| Age | Commit message (Collapse) | Author |
|---|
|
|
|
copypaste support to RNG source
|
|
Makes a given manual seed generate the same images across different
platforms, independently of the GPU architecture in use.
Fixes #9613.
|
|
Move most Mac related code to a separate file, don't even load it unless web UI is run under macOS.
|
|
|
|
Apparently I did not test with large enough images to trigger the bug with torch.narrow on MPS
|
|
Fix embeddings, upscalers, and refactor `--upcast-sampling`
|
|
The torch.narrow fix was required for nightly PyTorch builds for a while to prevent a hard crash, but newer nightly builds don't have this issue.
|
|
|
|
|
|
|
|
Adds "Upcast cross attention layer to float32" option in Stable Diffusion settings. This allows for generating images using SD 2.1 models without --no-half or xFormers.
In order to make upcasting cross attention layer optimizations possible it is necessary to indent several sections of code in sd_hijack_optimizations.py so that a context manager can be used to disable autocast. Also, even though Stable Diffusion (and Diffusers) only upcast q and k, unfortunately my findings were that most of the cross attention layer optimizations could not function unless v is upcast also.
|
|
This also handles type casting so that ROCm and MPS torch devices work correctly without --no-half. One cast is required for deepbooru in deepbooru_model.py, some explicit casting is required for img2img and inpainting. depth_model can't be converted to float16 or it won't work correctly on some systems (it's known to have issues on MPS) so in sd_models.py model.depth_model is removed for model.half().
|
|
Improve cumsum fix for MPS
|
|
The prior fix assumed that testing int16 was enough to determine if a fix is needed, but a recent fix for cumsum has int16 working but not bool.
|
|
|
|
|
|
|
|
When saving training results with torch.save(), an exception is thrown:
"RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead."
So for MPS, check if Tensor.requires_grad and detach() if necessary.
|
|
add support for adding upscalers in extensions
move LDSR, ScuNET and SwinIR to built-in extensions
|
|
|
|
|
|
Fix issue with 16xx cards
|
|
torch.manual_seed() already sets a CPU generator, so there is no reason to create a CPU generator manually. torch.randn_like also needs a MPS fix for k-diffusion, but a torch hijack with randn_like already exists so it can also be used for that.
|
|
Fixes for PyTorch 1.12.1 when using MPS
|
|
|
|
otherwise first. Updating torch_gc() to use the device set by --device-id if specified to avoid OOM edge cases on multi-GPU systems.
|
|
Fix typo "MasOS" -> "macOS"
If MPS is available and PyTorch is an earlier version than 1.13:
* Monkey patch torch.Tensor.to to ensure all tensors sent to MPS are contiguous
* Monkey patch torch.nn.functional.layer_norm to ensure input tensor is contiguous (required for this program to work with MPS on unmodified PyTorch 1.12.1)
|
|
This reverts commit 768b95394a8500da639b947508f78296524f1836.
|
|
|
|
|
|
|
|
thanks C43H66N12O12S2
|
|
|
|
cudnn
|
|
Get ESRGAN, SCUNet, and SwinIR working correctly on MPS by ensuring memory is contiguous for tensor views before sending to MPS device.
|
|
|
|
|
|
|
|
* Add 'interrogate' and 'all' choices to --use-cpu
* Change type for --use-cpu argument to str.lower, so that choices are case insensitive
|
|
|
|
|
|
just one for #1283
|
|
|
|
Remove MPS detection to use CPU for GFPGAN / CodeFormer and add a --use-cpu command line option.
|
|
|
|
|
|
GFPGAN will not work if the device is MPS, so default to CPU instead.
|
|
|
|
|
