Add support for `--upcast-sampling` with SD XL

modules/sd_hijack_unet.py

@@ -39,7 +39,10 @@ def apply_model(orig_func, self, x_noisy, t, cond, **kwargs):
 
     if isinstance(cond, dict):
         for y in cond.keys():
-            cond[y] = [x.to(devices.dtype_unet) if isinstance(x, torch.Tensor) else x for x in cond[y]]
+            if isinstance(cond[y], list):
+                cond[y] = [x.to(devices.dtype_unet) if isinstance(x, torch.Tensor) else x for x in cond[y]]
+            else:
+                cond[y] = cond[y].to(devices.dtype_unet) if isinstance(cond[y], torch.Tensor) else cond[y]
 
     with devices.autocast():
         return orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs).float()
@@ -77,3 +80,6 @@ first_stage_sub = lambda orig_func, self, x, **kwargs: orig_func(self, x.to(devi
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding', lambda orig_func, *args, **kwargs: orig_func(*args, **kwargs).float(), first_stage_cond)
+
+CondFunc('sgm.modules.diffusionmodules.wrappers.OpenAIWrapper.forward', apply_model, unet_needs_upcast)
+CondFunc('sgm.modules.diffusionmodules.openaimodel.timestep_embedding', lambda orig_func, timesteps, *args, **kwargs: orig_func(timesteps, *args, **kwargs).to(torch.float32 if timesteps.dtype == torch.int64 else devices.dtype_unet), unet_needs_upcast)

modules/sd_models.py

@@ -326,7 +326,7 @@ def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer
 
         timer.record("apply half()")
 
-    devices.dtype_unet = model.model.diffusion_model.dtype
+    devices.dtype_unet = torch.float16 if model.is_sdxl and not shared.cmd_opts.no_half else model.model.diffusion_model.dtype
     devices.unet_needs_upcast = shared.cmd_opts.upcast_sampling and devices.dtype == torch.float16 and devices.dtype_unet == torch.float16
 
     model.first_stage_model.to(devices.dtype_vae)