Skip to content

SUPPORT-MATRIX.md

Latest commit

 

History

History
122 lines (97 loc) · 10.5 KB

SUPPORT-MATRIX.md

File metadata and controls

122 lines (97 loc) · 10.5 KB
Raw

ODS Support Matrix

Last updated: 2026-10-04

What Works Today

Linux (NVIDIA, AMD Strix Halo), Windows (through Ubuntu on WSL2) and macOS (Apple Silicon) are supported. Only Linux + Strix Halo is Tier A; the others are Tier B. Intel Arc and AMD discrete GPUs are Tier C.

For the release gate behind these claims, see RELEASE_VALIDATION.md, VALIDATION-MATRIX.md, and TESTING.md. The validation docs are sanitized so they can be public without exposing private lab hostnames, LAN addresses, or paths. Support status means the project has an intended installer/runtime path for that platform; release evidence should still name the current run, enabled hardware classes, and any deferred or skipped phases.

Platform Status What you get today
Linux + AMD Strix Halo (llama.cpp, Vulkan) Fully supported Complete install and runtime. Primary development platform.
Linux + NVIDIA (CUDA) Supported Complete install and runtime. CI checks package-manager detection, prerequisite installation and installer syntax in 11 distro containers (not full installs); maintainers run full installs in a private distro lab, and GPU runtime is validated on real NVIDIA hardware.
Windows (Docker Desktop + WSL2) Supported .\install.ps1 prepares WSL2, Docker Desktop and Ubuntu, then runs the Linux installer inside Ubuntu with Portal (Pixel) and without Hermes. NVIDIA GPUs run in WSL; AMD GPUs run through llama.cpp's llama-server.exe (Vulkan) on Windows. Count Windows as current release-fleet evidence only when the Windows target is enabled and produces artifacts for that candidate.
macOS (Apple Silicon) Supported Complete install and runtime via ./install.sh. Native Metal inference + Docker services.
Linux + Intel Arc (SYCL) Experimental Hardware detection does not identify Arc yet: pass --tier ARC or --tier ARC_LITE to use docker-compose.arc.yml. Runtime validation on A770/A750 is still pending in this repo. See INTEL-ARC-GUIDE.md.
Linux + AMD discrete GPU Validation required Detected and routed to the AMD overlay; no validated tier/model results yet.

Support Tiers

  • Tier A — fully supported and actively tested in this repo
  • Tier B — supported (works end-to-end, broader validation ongoing)
  • Tier C — experimental or planned (installer diagnostics only, no runtime)

Agent selection matrix

The ODS platform matrix above is broader than Pixel's current qualification. Agent gating never changes whether the rest of ODS is supported.

Host/runtime Default agent result
Ubuntu 24.04/26.04 or Debian 12, PID1 systemd, model route through the authenticated ODS gateway Pixel preferred; Hermes remains available
WSL2 running a qualified distro with systemd, using the ODS Linux installer Same Pixel host path as Linux; Docker Desktop alone does not install it
Apple Silicon macOS, using the ODS macOS installer Native Pixel enabled by default, with Docker ingress/sandbox services; Hermes disabled while Pixel is selected
Qualified Pixel host using an external OpenAI-compatible endpoint (Lemonade Server included) Pixel uses the selected upstream through the authenticated ODS LiteLLM gateway
Other supported Linux distributions or WSL1 Hermes fallback
Legacy native Windows installer (installers/windows/install-windows.ps1, not used by .\install.ps1) Hermes; Pixel edge/relay are disabled because no Pixel host runtime is installed

Pixel source and its verified install bundle ship in the public ODS repository. Installation needs neither access to Osmantic/Pixel nor a separate license acknowledgement flag. See PIXEL.md for technical qualification and the bundled Pixel License for ODS.

Platform Matrix (detailed)

Platform GPU Path Installer Tier Notes
Linux (Ubuntu/Debian family) NVIDIA (llama-server/CUDA) Tier B Validated on real high-memory multi-GPU NVIDIA hardware; CI covers package-manager and syntax checks on 11 distros, and full distro installs run in a private lab
Linux (Strix Halo / AMD unified memory) AMD (llama-server, Vulkan; ROCm optional) Tier A Primary managed path via docker-compose.base.yml + docker-compose.amd.yml; validated on real Strix Halo hardware
Linux (Intel Arc A770/A750) Intel SYCL (llama-server/oneAPI) Tier C docker-compose.arc.yml; builds llama.cpp from intel/oneapi-basekit; see INTEL-ARC-GUIDE.md
Windows (Docker Desktop + WSL2) NVIDIA via Docker Desktop; AMD via llama-server.exe (Vulkan) on Windows Tier B .\install.ps1 runs the Linux installer inside Ubuntu on WSL2 (Portal, no Hermes); the runtime lives in ~/ods inside Ubuntu. A Windows laptop fleet target tracks Docker Desktop/WSL2 evidence
macOS (Apple Silicon) Metal (native llama-server) Tier B Standalone installer (./install.sh) with chip detection, native Metal inference, Docker services, and LaunchAgent auto-start; validated on constrained and high-memory Apple Silicon lab hosts

GPU Tier Map

The installer assigns these tiers, then the catalog selector picks the installed model for the measured memory (see the README hardware table). The models below are the tier-map fallback, used when the catalog is unavailable.

Installer Tier Hardware Fallback model VRAM Backend
NV_ULTRA NVIDIA 90 GB+ Qwen3-Coder-Next (x86_64); Qwen3.6 35B-A3B (arm64) ≥ 90 GB CUDA
SH_LARGE AMD Strix Halo 90+ Qwen3.6 35B-A3B ≥ 90 GB (unified) Vulkan
SH_COMPACT AMD Strix Halo < 90 GB Qwen3.6 35B-A3B < 90 GB (unified) Vulkan
4 NVIDIA 40 GB+ / multi-GPU Qwen3.6 35B-A3B ≥ 40 GB CUDA
3 NVIDIA 20 GB+ Qwen3.5 27B ≥ 20 GB CUDA
ARC Intel Arc ≥ 12 GB (A770; B580 not yet validated) Qwen3.5 9B ≥ 12 GB SYCL
2 NVIDIA 12 GB+ Qwen3.5 9B ≥ 12 GB CUDA
ARC_LITE Intel Arc < 12 GB (A750, A380) Qwen3.5 4B 6–11 GB SYCL
1 NVIDIA 4 GB+ Qwen3.5 9B ≥ 4 GB CUDA
0 NVIDIA < 4 GB, or --tier 0 (CPU-only hosts are usually tier 1) Qwen3.5 2B any CPU
CLOUD No local GPU Claude (API) — LiteLLM

Current Truth

  • Linux (NVIDIA, AMD Strix Halo), Windows (through Ubuntu on WSL2) and macOS (Apple Silicon) are supported. Only Linux + Strix Halo is Tier A; the others are Tier B. Intel Arc and AMD discrete GPUs are Tier C.
  • Pixel uses the native macOS path on Apple Silicon or the qualified Linux/WSL2 systemd path above. Agent eligibility is distinct from overall ODS support.
  • Linux + NVIDIA is supported and validated on real high-memory NVIDIA hardware. CI checks package-manager detection, prerequisite installation and installer syntax in 11 distro containers (Ubuntu 26.04/24.04/22.04, Debian 12, Mint 21.3, Fedora 41, Rocky 9, Arch, Manjaro, CachyOS, openSUSE Tumbleweed); it does not run full installs.
  • Windows installs via .\install.ps1, which runs the Linux installer inside Ubuntu on WSL2 with Docker Desktop. AMD local inference runs in llama.cpp's llama-server.exe (Vulkan) on Windows. Windows support is not inferred from Linux/macOS; treat it as release-current only when a Windows fleet target produces artifacts for that candidate.
  • The legacy native Windows installer (installers/windows/install-windows.ps1) is no longer used by .\install.ps1.
  • macOS installs via ./install.sh: native Metal inference, a native Pixel gateway and managed host helpers, plus Docker UI, ingress, sandbox and supporting services. Native Pixel does not require a separate Lima VM or Linux systemd.
  • AMD runtime diagnostics are explicit: .env records runtime, location, selected backend, supported backends, and whether ODS manages the process. AMD GPUs run on llama.cpp's llama-server; a Lemonade Server you run yourself connects as an external OpenAI-compatible server. See AMD GPUs now run on llama.cpp.
  • AMD discrete GPUs beyond the documented Strix Halo path should be treated as validation-required until the repo has tier/model benchmarks for that hardware.
  • Intel Arc (SYCL) is Tier C / experimental. Hardware detection does not identify Arc yet; pass --tier ARC or --tier ARC_LITE, which selects the Arc compose overlay. ComfyUI and Whisper GPU acceleration are not yet available for Arc. See INTEL-ARC-GUIDE.md for limitations.
  • Release-readiness claims should cite a matching version/tag, relevant distro-lab evidence, and a real-hardware fleet receipt from VALIDATION-MATRIX.md.
  • A supported platform can have code and installer support even when it is not included in every default private release-fleet run. Release notes should cite which hardware classes actually ran, which phases passed, and which surfaces were deferred or skipped.
  • Version baselines for triage are in docs/KNOWN-GOOD-VERSIONS.md.

Roadmap

Target Milestone
Now Linux (NVIDIA, Strix Halo), Windows (WSL2) and macOS supported; Strix Halo is Tier A
Now Intel Arc (SYCL) experimental: manual --tier, runtime validation pending
Ongoing CI smoke matrix expansion for all platforms
Planned Promote Intel Arc to Tier B after broader A770/B580 validation
Planned Arc-accelerated Whisper STT overlay

Next Milestones

  1. Keep the CI/container/Incus distro matrix green and add targeted full-install VM lanes where regressions justify the cost.
  2. Keep Windows laptop fleet evidence current for Docker Desktop/WSL2, NVIDIA mobile GPU, and Intel hybrid-GPU behavior.
  3. Expand macOS test coverage across more Apple Silicon generations and RAM tiers.
  4. Validate Intel Arc B580 (Battlemage 12 GB) on the ARC tier.
  5. Promote Intel Arc from Tier C to Tier B after A770 + B580 real-hardware validation.

See also