missing warpSum by @matthewdcong in #533
• Fix quaternion gradient accumulation in GaussianProjectionJaggedBackward by @swahtz in #534
• Fix publish workflow: Rocky Linux 8 containers and single unit test job by @harrism in #536
• Fix publish.yml python install action by @swahtz in #537
• Update publish.yml to include additional system dependencies by @swahtz in #538
• Align publish.yml build containers with tests.yml (Rocky Linux 8 / manylinux_2_28) by @swahtz in #540
• publish: dual S3 + PyPI publish on release (with GPU tests) by @swahtz in #545
• Improve mGPU partitioning for Gaussian projection operators by @matthewdcong in #547
• Improve mGPU partitioning for SH operators by @matthewdcong in #546
• nightly-publish fix errors for missing tools by @swahtz in #549
• Check for zero intersection case in tile intersection prefetch by @matthewdcong in #553
• Add automated doc version updates to release scripts by @harrism in #552
• finish-release-process.sh updates to preserve release/* branch integrity by @swahtz in #544
• Update NanoVDB Editor to latest. by @areidmeyer in #556
• Add CHANGES.md by @swahtz in #539
• Upgrade conda env files to gcc/gxx 14.3 by @swahtz in #557
• Replace Slack issue report with event-driven issue triage labels by @harrism in #551
• Added nanovdb-editor as an optional dependency by @swahtz in #559
• 0.4.2 Changelog and Docs update by @swahtz in #565
• Docs deployment workflow fix by @swahtz in #566
• Fuse computeGradientState into projectionBackwardsKernel by @matthewdcong in #560
• PyTorch 2.11 support for venv CI by @matthewdcong in #561
• Reduce shared memory usage in pinhole projection by re-arranging blocks by @matthewdcong in #555
• Centralize Github workflow and doc version configuration into shared config by @swahtz in #569
• Add camera_fov getter/setter to fvdb.viz.Scene by @swahtz in #558
• Remove torchsparse from all environment and CI configurations by @swahtz in #572
• Add cache clearing step in nightly publish workflow by @swahtz in #574
• Remove torch_scatter dependency in favor of built-in PyTorch scatter_reduce_ by @swahtz in #571
• Remove vestigial setup.py and GitLab CI config by @swahtz in #570
• Hotfix release process improvements by @swahtz in #563
• Update Gaussian splatting camera API and world-space parity by @fwilliams in #518
• Retire C++ GridBatch wrapper; add functional API and Grid class by @blackencino in #582
• scaled_dot_production_attention support for additional Torch backends (Flash Attention) by @swahtz in #365
• Fixes code samples in Sphinx docs that should have border by @swahtz in #577
• Reimplement JaggedReduce ops with PyTorch scatter_reduce_ by @swahtz in #578
• Eliminate .item() synchronization stalls in hot C++ paths by @swahtz in #586
• Pass current CUDA stream to all kernel launches by @swahtz in #587
• Add TEACHME interactive lesson documents for fvdb core API by @harrism in #584
• Fix missing CUDA device guards in kernel-launching functions by @swahtz in #589
• [CI] Get nanovdb-editor from pip instead from built whl by @phapalova in #581
• Use pip package of nanovdb-editor by @phapalova in #580
• Upgrade clang-tools to 21 to fix clangd SIGSEGV on CUDA files by @fwilliams in #491
• Move Gaussian splatting autograd and pipeline logic from C++ to Python by @fwilliams in #595
• Fix tutorial docs: move from wip, fix broken APIs, add CI testing by @harrism in #592
• Fix additional sync point introduced in autograd change by @matthewdcong in #599
• Materialize repeated opacities for compatibility with multiple splatting implementations by @matthewdcong in #600
• Restore useful comments from autograd/pipeline refactor by @swahtz in #603
• Fix weighted average in TSDF integration to apply pixelWeight to new samples by @jinhwanlazy in #588
• Improve tutorial content based on review feedback by @harrism in #598
• Environment/Docs/URL Updates by @swahtz in #605
• Initialize gradient accumulation tensors before UT projection path by @harrism in #608
• Versioned Documentation: Documentation integration with Read the Docs by @swahtz in #610
• CI: GH Actions Version Updates by @swahtz in #611
• Fixes to Sphinx Docs Build by @swahtz in #613
• Documentation: Add pre_build job to Read the Docs configuration for version generation by @swahtz in #615
• Documentation: Fix Read the Docs build and resolve Sphinx warnings by @swahtz in #618
• Refactor Gaussian splatting ops and extract utility functions by @fwilliams in #596
• Implement GitHub Actions workflow for Sphinx documentation build test by @swahtz in #622
• Documentation: Improve version label contrast in sidebar and fix reality-capture URLs by @swahtz in #623
• Documentation: Switch docs redirect to point to main RTD URL by @swahtz in #625
• CI: Revert failing drop cache step by @phapalova in #629
• Fix URLs in the README to point to the ReadtheDocs site by @swahtz in #626
• Promote GridBatchData to public header by @swahtz in #632
• Fix fvdb.viz.PointCloudView use of older API by @swahtz in #631
• Reorder Gaussian2D to improve field alignment by @matthewdcong in #624
• Remove unused SH function by @matthewdcong in #630
• Add CMake installation support for public headers and configuration files by @swahtz in #633
• Fix inject_from CUDA crash when source grid has 0 voxels by @harrism in #616
• Disambiguate CI job names and fix Torch CMake header path by @swahtz in #635
• Add sample_nearest operator for GridBatch and Grid by @swahtz in #628
• Generalize volume_render to N channels by @swahtz in #636
• Fix build issue with SampleNearest by @swahtz in #637
• Add __launch_bounds__ to forEach CUDA kernels by @swahtz in #638
• ci(nightly): anchor nightly version to upcoming release in pyproject.toml by @swahtz in #645
• CI: fix nightly wheel build by @phapalova in #634
• Add Vec2 and double fast paths to SampleGridTrilinear by @swahtz in #639
• Optimizations for volume_render and move its autograd layer to Python by @swahtz in #640
• Speed up builds with ccache, host PCH, and trimmed torch headers by @swahtz in #644
• NanoVDB loading: fix mixed grid type loading and add read_metadata API by @swahtz in #641
• [docs] Nightly build version numbering update for installation documentation by @swahtz in #646
• ci(publish): set short cache-control and invalidate CloudFront on index uploads by @swahtz in #647
• saveNVDB Optimizations by @swahtz in #650
• Shared memory optimizations for Gaussian rasterization by @matthewdcong in #554
• Avoid repeated delta computation by @matthewdcong in #651
• Add warp level early exit for forward rasterization by @matthewdcong in #658
• More optimal prefetching for mGPU Gaussian splatting by @matthewdcong in #657
• Fix Conda build failure by @matthewdcong in #661
• Improve parity with gsplat for dense rasterization by @matthewdcong in #659
• CMake: Link exported Torch target by @swahtz in #662
• fix: silence SyntaxWarning and tensor copy-construct UserWarning in tests by @mvanhorn in #654
• Improve mGPU Gaussian tile intersection by @matthewdcong in #664
• ray_implicit_intersection improvements by @swahtz in #663
• Improve prefetch granularity for rasterization kernels by @matthewdcong in #665
• Upgrade to PyTorch 2.11 by @swahtz in #573
• Add narrow-band SDF reinitialize/retopologize ops by @swahtz in #669
• Update dev_environment.yml to newer openusd version by @zlalena in #667
• CI: Removed the pytorch upper-bound version in pyproject.toml by @swahtz in #671
• CI Token Best Practices Sweep by @swahtz in #672
• Workflow Security: scope bundled shellcheck to real issues by @swahtz in #674
• CODEOWNERS: require NVIDIA maintainer review for governance/CI files by @harrism in #676
• CI: correct the change-detection gate so docs-only PRs skip cleanly by @harrism in #677
• v0.5 Release: Update CHANGES.md to reflect contributions since 0.4 by @swahtz in #675
• fix: TensorGrid blind-data lookup uses index 0 instead of loop counter by @mvanhorn in #652
• docs: fix marching_cubes return type (unique vertex indices, not normals) by @mvanhorn in #653
• Add CHANGES.md entries by @swahtz in #679
• CI: bump uv to 0.11.26 so Python 3.14 builds use stable CPython by @swahtz in #681

New Contributors

• @kmuseth made their first contribution in #305
• @iYuqinL made their first contribution in #325
• @zlalena made their first contribution in #350
• @jinhwanlazy made their first contribution in #588
• @mvanhorn made their first contribution in #654

Full Changelog: v0.3.0...v0.5.0