The Git Times

DeepSpeed has released version 0.18.8 with targeted fixes that improve reliability for distributed training and inference. The update suppresses noisy memory logs, resolves hangs in BLOOM model tests, and doubles reduction error messages for clearer debugging. It corrects async_io build errors on Huawei Ascend NPU hardware and adjusts ZeRO-3 gradient handling to use the current CUDA stream.

These changes arrive after the DeepSpeed team's March 2026 tutorial at ASPLOS on building efficient large-scale systems. Their SuperOffload work received an Honorable Mention for Best Paper. The library's core innovations—ZeRO, 3D parallelism, Ulysses sequence parallelism, and DeepSpeed-MoE—continue to support training at billion- and trillion-parameter scale while reducing memory and communication overhead.

Recent production use at LinkedIn leveraged ZeRO++ for large-scale LLM distillation in recommendation systems. Additional 2025 work on ZenFlow provides stall-free CPU offloading, while Arctic Long Sequence Training handles multi-million token contexts.

The Python library builds directly on PyTorch, offering data parallelism, model parallelism, pipeline parallelism, and compression techniques. Version 0.18.8 also adds PyTorch-style backward passes and low-precision master states to the Core API, easing integration for existing codebases.

(178 words)

NVIDIA has released GR00T N1.6, an updated open vision-language-action model for generalized humanoid robot skills. The update delivers improved performance and new features while maintaining the project's core approach to cross-embodiment learning.

The model accepts language instructions and images to generate actions for manipulation tasks across different environments. It was trained on a mixture of bimanual, semi-humanoid and large-scale humanoid datasets, allowing skill transfer between platforms. Its architecture pairs a vision-language foundation model with a diffusion transformer head that denoises continuous actions.

Implementation follows a defined workflow. Users first convert demonstration data—captured as (video, state, action) triplets—into the LeRobot compatible schema. The repository provides scripts to test zero-shot performance of the pretrained model, configuration examples for different embodiments, and tools for fine-tuning on custom datasets.

Benchmark support covers standard simulation suites including LIBERO, SimplerEnv and RoboCasa. The n1.5-release branch remains available for teams requiring the previous version.

This release matters as robotics developers seek foundation models that reduce the data and engineering effort needed to deploy capable systems on new hardware. By open-sourcing both the model and adaptation tooling, NVIDIA enables faster iteration between simulation and physical robots.

**

Peter Norvig has updated his pytudes repository with new Jupyter notebooks that extend its utility for Python skill development. The additions for 2025 and 2026 feature Advent of Code solutions alongside studies of truncatable primes and LLM applications to classic logic problems such as Cheryl's Birthday.

The project delivers short, difficult Python programs aimed at perfecting particular programming techniques. Practitioners can launch these notebooks in Colab, Deepnote, MyBinder, SageMaker or NBViewer to work through the material directly in the browser.

Norvig positions pytudes for those who approach programming as a craft demanding years of refinement rather than simple tool usage. Recent entries examine topics including cluster sizes in paint-by-numbers puzzles, number bracelet games, stubborn number endings and overtime rules in American football.

These updates keep the nine-year-old collection relevant by incorporating timely subjects like large language models and theory of mind. The notebooks demonstrate practical Python code for analyzing real-world and abstract problems alike.

By engaging with these études, developers build deeper competency in areas ranging from ancient numbering systems to modern computational challenges. The ongoing additions affirm the project's role in supporting deliberate practice.

RTAB-Map has issued version 0.23.1, refreshing its core dependencies to maintain compatibility with current robotics hardware and software ecosystems. The C++ library and standalone application, developed by IntRoLab in Sherbrooke, Québec, provides mature tools for RGB-D simultaneous localization and mapping.

The system constructs consistent 3D maps in real time while tracking camera pose. It relies on appearance-based loop closure detection and graph optimization to correct drift during extended operations, making it reliable for long-duration autonomous tasks.

This release updates core components to OpenCV 4.7.0 with xfeatures2d and nonfree modules, PCL 1.15.0 with VTK 9.3, and Qt 6.8.3. Hardware drivers now include librealsense 2.56.3 for RealSense devices, ZED SDK 5.0.3, Kinect for Azure, and updated Freenect2 support. Windows binaries bundle these libraries, simplifying deployment.

Integration with both ROS 1 Noetic and ROS 2 distributions from Humble to Rolling remains comprehensive, alongside Docker images and mobile support for Android and iOS. The standalone application offers visualization and processing capabilities independent of full ROS installations.

These updates address evolving sensor technology and library requirements, allowing existing users to maintain performance without major code changes. After more than a decade of development, RTAB-Map continues to serve as a stable foundation for mapping and localization projects.

Dora-rs has shipped version 0.5.0, bumping its workspace and dora-message components following an active period of feature additions. The Rust-based middleware continues to focus on low-latency, distributed dataflow for AI-driven robotic systems, with applications defined as directed graphs that connect reusable nodes.

The most notable change is dora.builder, a new Pythonic API that lets developers define dataflows imperatively rather than through configuration files alone. This joins existing Python, Rust, and C++ interfaces. Throughout 2025 the project added support for Microsoft Phi4 and Magma, Meta SAM2, Qwen2.5 LLMs and vision models, Kokoro and outetts TTS systems, plus Zenoh for distributed communication.

Pre-packaged nodes now cover Kornia-based V4L and Gstreamer camera pipelines, Sobel filtering, pytorch kinematics for forward and inverse kinematics, MediaPipe pose estimation, CoTracker point tracking, and AV1 encoding/decoding at up to 12-bit depth. Async Python support and git-based nodes further reduce integration time. Benchmarks still show the framework 10-17 times faster than ROS 2 when moving 40 MB of data through the Python API.

The project, accepted into GSoC 2025, maintains a growing hub of nodes that simplify connecting hardware, algorithms, and models while preserving deterministic low-latency performance required for embodied AI.

**

With the V0.2.8 release, XRobot has refined its automation for embedded robotics projects. The Python toolkit works with the LibXR hardware abstraction layer to handle module repositories and generate boilerplate code for microcontrollers such as STM32, ESP32 and Raspberry Pi.

The system pulls module repositories, resolves recursive dependencies and enforces version consistency. It scans header files to extract constructor and template parameters, then produces editable YAML configuration files. From these files XRobotMain() C++ entry functions are generated automatically, supporting multiple module instances and nested argument structures.

Additional capabilities include a manifest parser that reads descriptions and dependencies directly from headers, plus a module generator that creates standardized directories containing headers, README files and CMakeLists. The tool accepts both local and remote YAML sources.

These features matter as robotics applications incorporate growing numbers of sensors, actuators and control modules. By removing manual initialization work, XRobot reduces configuration errors and speeds iteration on real hardware and in Webots simulations. The latest updates improve parsing accuracy and dependency handling for larger projects.

**

SlimToolkit's version 1.40.11 brings targeted improvements to its container optimization suite. The update introduces two new flags for the build command—--include-dir-bins and --include-ssh-client—and a simple images command for listing container images. The xray command now handles OCI image formats and supplies richer reports that include object type information. Compatibility fixes ensure the tool works with the recently released Docker Engine 25.x.

Now a CNCF Sandbox project, SlimToolkit has for nearly a decade delivered automatic image minification without requiring changes to source Dockerfiles or application code. It executes the original container to observe actual runtime behavior, then constructs a minimal image that can be up to 30 times smaller. Compiled language applications often see even larger reductions. The toolkit automatically generates tailored Seccomp and AppArmor profiles, shrinking the attack surface without manual security engineering.

Core commands include xray for inspection, lint for Dockerfile best practices, build for optimization, debug for troubleshooting, and vulnerability for security analysis. Support extends beyond Docker to other cloud native runtimes. Integration with Root.io adds automated vulnerability remediation after optimization.

These incremental updates matter as organizations prioritize both deployment efficiency and supply-chain security. By preserving existing workflows and base images, SlimToolkit lowers the cost of maintaining lean, hardened containers.

Strix has enhanced its capabilities with native GitHub Actions integration, allowing automated security assessments in CI/CD pipelines. This update, part of v0.8.3, helps teams block vulnerabilities before they reach production.

The system's teams of AI agents collaborate to run applications dynamically, discover weaknesses, and validate them using real exploits rather than relying on static patterns. This results in accurate findings with minimal false positives.

Recent changes include adding a NestJS security testing module, implementing interactive agent mode, and introducing OpenTelemetry for observability with local JSONL traces. The platform supports multiple LLM providers including OpenAI, Anthropic and Google.

Developers can initiate scans using the strix CLI after a straightforward installation process that pulls a sandbox Docker image. Reports provide actionable insights and auto-fix recommendations to speed up remediation.

The integration marks a significant step in making advanced penetration testing accessible within everyday development processes, reducing the traditional timeline from weeks to hours.

GNU Radio continues to underpin open-source signal processing more than 14 years after its creation. The recent v3.10.12.0 release delivers stability updates to the C++ runtime and Python bindings, while the major 4.0 redesign moves ahead under the leadership of FAIR at GSI in Germany.

The toolkit lets users assemble flow graphs from reusable DSP blocks to build software-defined radio systems. Originally designed for wireless simulation, it now supports real-time data acquisition and analysis across demanding environments. Its modular architecture handles everything from simple receivers to complex multi-channel spectrum processors.

GNU Radio 4.0 focuses on modernizing the codebase for high-performance computing clusters used in particle physics and radio astronomy. Development emphasizes better scalability and tighter integration with contemporary SDR hardware.

The project sees active use in nuclear and high-energy physics experiments, astrophysics data collection, and cybersecurity protocol testing. Installation on Debian-based systems remains straightforward via sudo apt install gnuradio, with QT-based GNU Radio Companion available for visual flow-graph design.

Mailing lists, Matrix chat, and the project wiki continue to support both academic researchers and commercial developers working on wireless communications and scientific instrumentation.

(178 words)

frp has released version 0.68.0 with practical upgrades to its client management capabilities. The most significant change is a built-in store for frpc that supports persisted configuration from local files. This adds full CRUD admin APIs for proxies and visitors at endpoints such as /api/store/proxies*, with runtime reload support that applies changes without restarting the client.

The new store includes dedicated management pages in the frpc web dashboard, giving operators direct visibility and control over running tunnels. Configuration parsing has also been tightened: TOML syntax errors now report exact line and column positions immediately, eliminating confusing fallback behavior to YAML or JSON parsers.

OIDC authentication received an efficiency update, caching access tokens and refreshing them before expiry instead of requesting fresh tokens on every heartbeat. Build improvements include a noweb tag that lets developers compile without frontend assets, making fresh clones easier to build.

Now over ten years old, frp remains a reliable reverse proxy for exposing local services behind NAT or firewalls. It forwards TCP, UDP, HTTP and HTTPS traffic, supports custom domain routing, and includes optional P2P mode for direct connections. The latest changes particularly benefit teams managing dynamic proxy fleets in production environments.

(172 words)

Shmoergh Moduleur has updated its hardware documentation with the v1.1.1 release. The fully analog DIY modular synthesizer now uses rev x identifiers across all boards to follow established industry conventions. Previous v1.0 designs map to REV 1, v1.1 to REV 2, and so on, with no functional impact on existing builds.

The power supply unit received the only physical change. In rev 3, resistor R2 was relocated to improve mechanical compatibility with the enclosure's corner blocks.

Derived from the original Shmøergh Hog, Moduleur consists of discrete, self-contained modules including a VCO, mixer with sidechain compressor, VCF, ADSR and VCA, analog bitcrusher with S&H, LFO, output stage, and a digital Brain. Each module is designed for independent assembly or modification.

All circuits follow Eurorack standards with ±12 V rails, reverse polarity protection, and 10-pin power connectors. The repository supplies complete fabrication packages: schematics, Gerbers, BOMs, panel files, CircuitJS simulations, and test notes. Designs prioritize readily available components and straightforward signal paths.

The project continues incremental development, with new Brain firmware planned. Hardware is licensed under CC BY-NC 4.0 and software under MIT. Experienced builders are expected to follow the published safety and build guides.

(178 words)

Neural-Networks-on-Silicon has evolved. What began in 2016 as a focused list of neural network accelerator papers now functions as Fengbin Tu's personal selection of influential work at the intersection of deep learning and computer architecture.

The HKUST professor and ACCESS AI Chip Center faculty member continues to maintain the repository, with the most recent push adding 2026 papers from ISSCC and HPCA. The collection spans more than a decade, organizing notable contributions by conference and year. Early entries include the foundational DianNao (ASPLOS 2014), DaDianNao (MICRO 2014), and ShiDianNao (ISCA 2015) papers that shaped efficient machine-learning hardware.

Later sections prominently feature Tu's own research alongside other selected works from ISSCC, ISCA, MICRO, HPCA, ASPLOS, DAC, and HotChips. The structure provides a chronological view of how ideas progressed from basic accelerators to sophisticated systems that move computation closer to sensors and optimize for power and area.

For chip designers and systems architects, the value lies in its curation. Rather than an exhaustive database, it surfaces papers that caught the attention of an active researcher in the field. As demand grows for specialized silicon to run ever-larger models efficiently, the repository serves as a living timeline of hardware innovation.

**

Mach continues to refine its position as a practical Zig game engine and graphics toolkit four years into development. The system enables builders to create high-performance, truly cross-platform games, visualizations, and desktop or mobile GUI applications through a deliberately modular architecture.

Recent refinements emphasize composability. Developers assemble only the required graphics, windowing, and GUI modules, producing smaller binaries and faster compile times while retaining fine-grained control. This approach leverages Zig's compile-time execution and memory management to deliver robust performance without hidden overhead.

The toolkit maintains consistent behavior across operating systems through unified APIs for rendering and input. Games benefit from low-latency graphics pipelines, while GUI components integrate directly with the same rendering layer, simplifying tool development. Visualizations gain from efficient handling of large datasets in real time.

Development takes place primarily on code.hexops.com, with the GitHub repository operating as a mirror for visibility. The team directs community discussion to Discord, where participants focus on practical game creation in Zig while maintaining an inclusive environment.

As more programmers adopt Zig for systems work, Mach provides the graphics foundation that pairs language strengths with production-ready tooling.

**

Clockwork Labs has released SpacetimeDB v2.1.0, bringing Rust client SDK support for WebAssembly and full compatibility updates for C++ and Unreal Engine developers.

The Rust SDK can now compile and run in the browser after resolution of longstanding Wasm issues. This allows teams to write browser-based applications that connect directly to the database using the same language as their modules.

C++ module bindings and the Unreal SDK have been updated to match 2.0 APIs and codegen. Unreal developers no longer need workarounds to integrate real-time synchronization for player state, inventory or terrain data.

Several bugs were also fixed, including useTable incorrectly reverting to false after the first row event and v2 client disconnections dropping subscriptions for other connected clients.

SpacetimeDB merges relational database and application server into one system. Developers upload modules containing schema and business logic in Rust, C#, TypeScript or C++. The database executes this logic and automatically synchronizes state changes to all connected clients in real time.

All application state remains in memory for minimal latency while a commit log on disk delivers ACID guarantees and crash recovery. No separate web server, containers or orchestration layer is required.

The architecture continues to power demanding workloads, such as the complete backend of the MMORPG BitCraft Online.

**