DeepStream 7.1 on Jetson Orin Nano — nvvideoconvert (NVMM → system memory) silently fails in headless mode, producing 0-byte encoded files

Accelerated Computing Intelligent Video Analytics DeepStream SDK
,
1

Hardware / Software

  • Device: Jetson Orin Nano Super (8GB)

  • JetPack: 6.2.2

  • DeepStream: 7.1

  • CUDA: 12.6

  • TensorRT: 10.3

  • Pipeline framework: pyservicemaker (Python)

  • Power mode: MAXN_SUPER

Problem

I have a DeepStream pipeline that runs inference + NvDCF tracking and writes an H.264 .mkv file to disk.

The pipeline behaves differently depending on whether Xorg is running:

Mode systemd target DISPLAY Result
Desktop graphical.target (Xorg running) :0 Valid playable file with OSD overlays
Headless multi-user.target (no Xorg) unset 0-byte file, but pipeline runs at full FPS and inference/tracking are correct

Inference and tracking work correctly in both modes. The failure only affects the encode-to-file branch.

There are no bus errors:

  • Pipeline reaches PLAYING

  • FPS is normal

  • File handle is created

  • Output file remains 0 bytes for the entire run

Pipeline (save branch)

… → nvstreamdemux → queue → nvdsosd → tee

├─ queue → nv3dsink (optional display)

└─ queue → nvvideoconvert → x264enc → h264parse → matroskamux → filesink

In desktop mode, nv3dsink creates an NVIDIA EGL context which appears to be shared through GstContext. nvvideoconvert then successfully maps NVMM buffers to system memory.

In headless mode, there is no display sink and no EGL context provider in the pipeline.

Current understanding of the failure

nvvideoconvert converting from:

video/x-raw(memory:NVMM) → video/x-raw

appears to require EGLImage mapping.

On Jetson, EGLImage creation seems to depend on an NVIDIA EGL context backed by Xorg.

Without that context:

  • nvvideoconvert accepts buffers

  • no downstream buffers are produced

  • upstream queues eventually drop frames

  • filesink never receives encoded data

  • no explicit error is posted on the bus

CUDA-only paths (TensorRT inference + NvDCF tracking) continue working because they remain entirely in GPU memory.

What I tried

1. Removing nvdsosd

No change. Output file still remains 0 bytes.

2. compute-hw=1 on nvvideoconvert

Attempted to force CUDA compute instead of VIC.

Pipeline fails to start. The property does not appear valid for nvvideoconvert on Orin Nano under DS 7.1.

3. Xvfb (DISPLAY=:99)

This produces:

nvbufsurface: Failed to create EGLImage

Xvfb provides Mesa EGL rather than NVIDIA EGL. Once DISPLAY is set, NVIDIA plugins appear to attempt EGL initialization against the wrong provider.

4. External recorder path

Using OpenCV + FFmpeg separately from DeepStream works, but requires reading the source twice and bypasses the DeepStream encode path entirely.

Questions

1. Supported headless NVMM → system-memory conversion

Is there a supported way to make nvvideoconvert produce system-memory output in a fully headless environment on Jetson Orin Nano + DeepStream 7.1?

Specifically:

  • Can NVIDIA EGL be initialized without Xorg?

  • Is there a CUDA/VIC path that avoids EGLImage usage entirely?

2. Recommended headless encode topology

Is there a known-good save pipeline for headless Orin Nano deployments?

For example:

nvvideoconvert ! video/x-raw(memory:NVMM),format=I420 \

! nvv4l2h264enc ! h264parse ! matroskamux ! filesink

Would keeping everything in NVMM avoid the EGLImage dependency?

Orin Nano does not have NVENC, so I am not sure whether nvv4l2h264enc is actually usable on this platform.

If software encoding (x264enc) is the expected path, what is the supported way to perform NVMM → CPU memory conversion headlessly?

3. Documentation

Is the EGL/Xorg dependency for NVMM → system-memory conversion documented anywhere?

The failure mode is difficult to debug because:

  • pipeline reaches PLAYING

  • FPS remains healthy

  • no bus errors are emitted

  • only the output file is empty

A warning from nvvideoconvert when EGLImage creation fails would make this much easier to diagnose.

4. Recommended workaround

Is running a minimal Xorg session purely to host the NVIDIA EGL context the recommended workaround for headless Jetson deployments?

If yes:

  • is there an official minimal Xorg configuration?

  • is there a recommended systemd service/unit for this setup?

  • can this be done without a physical monitor attached?

Minimal repro

gst-launch-1.0 \

filesrc location=in.mp4 ! qtdemux ! h264parse ! nvv4l2decoder \

! m.sink_0 nvstreammux name=m batch-size=1 width=1280 height=720 \

! nvvideoconvert ! ‘video/x-raw,format=I420’ \

! x264enc tune=zerolatency speed-preset=ultrafast bitrate=4000 \

! h264parse ! matroskamux ! filesink location=out.mkv

Behavior:

  • graphical.target with DISPLAY=:0 → valid playable file

  • multi-user.target with no DISPLAY → 0-byte file with no errors

Any guidance on the supported headless encode path for Jetson Orin Nano + JetPack 6.2.2 + DeepStream 7.1 would be appreciated.

2

Hello @tech204!

Based on the title and content of your topic, it looks like it may receive better visibility and feedback in a different category. We took the liberty of moving it for you.

If this was an incorrect assessment, please send me a direct message.

Disclaimer: this moderation suggestion and message were generated with AI assistance.

6

I have tested this without connecting a monitor, and it works. However, I need to verify one point: after flashing JetPack, have you never connected a monitor to configure the Orin Nano?

Do not use nvv4l2h264enc on Orin Nano devices; it will not work due to the absence of the necessary hardware module. x264enc is recommended.

Please refer to this FAQ, for this sence, it provided a detailed resolution.

7

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks.