[model] support Aeva (#10214)

Co-authored-by: frozenleaves <frozen@Mac.local>
Co-authored-by: Yaowei Zheng <hiyouga@buaa.edu.cn>

.github/workflows/docs.yml

@@ -25,16 +25,16 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
-      
+
       - name: Set up Python
         uses: actions/setup-python@v5
         with:
           python-version: '3.10'
-          
+
       - name: Install dependencies
         run: |
           pip install -r docs/requirements.txt
-          
+
       - name: Build Sphinx
         run: |
           sphinx-build -b html docs/zh docs/_build/html/zh
@@ -56,10 +56,10 @@ jobs:
             > docs/_build/html/index.html
 
           touch docs/_build/html/.nojekyll
-          
+
       - name: Setup Pages
         uses: actions/configure-pages@v5
-        
+
       - name: Upload artifact
         uses: actions/upload-pages-artifact@v3
         with:

README.md

@@ -291,7 +291,7 @@ Read technical notes:
 | [GPT-2](https://huggingface.co/openai-community)                  | 0.1B/0.4B/0.8B/1.5B              | -                    |
 | [GPT-OSS](https://huggingface.co/openai)                          | 20B/120B                         | gpt_oss              |
 | [Granite 3-4](https://huggingface.co/ibm-granite)                 | 1B/2B/3B/7B/8B                   | granite3/granite4    |
-| [Hunyuan/Hunyuan1.5 (MT)](https://huggingface.co/tencent/)        | 0.5B/1.8B/4B/7B/13B              | hunyuan/hunyuan_small |
+| [Hunyuan/Hunyuan1.5 (MT)](https://huggingface.co/tencent/)        | 0.5B/1.8B/4B/7B/13B              | hunyuan/hunyuan_small|
 | [InternLM 2-3](https://huggingface.co/internlm)                   | 7B/8B/20B                        | intern2              |
 | [InternVL 2.5-3.5](https://huggingface.co/OpenGVLab)              | 1B/2B/4B/8B/14B/30B/38B/78B/241B | intern_vl            |
 | [Intern-S1-mini](https://huggingface.co/internlm/)                | 8B                               | intern_s1            |
@@ -319,6 +319,7 @@ Read technical notes:
 | [Pixtral](https://huggingface.co/mistralai)                       | 12B                              | pixtral              |
 | [Qwen2 (Code/Math/MoE/QwQ)](https://huggingface.co/Qwen)          | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen                 |
 | [Qwen3 (MoE/Instruct/Thinking/Next)](https://huggingface.co/Qwen) | 0.6B/1.7B/4B/8B/14B/32B/80B/235B | qwen3/qwen3_nothink  |
+| [Qwen3.5](https://huggingface.co/Qwen)                            | 27B/35B/122B/397B                | qwen3_5              |
 | [Qwen2-Audio](https://huggingface.co/Qwen)                        | 7B                               | qwen2_audio          |
 | [Qwen2.5-Omni](https://huggingface.co/Qwen)                       | 3B/7B                            | qwen2_omni           |
 | [Qwen3-Omni](https://huggingface.co/Qwen)                         | 30B                              | qwen3_omni           |

README_zh.md

@@ -293,7 +293,7 @@ https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc
 | [GPT-2](https://huggingface.co/openai-community)                  | 0.1B/0.4B/0.8B/1.5B              | -                    |
 | [GPT-OSS](https://huggingface.co/openai)                          | 20B/120B                         | gpt_oss              |
 | [Granite 3-4](https://huggingface.co/ibm-granite)                 | 1B/2B/3B/7B/8B                   | granite3/granite4    |
-| [Hunyuan/Hunyuan1.5 (MT)](https://huggingface.co/tencent/)        | 0.5B/1.8B/4B/7B/13B              | hunyuan/hunyuan_small |
+| [Hunyuan/Hunyuan1.5 (MT)](https://huggingface.co/tencent/)        | 0.5B/1.8B/4B/7B/13B              | hunyuan/hunyuan_small|
 | [InternLM 2-3](https://huggingface.co/internlm)                   | 7B/8B/20B                        | intern2              |
 | [InternVL 2.5-3.5](https://huggingface.co/OpenGVLab)              | 1B/2B/4B/8B/14B/30B/38B/78B/241B | intern_vl            |
 | [Intern-S1-mini](https://huggingface.co/internlm/)                | 8B                               | intern_s1            |
@@ -321,6 +321,7 @@ https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc
 | [Pixtral](https://huggingface.co/mistralai)                       | 12B                              | pixtral              |
 | [Qwen2 (Code/Math/MoE/QwQ)](https://huggingface.co/Qwen)          | 0.5B/1.5B/3B/7B/14B/32B/72B/110B | qwen                 |
 | [Qwen3 (MoE/Instruct/Thinking/Next)](https://huggingface.co/Qwen) | 0.6B/1.7B/4B/8B/14B/32B/80B/235B | qwen3/qwen3_nothink  |
+| [Qwen3.5](https://huggingface.co/Qwen)                            | 27B/35B/122B/397B                | qwen3_5              |
 | [Qwen2-Audio](https://huggingface.co/Qwen)                        | 7B                               | qwen2_audio          |
 | [Qwen2.5-Omni](https://huggingface.co/Qwen)                       | 3B/7B                            | qwen2_omni           |
 | [Qwen3-Omni](https://huggingface.co/Qwen)                         | 30B                              | qwen3_omni           |

docs/_static/css/lang-switcher.css

@@ -47,4 +47,3 @@
   border-color: rgba(255, 255, 255, 0.45);
   box-shadow: 0 0 0 3px rgba(255, 255, 255, 0.12);
 }
-

docs/conf.py

@@ -1,33 +1,31 @@
 # Configuration file for the Sphinx documentation builder.
 
-import os
-import sys
 
 # Define common settings here
-project = 'LlamaFactory'
-copyright = '2024, LlamaFactory Team'
-author = 'LlamaFactory Team'
+project = "LlamaFactory"
+copyright = "2024, LlamaFactory Team"
+author = "LlamaFactory Team"
 
 extensions = [
-    'sphinx.ext.autodoc',
-    'sphinx.ext.viewcode',
-    'sphinx.ext.napoleon',
-    'myst_parser',
+    "sphinx.ext.autodoc",
+    "sphinx.ext.viewcode",
+    "sphinx.ext.napoleon",
+    "myst_parser",
 ]
 
-templates_path = ['_templates']
-exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store']
+templates_path = ["_templates"]
+exclude_patterns = ["_build", "Thumbs.db", ".DS_Store"]
 
-html_theme = 'sphinx_rtd_theme'
+html_theme = "sphinx_rtd_theme"
 
-html_static_path = ['_static']
+html_static_path = ["_static"]
 
 html_js_files = [
-    'js/switcher.js',
+    "js/switcher.js",
 ]
 
 html_css_files = [
-    'css/lang-switcher.css',
+    "css/lang-switcher.css",
 ]
 
 myst_enable_extensions = [

docs/en/conf.py

@@ -1,20 +1,22 @@
 import os
 import sys
 
-# Add parent dir to path to allow importing conf.py
-sys.path.insert(0, os.path.abspath('..'))
 
-from conf import *
+# Add parent dir to path to allow importing conf.py
+sys.path.insert(0, os.path.abspath(".."))
+
+from conf import *  # noqa: F403
+
 
 # Language settings
-language = 'en'
-html_search_language = 'en'
+language = "en"
+html_search_language = "en"
 
 # Static files
 # Point to the root _static directory
-html_static_path = ['../_static']
+html_static_path = ["../_static"]
 
 # Add custom JS for language switcher
 html_js_files = [
-    'js/switcher.js',
+    "js/switcher.js",
 ]

docs/zh/conf.py

@@ -1,20 +1,22 @@
 import os
 import sys
 
-# Add parent dir to path to allow importing conf.py
-sys.path.insert(0, os.path.abspath('..'))
 
-from conf import *
+# Add parent dir to path to allow importing conf.py
+sys.path.insert(0, os.path.abspath(".."))
+
+from conf import *  # noqa: F403
+
 
 # Language settings
-language = 'zh_CN'
-html_search_language = 'zh'
+language = "zh_CN"
+html_search_language = "zh"
 
 # Static files
 # Point to the root _static directory
-html_static_path = ['../_static']
+html_static_path = ["../_static"]
 
 # Add custom JS for language switcher
 html_js_files = [
-    'js/switcher.js',
+    "js/switcher.js",
 ]

examples/v1/train_full/train_full_deepspeed.yaml

@@ -6,14 +6,14 @@ template: qwen3_nothink
 
 kernel_config:
     name: auto
-    include_kernels: auto 
+    include_kernels: auto
 
 dist_config:
     name: deepspeed
     config_file: examples/deepspeed/ds_z3_config.json
 
 ### data
-train_dataset: data/v1_sft_demo.yaml 
+train_dataset: data/v1_sft_demo.yaml
 
 ### training
 output_dir: outputs/Qwen3-0.6B-deepspeed
@@ -22,4 +22,3 @@ cutoff_len: 2048
 learning_rate: 1.0e-4
 bf16: true
 max_steps: 10
-

pyproject.toml

@@ -40,7 +40,7 @@ dependencies = [
     "torch>=2.4.0",
     "torchvision>=0.19.0",
     "torchaudio>=2.4.0",
-    "transformers>=4.51.0,<=5.0.0,!=4.52.0,!=4.57.0",
+    "transformers>=4.51.0,<=5.2.0,!=4.52.0,!=4.57.0",
     "datasets>=2.16.0,<=4.0.0",
     "accelerate>=1.3.0,<=1.11.0",
     "peft>=0.18.0,<=0.18.1",

src/llamafactory/data/collator.py

@@ -15,6 +15,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import inspect
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any, Literal, Optional
 
@@ -189,6 +190,16 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
                 "video_grid_thw": mm_inputs.get("video_grid_thw"),
                 "attention_mask": (features["attention_mask"] >= 1).float(),
             }
+            if "mm_token_type_ids" in inspect.signature(self.get_rope_func).parameters:
+                image_token_id = getattr(self.model.config, "image_token_id", None)
+                video_token_id = getattr(self.model.config, "video_token_id", None)
+                if image_token_id is not None or video_token_id is not None:
+                    mm_token_type_ids = torch.zeros_like(features["input_ids"])
+                    if image_token_id is not None:
+                        mm_token_type_ids[features["input_ids"] == image_token_id] = 1
+                    if video_token_id is not None:
+                        mm_token_type_ids[features["input_ids"] == video_token_id] = 2
+                    rope_index_kwargs["mm_token_type_ids"] = mm_token_type_ids
             if "second_per_grid_ts" in mm_inputs:  # for qwen2vl
                 rope_index_kwargs["second_per_grid_ts"] = mm_inputs.get("second_per_grid_ts")
             elif "video_second_per_grid" in mm_inputs:  # for qwen2.5 omni
@@ -219,6 +230,7 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
                 "qwen2_5_vl",
                 "qwen2_5_omni_thinker",
                 "qwen3_omni_moe_thinker",
+                "qwen3_5",
                 "qwen3_vl",
                 "qwen3_vl_moe",
             ]

src/llamafactory/data/template.py

@@ -2029,6 +2029,39 @@ register_template(
 )
 
 
+register_template(
+    name="qwen3_5",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_assistant=StringFormatter(slots=["{{content}}<|im_end|>\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_function=FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="qwen3_5"),
+    format_observation=StringFormatter(
+        slots=["<|im_start|>user\n<tool_response>\n{{content}}\n</tool_response><|im_end|>\n<|im_start|>assistant\n"]
+    ),
+    format_tools=ToolFormatter(tool_format="qwen3_5"),
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+    mm_plugin=get_mm_plugin(name="qwen3_vl", image_token="<|image_pad|>", video_token="<|video_pad|>"),
+    template_class=ReasoningTemplate,
+)
+
+
+register_template(
+    name="qwen3_5_nothink",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_assistant=StringFormatter(slots=["{{content}}<|im_end|>\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_function=FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="qwen3_5"),
+    format_observation=StringFormatter(
+        slots=["<|im_start|>user\n<tool_response>\n{{content}}\n</tool_response><|im_end|>\n<|im_start|>assistant\n"]
+    ),
+    format_tools=ToolFormatter(tool_format="qwen3_5"),
+    stop_words=["<|im_end|>"],
+    replace_eos=True,
+    mm_plugin=get_mm_plugin(name="qwen3_vl", image_token="<|image_pad|>", video_token="<|video_pad|>"),
+)
+
+
 register_template(
     name="sailor",
     format_user=StringFormatter(slots=["<|im_start|>question\n{{content}}<|im_end|>\n<|im_start|>answer\n"]),
@@ -2218,3 +2251,24 @@ register_template(
     format_system=StringFormatter(slots=["<|system|>\n{{content}}", {"eos_token"}]),
     default_system="You are Zephyr, a helpful assistant.",
 )
+
+
+# copied from glm4_7 template
+register_template(
+    name="aeva",
+    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_moe"),
+    format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>"]),
+    format_tools=ToolFormatter(tool_format="glm4_moe"),
+    format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
+    default_system=(
+        "You are an AI assistant named Aeva created by Zongzhi Lou. "
+        "Your answer should be friendly, unbiased, faithful, informative and detailed."
+    ),
+    stop_words=["<|user|>", "<|observation|>"],
+    thought_words=("<think>", "</think>"),
+    efficient_eos=True,
+    template_class=Glm47ReasoningTemplate,
+)

src/llamafactory/data/tool_utils.py

@@ -85,6 +85,21 @@ QWEN_TOOL_PROMPT = (
     """"arguments": <args-json-object>}}\n</tool_call>"""
 )
 
+QWEN35_TOOL_PROMPT = (
+    "\n\n# Tools\n\nYou have access to the following functions:\n\n<tools>{tool_text}"
+    "\n</tools>\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n"
+    "<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n"
+    "<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n"
+    "</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n"
+    "- Function calls MUST follow the specified format: "
+    "an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n"
+    "- Required parameters MUST be specified\n"
+    "- You may provide optional reasoning for your function call in natural language "
+    "BEFORE the function call, but NOT after\n"
+    "- If there is no function call available, answer the question like normal with your current knowledge "
+    "and do not tell the user about function calls\n</IMPORTANT>"
+)
+
 SEED_TOOL_PROMPT = (
     "system\nYou are Doubao, a helpful AI assistant. You may call one or more functions to assist with the user query."
     "Tool List:\nYou are authorized to use the following tools (described in JSON Schema format). Before performing "
@@ -453,6 +468,57 @@ class QwenToolUtils(ToolUtils):
         return results
 
 
+class Qwen35ToolUtils(ToolUtils):
+    r"""Qwen 3.5 tool using template."""
+
+    @override
+    @staticmethod
+    def tool_formatter(tools: list[dict[str, Any]]) -> str:
+        tool_text = ""
+        for tool in tools:
+            tool = tool.get("function", tool) if tool.get("type") == "function" else tool
+            tool_text += "\n" + json.dumps(tool, ensure_ascii=False)
+
+        return QWEN35_TOOL_PROMPT.format(tool_text=tool_text)
+
+    @override
+    @staticmethod
+    def function_formatter(functions: list["FunctionCall"]) -> str:
+        function_texts = []
+        for func in functions:
+            name, arguments = func.name, json.loads(func.arguments)
+            prompt = f"<tool_call>\n<function={name}>"
+            for key, value in arguments.items():
+                prompt += f"\n<parameter={key}>"
+                if not isinstance(value, str):
+                    value = json.dumps(value, ensure_ascii=False)
+                prompt += f"\n{value}\n</parameter>"
+            prompt += "\n</function>\n</tool_call>"
+            function_texts.append(prompt)
+
+        return "\n".join(function_texts)
+
+    @override
+    @staticmethod
+    def tool_extractor(content: str) -> Union[str, list["FunctionCall"]]:
+        results = []
+        regex = re.compile(r"<tool_call>\s*<function=\s*([^\s<>]+)\s*(.*?)\s*</function>\s*</tool_call>", re.DOTALL)
+        for func_name, params_block in re.findall(regex, content):
+            args_dict = {}
+            param_pattern = re.compile(r"<parameter=(.*?)>(.*?)</parameter>", re.DOTALL)
+            for key, raw_value in re.findall(param_pattern, params_block.strip()):
+                value = raw_value.strip()
+                try:
+                    parsed_value = json.loads(value)
+                except json.JSONDecodeError:
+                    parsed_value = raw_value.strip()
+                args_dict[key] = parsed_value
+
+            results.append(FunctionCall(func_name.strip(), json.dumps(args_dict, ensure_ascii=False)))
+
+        return results if results else content
+
+
 class GLM4MOEToolUtils(QwenToolUtils):
     r"""GLM-4-MOE tool using template."""
 
@@ -662,6 +728,7 @@ TOOLS = {
     "minimax2": MiniMaxM2ToolUtils(),
     "mistral": MistralToolUtils(),
     "qwen": QwenToolUtils(),
+    "qwen3_5": Qwen35ToolUtils(),
     "glm4_moe": GLM4MOEToolUtils(),
     "seed_oss": SeedToolUtils(),
     "ling": LingToolUtils(),

src/llamafactory/extras/constants.py

@@ -65,6 +65,7 @@ MCA_SUPPORTED_MODELS = {
     "qwen2_vl",
     "qwen2_5_vl",
     "qwen3_vl",
+    "qwen3_vl_moe",
     "qwen3",
     "qwen3_moe",
     "qwen3_next",
@@ -2809,6 +2810,29 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "Qwen3.5-27B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen3.5-27B",
+            DownloadSource.MODELSCOPE: "Qwen/Qwen3.5-27B",
+        },
+        "Qwen3.5-35B-A3B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen3.5-35B-A3B",
+            DownloadSource.MODELSCOPE: "Qwen/Qwen3.5-35B-A3B",
+        },
+        "Qwen3.5-122B-A10B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen3.5-122B-A10B",
+            DownloadSource.MODELSCOPE: "Qwen/Qwen3.5-122B-A10B",
+        },
+        "Qwen3.5-397B-A17B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen3.5-397B-A17B",
+            DownloadSource.MODELSCOPE: "Qwen/Qwen3.5-397B-A17B",
+        },
+    },
+    template="qwen3_5",
+)
+
+
 register_model_group(
     models={
         "Qwen2-Audio-7B": {
@@ -3450,3 +3474,35 @@ register_model_group(
     },
     template="zephyr",
 )
+
+
+register_model_group(
+    models={
+        "Aeva-Flash-Chat": {
+            DownloadSource.DEFAULT: "louzongzhi/Aeva-Flash",
+            DownloadSource.MODELSCOPE: "louzongktsi/Aeva-Flash",
+            DownloadSource.OPENMIND: "louzongzhi/Aeva-Flash",
+        },
+        "Aeva-Air-Chat": {
+            DownloadSource.DEFAULT: "louzongzhi/Aeva-Air",
+            DownloadSource.MODELSCOPE: "louzongktsi/Aeva-Air",
+            DownloadSource.OPENMIND: "louzongzhi/Aeva-Air",
+        },
+        "Aeva-Chat": {
+            DownloadSource.DEFAULT: "louzongzhi/Aeva",
+            DownloadSource.MODELSCOPE: "louzongktsi/Aeva",
+            DownloadSource.OPENMIND: "louzongzhi/Aeva",
+        },
+        "Aeva-Pro-Chat": {
+            DownloadSource.DEFAULT: "louzongzhi/Aeva-Pro",
+            DownloadSource.MODELSCOPE: "louzongktsi/Aeva-Pro",
+            DownloadSource.OPENMIND: "louzongzhi/Aeva-Pro",
+        },
+        "Aeva-Max-Chat": {
+            DownloadSource.DEFAULT: "louzongzhi/Aeva-Max",
+            DownloadSource.MODELSCOPE: "louzongktsi/Aeva-Max",
+            DownloadSource.OPENMIND: "louzongzhi/Aeva-Max",
+        },
+    },
+    template="aeva",
+)

src/llamafactory/extras/misc.py

@@ -94,7 +94,7 @@ def check_version(requirement: str, mandatory: bool = False) -> None:
 
 def check_dependencies() -> None:
     r"""Check the version of the required packages."""
-    check_version("transformers>=4.51.0,<=5.0.0")
+    check_version("transformers>=4.51.0,<=5.2.0")
     check_version("datasets>=2.16.0,<=4.0.0")
     check_version("accelerate>=1.3.0,<=1.11.0")
     check_version("peft>=0.18.0,<=0.18.1")

src/llamafactory/model/model_utils/moe.py

@@ -142,6 +142,11 @@ def add_z3_leaf_module(model: "PreTrainedModel") -> None:
 
         _set_z3_leaf_modules(model, [Qwen3OmniMoeThinkerTextSparseMoeBlock])
 
+    if model_type == "qwen3_next":
+        from transformers.models.qwen3_next.modeling_qwen3_next import Qwen3NextSparseMoeBlock
+
+        _set_z3_leaf_modules(model, [Qwen3NextSparseMoeBlock])
+
 
 def configure_moe(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
     if not is_trainable or not model_args.moe_aux_loss_coef:

src/llamafactory/train/mca/workflow.py

@@ -82,7 +82,33 @@ def _check_model_support(model_args: "ModelArguments"):
         model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
     )
     if config.model_type not in MCA_SUPPORTED_MODELS:
-        raise ValueError(f"Model {config.model_type} is not supported by MCA.")
+        raise ValueError(
+            f"Model {config.model_type} is not supported by mcore_adapter."
+            "You can try to upgrade mcore_adapter to the latest version for more supported models."
+        )
+
+
+def _freeze_model_parameters(model: Any, finetuning_args: "FinetuningArguments"):
+    """Freeze model parameters for qwen_vl series models based on finetuning arguments."""
+    if getattr(model.config, "hf_model_type", None) not in ["qwen2_vl", "qwen2_5_vl", "qwen3_vl", "qwen3_vl_moe"]:
+        return
+
+    params_to_freeze = []
+    if finetuning_args.freeze_vision_tower:
+        params_to_freeze.extend(["vision_model.blocks", "vision_model.patch_embed"])
+        if getattr(model.config, "hf_model_type", None) in ["qwen3_vl", "qwen3_vl_moe"]:
+            params_to_freeze.extend(["vision_model.pos_embed"])
+
+    if finetuning_args.freeze_multi_modal_projector:
+        params_to_freeze.extend(["multi_modal_projector"])
+
+    if finetuning_args.freeze_language_model:
+        params_to_freeze.extend(["embedding", "decoder", "output_layer"])
+
+    if params_to_freeze:
+        for name, p in model.named_parameters():
+            if any(name.startswith(k) for k in params_to_freeze):
+                p.requires_grad_(False)
 
 
 def run_pt(
@@ -161,22 +187,8 @@ def run_sft(
     _check_model_support(model_args)
     model = AutoModel.from_pretrained(model_args.model_name_or_path, training_args)
 
-    # optional freezing for qwen2_vl, qwen2_5_vl
-    if getattr(model.config, "hf_model_type", None) in ["qwen2_vl", "qwen2_5_vl", "qwen3_vl"]:
-        params_to_freeze = []
-        if finetuning_args.freeze_vision_tower:
-            params_to_freeze.extend(["vision_model.blocks", "vision_model.patch_embed"])
-
-        if finetuning_args.freeze_multi_modal_projector:
-            params_to_freeze.extend(["multi_modal_projector"])
-
-        if finetuning_args.freeze_language_model:
-            params_to_freeze.extend(["embedding", "decoder", "output_layer"])
-
-        if params_to_freeze:
-            for name, p in model.named_parameters():
-                if any(name.startswith(k) for k in params_to_freeze):
-                    p.requires_grad_(False)
+    # optional freezing for qwen_vl series
+    _freeze_model_parameters(model, finetuning_args)
 
     pad_to_max = training_args.expert_model_parallel_size is not None and training_args.expert_model_parallel_size > 1
     data_collator = SFTDataCollatorWith4DAttentionMask(
@@ -229,6 +241,8 @@ def run_dpo(
     _check_model_support(model_args)
     model = AutoModel.from_pretrained(model_args.model_name_or_path, training_args)
 
+    _freeze_model_parameters(model, finetuning_args)
+
     if finetuning_args.use_ref_model:
         ref_config = AutoConfig.from_pretrained(model_args.model_name_or_path, training_args)
         ref_model = AutoModel.from_config(ref_config)

src/llamafactory/train/tuner.py

@@ -24,7 +24,7 @@ from ..data import get_template_and_fix_tokenizer
 from ..extras import logging
 from ..extras.constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from ..extras.misc import find_available_port, get_device_name, get_torch_device, infer_optim_dtype
-from ..extras.packages import is_mcore_adapter_available, is_ray_available
+from ..extras.packages import is_mcore_adapter_available, is_ray_available, is_transformers_version_greater_than
 from ..hparams import RayArguments, get_infer_args, get_ray_args, get_train_args, read_args
 from ..model import load_model, load_tokenizer
 from .callbacks import LogCallback, PissaConvertCallback, ReporterCallback
@@ -160,17 +160,28 @@ def export_model(args: Optional[dict[str, Any]] = None) -> None:
         model = model.to(output_dtype)
         logger.info_rank0(f"Convert model dtype to: {output_dtype}.")
 
-    model.save_pretrained(
-        save_directory=model_args.export_dir,
-        max_shard_size=f"{model_args.export_size}GB",
-        safe_serialization=(not model_args.export_legacy_format),
-    )
+    # Prepare save arguments (safe_serialization removed in transformers v5.0.0)
+    save_kwargs = {
+        "save_directory": model_args.export_dir,
+        "max_shard_size": f"{model_args.export_size}GB",
+    }
+    if not is_transformers_version_greater_than("5.0.0"):
+        save_kwargs["safe_serialization"] = not model_args.export_legacy_format
+
+    model.save_pretrained(**save_kwargs)
+
     if model_args.export_hub_model_id is not None:
+        # Prepare push arguments (safe_serialization removed in transformers v5.0.0)
+        push_kwargs = {
+            "max_shard_size": f"{model_args.export_size}GB",
+        }
+        if not is_transformers_version_greater_than("5.0.0"):
+            push_kwargs["safe_serialization"] = not model_args.export_legacy_format
+
         model.push_to_hub(
             model_args.export_hub_model_id,
             token=model_args.hf_hub_token,
-            max_shard_size=f"{model_args.export_size}GB",
-            safe_serialization=(not model_args.export_legacy_format),
+            **push_kwargs,
         )
 
     if finetuning_args.stage == "rm":

tests/data/test_collator.py

@@ -22,6 +22,7 @@ from transformers import AutoConfig, AutoModelForImageTextToText
 from llamafactory.data import get_template_and_fix_tokenizer
 from llamafactory.data.collator import MultiModalDataCollatorForSeq2Seq, prepare_4d_attention_mask
 from llamafactory.extras.constants import IGNORE_INDEX
+from llamafactory.extras.packages import is_transformers_version_greater_than
 from llamafactory.hparams import get_infer_args
 from llamafactory.model import load_tokenizer
 
@@ -116,14 +117,16 @@ def test_multimodal_collator():
         "labels": [
             [0, 1, 2, 3, q, q, q, q, q, q, q, q],
         ],
-        "position_ids": [
-            [[0, 1, 2, 3, 1, 1, 1, 1, 1, 1, 1, 1]],
-            [[0, 1, 2, 3, 1, 1, 1, 1, 1, 1, 1, 1]],
-            [[0, 1, 2, 3, 1, 1, 1, 1, 1, 1, 1, 1]],
-        ],
-        "rope_deltas": [[-8]],
+        "position_ids": [[[0, 1, 2, 3, 0, 0, 0, 0, 0, 0, 0, 0]]] * 3,
+        "rope_deltas": [[0]],
         **tokenizer_module["processor"].image_processor(fake_image),
     }
+    if not is_transformers_version_greater_than("5.0.0"):
+        # adapt position_ids and rope_deltas for transformers < 5.0.0
+        # https://github.com/huggingface/transformers/pull/43972
+        expected_input["position_ids"] = [[[0, 1, 2, 3, 1, 1, 1, 1, 1, 1, 1, 1]]] * 3
+        expected_input["rope_deltas"] = [[-8]]
+
     assert batch_input.keys() == expected_input.keys()
     for k in batch_input.keys():
         assert batch_input[k].eq(torch.tensor(expected_input[k])).all()

tests/version.txt

@@ -1,2 +1,2 @@
 # change if test fails or cache is outdated
-0.9.5.106
+0.9.5.107