StableHLO

Higgsfield is an open-source GPU workload manager and machine-learning framework for training models ranging from billions to trillions of parameters, including large language models. It allocates exclusive or non-exclusive access to nodes, queues experiments to manage resource contention, and provides interfaces for distributed training with DeepSpeed ZeRO-3 and PyTorch fully sharded data parallelism. Its framework supports initiating, running, monitoring, and checkpointing experiments on allocated nodes, while its LLM training API includes model, data-loader, and experiment components. Higgsfield integrates with GitHub and GitHub Actions to generate deployment and run workflows: required tools and project deployment keys are installed on servers, code is deployed to configured nodes, and experiments are launched and monitored through GitHub. The repository describes compatibility with Ubuntu nodes offering SSH access and passwordless sudo, and reports testing with Azure, LambdaLabs, and FluidStack.

Halofy is an open-source governance layer between organizational knowledge and AI agents. It resolves actor identity, roles, namespaces, and source scope on the server; enforces access policies; and records provenance, supersedence, append-only audit history, policy decisions, and signed erasure certificates for reads and writes. The repository exposes governed memory and context operations over MCP and HTTP, including writing, reading, searching, assembling, faulting, statistics, forgetting, policy, sharing, pinning, export, and manifest operations. Postgres with pgvector is the operational authority, while retrieval engines connect through an ACL-scoped read-only driver interface; encrypted, Git-versioned knowledge provides a cold tier, and embedded PGlite is the zero-setup local default. It includes connectors for filesystem, Postgres, Obsidian, and manual CSV import, plus an offline demo and hermetic test lane using deterministic stub components.

MLX is an array framework for machine learning on Apple silicon, developed by Apple machine learning research. It provides Python, C++, C, and Swift APIs modeled on NumPy, plus higher-level packages such as mlx.nn and mlx.optimizers with PyTorch-like interfaces. Its composable transformations support automatic differentiation, automatic vectorization, and computation-graph optimization; computations are lazy, graphs are built dynamically, and operations can run on supported CPUs and GPUs using shared unified memory rather than explicit data transfers. MLX is distributed through PyPI and supports model training and deployment workflows including language-model training and generation, LoRA fine-tuning, image generation, and speech recognition.

Lithos Metal is an open-source LLM inference engine for Apple silicon that generates specialized Metal megakernels and serves models locally for coding agents and other applications. Its compiler decomposes operators into matrix tiles, attention partitions, and recurrent-state slices, records their dependencies, and generates layer-wise fused GPU tasks. The runtime combines mixer megakernels, GPU-resident generation control, pre-encoded Metal indirect command buffers, and DSpark speculative decoding, in which a target-conditioned draft proposes token blocks and the target model verifies and commits the accepted prefix. The server provides OpenAI-compatible Chat Completions and Responses adapters plus an Anthropic Messages adapter, with text generation, tool calls, and SSE streaming; launchers configure clients including OpenCode, Claude Code, Codex, and Hermes. It runs on Apple silicon with macOS 26 or later, is distributed through a Lithos Homebrew tap or source installation, and is licensed under Apache 2.0.

vLLM is an open-source library and inference-serving engine for deploying large language models on GPUs and other hardware accelerators. Originally developed in UC Berkeley's Sky Computing Lab, it is maintained by a broad community and provides an OpenAI-compatible API server, as well as Anthropic Messages API and gRPC support. Its serving stack manages attention key-value memory with PagedAttention, continuous batching, chunked prefill, prefix caching, streaming generation, and disaggregated prefill, decode, and encode. It also supports speculative decoding, quantization, optimized attention and GEMM/MoE kernels, automatic kernel generation and graph-level transformations, structured outputs, tool calling, multiple decoding algorithms, and distributed inference through tensor, pipeline, data, expert, and context parallelism. The project integrates with Hugging Face models and supports decoder-only, mixture-of-experts, hybrid state-space, multimodal, embedding, retrieval, reward, and classification models. It can run across NVIDIA, AMD, Intel, and other supported accelerators, as well as x86, ARM, and PowerPC CPUs, and can be installed with uv or pip or built from source.