Merge pull request #16078 from huchenlei/fix_sd2

Fix SD2 loading

modules/sd_hijack_unet.py

@@ -138,6 +138,7 @@ CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.decode_first_stage', first_s
 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)
 
+# Always make sure inputs to unet are in correct dtype
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model)
 CondFunc('sgm.modules.diffusionmodules.wrappers.OpenAIWrapper.forward', apply_model)
 
@@ -150,5 +151,6 @@ def timestep_embedding_cast_result(orig_func, timesteps, *args, **kwargs):
     return orig_func(timesteps, *args, **kwargs).to(dtype=dtype)
 
 
+# Always make sure timestep calculation is in correct dtype
 CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', timestep_embedding_cast_result)
 CondFunc('sgm.modules.diffusionmodules.openaimodel.timestep_embedding', timestep_embedding_cast_result)

modules/sd_models_config.py

@@ -56,14 +56,19 @@ def is_using_v_parameterization_for_sd2(state_dict):
         unet.eval()
 
     with torch.no_grad():
+        unet_dtype = torch.float
+        original_unet_dtype = devices.dtype_unet
+
         unet_sd = {k.replace("model.diffusion_model.", ""): v for k, v in state_dict.items() if "model.diffusion_model." in k}
         unet.load_state_dict(unet_sd, strict=True)
-        unet.to(device=device, dtype=torch.float)
+        unet.to(device=device, dtype=unet_dtype)
+        devices.dtype_unet = unet_dtype
 
         test_cond = torch.ones((1, 2, 1024), device=device) * 0.5
         x_test = torch.ones((1, 4, 8, 8), device=device) * 0.5
 
         out = (unet(x_test, torch.asarray([999], device=device), context=test_cond) - x_test).mean().item()
+        devices.dtype_unet = original_unet_dtype
 
     return out < -1