Skip to content

ONNX Runtime GPU 1.26.0 for DeepSNR (Linux, NVIDIA)

Latest
Latest

Choose a tag to compare

View all tags
@dustinspace217 dustinspace217 released this 20 Jun 06:36
ort1.26.0-r1
f833ecc

Prebuilt official ONNX Runtime GPU 1.26.0 libraries that re-enable CUDA acceleration for the current ONNX-backed DeepSNR PixInsight module on Linux — including Blackwell (RTX 50-series). No fork, no source build.

What's in the bundle

ort-gpu-1.26.0-linux-x86_64.tar.gz contains the three libraries StarNet's Linux package omits, extracted unmodified from the official onnxruntime-gpu 1.26.0 PyPI wheel:

  • libonnxruntime.so.1.26.0
  • libonnxruntime_providers_shared.so
  • libonnxruntime_providers_cuda.so

plus ONNX Runtime's LICENSE and a PROVENANCE.txt recording the source wheel's sha256 so anyone can reproduce and diff it.

Do you need to download this?

Most people don't — install-deepsnr-gpu.sh fetches the same libraries from PyPI for you (integrity-checked, no pip required). This bundle is for systems without python3, or anyone who prefers grabbing a checked artifact from here.

Verified

  • glibc symbol floor ≤ 2.28 → x86_64 Debian 10+ (incl. 13 "trixie"), Ubuntu 20.04+, RHEL/AlmaLinux/Rocky 8+, Fedora, Arch, openSUSE.
  • Loads on Debian 13 and AlmaLinux 8 containers; CPU inference of the DeepSNR model verified inside Debian 13.
  • GPU end-to-end confirmed in PixInsight on an RTX 5080 / Fedora 44 (~136× per tile vs CPU).

Requirements / scope

NVIDIA GPU + driver, CUDA 12.x runtime, cuDNN 9 for CUDA 12. Without them DeepSNR still runs on CPU — the module announces the fallback itself in the process console. NVIDIA only: AMD is blocked at the vendor level, since the signed DeepSNR module registers ONNX Runtime's CUDA execution provider specifically.

Other PixInsight AI tools?

RC Astro's NoiseXTerminator / StarXTerminator / BlurXTerminator use TensorFlow today, not ONNX — so these libraries don't accelerate them yet. The developer has confirmed a move to ONNX+CUDA on Linux in a future version, at which point the same drop-in will cover them too. For GPU acceleration of the TensorFlow-based RC Astro tools right now, see the companion project: pixinsight-blackwell-tensorflow.

Verify the download:

sha256sum -c ort-gpu-1.26.0-linux-x86_64.tar.gz.sha256

See the README for full install steps.

Assets 4
Loading