Fixes#882
If the load of the model fails during the initialization, an attempt to render an
image using the same model fails because ED doesn't notice that the model has to
be loaded. This PR ensures that the model is being reloaded if the initial load
fails. If the second load attempt fails as well, the user will get a more helpful
error message than 'model not loaded yet'.
The old code had a 1.6% chance of name collisions (AAb/345 and AAB3/45 would generate the same filename) on Linux and 39% on Windows (base64 is case sensitive, Windows isn't).
This code uses base36 (0-9, A-Z) to avoid case issues. To avoid collisions on fast computers or multi GPU computers, the time resolution of the timestamp is changed from seconds to 0.1ms, and the image number isn't added to the timestamp but appended as an extra character. Due to the limitation of the timestamps to 7 characters, the timecode will repeat every 90 days. This shouldn't be an issue since most sessions will not last this long.
By default, uvicorn uses the 'Content Type' value from 'Computer\HKEY_CLASSES_ROOT\.css' for the Content-Type header in its responses. Some systems have this set to 'application/x-css', an outdated content type used in the early days of CSS. Modern browsers ignore stylesheets that don't have the content-type 'text/css'. This change hardcodes the Content-Type to 'text/css', ignoring the registry.
* Reverts the recent 'torch.mps' changes since .mps is only available in torch v2.0, which isn't yet released.
* Includes Hotfitx 984
* Enables 'cpu-only' option when running on Apple silicon.
Changes:
* autodetect if MPS is available and the pytorch version has MPS support.
* change logic from "is the device CPU?" to "is the device not CUDA?".
* set PYTORCH_ENABLE_MPS_FALLBACK=1
Known issues:
* Some samplers (eg DDIM) will fail on MPS unless forced to CPU-only mode
* Make stream_image_progress accept an integer
for the rate the progress frames should be generated.
* Use a different field for the progress interval.
In newer versions of the picklescanner, scanning of .safetensors files creates an error:
21:28:01.067 ERROR MainThread ERROR: parsing pickle in D:\2.35\dev\models\stable-diffusion\dantionrealmix_10.safetensors: at position 1, opcode b'\xce' unknown
To avoid these entries in the logs, skip scanning of safetensors files.
\r\n creates CR CR LF in python, which confuses the Windows batch processor.
With only \n, adding the config line for FP32 works as expected:
10:50:43.659 WARNING cuda:0 forcing full precision on this GPU, to avoid green images. GPU detected: NVIDIA GeForce GTX 1060 6GB
