[model] add liger kernel support for Qwen3-Next (#10176)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

requirements/liger-kernel.txt

@@ -1 +1 @@
-liger-kernel>=0.5.5
+liger-kernel>=0.6.3

src/llamafactory/data/collator.py

@@ -213,6 +213,7 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
             and getattr(self.model.config, "model_type", None)
             in [
                 "glm4v",
+                "glm_ocr",
                 "Keye",
                 "qwen2_vl",
                 "qwen2_5_vl",

src/llamafactory/data/template.py

@@ -1061,6 +1061,22 @@ register_template(
 )
 
 
+# copied from glm4 template
+register_template(
+    name="glm_ocr",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>"]),
+    format_assistant=StringFormatter(slots=["\n{{content}}"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}"]),
+    format_function=FunctionFormatter(slots=["{{content}}"], tool_format="glm4"),
+    format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>"]),
+    format_tools=ToolFormatter(tool_format="glm4"),
+    format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
+    stop_words=["<|user|>", "<|observation|>"],
+    efficient_eos=True,
+    mm_plugin=get_mm_plugin(name="glm4v", image_token="<|image|>", video_token="<|video|>"),
+)
+
+
 # copied from glm4_moe template
 register_template(
     name="glm4_7",

src/llamafactory/extras/constants.py

@@ -950,6 +950,18 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "GLM-OCR": {
+            DownloadSource.DEFAULT: "zai-org/GLM-OCR",
+            DownloadSource.MODELSCOPE: "ZhipuAI/GLM-OCR",
+        },
+    },
+    template="glm_ocr",
+    multimodal=True,
+)
+
+
 register_model_group(
     models={
         "GLM-Z1-0414-9B-Chat": {

src/llamafactory/model/model_utils/liger_kernel.py

@@ -77,6 +77,8 @@ def apply_liger_kernel(
         from liger_kernel.transformers import apply_liger_kernel_to_qwen3 as apply_liger_kernel
     elif model_type == "qwen3_moe":
         from liger_kernel.transformers import apply_liger_kernel_to_qwen3_moe as apply_liger_kernel
+    elif model_type == "qwen3_next":
+        from liger_kernel.transformers import apply_liger_kernel_to_qwen3_next as apply_liger_kernel
     elif model_type == "gpt_oss":
         try:
             from liger_kernel.transformers import apply_liger_kernel_to_gpt_oss as apply_liger_kernel

src/llamafactory/model/model_utils/visual.py

@@ -239,6 +239,15 @@ _register_composite_model(
 )
 
 
+_register_composite_model(
+    model_type="glm_ocr",
+    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",
 )