fix vae upscale

CHANGELOG.md

@@ -3,7 +3,7 @@
 ## Update for 2023-09-13
 
 Started as a mostly a service release with quite a few fixes, but then...  
-Major changes how **hires** works as well as support for a very interesting new model [wuerstchen](https://huggingface.co/blog/wuertschen)  
+Major changes how **hires** works as well as support for a very interesting new model [Wuerstchen](https://huggingface.co/blog/wuertschen)  
 
 - tons of fixes  
 - changes to **hires**  
@@ -46,8 +46,6 @@ Major changes how **hires** works as well as support for a very interesting new
   - cleaner server restart  
   - custom exception handling
 
-Note: **wuerstchen** is pending critical bugfix in `diffusers=0.21.1`
-
 ## Update for 2023-09-06
 
 One week later, another large update!

modules/processing_diffusers.py

@@ -77,6 +77,12 @@ def process_diffusers(p: StableDiffusionProcessing, seeds, prompts, negative_pro
         if not shared.cmd_opts.lowvram and not shared.opts.diffusers_seq_cpu_offload:
             model.vae.to(devices.device)
         latents.to(model.vae.device)
+
+        needs_upcasting = model.vae.dtype == torch.float16 and model.vae.config.force_upcast
+        if needs_upcasting: # this is done by diffusers automatically if output_type != 'latent'
+            model.upcast_vae()
+            latents = latents.to(next(iter(model.vae.post_quant_conv.parameters())).dtype)
+
         decoded = model.vae.decode(latents / model.vae.config.scaling_factor, return_dict=False)[0]
         if shared.opts.diffusers_move_unet and not model.has_accelerate:
             model.unet.to(unet_device)

requirements.txt

@@ -47,7 +47,7 @@ requests==2.31.0
 tqdm==4.66.1
 accelerate==0.20.3
 opencv-python-headless==4.7.0.72
-diffusers==0.20.2
+diffusers==0.21.1
 einops==0.4.1
 gradio==3.43.2
 huggingface_hub==0.17.1