Troubleshooting#
1. DL Streamer Optimizer takes a long time#
Optimization runs are long‑running: it may take 5–6 minutes (or more, depending on pipeline complexity and hardware) for DL Streamer Optimizer to explore variants and return an optimized pipeline.
The optimizer runs as a job on DL Streamer Pipeline Server, so the candidate pipelines run on the server’s runtime and
devices. A candidate that crashes fails only the optimization job; the ViPPET application keeps running. A running
optimization job can be cancelled with DELETE /api/v1/jobs/optimization/{job_id}.
For more information about DL Streamer Optimizer behavior and limitations, see the DL Streamer Optimizer limitations section in the DL Streamer documentation: DL Streamer Optimizer limitations.
If ViPPET or the server is restarted while DL Streamer Optimizer is running, the optimization job is interrupted and its results are lost. When ViPPET starts or stops, it cancels any optimizer job still running on the server.
2. NPU metrics are not visible in the UI#
ViPPET currently does not support displaying NPU‑related metrics:
NPU utilization, throughput, and latency are not exposed in the ViPPET UI.
Metrics and visualizations are limited to what is currently integrated for other devices.
As a result, even if pipelines use an NPU, you will not see NPU‑specific telemetry in ViPPET.
3. Occasional “Connection lost” message in the UI#
The ViPPET UI is a web application that communicates with backend services. Under transient network interruptions or short service unavailability, the UI may show a “Connection lost” message.
It typically appears sporadically.
It is often related to short‑lived connectivity issues between the browser and the backend.
If the “Connection lost” message appears occasionally: Refresh the browser page to re‑establish the connection to the backend.
4. Application restart removes user-created pipelines and jobs#
In the current release, restarting the ViPPET application removes:
All pipelines created by the user, and
All types of jobs (tests, optimization runs, validation runs, and similar).
After a restart, only predefined pipelines remain available. If a restart happens during a long‑running operation (for example, during DL Streamer Optimizer runs), the in‑progress job is lost, and you need to recreate or reimport your custom pipelines and rerun the jobs.
5. Support limited to DL Streamer 2026.3.0 pipelines and models#
ViPPET currently supports only pipelines and models that are supported by DL Streamer 2026.3.0.
For the full list of supported models, elements, and other details, see the DL Streamer release notes: DL Streamer release notes
If a custom pipeline works correctly with DL Streamer 2026.3.0, it is expected to also work in ViPPET (see also the “Limited validation scope” limitation below).
6. Limited validation scope#
Validation and testing in this release focused mainly on sanity checks for predefined pipelines.
For custom pipelines:
Their behavior in ViPPET is less explored and may vary.
However, if a custom pipeline is supported and works correctly with DL Streamer 2026.3.0, it is expected to behave similarly when run via ViPPET (see also “Support limited to DL Streamer 2026.3.0 pipelines and models” above).
7. Recommended to run only one operation at a time#
Currently, it is recommended to run a single operation at a time from the following set:
Tests,
Optimization,
Validation.
In this release:
New jobs are not rejected or queued when another job is already running.
Starting more than one job at the same time launches multiple GStreamer instances.
This can significantly distort performance results (for example, CPU/GPU utilization and FPS).
For accurate and repeatable measurements, run these operations one by one.
8. Some GStreamer / DL Streamer elements may not be displayed correctly in the UI#
Some GStreamer or DL Streamer elements used in a pipeline may not be displayed correctly by the ViPPET UI.
Even if some elements are not shown as expected in the UI, the underlying pipeline is still expected to run.
9. Supported models list is limited and extending it is not guaranteed to work#
ViPPET currently supports only models defined under:
vippet/models/— one YAML file per model. This catalog is synced into the application database at startup (insert-only: existing rows are never updated or removed when a file changes).
A user can try to add a new model by creating a new YAML file in
vippet/models/ (following the shape of an existing entry), but there is
no guarantee that such models will work out of the box. New entries
must point to a source/hub that the model-download microservice
knows how to handle (for example huggingface, ultralytics, openvino,
pipeline-zoo-models, geti, hls).
vippet/models/ is copied into the image at build time, not mounted as a
runtime volume, so after adding or editing a model file you need to
rebuild the image, not just restart the stack:
make build run
(If you are running with the dev override, make run-dev, the whole
vippet/ source tree is bind-mounted, so a restart — make stop
then make run-dev — is enough, no rebuild required.)
After the stack comes back up, install the new models from the Models
page in the UI (or via the /api/v1/models API).
10. Pipelines cannot depend on files other than models or videos#
Currently, ViPPET does not support pipelines that require additional files beyond:
Model files,
Video files, and
User-defined Python scripts.
Pipelines that depend on other external artifacts (for example, configuration files, custom resources, etc.) are not supported in this release.
11. Application containers fail to start#
In some environments, ViPPET services may fail to start correctly and the UI may not be reachable. In such cases, stop the currently running containers and start them again with the default configuration:
Check container logs:
docker compose logs
Restart the stack using the provided Makefile:
make stop run
12. Port conflicts for vippet-ui#
If the vippet-ui service cannot be accessed in the browser, it may be caused by a port
conflict on the host. If that is the case, restart the stack and access ViPPET using the new
port, e.g., http://localhost:8082:
In the Compose file (
compose.yml), find thevippet-uiservice and itsportssection:services: vippet-ui: ports: - "80:80"
Change the host port (left side) to an available one, for example:
services: vippet-ui: ports: - "8082:80"
13. Network Camera Discovery does not find cameras#
If the Sensor Manager service does not find any cameras on the network, check the following:
Ensure that the cameras are ONVIF‑compliant and support discovery.
Verify that the cameras have ONVIF services enabled.
Confirm that the cameras are on the same network segment as the ViPPET application.
Check for any firewall rules or network configurations that may block discovery traffic (UDP multicast
239.255.255.250:3702).Check that the
sensor-managercontainer is running and reachable from thevippetcontainer (docker logs sensor-manager,curl http://localhost:8090/api/v1/sensorson the host).
14. Network Camera Authentication fails#
If you are able to discover network cameras but cannot authenticate to them, check the following:
Verify that the correct username and password are being used for each camera.
Ensure time synchronization between the ViPPET host and the cameras, as some ONVIF implementations require closely synchronized clocks for authentication.
Check for any specific ONVIF profiles or settings required by the cameras for authentication.
15. GigE Vision camera is discovered, but the pipeline receives no frames#
If a GigE Vision camera appears in the Cameras tab, but a pipeline that uses it fails or does not produce frames, check the following:
If validation fails with an unknown element (
aravissrc), the server could not load the Aravis plugin that ViPPET ships in thevippet-gst-pluginsvolume. Rundocker exec dlstreamer-pipeline-server gst-inspect-1.0 aravissrcand checkdocker logs dlstreamer-pipeline-serverfor aFailed to load pluginwarning. Rebuild thevippetimage if it predates the plugin, and make sure the server image uses the same Ubuntu 24.04 base asvippet(seeDLSPS2_IMAGEandDLSPS2_BASE_IMAGE).Ensure that ViPPET was started with
make run GIGE=1. Without host networking, thedlstreamer-pipeline-servercontainer cannot receive the UDP video stream (GVSP) sent by the camera.Make sure that the camera is not streaming to another application. A GigE Vision camera can be controlled by only one application at a time.
Check for firewall rules on the host that may block incoming UDP traffic from the camera.
Verify that the IP address shown in the Cameras tab matches the one used in the pipeline (
gige://<IP address>). If the camera IP address has changed, select the camera again in the Pipeline Builder.For high-resolution cameras, enable jumbo frames (for example, MTU 9000) on the network interface connected to the camera to reduce packet loss.
16. Known limitations of pipeline execution on DL Streamer Pipeline Server 2.0#
Every pipeline runs on DL Streamer Pipeline Server 2.0 (the dlstreamer-pipeline-server container). See
Pipeline execution on DL Streamer Pipeline Server 2.0
for the setup. The following limitations apply:
The Aravis plugin for GigE Vision cameras comes from the
vippetimage. The server image does not includearavissrc, so ViPPET ships the plugin andlibaravisfrom its Ubuntu 24.04 packages in thevippet-gst-pluginsvolume. It loads on a server image with the same base, which the default deployment provides; with another server image (DLSPS2_IMAGE,DLSPS2_BASE_IMAGE), a GigE Vision pipeline can fail validation with an unknown element (aravissrc).Error details of a failed run are short. A failed run reports only the server’s one-line status message and the failing element. For the full GStreamer log, run
docker logs dlstreamer-pipeline-server, or open a shell in the server withmake shell-dlsps.The latency tracer runs for every pipeline. Tracers are process-wide, so the
latency_tracerthat collects latency metrics also runs for tests that do not ask for them, and can lower their FPS. For FPS measurements without the tracer, setDLSPS2_GST_TRACERS=(empty) beforemake run; tests that ask for latency metrics then run without them, and the job details say so.ViPPET is the server’s only client. When the
vippetcontainer starts or stops, it stops every pipeline still running on the server, including pipelines started on the server by other means.Stopping a run can take a few seconds. If a run is stopped while the server is still validating it, the server stops it only after the validation ends.
Known issue: a VLM density test on CPU can make every later run fail. The server runs all pipelines in one shared worker process, and each CPU VLM stream needs several gigabytes of memory. A density test of a VLM pipeline on CPU, such as Alerting VLM, can run that process out of memory, and every test, validation or optimization started after it fails. The server also keeps part of the memory of every finished pipeline, so its memory use grows over a long series of tests. Restart the server with
docker restart dlstreamer-pipeline-serverwhen runs start failing, and between long series of tests.