PaddleOCR-VL 微调后的模型无法使用 FastDeploy 加载和推理

[megemini]

在 AI Studio 中微调了一个 PaddleOCR-VL 模型，参考 https://aistudio.baidu.com/projectdetail/9857242

使用如下命令进行部署，报错

aistudio@jupyter-942478-9857242:~$ python -m fastdeploy.entrypoints.openai.api_server \
>     --model /home/aistudio/paddleocr_vl/PaddleOCR-VL-SFT \
>     --port 8185 \
>     --metrics-port 8186 \
>     --engine-worker-queue-port 8187 \
>     --max-model-len 16384 \
>     --max-num-batched-tokens 16384 \
>     --gpu-memory-utilization 0.7 \
>     --max-num-seqs 256
/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/paddle/utils/cpp_extension/extension_utils.py:718: UserWarning: No ccache found. Please be aware that recompiling all source files may be required. You can download and install ccache from: https://github.com/ccache/ccache/blob/master/doc/INSTALL.md
  warnings.warn(warning_message)
INFO     2025-12-12 12:54:19,655 43112 api_server.py[line:86] Number of api-server workers: 1.
[2025-12-12 12:54:19,665] [    INFO] - Using download source: huggingface
[2025-12-12 12:54:19,666] [    INFO] - Loading configuration file /home/aistudio/paddleocr_vl/PaddleOCR-VL-SFT/config.json
[2025-12-12 12:54:19,666] [ WARNING] - You are using a model of type paddleocr_vl to instantiate a model of type . This is not supported for all configurations of models and can yield errors.
[2025-12-12 12:54:19,667] [ WARNING] - You are using a model of type paddleocr_vl to instantiate a model of type . This is not supported for all configurations of models and can yield errors.
/home/aistudio/external-libraries/lib/python3.10/site-packages/_distutils_hack/__init__.py:30: UserWarning: Setuptools is replacing distutils. Support for replacing an already imported distutils is deprecated. In the future, this condition will fail. Register concerns at https://github.com/pypa/setuptools/issues/new?template=distutils-deprecation.yml
  warnings.warn(
/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/paddleslim/common/load_model.py:20: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
  import pkg_resources as pkg
/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/model_executor/graph_optimization/utils.py:21: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
None of PyTorch, TensorFlow >= 2.0, or Flax have been found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
[2025-12-12 12:54:22,422] [    INFO] - Using download source: huggingface
[2025-12-12 12:54:22,423] [    INFO] - Loading configuration file /home/aistudio/paddleocr_vl/PaddleOCR-VL-SFT/generation_config.json
Traceback (most recent call last):
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/preprocess.py", line 73, in create_processor
    Processor = load_input_processor_plugins()
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/plugins/input_processor/__init__.py", line 26, in load_input_processor_plugins
    assert len(plugins) == 1, "Only one plugin is allowed to be loaded."
AssertionError: Only one plugin is allowed to be loaded.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/entrypoints/openai/api_server.py", line 715, in <module>
    main()
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/entrypoints/openai/api_server.py", line 698, in main
    if not load_engine():
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/entrypoints/openai/api_server.py", line 124, in load_engine
    if not engine.start(api_server_pid=args.port):
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/engine/engine.py", line 134, in start
    self.engine.create_data_processor()
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/engine/common_engine.py", line 161, in create_data_processor
    self.data_processor = self.input_processor.create_processor()
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/preprocess.py", line 116, in create_processor
    self.processor = PaddleOCRVLProcessor(
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/paddleocr_vl_processor/paddleocr_vl_processor.py", line 65, in __init__
    super().__init__(model_name_or_path, reasoning_parser_obj, tool_parser_obj)
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/text_processor.py", line 179, in __init__
    self.tokenizer = self._load_tokenizer()
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/text_processor.py", line 628, in _load_tokenizer
    return AutoTokenizer.from_pretrained(self.model_name_or_path, padding_side="left", use_fast=True)
  File "/home/aistudio/external-libraries/lib/python3.10/site-packages/paddleformers/transformers/auto/tokenizer.py", line 329, in from_pretrained
    raise ValueError(
ValueError: Tokenizer class Ernie4_5_Tokenizer does not exist or is not currently imported.
/opt/conda/envs/python35-paddle120-env/lib/python3.10/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to be 1 leaked shared_memory objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '

尝试使用 python 脚本进行推理，同样错误

from fastdeploy.entrypoints.llm import LLM
# 加载模型
llm = LLM(model="/home/aistudio/paddleocr_vl/PaddleOCR-VL-SFT")

outputs = llm.chat(
    messages=[
        {"role": "user", "content": [ {"type": "image_url", "image_url": {"url": "https://ai-studio-static-online.cdn.bcebos.com/dc31c334d4664ca4955aa47d8e202a53a276fd0aab0840b09abe953fe51207d0"}},
                                     {"type": "text", "text": "OCR:{}"}]}
    ],
    chat_template_kwargs={"enable_thinking": False})

# 输出结果
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs.text
    reasoning_text = output.outputs.reasoning_content

输出

[2025-12-12 12:58:45,335] [    INFO] - Using download source: huggingface
[2025-12-12 12:58:45,336] [    INFO] - Loading configuration file /home/aistudio/paddleocr_vl/PaddleOCR-VL-SFT/config.json
[2025-12-12 12:58:45,337] [ WARNING] - You are using a model of type paddleocr_vl to instantiate a model of type . This is not supported for all configurations of models and can yield errors.
[2025-12-12 12:58:45,338] [ WARNING] - You are using a model of type paddleocr_vl to instantiate a model of type . This is not supported for all configurations of models and can yield errors.
[2025-12-12 12:58:45,420] [    INFO] - Using download source: huggingface
[2025-12-12 12:58:45,422] [    INFO] - Loading configuration file /home/aistudio/paddleocr_vl/PaddleOCR-VL-SFT/generation_config.json
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/preprocess.py:73, in InputPreprocessor.create_processor(self)
     71 from fastdeploy.plugins.input_processor import load_input_processor_plugins
---> 73 Processor = load_input_processor_plugins()
     74 self.processor = Processor(
     75     model_name_or_path=self.model_name_or_path,
     76     reasoning_parser_obj=reasoning_parser_obj,
     77     tool_parser_obj=tool_parser_obj,
     78 )

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/plugins/input_processor/__init__.py:26, in load_input_processor_plugins()
     25 plugins = load_plugins_by_group(group=PLUGINS_GROUP)
---> 26 assert len(plugins) == 1, "Only one plugin is allowed to be loaded."
     27 return next(iter(plugins.values()))()

AssertionError: Only one plugin is allowed to be loaded.

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
Cell In[5], line 3
      1 from fastdeploy.entrypoints.llm import LLM
      2 # 加载模型
----> 3 llm = LLM(model="/home/aistudio/paddleocr_vl/PaddleOCR-VL-SFT")
      5 outputs = llm.chat(
      6     messages=[
      7         {"role": "user", "content": [ {"type": "image_url", "image_url": {"url": "https://ai-studio-static-online.cdn.bcebos.com/dc31c334d4664ca4955aa47d8e202a53a276fd0aab0840b09abe953fe51207d0"}},
      8                                      {"type": "text", "text": "OCR:{}"}]}
      9     ],
     10     chat_template_kwargs={"enable_thinking": False})
     12 # 输出结果

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/entrypoints/llm.py:100, in LLM.__init__(self, model, revision, tokenizer, enable_logprob, chat_template, **kwargs)
     96 self.llm_engine = LLMEngine.from_engine_args(engine_args=engine_args)
     98 self.default_sampling_params = SamplingParams(max_tokens=self.llm_engine.cfg.model_config.max_model_len)
--> 100 self.llm_engine.start()
    102 self.mutex = threading.Lock()
    103 self.req_output = dict()

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/engine/engine.py:134, in LLMEngine.start(self, api_server_pid)
    131 self.launch_components()
    133 self.engine.start()
--> 134 self.engine.create_data_processor()
    135 self.data_processor = self.engine.data_processor
    137 # If block numer is specified and model is deployed in mixed mode, start cache manager first

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/engine/common_engine.py:161, in EngineService.create_data_processor(self)
    153 def create_data_processor(self):
    154     self.input_processor = InputPreprocessor(
    155         self.cfg.model_config,
    156         self.cfg.structured_outputs_config.reasoning_parser,
   (...)
    159         self.cfg.tool_parser,
    160     )
--> 161     self.data_processor = self.input_processor.create_processor()

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/preprocess.py:116, in InputPreprocessor.create_processor(self)
    111 elif "PaddleOCRVL" in architecture:
    112     from fastdeploy.input.paddleocr_vl_processor import (
    113         PaddleOCRVLProcessor,
    114     )
--> 116     self.processor = PaddleOCRVLProcessor(
    117         config=self.model_config,
    118         model_name_or_path=self.model_name_or_path,
    119         limit_mm_per_prompt=self.limit_mm_per_prompt,
    120         mm_processor_kwargs=self.mm_processor_kwargs,
    121         reasoning_parser_obj=reasoning_parser_obj,
    122     )
    123 elif "PaddleOCRVL" in architecture:
    124     from fastdeploy.input.paddleocr_vl_processor import (
    125         PaddleOCRVLProcessor,
    126     )

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/paddleocr_vl_processor/paddleocr_vl_processor.py:65, in PaddleOCRVLProcessor.__init__(self, config, model_name_or_path, limit_mm_per_prompt, mm_processor_kwargs, reasoning_parser_obj, tool_parser_obj, enable_processor_cache)
     44 def __init__(
     45     self,
     46     config,
   (...)
     52     enable_processor_cache=False,
     53 ):
     54     """
     55     Initialize PaddleOCRVLProcessor instance.
     56 
   (...)
     63         tool_parser_obj: Tool parser instance
     64     """
---> 65     super().__init__(model_name_or_path, reasoning_parser_obj, tool_parser_obj)
     66     data_processor_logger.info(f"model_name_or_path: {model_name_or_path}")
     67     processor_kwargs = self._parse_processor_kwargs(mm_processor_kwargs)

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/text_processor.py:179, in DataProcessor.__init__(self, model_name_or_path, reasoning_parser_obj, tool_parser_obj)
    177 self.decode_status = dict()
    178 self.tool_parser_dict = dict()
--> 179 self.tokenizer = self._load_tokenizer()
    180 data_processor_logger.info(
    181     f"tokenizer information: bos_token is {self.tokenizer.bos_token}, {self.tokenizer.bos_token_id}, \
    182                         eos_token is {self.tokenizer.eos_token}, {self.tokenizer.eos_token_id} "
    183 )
    185 from paddleformers.trl.llm_utils import get_eos_token_id

File /opt/conda/envs/python35-paddle120-env/lib/python3.10/site-packages/fastdeploy/input/text_processor.py:628, in DataProcessor._load_tokenizer(self)
    625 else:
    626     from paddleformers.transformers import AutoTokenizer
--> 628     return AutoTokenizer.from_pretrained(self.model_name_or_path, padding_side="left", use_fast=True)

File ~/external-libraries/lib/python3.10/site-packages/paddleformers/transformers/auto/tokenizer.py:329, in AutoTokenizer.from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs)
    327     tokenizer_class = tokenizer_class_from_name(tokenizer_class_candidate)
    328 if tokenizer_class is None:
--> 329     raise ValueError(
    330         f"Tokenizer class {tokenizer_class_candidate} does not exist or is not currently imported."
    331     )
    333 # Bind PaddleTokenizerMixin
    334 tokenizer_class = _bind_paddle_mixin_if_available(tokenizer_class)

ValueError: Tokenizer class Ernie4_5_Tokenizer does not exist or is not currently imported.

@jzhang533

[zhang-prog]

这是缺一些文件，可以拉 hf 的 repo，其他文件保持不变，只用你微调后的 safetensors 替换一下原来的 safetensors 就可以运行了。

[megemini]

收到 ～ 感谢！

[megemini]

@zhang-prog 还是有点问题

hf 那边是 processing_paddleocr_vl.py，这边提示缺少 processing_ppocrvl.py ～

手动把文件名改过来之后，又提示错误

(venv310)  ✘ shun@shun-B660M-Pro-RS  ~/workspace/Projects/erniekit_paddleocr_vl_ner   master ±✚  python paddleocr_vl_transformers.py
The repository /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt contains custom code which must be executed to correctly load the model. You can inspect the repository content at /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt .
 You can inspect the repository content at https://hf.co//media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt.
You can avoid this prompt in future by passing the argument `trust_remote_code=True`.

Do you wish to run the custom code? [y/N] y
Traceback (most recent call last):
  File "/home/shun/workspace/Projects/erniekit_paddleocr_vl_ner/paddleocr_vl_transformers.py", line 22, in <module>
    processor = AutoProcessor.from_pretrained("/media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt")
  File "/home/shun/workspace/Projects/github/transformers/src/transformers/models/auto/processing_auto.py", line 382, in from_pretrained
    processor_class = get_class_from_dynamic_module(
  File "/home/shun/workspace/Projects/github/transformers/src/transformers/dynamic_module_utils.py", line 583, in get_class_from_dynamic_module
    return get_class_in_module(class_name, final_module, force_reload=force_download)
  File "/home/shun/workspace/Projects/github/transformers/src/transformers/dynamic_module_utils.py", line 311, in get_class_in_module
    return getattr(module, class_name)
AttributeError: module 'transformers_modules.PaddleOCR_VL_Prompt.processing_ppocrvl' has no attribute 'PPOCRVLProcessor'. Did you mean: 'PaddleOCRVLProcessor'?

[megemini]

是不是 tokenizer_config.json 中的配置错了

	"auto_map": {
		"AutoProcessor": "processing_ppocrvl.PPOCRVLProcessor"
	},

[megemini]

改了之后还是有错误 🫠

(venv310)  shun@shun-B660M-Pro-RS  ~/workspace/Projects/erniekit_paddleocr_vl_ner   master ±✚  python paddleocr_vl_transformers.py
The repository /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt contains custom code which must be executed to correctly load the model. You can inspect the repository content at /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt .
 You can inspect the repository content at https://hf.co//media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt.
You can avoid this prompt in future by passing the argument `trust_remote_code=True`.

Do you wish to run the custom code? [y/N] y
Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`.
Traceback (most recent call last):
  File "/home/shun/workspace/Projects/erniekit_paddleocr_vl_ner/paddleocr_vl_transformers.py", line 22, in <module>
    processor = AutoProcessor.from_pretrained("/media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR_VL_Prompt")
  File "/home/shun/workspace/Projects/github/transformers/src/transformers/models/auto/processing_auto.py", line 387, in from_pretrained
    return processor_class.from_pretrained(
  File "/home/shun/workspace/Projects/github/transformers/src/transformers/processing_utils.py", line 1396, in from_pretrained
    args = cls._get_arguments_from_pretrained(pretrained_model_name_or_path, **kwargs)
  File "/home/shun/workspace/Projects/github/transformers/src/transformers/processing_utils.py", line 1482, in _get_arguments_from_pretrained
    sub_processor = auto_processor_class.from_pretrained(pretrained_model_name_or_path, **kwargs)
  File "/home/shun/workspace/Projects/github/transformers/src/transformers/models/auto/tokenization_auto.py", line 736, in from_pretrained
    if tokenizer_class.__name__ == "PythonBackend":  # unless you inherit from it?
AttributeError: 'NoneType' object has no attribute '__name__'. Did you mean: '__ne__'?

[zhang-prog]

请确认是否拉的最新代码呢？贴一下config.json的内容我看看～

[megemini]

transformers 是最新的代码 ～ 微调后生成的模型文件是不是有点问题？

根据 https://github.com/PaddlePaddle/ERNIE/blob/release/v1.4/docs/paddleocr_vl_sft_zh.md 里面用的是 git clone https://github.com/PaddlePaddle/ERNIE -b release/v1.4 这个版本 ～

微调后生成的模型文件中 config.json

{
  "architectures": [
    "PaddleOCRVLForConditionalGeneration"
  ],
  "attention_probs_dropout_prob": 0.0,
  "auto_map": {
    "AutoConfig": "configuration_paddleocr_vl.PaddleOCRVLConfig",
    "AutoModel": "modeling_paddleocr_vl.PaddleOCRVLForConditionalGeneration",
    "AutoModelForCausalLM": "modeling_paddleocr_vl.PaddleOCRVLForConditionalGeneration"
  },
  "compression_ratio": 1.0,
  "disable_pipeline_warmup": false,
  "enable_mtp_magic_send": false,
  "fp16_opt_level": "O2",
  "freq_allocation": 20,
  "fuse_ln": false,
  "fuse_rms_norm": true,
  "head_dim": 128,
  "hidden_act": "silu",
  "hidden_dropout_prob": 0.0,
  "hidden_size": 1024,
  "ignored_index": -100,
  "im_patch_id": 100295,
  "image_token_id": 100295,
  "intermediate_size": 3072,
  "max_position_embeddings": 131072,
  "max_text_id": 100295,
  "model_type": "paddleocr_vl",
  "moe_dropout_prob": 0.0,
  "moe_multimodal_dispatch_use_allgather": "v2-alltoall-unpad",
  "num_attention_heads": 16,
  "num_hidden_layers": 18,
  "num_key_value_heads": 2,
  "paddleformers_version": "0.3.2",
  "pixel_hidden_size": 1152,
  "rms_norm_eps": 1e-05,
  "rope_is_neox_style": true,
  "rope_scaling": {
    "mrope_section": [
      16,
      24,
      24
    ],
    "rope_type": "default",
    "type": "default"
  },
  "rope_theta": 500000,
  "scale_qk_coeff": 1.0,
  "seqlen": 16384,
  "sliding_window": null,
  "tie_word_embeddings": false,
  "token_balance_loss": false,
  "token_balance_seqlen": 16384,
  "torch_dtype": "bfloat16",
  "use_3d_rope": true,
  "use_bias": false,
  "use_flash_attn_with_mask": true,
  "use_fp8": false,
  "use_mem_eff_attn": true,
  "use_recompute_moe": false,
  "use_rmsnorm": true,
  "video_token_id": 101307,
  "vision_config": {
    "_attn_implementation": "eager",
    "_name_or_path": "",
    "_save_to_hf": false,
    "add_cross_attention": false,
    "add_tail_layers": 0,
    "architectures": [
      "SiglipVisionModel"
    ],
    "attention_dropout": 0.0,
    "auto_map": {
      "AutoConfig": "configuration_paddleocr_vl.PaddleOCRVLConfig",
      "AutoModel": "modeling_paddleocr_vl.SiglipVisionModel"
    },
    "bad_words_ids": null,
    "begin_suppress_tokens": null,
    "bos_token_id": null,
    "chunk_size_feed_forward": 0,
    "classifier_dropout": null,
    "context_parallel_degree": 1,
    "cross_attention_hidden_size": null,
    "decoder_start_token_id": null,
    "diversity_penalty": 0.0,
    "do_sample": false,
    "dpo_config": null,
    "dtype": "bfloat16",
    "early_stopping": false,
    "encoder_no_repeat_ngram_size": 0,
    "eos_token_id": null,
    "exponential_decay_length_penalty": null,
    "finetuning_task": null,
    "forced_bos_token_id": null,
    "forced_eos_token_id": null,
    "fuse_attention_ffn": false,
    "fuse_attention_qkv": false,
    "fuse_linear": false,
    "fuse_rope": false,
    "fuse_sequence_parallel_allreduce": false,
    "fuse_swiglu": false,
    "hidden_act": "gelu_new",
    "hidden_size": 1152,
    "id2label": {
      "0": "LABEL_0",
      "1": "LABEL_1"
    },
    "image_size": 384,
    "intermediate_size": 4304,
    "is_decoder": false,
    "is_encoder_decoder": false,
    "kto_config": null,
    "label2id": {
      "LABEL_0": 0,
      "LABEL_1": 1
    },
    "layer_norm_eps": 1e-06,
    "length_penalty": 1.0,
    "loss_subbatch_sequence_length": -1,
    "max_length": 20,
    "min_length": 0,
    "model_type": "paddleocr_vl",
    "moe_subbatch_token_num": 0,
    "no_recompute_layers": null,
    "no_repeat_ngram_size": 0,
    "num_attention_heads": 16,
    "num_beam_groups": 1,
    "num_beams": 1,
    "num_channels": 3,
    "num_choices": null,
    "num_hidden_layers": 27,
    "num_return_sequences": 1,
    "offload_recompute_inputs": false,
    "output_attentions": false,
    "output_hidden_states": false,
    "output_scores": false,
    "pad_token_id": 0,
    "patch_size": 14,
    "pipeline_parallel_degree": 1,
    "pp_recompute_interval": 1,
    "prefix": null,
    "problem_type": null,
    "pruned_heads": {},
    "quantization_config": {
      "act_quant_method": "abs_max",
      "activation_scheme": null,
      "actscale_moving_rate": 0.01,
      "apply_hadamard": false,
      "apply_online_actscale_step": 200,
      "dense_quant_type": "",
      "dtype": null,
      "fmt": null,
      "fp8_format_type": "hybrid",
      "group_size": -1,
      "hadamard_block_size": 32,
      "ignore_modules": null,
      "llm_int8_threshold": 6.0,
      "moe_quant_type": "",
      "qlora_weight_blocksize": 64,
      "qlora_weight_double_quant": false,
      "qlora_weight_double_quant_block_size": 256,
      "quant_input_grad": false,
      "quant_method": null,
      "quant_round_type": 0,
      "quant_type": null,
      "quant_weight_grad": false,
      "quantization": "",
      "scale_epsilon": 1e-08,
      "shift": false,
      "shift_smooth_all_linears": false,
      "smooth": false,
      "weight_block_size": null,
      "weight_quant_method": "abs_max_channel_wise",
      "weight_quantize_algo": null
    },
    "recompute": true,
    "recompute_granularity": "full",
    "recompute_use_reentrant": false,
    "refined_recompute": "",
    "remove_invalid_values": false,
    "repetition_penalty": 1.0,
    "return_dict": false,
    "return_dict_in_generate": false,
    "sep_parallel_degree": 1,
    "sep_token_id": null,
    "sequence_parallel": false,
    "spatial_merge_size": 2,
    "suppress_tokens": null,
    "task_specific_params": null,
    "temperature": 1.0,
    "temporal_patch_size": 2,
    "tensor_parallel_degree": 1,
    "tensor_parallel_output": true,
    "tensor_parallel_rank": 0,
    "tie_encoder_decoder": false,
    "tie_word_embeddings": true,
    "tokenizer_class": null,
    "tokens_per_second": 2,
    "top_k": 50,
    "top_p": 1.0,
    "typical_p": 1.0,
    "use_cache": false,
    "use_filtered_label_loss": false,
    "use_flash_attention": true,
    "use_fused_dropout_add": false,
    "use_fused_head_and_loss_fn": false,
    "use_fused_linear": false,
    "use_fused_linear_cross_entropy": false,
    "use_fused_rms_norm": false,
    "use_fused_rope": false,
    "use_sparse_flash_attn": true,
    "use_sparse_head_and_loss_fn": false,
    "virtual_pp_degree": 1
  },
  "vision_start_token_id": 101305,
  "vocab_size": 103424,
  "weight_share_add_bias": true
}

[zhang-prog]

这是缺一些文件，可以拉 hf 的 repo，其他文件保持不变，只用你微调后的 safetensors 替换一下原来的 safetensors 就可以运行了。

您是不是只是copy了一些py文件到微调目录下了呢？这里可能我表达有误，我的意思是全部的文件（包括config.json）都用huggingface上的，也就是这里：
https://huggingface.co/PaddlePaddle/PaddleOCR-VL/tree/main

然后只是替换其中的 safetensors，也就是说只需要用到您微调之后的safetensors文件。

[megemini]

这是缺一些文件，可以拉 hf 的 repo，其他文件保持不变，只用你微调后的 safetensors 替换一下原来的 safetensors 就可以运行了。

您是不是只是copy了一些py文件到微调目录下了呢？这里可能我表达有误，我的意思是全部的文件（包括config.json）都用huggingface上的，也就是这里： https://huggingface.co/PaddlePaddle/PaddleOCR-VL/tree/main

然后只是替换其中的 safetensors，也就是说只需要用到您微调之后的safetensors文件。

收到 ～

我之前的模型文件是从 modelscope https://modelscope.cn/models/PaddlePaddle/PaddleOCR-VL/files 下载的（速度快），这么看，那边的模型好像是旧的 ... ...

[zhang-prog]

好的，这块儿我反应相关同学更新下～