Engine and hardware compatibility¶
This page describes the repository capability catalog, not successful tests on every device. Availability also requires live driver/resource readiness, free slots and usable memory telemetry.
Source: compute-target catalog.
Ordinary engine targets¶
| Compute target | Architecture | Configured engines |
|---|---|---|
| CPU | amd64, arm64 | vLLM, Ollama |
| NVIDIA GPU | amd64, arm64 | vLLM, Ollama, FreeToken |
| AMD GPU (ROCm) | amd64 | vLLM, Ollama |
| Intel GPU (XPU) | amd64 | vLLM |
FreeToken admission policy¶
- Upstream version:
0.1.3. - OS / host architecture: linux / amd64.
- Vendor: nvidia; NVIDIA driver major at least 580; CUDA major 13.
- Admitted compute capabilities: 8.6, 8.9, 12.0.
- Allocation contract:
whole-gpus-single-node; whole homogeneous GPUs, not MIG or time-slicing slots. - Memory strategies: auto, fused, offload, cpu, hybrid; start with
auto. - Precision choices: auto, float16, bfloat16, float32.
- Model IDs are checked against the catalog policy; a family name is not blanket approval.
See FreeToken configuration and the user guide.
Experimental Realtime profiles¶
These are runtime selections, not hardware or model certifications. An empty image needs an explicitly compatible runtime.
| Profile | Compute target | Requested device counts | Bundled image |
|---|---|---|---|
qwen3-omni |
nvidia-gpu | 1, 2 | Pinned |
qwen3-omni-rocm |
amd-gpu | 1, 2 | Not supplied |
qwen3-omni-xpu |
intel-gpu | 1, 2 | Not supplied |
qwen3-omni-cpu |
cpu | 1 | Not supplied |
A CPU profile has no physical GPU despite its stage/device-count convention. See Realtime and GPU operator support for additional boundaries.