Update app.py

app.py

@@ -4,6 +4,7 @@ import numpy as np
 import spaces
 import torch
 from diffusers import AutoPipelineForText2Image, AutoencoderKL #,EulerDiscreteScheduler
+from transformers import CLIPTokenizer, CLIPTextModel
 
 if not torch.cuda.is_available():
     DESCRIPTION += "\n<p>你现在运行在CPU上 但是只支持GPU.</p>"
@@ -21,19 +22,13 @@ if torch.cuda.is_available():
         add_watermarker=False
     )
     #pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
-    # 设置 tokenizer 的最大长度
-    max_token_length = 512
-    pipe.tokenizer.model_max_length = max_token_length
+    # 加载支持更长序列的 Tokenizer 和文本编码器
+    tokenizer = CLIPTokenizer.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", model_max_length=512)
+    text_encoder = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", trust_remote_code=True)
     
-    # 调整文本编码器的配置
-    pipe.text_encoder.config.max_position_embeddings = max_token_length
-    
-    # 如果需要，重新初始化位置嵌入
-    import torch.nn as nn
-    old_emb = pipe.text_encoder.text_model.embeddings.position_embedding.weight.data
-    new_emb = nn.Parameter(torch.zeros(max_token_length, old_emb.shape[1]))
-    new_emb.data[:old_emb.shape[0], :] = old_emb
-    pipe.text_encoder.text_model.embeddings.position_embedding.weight = new_emb
+    # 更新 pipeline
+    pipe.tokenizer = tokenizer
+    pipe.text_encoder = text_encoder
     pipe.to("cuda")
 
 def randomize_seed_fn(seed: int, randomize_seed: bool) -> int: