This repository is a high-performance computing library optimized for large language model (LLM) inference and training on NVIDIA GPUs. It provides accelerated implementations of core operations including attention mechanisms, GEMM, fused Mixture of Experts (MoE), all-reduce with normalization, sampling, and rotary position embeddings. The architecture combines C++/CUDA kernels for performance-critical operations with Python bindings via PyTorch, enabling easy integration into deep learning frameworks. The library emphasizes kernel fusion, efficient memory access patterns, and specialized optimizations for modern GPU architectures (particularly Hopper and Blackwell).
AdaptiveCpp is a SYCL and C++ standard parallelism implementation targeting CPUs and GPUs from multiple vendors. It provides a complete heterogeneous computing stack including compiler infrastructure (LLVM-to-backend translators, optimization passes), runtime system (memory management, scheduling, event handling), and SYCL API headers. The project supports multiple backends (CUDA/HIP, OpenCL, Level Zero, Metal, Vulkan, OpenMP) and features compilation flows including LLVM JIT, source-level interop, and library-only modes. While primarily a production SYCL implementation, it contains research elements reflected in academic citations and experimental features like PCUDA and stdpar support. The codebase demonstrates strong modularity with clear separation between compiler, runtime, and API layers.
This repository is a comprehensive personal R&D platform implementing a polyglot microservices architecture. It hosts a personal website/web application (React/Vite), multiple API servers in different languages (Rust/Axum, Go/Gin/gRPC, Python/FastAPI, Node.js/Express), mobile applications (iOS/Android/React Native), and desktop applications (Qt). The platform includes extensive data processing infrastructure (Kafka, Pulsar, NATS, Spark, Flink), machine learning capabilities (model training/serving/fine-tuning), monitoring and observability tools, and infrastructure as code (Terraform/OpenTofu, Kubernetes). Services communicate via messaging systems and APIs, with a clear separation of concerns and standardized patterns for configuration, health checks, logging, and error handling across language implementations.
Hermit-rs is a Rust-based unikernel (library OS) that enables applications to run directly on hardware or hypervisors without a traditional operating system. The repository provides the build infrastructure (the `hermit` crate), the system call ABI (`hermit-abi`), and a collection of example applications demonstrating networking, file systems, TLS, WASM execution, and more. The actual kernel is maintained in a separate submodule. The project is actively developed with CI testing across x86_64, aarch64, and riscv64 architectures, and supports running under QEMU, Uhyve, and KVM. It is a hybrid of a system library and an application framework, but its dominant character is an application platform for building unikernel images.
ASPECT (Advanced Solver for Problems in Earth's ConvecTion) is a massively parallel, finite-element geodynamics simulation code built on the deal.II library. It solves the Stokes and advection-diffusion equations for mantle convection, lithosphere dynamics, melt migration, and related tectonic processes. The code features a highly modular plugin architecture with registration macros for material models, heating models, boundary conditions, postprocessors, mesh refinement, particle tracking, and more. It includes an extensive benchmark suite for code verification and numerous cookbooks demonstrating real geodynamic applications. The repository also bundles the Geodynamic World Builder (contrib/world_builder), a standalone library for constructing initial temperature, composition, and velocity fields from structured input files. ASPECT is actively developed with continuous integration, Docker support, and coupling to external tools like Landlab and FastScape for surface processes.
KaHIP (Karlsruhe High Quality Partitioning) is a high-performance C++ library and command-line toolkit for graph partitioning, node ordering, and mapping. It implements multilevel partitioning with advanced refinement techniques including Fiduccia-Mattheyses local search, flow-based min-cut refinement, cycle improvements, and tabu search. The library provides a C API for integration into external applications, along with Python bindings and a Java JNI wrapper. The codebase includes both a sequential version and a distributed-memory parallel version (ParHIP) using MPI, enabling partitioning of large graphs on clusters. Core algorithms are well-structured with modular coarsening, initial partitioning, and uncoarsening/refinement phases, supporting multiple matching strategies (GPA, random, label propagation) and quality/speed trade-offs via preconfigured modes (Strong, Eco, Fast, Social). Despite research-grade complexity and some code duplication between serial and parallel variants, the library is actively maintained and widely used in scientific computing and graph analytics.
DeepEP is a high-performance communication library developed by DeepSeek for distributed Mixture-of-Experts (MoE) model training and inference. Its primary function is implementing expert-parallel (EP) all-to-all dispatch and combine operations, routing tokens to the correct GPU holding an expert and gathering the results. It also provides general collective communication primitives (all-reduce, reduce-scatter, all-gather) optimized for multi-node multi-GPU clusters with NVLink and RDMA interconnects. The architecture is a hybrid Python/C++/CUDA library. The Python layer (deep_ep) provides user-facing buffer classes (EPBuffer, BucketBuffer, EngramBuffer, PPBuffer) and orchestration utilities. The C++/CUDA layer (csrc/, deep_ep/include/) contains JIT-compiled kernels that leverage cutting-edge NVIDIA Hopper/Blackwell features (TMA, mbarriers, NCCL GIN) for maximum performance. The library is designed to overlap communication with computation through deferred epilogue patterns. The codebase exhibits exceptional craftsmanship in its low-level CUDA kernels, particularly the ptx.cuh utility layer and the distributed barrier implementations. The Python utilities (EventOverlap, weak_lru, EnvOverride, BufferAllocator) are well-isolated, elegant patterns that are genuinely reusable in other GPU or distributed computing projects. However, the library has a high technical drift risk due to its deep reliance on experimental and rapidly evolving NVIDIA APIs (NCCL device-side, specific CUDA architecture features). It is tightly coupled to a specific hardware stack (Hopper+ GPUs, NVLink, RDMA) and software versions (CUDA 13.1+, NCCL 2.32.3+). While actively maintained, its specialized nature and complex build system limit its applicability outside of large-scale MoE training on cutting-edge NVIDIA clusters.