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

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()