A list of known issues with the SyQon Starless script, and the status: * Blocks of unprocessed stars to the left hand side of the image. Believed to affect specific aspect ratios. Improved tiling management has been implemented. The issue can also be caused by CUDA 11.8: the script below which changes Torch to one built against an older CUDA version is reported to fix this. **Detection and reversion of too-recent CUDA versions will be built into the next version** * Doesn't work on older GPUs (e.g. GeForce 10xx). The sirilpy TorchHelper class appears to be too aggressive in installing recent versions of Torch: it defaults to the cu128 build which dropped support for Geforce 10xx GPUs. If you run the following script, you should be able to restore support by forcing installation of the cu126 Torch build: ``` import sirilpy as s th = s.TorchHelper() th.install_torch(cuda_version='cu126', force_reinstall=True) th.test_torch() ``` If you have a **very** old GPU such as a Geforce 900 series, replace 'cu126' in the script above with 'cu118' but note you will need CUDA 11 libraries installed on your system. An update to the TorchHelper will be required to provide better autodetection, this will hopefully appear in Siril 1.4.2. **This is now autodetected and the installed version of torch adjusted automatically if necessary to ensure compatibility.** * A "disable GPU" option will be added in an update to force CPU mode. Inference will clearly be much slower. **Fix is done, will be in the next update** * Image dimensions for RGB images are printed incorrectly. **Fix is done, will be in the next update** * numpy issues with Intel Apple Macs. The Torch wheel for Intel Apple Macs is still built against numpy 1.x whereas almost certainly numpy 2.x is installed in the Siril venv. A workaround has been proposed, which is awaiting feedback: the following script will downgrade numpy to 1.x which may allow the script to work. The downside is numpy 1.x has poorer performance than 2.x so this will nerf performance across all other scripts. We encourage submission of bug reports to the Torch project to ask them to provide an updated wheel for Intel Macs built with numpy 2.x. ``` import subprocess import sys subprocess.check_call([sys.executable, "-m", "pip", "install", "numpy<2.0"]) ``` **A fix is complete in the next version that will detect when running on an Intel Apple Mac and automatically downgrade numpy** * On Intel Macs the MPS backend is detected but doesn't work. Sadly there is no accelerated Torch backend that works on these, so we need to fall back to CPU. **Fix is done, will be in the next update** * Issue updating FITS metadata with certain invalid FITS header cards. astropy reports the error here and also suggests the fix, so this will be implemented in the next update. **Fix is done, will be in the next update.** * Issues downloading the model file. This is likely due to network issues, as it downloads reliably for the vast majority of users. I will look at adding SHA256 validation in an update so you can be confident you have a complete, uncorrupted model file. **SHA256 verification is added, this will be in the next update. Partially downloaded files will be retried, but if it fails a second time advice will be printed on how to download the necessary files separately and where to put them** The issues above do not need reporting again: we are aware of them and will aim to fix them. If you come across any other issues, please report them in the comments.