[template] add GLM4_5Template

Author: laocheyujie

swift/llm/template/base.py

@@ -1135,8 +1135,7 @@ def _swift_encode(self, inputs: StdTemplateInputs):
                 context_list.append('{{RESPONSE}}')
                 # self.is_training needed because we may want to continue generation from
                 # the current response
-                string_stop_words = tuple(s for s in template_meta.stop_words if isinstance(s, str))
-                if self.is_training and not sep_token and not response.endswith(string_stop_words) or self.task_type == 'embedding':
+                if self.is_training and not sep_token or self.task_type == 'embedding':
                     extra_context_list = template_meta.suffix
                     extra_context_type = ContextType.SUFFIX
             elif template_meta.response_prefix:

swift/llm/template/template/glm.py

@@ -38,6 +38,17 @@ class GLM4_0414Template(ThinkingTemplate, GLM4Template):
     pass
 
 
+class GLM4_5Template(ThinkingTemplate):
+    def _swift_encode(self, inputs: StdTemplateInputs):
+        res_context_list, loss_scale_list, answer_len = super()._swift_encode(inputs)
+        # When it's a tool_call, avoid generating <|observation|><|user|>
+        penultimate_content = res_context_list[-2] if len(res_context_list) >= 2 else None
+        if isinstance(penultimate_content, str) and penultimate_content.endswith('<|observation|>') and res_context_list[-1] == '<|user|>':
+            res_context_list = res_context_list[:-1]
+        answer_len -= 1
+        return res_context_list, loss_scale_list, answer_len
+    
+
 register_template(
     GLMTemplateMeta(
         LLMTemplateType.chatglm2,
@@ -241,7 +252,7 @@ def _data_collator_mm_data(self, batch: List[Dict[str, Any]]) -> Dict[str, Any]:
 
 register_template(GLM4_0414TemplateMeta(LLMTemplateType.glm4_0414, template_cls=GLM4_0414Template))
 
-register_template(GLM4_5TemplateMeta(LLMTemplateType.glm4_5, template_cls=ThinkingTemplate))
+register_template(GLM4_5TemplateMeta(LLMTemplateType.glm4_5, template_cls=GLM4_5Template))
 
 register_template(GLM4_1VTemplateMeta(MLLMTemplateType.glm4_1v, template_cls=GLM4_1VTemplate))