Capturas de pantalla del proyecto
Resumen
hal0 is a self-hosted AI inference platform that transforms a Linux machine into a polished, OpenAI-compatible inference appliance. It is natively optimized for AMD Ryzen AI Max+ 395 (Strix Halo) with 128 GB of unified memory but supports fallback deployments on other AMD, NVIDIA, and CPU-only systems. The platform uses a containerized architecture where every inference workload runs as an isolated Podman container under dedicated systemd units. A single control plane (`hal0-api` on port 8080) manages slot state machines, dispatches OpenAI-compatible `/v1/*` requests to the appropriate slot ports, and serves the web dashboard. It features declarative configuration, atomic self-updates, and a built-in hardware-aware probe.
Características principales
- OpenAI-compatible /v1/* API gateway
- Isolated Podman containers per inference slot
- Native AMD Strix Halo and XDNA NPU support with FLM trio packing
- Hardware-aware probe with unified memory and VRAM fit warnings
- Declarative Stacks for atomic model/slot layout changes with rollback
- OmniRouter with 8 client-side tool-calling tools (image generation/editing, TTS, transcription, vision, embed, rerank, route_to_chat)
- Cosign-verified atomic self-updates with one-flag rollback
- Built-in dashboard with React 18, live telemetry, and journald log streaming
- Companion service management (Open WebUI, ComfyUI, Hermes, Hindsight, n8n) with mDNS discovery
- hal0-brain steward agent with 74-tool admin catalog and per-persona tool policy
Requisitos, instalación y guía rápida
sh
curl -fsSL https://hal0.dev/install.sh | bash
For Proxmox VE, a dedicated script creates an unprivileged Debian 13 LXC and runs the standard bootstrap:
sh
bash -c "$(curl -fsSL https://raw.githubusercontent.com/Hal0ai/hal0/main/scripts/proxmox-ve/hal0.sh)"
Environment variables like `HAL0_SKIP_SETUP=1`, `HAL0_SKIP_BRAIN_MODEL=1`, or `HAL0_NONINTERACTIVE=1` can be set to skip specific installer steps. Re-running the installer is safe as all provisioning steps are idempotent.
Uso
sh
# backend
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
hal0 serve --reload
# frontend
cd ui
npm install
npm run dev
Day-2 operational commands include:
sh
systemctl status hal0-api
systemctl list-units 'hal0-slot@*'
journalctl -fu hal0-api
systemctl restart hal0-slot@agent
Use `hal0 doctor` to check pre-flight requirements and `hal0 uninstall [--keep-data]` to tear down an installation.
Compatibilidad de modelos y casos de uso
Supports local models pulled from Hugging Face via `hal0 model pull <ref>` (stored in `registry.toml`) and external upstream models (OpenRouter, Anthropic, OpenAI, Google AI Studio, Ollama, custom). Local inference backends include ROCm (FP4 fork), Vulkan, experimental CUDA, CPU, and AMD XDNA NPU (via FastFlowLM). Specific model configurations depend on the slot profile (e.g., `rocm`, `rocm-dnse`, `vulkan`, `cuda`, `flm`, `comfyui`).
Notas sobre la licencia y los riesgos
Apache-2.0
Editorial verification 2026-08-02: repository URL, owner, description, license and repository statistics were reviewed. License metadata: Apache-2.0. README was fetched for the channel draft; re-check repository dependencies, releases and model terms before production use.
Lanzamiento y mantenimiento
v0.9.4 — public beta. Container-runtime era with declarative config. Features the hal0-brain steward, upstream model controls, in-dashboard benchmarks, dashboard redesign with live telemetry, and companion service management.