DeepStream 9.1 is here — Agentic Skills + a unified monorepo

Accelerated Computing Intelligent Video Analytics DeepStream SDK
,
1

Hey all,

DeepStream 9.1 just dropped, and there are a few key things worth your attention:

[1] DeepStream has a new address! It’s all in one repo now, github.com/NVIDIA/DeepStream: source, packages, samples, tools such as AMC and Inference Builder, and Skills, across x86/Jetson/SBSA.

  • SDK packages (tar/deb) → are now available as part of the GitHub release assets, and are no longer on NGC.
  • Pre-built containers → are still on NGC under a multi-arch unified deepstream repo. The legacy DeepStream-l4t is no longer maintained. The Dockerfiles are now available on the DeepStream monorepo.
  • Building from source? Make sure you grab the deepstream-binaries packages from the monorepo.

[2] A set of agentic skills you can use to accelerate building your pipelines: DeepStream/skills at main · NVIDIA/DeepStream · GitHub

We will continue to push additional skills later this month, but two worth looking at are:

  • MV3DT — accurate, end-to-end 3D tracking across multiple cameras.
  • AMC (AutoMagicCalib) — automates camera calibration and skips the manual, error-prone step.

[3] DeepStream 9.1 with JetPack 7.2 brings Orin support back! Grab it from the repo or use the SDK Manager to install it.

Full list of supported platforms: x86 (dGPU), ARM SBSA (DGX Spark, DGX Station), and Jetson via JetPack 7.2 (Thor + Orin).

Want to learn more? Check the FAQ and the DeepStream documentation.

Give the skills a spin and tell us what you build — questions and feedback welcome in the thread.

As always, happy DeepStreaming!

3

I find it hard, that the default Deepstream Api is now deprecated, only supporting py service maker from now on further.

This is especially problematic, when the python servicemaker api cant fully replace the Gst Api. It does not enable access to child elements for example.
How can i work with both Gst Api to build the pipelines and fully access element properties during runtime and python service maker to access NvDsMetaData?
related issue:
servicemaker-python-access-childproxy-element-properties

4

also i noticed the following bug when feeding large numpy arrays to the Buffer() used in the generate function:

import numpy as np, time
from pyservicemaker import Buffer

frame = np.zeros((2062, 2472, 3), dtype=np.uint8)

t0 = time.perf_counter()
Buffer(bytearray(frame))                 # ~2.5 s  (list[int] per-element path)
print("bytearray:", time.perf_counter() - t0)

b = Buffer(frame.tobytes())              # ~4 ms — but silently EMPTY:

feeding this from BufferProvider.generate() terminates the stream as EOS

to work around this i used
Buffer(Gst.Buffer.new_wrapped(...))
instead

5

Finally!! I appreciate it !!! DeepStream C/C++ sources on GitHub — feature request