fVDB Version History

Version 0.5.0 - In Development

Version 0.4.2 - March 25, 2026

Bug Fixes:

  • Build warning fix for GCC 14.

Version 0.4.1 - March 25, 2026

Bug Fixes:

  • Updated nanovdb-editor dependency to use version 0.0.23.

Version 0.4.0 - March 12, 2026

140 commits, 300+ files changed, 10 contributors.

This release delivers major advances across the Gaussian splatting pipeline, sparse convolution, multi-GPU performance, and build/release infrastructure. fVDB now supports PyTorch 2.10 and CUDA 12.8/13.0, and ships its first formal release process with automated nightly builds.

Highlights:

  • Gaussian splatting gains a new rasterize-from-world path that renders directly from 3D Gaussians, Unscented Transform projection for non-pinhole camera models, full MCMC splatting support, sparse rendering, per-pixel/per-tile masking, and a composable camera model that decouples kernels from camera internals. Numerous gradient correctness fixes harden the backward pass.

  • Sparse convolution has been consolidated into a single GatherScatterDefault backend with full feature support including transposed convolution and arbitrary strides. A new PredGatherIGemm backend using CUTLASS/CuTe implicit-GEMM with TF32 tensor cores delivers significantly faster convolution on dense grids.

  • A new multi-axis dispatch framework provides flexible kernel execution across multiple dimensions with typed views and for_each iteration.

  • SampleGridTrilinear is roughly 2x faster via vectorized float4 loads and a fused stencil-plus-sample optimization. Morton and Hilbert space-filling curve ordering is now available for grid coordinates.

  • Multi-GPU scaling is significantly improved through batched prefetching, device-centric synchronization, and radix sort optimizations. All tensor index accessors are now 64-bit, enabling larger datasets.

  • A fully automated nightly wheel build and publish pipeline, a formal OneFlow release process with automation scripts, and GPU-validated publish workflows are all new in this release.

Contributors: @blackencino, @fwilliams, @harrism, @iYuqinL, @kmuseth, @matthewdcong, @phapalova, @areidmeyer, @swahtz, @zlalena

Gaussian Splatting & Rendering

New Features:

  • Added a new rasterization pathway that operates directly on 3D Gaussians (#444 - @fwilliams).

  • Added Gaussian projection via the Unscented Transform, providing an alternative to the EWA splatting approximation (#420 - @fwilliams).

  • Added full MCMC Gaussian Splatting support, including relocation (#374) and add-noise (#377) kernels, Python bindings (#394), and tunable min_opacity (#396) and k/t (#402) parameters (@harrism, @fwilliams).

  • Added end-to-end sparse Gaussian rendering with sparse rendering functions (#348) and sparse tile intersection (#401) (@fwilliams, @swahtz).

  • Rasterization can now render all contributing Gaussian IDs and weights per pixel (#340 - @swahtz).

  • Gaussian rasterization now supports background colors (#343 - @harrism).

  • All Gaussian render methods now accept masks and backgrounds (#480 - @swahtz).

  • The evaluate_spherical_harmonics function is now exposed in the Python API (#431 - @swahtz).

  • Refactored the rendering pipeline around composable camera operation classes that encapsulate camera-space transform and projection, decoupling kernels from camera internals (#485 - @fwilliams), with a CameraIntrinsics constructor fix for host/device compatibility (#489 - @blackencino).

Optimizations:

  • Switched the GaussianTileIntersection cumulative sum to use CUB for better performance (#427 - @swahtz).

  • Optimized the computeSparseInfo path to reduce overhead in sparse rendering (#428 - @swahtz).

  • Improved the contributing Gaussian ID kernels with shared-memory and loop optimizations (#429 - @swahtz).

  • Optimized tile intersection for multi-GPU execution with better prefetching (#446 - @matthewdcong).

  • Removed an unnecessary stream synchronization in GaussianTileIntersection (#370 - @harrism).

  • ProjectedGaussianSplats opacities now use an efficient expand/view instead of per-element copy (#457 - @swahtz).

Bug Fixes:

  • Fixed a shared memory alignment issue in the Gaussian rasterization kernel (#342 - @swahtz).

  • Fixed inverted abs(gradient) logic in the backward rasterization pass that produced incorrect gradients (@harrism).

  • Fixed NaN outputs in the top-contributing Gaussian IDs weights computation (#400 - @swahtz).

  • Fixed camera data loading that could exceed blockDim when using many cameras (#345 - @swahtz).

  • Fixed incorrect derivation of the number of cameras in packed rasterization mode (#414 - @swahtz).

  • Fixed the chain rule for the log_scale gradient in the projection backward pass (#433 - @harrism).

  • Fixed a race condition in the spherical harmonics backward pass when using multiple cameras or large batch sizes (#437 - @swahtz).

  • Fixed the dLossDQuat quaternion gradient missing a warp-level reduction in the projection backward pass (#435, #533 - @swahtz, @matthewdcong).

  • Fixed a multi-GPU race condition in the multibatch spherical harmonics backward pass (#484 - @matthewdcong).

  • Fixed the ProjectionForward kernel double-initializing accessors, which caused correctness issues (#453 - @swahtz).

  • Fixed a crash when loading GaussianPly files to a CPU device (#417 - @swahtz).

  • Fixed handling of duplicate pixels in sparse pixel Gaussian rendering (#488 - @harrism).

  • Fixed an incorrect datatype in the backward projection test (#486 - @matthewdcong).

Sparse Convolution (Major)

  • Consolidated all legacy sparse convolution backends into a single GatherScatterDefault backend with full feature support, including transposed convolution, arbitrary strides, and all float types (#473 - @blackencino).

  • Added a new PredGatherIGemm sparse convolution backend using CUTLASS/CuTe implicit-GEMM with TF32 tensor cores, significantly faster than GatherScatterDefault for dense or near-dense grids (#508 - @blackencino).

  • Fixed the default convolution behavior and added extensive correctness tests (#321 - @blackencino).

  • Added gradient and backward pass tests to the convolution unit test suite (#358, #361 - @blackencino).

  • Removed unused legacy sparse convolution backends (ImplicitGEMM, CUTLASS, LGGS, ME), deleting approximately 22,500 lines of code (#454 - @blackencino).

  • Moved all op dispatch and precondition code into each op’s C++ implementation files, making ops self-contained and reducing compile-time interconnectivity (#492 - @blackencino).

Multi-Axis Dispatch Framework (New)

  • Introduced a new multi-axis dispatch framework for flexible kernel execution across multiple dimensions (#418 - @blackencino).

  • Extended the dispatch framework with for_each iteration, typed views, and tag canonicalization (#452 - @blackencino).

  • The framework ships as a full C++ library under src/dispatch/ with comprehensive tests and benchmarks.

Grid Operations & Spatial Indexing

  • Added Morton and Hilbert space-filling curve ordering for Grid and GridBatch ijk coordinates, with module-level standalone functions (#311, #316, #323 - @blackencino).

  • SampleGridTrilinear now uses vectorized float4 loads, yielding roughly a 2x throughput improvement (#430 - @swahtz).

  • SampleGridTrilinear received a second optimization pass using a fused stencil-plus-sample approach (#474 - @swahtz).

  • Cleaned up the active grid coordinate generation code for clarity and consistency (#318 - @blackencino).

JaggedTensor

  • JaggedTensor reduce operators now support bfloat16 (#501 - @swahtz).

  • Fixed a binary search edge case in JIdxForJOffsets that returned incorrect indices when joffsets contained duplicate values (#325 - @iYuqinL).

  • Fixed from_*_and_list_ids producing incorrect results with ldim=2 (#357 - @swahtz).

  • Fixed concatenation errors in JaggedTensor.jcat (#352 - @blackencino).

  • Reduced the number of blocking GPU-to-CPU copies in the unbind* methods, improving throughput (#363 - @swahtz).

  • Fixed the single-element JaggedTensor constructor unconditionally initializing CUDA even for CPU tensors (#469 - @swahtz).

Performance & Multi-GPU

  • Optimized joffsets construction by using pinned memory to overlap CPU/GPU transfers (#403 - @matthewdcong).

  • Significantly improved multi-GPU scaling through batched prefetching and sorting changes (#499 - @matthewdcong).

  • Switched to device-centric synchronization for the forEach multi-GPU codepath (#440 - @matthewdcong).

  • Fixed and improved radix sort synchronization across multiple rounds of improvements (#315, #409, #415 - @matthewdcong).

  • Fused SSIM outputs now prefetch to avoid write page faults that degraded performance (#407 - @matthewdcong).

  • MCMC kernels now support PrivateUse1 for multi-GPU execution (#421 - @harrism).

  • Switched from torch.inverse to torch.linalg.inv_ex to avoid an unnecessary device synchronization (#487 - @matthewdcong).

  • All 32-bit tensor index accessors have been upgraded to 64-bit across every op, enabling support for larger datasets (#505 - @harrism).

PyTorch & CUDA Compatibility

  • fVDB now builds and runs with PyTorch 2.10 (#423, #521 - @matthewdcong, @swahtz).

  • Added support for CUDA 12.8 and 13.0 toolkits (#521 - @swahtz).

  • Replaced the custom scaled dot-product attention implementation with PyTorch’s native torch.scaled_dot_product_attention (#364 - @swahtz).

  • Fixed the CCCL version check macro that could cause build failures with newer CUDA toolkits (#509 - @matthewdcong).

  • Improved PyTorch build configuration time by streamlining CMake detection (#441 - @matthewdcong).

NanoVDB

  • Updated the bundled NanoVDB dependency to version 32.9.1 (#475, #483, #493 - @swahtz).

  • Fixed voxel size and origin metadata not being preserved when serializing index grids (#490 - @swahtz).

Neural Network Modules

  • Fixed several bugs in SimpleUnet: NaN propagation from -inf values entering BatchNorm after max-pooling, incorrect ConvolutionPlan source/target grid assignments, and a crash on non-contiguous grad_output in the convolution backward pass (#496 - @swahtz).

  • Added dedicated unit tests for all fvdb.nn modules to improve coverage and prevent regressions (#497 - @swahtz).

Visualization / Viewer

  • The viewer now supports displaying multiple scenes simultaneously with a scene-switching UI (#308 - @phapalova).

  • Added viz bindings for wait and add_image to enable blocking display and image overlays (#332 - @phapalova).

  • Fixed the viewer so it works correctly inside Jupyter notebooks (#350 - @zlalena).

Build & Packaging

  • Renamed the Python extension binary from _Cpp to _fvdb_cpp for clarity and to avoid naming conflicts (#317, #322 - @harrism, @blackencino).

  • Improved build times with compilation speedups and added build tracing support (#443 - @blackencino).

  • Fixed potential oversubscription when nvcc and cmake parallelism combined to exceed available cores (#351 - @swahtz).

  • Added a lineinfo build option to include source-line debug info for GPU profiling (#367 - @harrism).

  • Added a getMaxSharedMemory utility to centralize shared memory limit queries across kernels (#368 - @harrism).

  • Added a Version class that provides structured version information at runtime (#507 - @swahtz).

Nightly Builds & Release Infrastructure (New)

  • Added a fully automated nightly wheel build and publish pipeline that builds across a matrix of Python, PyTorch, and CUDA versions and publishes to an S3 simple index (#477, #478 - @swahtz).

  • Established a formal release process based on the OneFlow branching model, with start-release.sh and finish-release.sh automation scripts (#512, #525 - @harrism, @swahtz).

  • The publish workflow now includes GPU validation with smoke tests and full unit tests on built wheels, an S3 staging index with automatic 30-day pruning, and support for release branch pushes triggering builds automatically.

CI / DevOps

  • Documentation-only PRs now auto-pass CI instead of showing a perpetual “waiting for status” indicator (#462 - @swahtz).

  • Draft PRs now skip test runs entirely, saving compute resources (#339 - @swahtz).

  • CI checkout references are pinned to immutable commit SHAs to prevent build/test skew between checkout and merge steps (#503 - @swahtz).

  • Nightly workflows are now restricted to the upstream openvdb/fvdb-core repository and no longer run on forks (#319 - @harrism).

  • Runner stop jobs are now skipped when the corresponding start job was skipped, avoiding spurious failures (#471, #472 - @harrism).

  • Unit tests now only run for the matrix entry matching the test_environment.yml configuration (#531 - @harrism).

Developer Tooling (New)

  • Added git worktree tools (fvdb-open, fvdb-close, fvdb-issue) that make it easy to work on multiple branches simultaneously (#445 - @harrism).

  • Added an unanswered external issues reporter with Slack output and a daily CI workflow to help the team stay on top of community questions (#510, #513 - @harrism).

  • Added an AGENTS.md file providing persistent coding guidelines for AI agents working on the codebase (#455 - @harrism).

Documentation

  • Added and updated introductory, neural network, and convolution notebooks (#504 - @swahtz).

  • Applied NVIDIA branding to the documentation site (#405 - @fwilliams).

  • Added documentation for installing nightly builds from the S3 package index (#481 - @swahtz).

  • Integrated Google Analytics into the documentation site for usage tracking (#312 - @fwilliams).