Fix llama4 (#5105)

Author: Jintao-Huang

swift/llm/model/model/llama.py

@@ -232,7 +232,7 @@ def get_model_tokenizer_llama4(*args, **kwargs):
         get_model_tokenizer_llama4,
         requires=['transformers>=4.51'],
         architectures=['Llama4ForConditionalGeneration'],
-        model_arch=ModelArch.llama3_2_vision,
+        model_arch=ModelArch.llama4,
         tags=['vision'],
     ))
 

swift/llm/model/model_arch.py

@@ -39,6 +39,7 @@ class MLLMModelArch:
 
     llama3_1_omni = 'llama3_1_omni'
     llama3_2_vision = 'llama3_2_vision'
+    llama4 = 'llama4'
 
     llava_hf = 'llava_hf'
     llava_hf_legacy = 'llava_hf_legacy'  # transformers<4.52
@@ -573,6 +574,14 @@ def register_model_arch(model_arch: ModelKeys, *, exist_ok: bool = False) -> Non
             vision_tower='vision_model',
         ))
 
+register_model_arch(
+    MultiModelKeys(
+        MLLMModelArch.llama4,
+        language_model='language_model',
+        aligner='multi_modal_projector',
+        vision_tower='vision_model',
+    ))
+
 register_model_arch(MultiModelKeys(
     MLLMModelArch.ovis1_6,
     language_model='llm',