bugfix

app.py

@@ -365,31 +365,40 @@ def run_lora(prompt_bg, character_prompts_json, character_positions_json, lora_s
         with calculateDuration("Set random seed"):
             seed = random.randint(0, MAX_SEED)
     
-    # 编码提示词
     with calculateDuration("Encoding prompts"):
         # 编码背景提示词
+        # 使用 tokenizer_2 和 text_encoder_2
+        bg_text_input_2 = pipe.tokenizer_2(prompt_bg, return_tensors="pt").to(device)
+        bg_prompt_embeds = pipe.text_encoder_2(bg_text_input_2.input_ids.to(device))[0]
+
+        # 使用 tokenizer 和 text_encoder
         bg_text_input = pipe.tokenizer(prompt_bg, return_tensors="pt").to(device)
-        bg_prompt_embeds = pipe.text_encoder_2(bg_text_input.input_ids.to(device))[0]
         bg_pooled_embeds = pipe.text_encoder(bg_text_input.input_ids.to(device)).pooler_output
 
         # 编码角色提示词
         character_prompt_embeds = []
         character_pooled_embeds = []
         for prompt in character_prompts:
+            # 使用 tokenizer_2 和 text_encoder_2
+            char_text_input_2 = pipe.tokenizer_2(prompt, return_tensors="pt").to(device)
+            char_prompt_embeds = pipe.text_encoder_2(char_text_input_2.input_ids.to(device))[0]
+            # 使用 tokenizer 和 text_encoder
             char_text_input = pipe.tokenizer(prompt, return_tensors="pt").to(device)
-            char_prompt_embeds = pipe.text_encoder_2(char_text_input.input_ids.to(device))[0]
             char_pooled_embeds = pipe.text_encoder(char_text_input.input_ids.to(device)).pooler_output
+
             character_prompt_embeds.append(char_prompt_embeds)
             character_pooled_embeds.append(char_pooled_embeds)
 
         # 编码互动细节提示词
+        details_text_input_2 = pipe.tokenizer_2(prompt_details, return_tensors="pt").to(device)
+        details_prompt_embeds = pipe.text_encoder_2(details_text_input_2.input_ids.to(device))[0]
+
         details_text_input = pipe.tokenizer(prompt_details, return_tensors="pt").to(device)
-        details_prompt_embeds = pipe.text_encoder_2(details_text_input.input_ids.to(device))[0]
         details_pooled_embeds = pipe.text_encoder(details_text_input.input_ids.to(device)).pooler_output
 
         # 合并背景和互动细节的嵌入
         prompt_embeds = torch.cat([bg_prompt_embeds, details_prompt_embeds], dim=1)
-        pooled_prompt_embeds = torch.cat([bg_pooled_embeds, details_pooled_embeds], dim=1)
+        pooled_prompt_embeds = torch.cat([bg_pooled_embeds, details_pooled_embeds], dim=-1)
     
     # 解析角色位置
     character_infos = []