Radiance-LLM stilldeadcode/radiance:1.17.0 https://hub.docker.com/r/stilldeadcode/radiance 1.17.0 Pinned release (recommended) latest Follows the latest release 1.16.0 Previous release bridge sh false https://github.com/Aurial/Unraid-Community-Apps/issues https://codeberg.org/StillDeadcode/radiance https://raw.githubusercontent.com/Aurial/Unraid-Community-Apps/main/README.md Radiance is a modular inference engine for AMD GPUs: it serves language models (OpenAI chat/completions/Responses, Anthropic Messages and llama-server APIs, so their clients work unchanged) plus image, video, speech and music generation from one server, and ships a live dashboard at the root URL. This template runs the official image, stilldeadcode/radiance, Apache-2.0, maintained by StillDeadcode (Codeberg). Requirements: the amdgpu kernel driver, which Unraid ships as standard (Linux 6.8+), and an RDNA 3 / RDNA 4 / Ryzen AI Max card: RX 9000 series and Radeon AI PRO R9700 (RDNA4), RX 7000 series (RDNA3) or Strix Halo. The host needs no ROCm, no compiler, no driver plugin. Resizable BAR (Above 4G Decoding) should be on in the firmware. Model files must sit on a fast local pool (ext4/XFS/btrfs, e.g. /mnt/cache/...), never /mnt/user: the engine reads the model by direct I/O while it serves. Do not set a container memory limit. Models are .rad packs the author publishes on Hugging Face: qwen3.8-27b-mxfp4.rad (19 GiB) or qwen3.8-27b-fp8.rad (29 GiB) are the natural single-card starting points; Qwen3.8-Flash-Next (114 GiB) is the flagship MoE and wants two cards (a one-card profile exists but pins about 49 GiB of host RAM). Download one into the Models folder, then set its exact filename after the /models/ in Post Arguments. API key: unlike some servers this one NEVER generates a key. Every request but /health, /ping and the dashboard needs Authorization: Bearer KEY (or X-Api-Key), and the server REFUSES TO START if the key variable is empty (fail closed). Generate one with: openssl rand -hex 24. To run keyless on a trusted network, delete the api-key-env flag from Post Arguments. Launch flags: the Post Arguments field holds the engine command line; it starts safe for one card. The README explains every flag (tp, placement, host pool, kv cache, speculation, batch sizes) with single- and two-card recommendations measured on a live two-R9700 box. Template and Unraid problems: the Support link. Radiance itself: the Project link. AI: llm qwen local ai openai anthropic api inference chat amd rocm radeon rdna gpu image generation video speech music dashboard rad flashnext mxfp4 moe speculative decoding http://[IP]:[PORT:8000]/ https://raw.githubusercontent.com/Aurial/Unraid-Community-Apps/main/templates/radiance.xml https://raw.githubusercontent.com/Aurial/Unraid-Community-Apps/main/radiance-icon.png Apache-2.0 --device /dev/kfd --device /dev/dri --group-add video --ulimit memlock=-1:-1 --security-opt seccomp=unconfined --stop-timeout 120 --model /models/qwen3.8-27b-mxfp4.rad --api-key-env RADIANCE_API_KEY **amdgpu** kernel driver, which Unraid ships (Linux 6.8+): no driver plugin, no host ROCm. An RDNA 3 / RDNA 4 / Ryzen AI Max card (RX 7000, RX 9000, Radeon AI PRO R9700, Strix Halo); other AMD cards need a kernel plugin (see README). Enable **Resizable BAR** / Above 4G Decoding in the firmware. The Models folder must be on a fast local pool with direct I/O (ext4/XFS/btrfs, e.g. /mnt/cache/...), not /mnt/user. The container runs **seccomp unconfined**: Docker's default profile blocks io_uring and the NUMA calls radiance needs, and without it models that stream from disk cannot serve. Do not set a memory limit. One model at a time: the engine fills the cards on purpose; stop other GPU workloads first. 8180 /mnt/cache/appdata/radiance/models