Big library? Hand the analysis to your GPU

This one came straight from a listener request. Deep analysis, the sounds-like fingerprint and the vocal detection, chews through a big library one track at a time on the CPU. If that CPU is also running your local LLM and your voices, the scan and the station end up fighting over the same cores. Two fixes shipped together.

What's new

There is a third analyzer image, subwave-analyzer-cuda. It has everything the heavy image has, but CLAP and Demucs run on an NVIDIA GPU instead of the CPU, and Demucs especially is much faster there.

And for everyone, GPU or not, there is a new quiet-times switch. Turn it on and any analysis run pauses while someone is listening, then resumes once the stream has been empty for a while.

Put the card to work

One command, no rebuild:

docker compose -f docker-compose.yml -f docker-compose.analyzer-gpu.yml up -d

The overlay swaps the analyzer to the CUDA image and hands it the GPU. The host needs the NVIDIA driver and the Container Toolkit, nothing else. If the card is not visible, the analyzer notes it in the log and runs on CPU, so nothing breaks while you sort the toolkit out. You can drop ANALYZER_HEAVY from .env too, the overlay covers it.

Analyse at quiet times

Open the Library page in admin. Next to the sounds-like and vocal controls there is a new row, Quiet times. Enable it and set the idle window, ten minutes by default. A running pass pauses between tracks the moment someone tunes in and shows "Waiting for quiet". Once the room has been empty long enough, it picks up where it left off. It applies to manual runs too, so turn it off if you want a scan right now regardless.

Why it helps

A big collection gets its fingerprints without a week of pegged cores, and the scan never steals cycles from the broadcast. Turn on quiet times, kick off a full rescan, and the station does the homework in its own gaps.