update

Author: Jintao-Huang

swift/llm/model/constant.py

@@ -144,6 +144,7 @@ class MLLMModelType:
     ovis1_6 = 'ovis1_6'
     ovis1_6_llama3 = 'ovis1_6_llama3'
     ovis2 = 'ovis2'
+    mimo_vl = 'mimo_vl'
 
     glm4v = 'glm4v'
     glm_edge_v = 'glm_edge_v'

swift/llm/model/model/qwen.py

@@ -677,19 +677,31 @@ def get_model_tokenizer_qwen2_5_vl(*args, **kwargs):
                 Model('Qwen/Qwen2.5-VL-32B-Instruct-AWQ', 'Qwen/Qwen2.5-VL-32B-Instruct-AWQ'),
                 Model('Qwen/Qwen2.5-VL-72B-Instruct-AWQ', 'Qwen/Qwen2.5-VL-72B-Instruct-AWQ'),
             ]),
+        ],
+        TemplateType.qwen2_5_vl,
+        get_model_tokenizer_qwen2_5_vl,
+        model_arch=ModelArch.qwen2_vl,
+        architectures=['Qwen2_5_VLForConditionalGeneration'],
+        requires=['transformers>=4.49', 'qwen_vl_utils>=0.0.6', 'decord'],
+        tags=['vision', 'video']))
+
+register_model(
+    ModelMeta(
+        MLLMModelType.mimo_vl, [
             ModelGroup([
                 Model('XiaomiMiMo/MiMo-VL-7B-SFT', 'XiaomiMiMo/MiMo-VL-7B-SFT'),
                 Model('XiaomiMiMo/MiMo-VL-7B-RL', 'XiaomiMiMo/MiMo-VL-7B-RL'),
             ])
         ],
-        TemplateType.qwen2_5_vl,
+        TemplateType.mimo_vl,
         get_model_tokenizer_qwen2_5_vl,
         model_arch=ModelArch.qwen2_vl,
         architectures=['Qwen2_5_VLForConditionalGeneration'],
         requires=['transformers>=4.49', 'qwen_vl_utils>=0.0.6', 'decord'],
         tags=['vision', 'video']))
 
 
+
 def get_model_tokenizer_qwen2_5_omni(model_dir, *args, **kwargs):
     from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor, Qwen2_5OmniConfig
     from qwen_omni_utils import vision_process

swift/llm/template/constant.py

@@ -103,6 +103,7 @@ class MLLMTemplateType:
     ovis1_6 = 'ovis1_6'
     ovis1_6_llama3 = 'ovis1_6_llama3'
     ovis2 = 'ovis2'
+    mimo_vl = 'mimo_vl'
 
     llama3_1_omni = 'llama3_1_omni'
     llama3_2_vision = 'llama3_2_vision'

swift/llm/template/template/qwen.py

@@ -397,6 +397,8 @@ class Qwen2_5VLTemplate(Qwen2VLTemplate):
 
 register_template(QwenTemplateMeta(MLLMTemplateType.qwen2_5_vl, template_cls=Qwen2_5VLTemplate))
 
+register_template(QwenTemplateMeta(MLLMTemplateType.mimo_vl, template_cls=Qwen2_5VLTemplate,
+                                   default_system='You are MiMo, an AI assistant developed by Xiaomi.'))
 
 class Qwen2_5OmniTemplate(Qwen2_5VLTemplate):
     version = 'omni'