Fix GLM4 agent toolcall (modelscope#1767)

swift/llm/agent/utils.py

@@ -64,6 +64,12 @@
 2.Do not use origin tool names, use only subfunctions' names.
 Specifically, you have access to the following APIs: {tool_list}'''
 
+GLM4_PROMPT = '''你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。
+
+# 可用工具
+
+{tool_list}'''
+
 
 def calculate_loss_scale(query: str,
                          response: str,
@@ -186,10 +192,13 @@ def get_tools_prompt(TOOLS: List[Dict[str, Union[str, dict]]], prompt_format: st
             print('invalid tools format, please check'
                   'https://github.com/modelscope/swift/blob/main/docs/source_en/LLM/Agent-deployment-best-practice.md')
             return None
-    tool_descs = '\n\n'.join(tool_descs)
-    tool_names = ','.join(tool_names)
     if prompt_format == 'react_en':
-        return REACT_PROMPT.format(tool_list=tool_descs, tool_names=tool_names)
+        return REACT_PROMPT.format(tool_list='\n\n'.join(tool_descs), tool_names=','.join(tool_names))
     elif prompt_format == 'react_zh':
-        return REACT_ZH_PROMPT.format(tool_list=tool_descs, tool_names=tool_names)
-    return TOOLBENCH_PROMPT.format(tool_list=tool_descs)
+        return REACT_ZH_PROMPT.format(tool_list='\n\n'.join(tool_descs), tool_names=','.join(tool_names))
+    elif prompt_format == 'glm4':
+        tool_list = ''
+        for name, tool in zip(tool_names, tool_descs):
+            tool_list += f'## {name}\n\n{tool}\n\n'
+        return GLM4_PROMPT.format(tool_list=tool_list)
+    return TOOLBENCH_PROMPT.format(tool_list='\n\n'.join(tool_descs))

swift/llm/utils/dataset.py

@@ -2321,15 +2321,13 @@ def _preprocess_toolbench(dataset: DATASET_TYPE) -> DATASET_TYPE:
 
     def reorganize_row(row):
         convs = row['conversations']
-        sys = convs[0]['value']
         history = []
         history_roles = []
         for idx in range(1, len(convs) - 2, 2):
             history.append((convs[idx]['value'], convs[idx + 1]['value']))
             history_roles.append((convs[idx]['from'], convs[idx + 1]['from']))
 
         return {
-            'system': sys,
             'history': history,
             'history_roles': history_roles,
             'query': convs[-2]['value'],

swift/llm/utils/model.py

@@ -1653,7 +1653,7 @@ def cross_entropy_forward(self, inputs: Tensor, target: Tensor) -> Tensor:
     ModelType.glm4_9b_chat,
     'ZhipuAI/glm-4-9b-chat',
     LoRATM.chatglm,
-    TemplateType.chatglm3,
+    TemplateType.chatglm4,
     support_flash_attn=True,
     support_vllm=True,
     support_lmdeploy=True,
@@ -1663,7 +1663,7 @@ def cross_entropy_forward(self, inputs: Tensor, target: Tensor) -> Tensor:
     ModelType.glm4_9b_chat_1m,
     'ZhipuAI/glm-4-9b-chat-1m',
     LoRATM.chatglm,
-    TemplateType.chatglm3,
+    TemplateType.chatglm4,
     support_flash_attn=True,
     support_vllm=True,
     support_lmdeploy=True,

swift/llm/utils/template.py

@@ -50,6 +50,7 @@ class TemplateType:
     baichuan = 'baichuan'
     chatglm2 = 'chatglm2'
     chatglm3 = 'chatglm3'
+    chatglm4 = 'chatglm4'
     codegeex4 = 'codegeex4'
     llama = 'llama'  # llama2
     llama3 = 'llama3'
@@ -1397,6 +1398,13 @@ def data_collator(self, batch: List[Dict[str, Any]], padding_to: Optional[int] =
     TemplateType.chatglm3,
     GLMTemplate([], ['<|user|>\n{{QUERY}}<|assistant|>\n'], [], ['<|user|>'], None, ['<|system|>\n{{SYSTEM}}']))
 
+register_template(
+    TemplateType.chatglm4,
+    GLMTemplate([], ['<|user|>\n{{QUERY}}<|assistant|>\n'], [], ['<|user|>'],
+                None, ['<|system|>\n{{SYSTEM}}'],
+                tools_prompt='glm4',
+                tool_prompt=['<|observation|>\n{{QUERY}}<|assistant|>\n']))
+
 codegeex4_system = '你是一位智能编程助手，你叫CodeGeeX。你会为用户回答关于编程、代码、计算机方面的任何问题，并提供格式规范、可以执行、准确安全的代码，并在必要时提供详细的解释。'
 
 register_template(