fix

src/llamafactory/data/collator.py

@@ -194,7 +194,7 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
             elif "video_second_per_grid" in mm_inputs:  # for qwen2.5 omni
                 rope_index_kwargs["second_per_grids"] = mm_inputs.get("video_second_per_grid")
 
-            if getattr(self.model.config, "model_type", None) == "qwen2_5_omni_thinker":  # for qwen2.5 omni
+            if getattr(self.model.config, "model_type", None) in ["qwen2_5_omni_thinker", "qwen3_omni_moe_thinker"]:  # for qwen2.5 omni
                 rope_index_kwargs["use_audio_in_video"] = getattr(self.processor, "use_audio_in_video", False)
                 feature_attention_mask = mm_inputs.get("feature_attention_mask", None)
                 if feature_attention_mask is not None:  # FIXME: need to get video image lengths
@@ -211,7 +211,7 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
         if (
             self.model is not None
             and getattr(self.model.config, "model_type", None)
-            in ["glm4v", "Keye", "qwen2_vl", "qwen2_5_vl", "qwen2_5_omni_thinker"]
+            in ["glm4v", "Keye", "qwen2_vl", "qwen2_5_vl", "qwen2_5_omni_thinker", "qwen3_omni_moe_thinker"]
             and ("position_ids" not in features or features["position_ids"].dim() != 3)
         ):
             raise ValueError(f"{self.model.config.model_type} requires 3D position ids for mrope.")

src/llamafactory/model/model_utils/visual.py

@@ -350,6 +350,32 @@ _register_composite_model(
     lora_conflict_keys=["patch_embed"],
 )
 
+_register_composite_model(
+    model_type="qwen3_vl",
+    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="qwen3_vl_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="qwen3_omni_moe_thinker",
+    projector_key="visual.merger",
+    vision_model_keys=["visual.patch_embed", "visual.blocks", "audio_tower"],
+    language_model_keys=["model", "lm_head"],
+    lora_conflict_keys=["patch_embed"],
+)
+
 
 _register_composite_model(
     model_type="video_llava",

tests/data/test_mm_plugin.py

@@ -332,7 +332,14 @@ def test_qwen2_omni_plugin():
     image_seqlen, audio_seqlen = 4, 2
     tokenizer_module = _load_tokenizer_module(model_name_or_path="Qwen/Qwen2.5-Omni-7B")
     qwen2_omni_plugin = get_mm_plugin(
-        name="qwen2_omni", audio_token="<|AUDIO|>", image_token="<|IMAGE|>", video_token="<|VIDEO|>"
+        name="qwen2_omni",
+        audio_token="<|AUDIO|>",
+        image_token="<|IMAGE|>",
+        video_token="<|VIDEO|>",
+        vision_bos_token="<|vision_bos|>",
+        vision_eos_token="<|vision_eos|>",
+        audio_bos_token="<|audio_bos|>",
+        audio_eos_token="<|audio_eos|>",
     )
     check_inputs = {"plugin": qwen2_omni_plugin, **tokenizer_module}
     check_inputs["expected_mm_messages"] = [