Skip to content
dsp-kitchen

2026-11-06: Kilosort4 in the container and through the server

Question. Can a recording be sorted on a GPU inside the hardened container, and started from another client?

Setup

Section titled “Setup”
GPU NVIDIA RTX 2070, driver 615.71
Container Rootless Podman 6.1, Fedora 45, the nvidia image target
Runtime CUDA
Recording The Kilosort4 test recording: Neuropixels, 45 s at 30 kHz

In the container, 30 s

Section titled “In the container, 30 s”
Terminal window
podman run --rm --device nvidia.com/gpu=all --security-opt=label=disable --userns=keep-id:uid=10001,gid=10001 \
  -v ~/dspk-data:/data:Z dspk-serve:nvidia \
  sort --runtime cuda kilosort4 /data/recording.ap.bin --probe neuropixels1 \
  --output /data/recording.kilosort4.sorting.zarr --duration-sec 30

Result: 136 units, 63,014 spikes, in 28.6 s.

dspk doctor in the same image listed wgpu and CUDA as usable.

Through the server, whole recording

Section titled “Through the server, whole recording”

A client enrolled with --control asked the server to sort the whole 45 s recording on CUDA.

Result: 287 units, 133,544 spikes, in 46 s. The sorting, fetched back to the client, was identical.

The same recording sorted locally with dspk sort kilosort4 also gave 287 units and 133,544 spikes.

Found on the way, and fixed the same day: a stopped or failed job left its partial output and blocked reruns, and the server’s runtime wasn’t visible to clients.