keep textual inversion dataset latents in CPU memory to save a bit of VRAM

modules/textual_inversion/dataset.py

@@ -8,6 +8,7 @@ from torchvision import transforms
 
 import random
 import tqdm
+from modules import devices
 
 
 class PersonalizedBase(Dataset):
@@ -47,6 +48,7 @@ class PersonalizedBase(Dataset):
             torchdata = torch.moveaxis(torchdata, 2, 0)
 
             init_latent = model.get_first_stage_encoding(model.encode_first_stage(torchdata.unsqueeze(dim=0))).squeeze()
+            init_latent = init_latent.to(devices.cpu)
 
             self.dataset.append((init_latent, filename_tokens))
 

modules/textual_inversion/textual_inversion.py

@@ -212,7 +212,10 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps,
 
         with torch.autocast("cuda"):
             c = cond_model([text])
+
+            x = x.to(devices.device)
             loss = shared.sd_model(x.unsqueeze(0), c)[0]
+            del x
 
             losses[embedding.step % losses.shape[0]] = loss.item()