update

inference.py

@@ -14,6 +14,7 @@ import sys
 import gc
 from transformers import CLIPTextModel, CLIPTokenizer, BertModel, BertTokenizer
 
+from hf_demo import dtype
 # import train_util
 
 from utils.train_util import get_noisy_image, encode_prompts
@@ -319,8 +320,8 @@ def inference(network: LoRANetwork, tokenizer: CLIPTokenizer, text_encoder: CLIP
             uncond_embeddings = uncond_embed.repeat(bcz, 1, 1)
         else:
             uncond_embeddings = uncond_embed
-        style_text_embeddings = torch.cat([uncond_embeddings, style_embeddings])
-        original_embeddings = torch.cat([uncond_embeddings, original_embeddings])
+        style_text_embeddings = torch.cat([uncond_embeddings, style_embeddings], dtype=weight_dtype)
+        original_embeddings = torch.cat([uncond_embeddings, original_embeddings], dtype=weight_dtype)
 
         generator = torch.manual_seed(single_seed) if single_seed is not None else None
         noise_scheduler.set_timesteps(steps)