Ultralytics v8.3.253 is out! 
TL;DR: v8.3.253 adds explicit Vulkan GPU device selection for NCNN inference (plus safer PaddlePaddle/NCNN dependency handling) to improve cross-vendor GPU acceleration and reliability.
If you’re running Ultralytics YOLO on AMD/Intel GPUs, edge devices, or multi-GPU Vulkan setups, this release is for you.
New Features
NCNN + Vulkan device targeting (main feature)
You can now choose exactly which Vulkan-capable GPU NCNN uses by passing:
device="vulkan:0"device="vulkan:1"- etc.
This is especially useful for:
- Non-NVIDIA GPUs (AMD/Intel where CUDA isn’t available)
- Multi-GPU machines where you want predictable device selection
NCNN now enables Vulkan compute when the device starts with "vulkan", calls net.set_vulkan_device(<id>), and keeps the rest of the pipeline consistent by continuing on CPU from PyTorch’s perspective (NCNN handles the GPU work internally).
Minimal example:
from ultralytics import YOLO
model = YOLO("yolo11n_ncnn_model")
results = model("image.jpg", device="vulkan:0") # pick Vulkan GPU 0
Docs
Updated Vulkan acceleration documentation
The NCNN integration docs now clearly document Vulkan GPU acceleration and include CLI/Python examples for device=vulkan:<id>.
Improvements (reliability & installs)
PaddlePaddle guardrail
We now exclude PaddlePaddle 3.3.0 (!=3.3.0) in both export and inference paths due to an upstream breakage—helping avoid confusing runtime failures 
ARM64 NCNN installation fixes
On ARM64, NCNN now installs from Git source instead of PyPI (current PyPI packages are broken), improving install success on devices like Raspberry Pi.
Bug Fixes / Maintenance
This release is also about reducing “mysterious failures” by hardening dependency behavior across export + inference paths.
What’s Changed (PRs)
- NCNN source build and Paddle 3.3.0 exclusions by @glenn-jocher in PR #23173
- Add Vulkan device selection support for NCNN by @Faerbit in PR #23164
New Contributors
- Huge thanks to @Faerbit for their first contribution in PR #23164
(and thanks to the whole YOLO community for helping push things forward!)
Try it now & share feedback
Update with pip install -U ultralytics, then test Vulkan targeting with your NCNN export (or existing NCNN model) and let us know how it performs on your hardware.
You can review the full release details in the official v8.3.253 release page and browse the complete changes in the v8.3.252…v8.3.253 changelog.