Agent Known Issues

This page is for version 3.3.0 (default). For other versions, use one of these documentation indexes: - 3.3.0 (default): https://docs.nvidia.com/vss/3.3.0/llms.txt - Dev: https://docs.nvidia.com/vss/dev/llms.txt - Archives: https://docs.nvidia.com/vss/archives/llms.txt

VLM Clip Duration and Report Accuracy

• For optimal results, use video clips between 5 and 300 seconds for VLM analysis.
• VLM-generated reports may contain inaccurate timestamps.

Shared-GPU NIM Deployment

On H100 and RTX PRO 6000 Blackwell, only the default LLM and VLM NIMs have been tested on shared GPUs. Shared-GPU mode is not supported on L40S. Use dedicated GPUs or remote NIMs for other models.

Nemotron Omni and ENABLE_AUDIO

Non-Omni VLM with ENABLE_AUDIO=true

If ENABLE_AUDIO=true but VLM_NAME does not contain omni, the agent logs a warning and falls back to JPEG frame sampling (audio is dropped). Non-Omni remote VLMs cannot use VST video_url the way a co-located NIM can, so enabling audio without an Omni model does not preserve the audio track.

Omni model id

Audio-aware prompts and full-MP4 delivery apply only when omni appears in VLM_NAME. A mismatch between VLM_NAME and the model id from GET /v1/models on your vLLM endpoint causes HTTP 404 on chat completions. See Using Nemotron Omni (audio-enabled VLM).

Firewall ports (remote VLM)

• Remote VLM with model-type nim (not Omni audio path): the VLM must reach VST on port 30888 on the host. See How does a remote nim VLM access videos?.
• Remote Nemotron Omni with ENABLE_AUDIO=true on the base profile: the agent inlines the MP4 in the API request; ensure port 30082 (OpenAI-compatible VLM API) is reachable from the VSS host. See Using Nemotron Omni (audio-enabled VLM) and How does a remote nim VLM access videos?.

Snapshot Timestamp Accuracy (Video Summarization Profile)

Symptom:

When requesting a snapshot at a specific timestamp (e.g., “take a snapshot at 3 seconds from <sensor_id>”), the actual frame captured may not precisely match the requested timestamp.

Cause:

This is a known limitation in the current frame extraction implementation. The snapshot tool may select a nearby frame rather than the exact frame at the requested timestamp.

Solution:

We expect to fix this in the next release.

Empty Chart for Metrics Queries When No Results

Symptom:

For natural language metrics/occupancy count queries (e.g., How many people are there in camera 1?), when the count is 0 the chart appears empty, which can be confusing.

Cause:

This is a known limitation in the current chart rendering for metrics queries.

Solution:

This will be resolved in a future release.