[model] support GLM4.5V (#8876)

src/llamafactory/model/model_utils/moe.py

@@ -62,6 +62,11 @@ def add_z3_leaf_module(model: "PreTrainedModel") -> None:
 
         _set_z3_leaf_modules(model, [Glm4MoeMoE])
 
+    if model_type == "glm4v_moe":
+        from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeTextMoE
+
+        _set_z3_leaf_modules(model, [Glm4vMoeTextMoE])
+
     if model_type == "jamba":
         from transformers.models.jamba.modeling_jamba import JambaSparseMoeBlock
 

src/llamafactory/model/model_utils/visual.py

@@ -221,6 +221,15 @@ _register_composite_model(
 )
 
 
+_register_composite_model(
+    model_type="glm4v_moe",
+    projector_key="visual.merger",
+    vision_model_keys=["visual.patch_embed", "visual.blocks"],
+    language_model_keys=["language_model", "lm_head"],
+    lora_conflict_keys=["patch_embed"],
+)
+
+
 _register_composite_model(
     model_type="internvl",
 )