libtensorflow 2.19.0 — multi-arch GPU (sm_80-sm_120), portable, for PixInsight on Linux

Portable, prebuilt libtensorflow 2.19 C library with native GPU kernels for Ampere through Blackwell (sm_80, 86, 89, 90, 120 + PTX), for GPU-accelerating PixInsight's TensorFlow-based AI tools on Linux — RC Astro's NoiseXTerminator / StarXTerminator / BlurXTerminator and StarNet — including RTX 50-series (no CUDA_ERROR_INVALID_PTX).

GPU if possible, announced CPU fallback otherwise

CUDA is loaded lazily: with an NVIDIA driver + CUDA 12.x + cuDNN 9 present, the tools run on the GPU. Without them they still run on CPU, and the launch log says why (Could not find cuda drivers on your machine, GPU will not be used). An incomplete GPU stack never breaks the tools — it only makes them slower.

Runs across distros

Built in a manylinux_2_28 container → glibc symbol floor ≤ 2.28: x86_64 Debian 10+ (incl. 13 "trixie"), Ubuntu 20.04+, RHEL/AlmaLinux/Rocky 8+, Fedora, Arch, openSUSE. (musl/Alpine not supported — glibc artifact.)

Verification status — please read

• ✅ glibc symbol floor ≤ 2.28 audited; no CUDA in DT_NEEDED (lazy-load contract); native SASS for sm_80/86/89/90/120 + PTX confirmed via cuobjdump.
• ✅ Loads and creates a TensorFlow session in Debian 13 and AlmaLinux 8 containers; the announced CPU fallback verified.
• ⚠️ GPU acceleration in PixInsight was confirmed on the sm_120-only predecessor build (RTX 5080 / Fedora 44). A host GPU run on this exact multi-arch artifact is pending — it's expected to behave identically, but it has not been run yet. Reports welcome.

Requirements / scope

NVIDIA GPU + driver (570+ for Blackwell), CUDA 12.x runtime, cuDNN 9 for CUDA 12, on PixInsight's library path. NVIDIA only — AMD/ROCm is out of scope (this is CUDA machine code).

Install

See the README. Back up /usr/local/libtensorflow, install this build, point PixInsight.sh at your CUDA 12 + cuDNN 9 directories, then run NoiseXTerminator and watch nvidia-smi.

Using DeepSNR? (it's ONNX, not TensorFlow)

DeepSNR moved off TensorFlow to ONNX Runtime, so this libtensorflow won't accelerate it. Use the companion release: deepsnr-gpu-linux. (RC Astro's tools are also moving to ONNX+CUDA on Linux in a future version, per the developer — at which point that release will cover them too.)

Verify the download:

sha256sum -c libtensorflow-2.19.0-gpu-cuda12.8-sm80_120-linux-x86_64.tar.xz.sha256