fVDB Version History

Version 0.5.0 - July 1, 2026

115 commits, 500+ files changed, 10 contributors.

This release eliminates the C++ implementation behind GridBatch in favor of a new torch.nn.functional-style fvdb.functional module, with GridBatch now a thin pure-Python class delegating to it, and adds a complementary single-grid Grid class alongside it. It also moves the entire Gaussian splatting autograd and rendering pipeline from C++ into pure Python. Multi-GPU Gaussian splatting rasterization and projection get another round of performance tuning, and fVDB gains its first SDF reinitialization/retopologization operators alongside NanoVDB loading fixes and PyTorch 2.11/Python 3.14 support. Documentation has moved to a fully versioned Read the Docs site.

Highlights:

  • Eliminated the C++ GridBatch implementation for a new functional API (fvdb.functional), with GridBatch reimplemented as a thin Python wrapper and a new complementary Grid class added, and moved Gaussian splatting’s autograd/pipeline logic from C++ to pure Python.

  • Continued multi-GPU Gaussian splatting performance work: repartitioned SH/projection kernels, smarter prefetching, shared-memory rasterization optimizations, and improved parity with gsplat.

  • Added the first SDF reinitialization/retopologization operators and a sample_nearest grid sampling operator, and generalized volume_render to N channels.

  • Fixed NanoVDB loading of mixed grid types and added a read_metadata API; optimized saveNVDB/to_nanovdb.

  • fVDB now supports PyTorch 2.11, Python 3.14, and SM 8.6, with the torch_scatter and torchsparse dependencies removed.

  • Documentation moved to a fully versioned Read the Docs site, with new interactive TEACHME lessons for the core API.

Contributors: @swahtz, @matthewdcong, @harrism, @fwilliams, @phapalova, @blackencino, @areidmeyer, @zlalena, @mvanhorn, @jinhwanlazy

Core Library Architecture (Major)

  • Eliminated the legacy C++ GridBatch wrapper class and its Vec3*OrScalar type system (~2,600 lines), replacing them with a frozen GridBatchData struct and a torch.nn.functional-style fvdb.functional module (~120 functions with explicit _batch/_single variants). GridBatch is now a thin pure-Python class whose methods delegate to fvdb.functional, and a new complementary Grid class was added for the single-grid case (grid_count == 1), with plain torch.Tensor I/O instead of JaggedTensor (#582 - @blackencino).

  • Promoted GridBatchData, TorchDeviceBuffer, VoxelCoordTransform, and related type traits from private detail headers to the public fvdb namespace, and added CMake installation rules for public headers so downstream C++ projects can link against fVDB (#632, #633 - @swahtz).

Gaussian Splatting & Rendering

New Features:

  • Updated the Gaussian splatting camera API to separate camera semantics from projection implementation, replacing the old ProjectionType/DistortionModel split with explicit CameraModel/ProjectionMethod controls, and added world-space render parity for depth and RGBD paths (#518 - @fwilliams).

Architecture:

  • Moved the entire Gaussian splatting autograd and pipeline from C++ into pure Python torch.autograd.Function implementations, eliminating the C++ GaussianSplat3d class (#595 - @fwilliams), with a follow-up fixing a synchronization regression it introduced (#599 - @matthewdcong) and restoring comments lost in the migration (#603 - @swahtz).

  • Renamed 21 Gaussian splatting operator file pairs for clarity and extracted shared CUDA utilities (BinSearch, CubWrapper, Prefetch, WarpReduce) with no functional changes (#596 - @fwilliams).

Optimizations:

  • Repartitioned multi-GPU spherical harmonics and projection kernels across Gaussians instead of camera-Gaussian pairs, removing the system-scope atomics that bottlenecked the interconnect (#546, #547 - @matthewdcong).

  • Added shared-memory feature caching and opacity-threshold culling to forward/backward rasterization, and reduced shared memory usage in pinhole projection by parameterizing blocks per-camera (#554, #555 - @matthewdcong).

  • Fused computeGradientState into projectionBackwardKernel to avoid redundant global memory reads (#560 - @matthewdcong).

  • Improved multi-GPU prefetching with batched cudaMemPrefetchBatchAsync calls and finer per-tile (rather than per-camera) granularity, and iterated multi-GPU tile intersection to reduce cross-device communication during the radix sort (#657, #664, #665 - @matthewdcong).

  • Added a warp-level early exit to forward rasterization on top of the existing block-level exit (#658 - @matthewdcong).

  • Improved rasterization parity with gsplat by tuning the alpha threshold and sigma clamp, and using per-axis 2D radii for tile intersection (#659 - @matthewdcong).

  • Reordered Gaussian2D fields for better memory alignment, removed an unused spherical harmonics function, and eliminated a redundant delta computation (#624, #630, #651 - @matthewdcong).

  • Materialized repeated opacities so image-space multi-GPU and world-space rasterization implementations always see a contiguous tensor (#600 - @matthewdcong).

Bug Fixes:

  • Fixed a missing warp-level reduction causing incorrect quaternion gradient accumulation across multiple cameras in the projection backward pass (#534 - @swahtz).

  • Fixed a cudaErrorIllegalArgument crash in tile intersection prefetch when a subset of Gaussians has zero intersections (#553 - @matthewdcong).

  • Fixed gradient accumulation tensors not being initialized on the Unscented Transform projection’s early-return path (#608 - @harrism).

Grid Operations, Sampling & SDF

  • Added a sample_nearest nearest-neighbor sampling operator for GridBatch and Grid (#628, #637 - @swahtz).

  • Added Vec2 and double-precision vectorized fast paths to SampleGridTrilinear (#639 - @swahtz).

  • Generalized volume_render to support up to 16 radiance channels for spectral rendering pipelines, rewrote its forward pass around per-ray register accumulation, and moved its autograd layer to Python (#636, #640 - @swahtz).

  • Added narrow-band SDF reinitialize and retopologize operators built around a ported VoxelBlockManager eikonal solver — the first component of a NanoVDB port of OpenVDB’s VolumeToMesh (#669 - @swahtz).

  • Optimized the per-ray SDF zero-crossing kernel ray_implicit_intersection and its underlying HDDA traversal layer, fixing several correctness issues along the way (#663 - @swahtz).

  • Optimized saveNVDB/to_nanovdb to build grids directly on-device (or via a host-only fast path) instead of rebuilding through CPU setValue calls (#650 - @swahtz).

  • Fixed a CUDA crash in Grid.inject_from() when the source grid has 0 active voxels, which surfaced in narrow-band level-set simulations (#616 - @harrism).

  • Fixed the TSDF/feature-blending weighted average formula to correctly apply pixelWeight to new samples (#588 - @jinhwanlazy).

NanoVDB

  • Fixed wrong-type dispatch when loading multi-grid NanoVDB handles with mixed grid types, and added a read_metadata API (#641 - @swahtz).

  • Fixed TensorGrid blind-data loading reading channel 0 for every channel, which silently returned incorrect data for all channels of multi-channel grids (#652 - @mvanhorn).

  • Updated the bundled NanoVDB Editor dependency, moving off a version with ABI issues (#556 - @areidmeyer).

  • Added nanovdb-editor as an optional dependency and switched to consuming it from pip instead of building it from source (#559, #580, #581 - @swahtz, @phapalova).

JaggedTensor

  • Reimplemented jsum/jmin/jmax reductions on top of PyTorch’s built-in scatter_reduce_, removing ~450 lines of custom CUDA/autograd code and the torch_scatter test dependency it required (#578, #571 - @swahtz).

Neural Network Modules

  • Added support for additional PyTorch attention backends (Flash, memory-efficient, math) in scaled_dot_product_attention by wrapping JaggedTensor data in nested tensors, selected the same way as PyTorch’s built-in sdpa_kernel context manager (#365 - @swahtz).

Core Library: Correctness & Performance

  • Eliminated redundant .item() calls on CUDA tensors that triggered implicit device syncs, consolidating them into single bulk .cpu() transfers (#586 - @swahtz).

  • Passed the current PyTorch CUDA stream to 15 kernel launches that previously used the implicit default stream, and added missing CUDA device guards across kernel-launching functions (#587, #589 - @swahtz).

  • Replaced the runtime numThreads argument in the forEach CUDA dispatch framework with a compile-time template parameter, enabling __launch_bounds__ on all forEach kernels for better register allocation (#638 - @swahtz).

Viewer

  • Added a camera_fov getter/setter to fvdb.viz.Scene, exposed through both C++ and Python (#558 - @swahtz).

  • Fixed fvdb.viz.PointCloudView using an outdated add_gaussian_splat_3d_view signature (#631 - @swahtz).

PyTorch & CUDA Compatibility

  • fVDB now builds and runs with PyTorch 2.11 and adds Python 3.14 support, including SM 8.6 in the published wheels’ CUDA architecture list to match PyTorch’s support (#573, #561 - @swahtz, @matthewdcong).

  • Removed the torchsparse dependency from all environment and CI configurations (#572 - @swahtz).

Build & Packaging

  • Upgraded clang-tools to 21 to fix a clangd SIGSEGV on CUDA files, and conda environments to gcc/g++ 14.3 (#491, #557 - @fwilliams, @swahtz).

  • Removed the vestigial setup.py and GitLab CI configuration (#570 - @swahtz).

  • Refactored CMake to consume libtorch through the canonical torch imported target instead of legacy TORCH_INCLUDE_DIRS/TORCH_LIBRARIES variables, retiring the deprecated THC headers and fixing a Conda build failure along the way (#661, #662 - @matthewdcong, @swahtz).

  • Fixed a Torch CMake header path issue and disambiguated CI job names (#635 - @swahtz).

  • Sped up incremental builds with ccache/sccache auto-detection, host-side precompiled headers, and trimmed Torch header includes (#644 - @swahtz).

  • Updated the dev environment’s OpenUSD version (#667 - @zlalena).

Documentation

  • Migrated documentation hosting to Read the Docs with full versioned-docs support: a pre-build hook for version generation, a dedicated Sphinx-build CI workflow, and sidebar/redirect/URL fixes (#610, #613, #615, #618, #622, #623, #625, #646 - @swahtz).

  • Added docs/TEACHME, a set of LLM-loadable interactive lesson documents that teach the fVDB API through an AI coding assistant (#584 - @harrism).

  • Fixed and expanded the tutorial notebooks — moved them out of WIP status, corrected broken API calls, added CI testing, and incorporated review feedback (#592, #598 - @harrism).

  • Fixed the docs deployment workflow, Sphinx code-sample borders, and README/docs redirect URLs (#566, #577, #605, #626 - @swahtz).

  • Fixed the marching_cubes docstrings to describe the third return value correctly as unique vertex indices (int64), not vertex normals (#653 - @mvanhorn).

CI / DevOps / Release Infrastructure

  • Hardened the release process scripts: idempotent start-release.sh re-runs, draft release PRs with fixed smoke-test Python setup, branch-integrity preservation in finish-release-process.sh, and automated doc-version updates (#528, #529, #544, #552, #563 - @harrism, @swahtz).

  • Fixed the publish workflow across several iterations: Rocky Linux 8 / manylinux_2_28 containers, the Python install action, additional system dependencies, and dual S3 + PyPI publishing with GPU-validated tests (#536, #537, #538, #540, #545 - @harrism, @swahtz).

  • Fixed several nightly wheel build and publish issues: missing tool errors, stale caches, CloudFront invalidation, and anchoring the nightly version to the upcoming release in pyproject.toml (#549, #574, #634, #645, #647 - @swahtz, @phapalova).

  • Replaced the Slack unanswered-issues report with event-driven issue triage labels, after first fixing its insider-issue filtering (#522, #551 - @harrism).

  • Centralized GitHub workflow and doc-version configuration into shared config, updated CI Actions versions, added git installation to CI system dependencies, and reverted a failing drop-cache step (#569, #611, #526, #629 - @swahtz, @phapalova).

  • Swept CI tokens to least-privilege scopes and scoped the bundled shellcheck security check to real issues (#672, #674 - @swahtz).

  • Removed the PyTorch upper-bound pin in pyproject.toml, updated CONDA_OVERRIDE_CUDA to 13.0, fixed a flaky bfloat16 JaggedTensor test, and silenced spurious warnings in the test suite (#671, #519, #517, #654 - @swahtz, @fwilliams, @mvanhorn).

Repository Governance

  • Split CODEOWNERS into two review tiers — any maintainer may review general code, while governance, legal, and CI/CD infrastructure files (.github/, LICENSE, MAINTAINERS.md, SECURITY.md, etc.) require sign-off from an NVIDIA maintainer. Kept identical across fvdb-core, fvdb-reality-capture, and fvdb-examples (#676 - @harrism).

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).