Hardware
Leyla plays a file as it is whenever the screen can show it, and that needs no hardware at all. This page is for the times something must be converted, and for keeping an eye on the disks your films live on.
When hardware matters
Every time a screen presses play it tells Leyla what it can decode, and Leyla picks the lightest of three ways to send the film:
- Play
- The file as it is. Nothing runs on the server.
- Copy
- The picture goes untouched and only the wrapping changes. This costs almost nothing, even on a Raspberry Pi.
- Convert
- Only what the screen cannot play is re-encoded. This is the one that needs a graphics chip, or a fast CPU, for 4K.
A server with no graphics chip still works. Conversions run on the CPU, which is fine for 1080p on most machines. How a film reaches a screen has the whole story.
What Leyla can use
| Machine | What Leyla uses | How it gets there |
|---|---|---|
| Intel with graphics (most “+” NAS models, mini PCs) | Quick Sync and VAAPI | /dev/dri |
| AMD with graphics | VAAPI | /dev/dri |
| NVIDIA card | NVENC | The NVIDIA runtime |
| Raspberry Pi, ARM NAS boxes, Intel without graphics | The CPU | Nothing to set |
Turn it on
Intel or AMD
Check the box has the device
$ ls /dev/dri # renderD128 means it is thereUncomment
devicesin the compose filedevices: - /dev/dri:/dev/driLeyla joins the device’s group itself. On unRAID, keep
/dev/driin the template.Start the container again
$ docker compose up -d
Important On a box without /dev/dri, such as AMD Ryzen Synology models and most ARM boxes, the container will not start with devices uncommented. Comment the block out again and conversions run on the CPU.
NVIDIA
Install NVIDIA’s drivers and container runtime on the host first. On TrueNAS that is under Apps › Settings; on unRAID it is the Nvidia-Driver plugin. Then add this to the service:
runtime: nvidia
environment:
NVIDIA_VISIBLE_DEVICES: all
NVIDIA_DRIVER_CAPABILITIES: compute,video,utilityOn unRAID, put --runtime=nvidia in Extra Parameters and set NVIDIA_VISIBLE_DEVICES in the template.
How Leyla decides to trust an encoder
Leyla tests each encoder on your own machine when it starts, with the shapes of video it will actually be asked to convert. Only what passes is used. An encoder that fails its test is never used, Leyla records why, and the CPU takes over.
HDR10, HLG and Dolby Vision go untouched to screens that show them. For a screen that cannot, Leyla tone-maps on the server, so a dark scene stays readable instead of grey.
The ffmpeg inside is Leyla’s own build (8.1.3). It has no network protocols: it reads files and pipes and nothing else, so a crafted file cannot make it fetch anything. Every ffmpeg runs as a separate, supervised process with a clean environment.
Disk health
Leyla warns about failing drives from the read errors it meets anyway, with nothing to set up. To add each drive’s own SMART counters, give the container the disks and the right to read them:
devices:
- /dev/sda:/dev/sda # one line per disk under a library
cap_add:
- SYS_RAWIO
Leyla reads the counters once a day and never wakes a sleeping disk. Network shares and pooled drives, such as mergerfs, have no disk to read. Leyla says so, and everything else works.
If something is wrong
- The container will not start after you turned hardware transcoding on. The box has no
/dev/dri. Comment thedevicesblock out again. - Picture or sound stutters. The film is being converted and the CPU cannot keep up, usually 4K on a box without a graphics chip. A screen that plays the file as it is never stutters this way.
Something missing or wrong here? Tell us.