HF配置迁移：copy HF模型配置py文件，实现装饰器机制动态注册MF独有参数功能和拦截无需在MF中使用的HF参数

设计思路：
1. copy HF ModelConfig源码，修改PretrainedConfig和logger导入信息，删除不必要的代码逻辑，参数及定义均保持不变
2. 显示定义装饰器动态注册MF中MLATransformerConfig支持的自定义参数
2. 显示定义装饰器对MF使用不到的参数进行统一拦截并打印，定义拦截参数的类型，可默认预置不同模型的默认拦截参数，支持自定义增加拦截参数

要求规范：
1. 显示打印不支持的HF Config参数，传入预置模型过滤类型和自定义过滤参数和对应过滤说明，不传入时默认过滤HF已知的不支持参数；
2. 动态注册MFModelConfig，显示写明额外支持的MF入参
3. 装饰器在__init__处装饰，同时如果不传任何参数的话，要注意加()，如@register_mf_parameter()或@IgnoreAndDeleteParameterDecorator()

装饰器实现demo，命名model_config_demo.py
代码样例
```
import inspect
from dataclasses import dataclass, asdict
from typing import List, Tuple
from functools import wraps

def scatter_multi_mapping_keys_to_mapping(mapping):
    """
    Expand multiple mapping relationships contained in `convert_map`.

Args:
        mapping (dict): A dict contains the mapping to convert the keys of `model_config`.

Returns:
        A dict with expand of all mappings.
    """
    new_mapping = {}

for k, v in mapping.items():
        if not isinstance(k, tuple):
            new_mapping[k] = v
            continue

for multi_key in k:
            new_mapping[multi_key] = v

return new_mapping

def validate_ignore_parameter_format(ignore_list: List[Tuple[str, str]] = None):
    """
    校验忽略参数列表的格式是否符合规范

参数:
        ignore_list: 要校验的忽略参数列表

异常:
        TypeError: 当数据类型不符合规范时抛出
        ValueError: 当数据结构不符合规范时抛出
    """
    if ignore_list is None:
        return []

# 1. 检查是否为列表
    if not isinstance(ignore_list, list):
        raise TypeError(f"IGNORE_COMMON_HF_PARAMETER must be a list, got {type(ignore_list).__name__}")

# 2. 检查每个元素是否为元组
    for i, item in enumerate(ignore_list):
        if not isinstance(item, tuple):
            raise TypeError(
                f"Item at index {i} must be a tuple, got {type(item).__name__}. "
                f"Expected format: [('param_name', 'reason'), ...]"
            )

# 3. 检查元组长度是否为2
    for i, item in enumerate(ignore_list):
        if len(item) != 2:
            raise ValueError(
                f"Tuple at index {i} must have exactly 2 elements, got {len(item)}. "
                f"Expected format: ('param_name', 'reason')"
            )

# 4. 检查第一个元素是否为字符串
    for i, (param_name, _) in enumerate(ignore_list):
        if not isinstance(param_name, str):
            raise TypeError(
                f"First element in tuple at index {i} must be a string (parameter name), "
                f"got {type(param_name).__name__}"
            )

# 5. 检查第二个元素是否为有效类型
    for i, (_, reason) in enumerate(ignore_list):
        if not isinstance(reason, str):
            raise TypeError(
                f"Second element in tuple at index {i} must be a string or have a string representation, "
                f"got {type(reason).__name__}"
            )

return ignore_list

@dataclass
class NOT_SUPPORTED_INFO:
    useless = "Useless"  # useless for MF
    not_implemented = "NotImplemented"  # not supported for now

# 随意写的过滤参数，需要重新梳理
IGNORE_COMMON_HF_PARAMETER = \
    [('torch_dtype', 'Useless, replace by compute_dtype'),

('use_cache', 'Useless, enable kv_cache by default'),

('transformers_version', NOT_SUPPORTED_INFO.useless),

('use_sliding_window', NOT_SUPPORTED_INFO.not_implemented), ]

# 随意写的过滤参数，需要重新梳理
IGNORE_HF_MODEL_CONFIG_MAPPING = scatter_multi_mapping_keys_to_mapping({
    ("qwen2", "qwen2.5", "qwen3"):
        [('decoder_sparse_step', NOT_SUPPORTED_INFO.not_implemented),

('mlp_only_layers', NOT_SUPPORTED_INFO.not_implemented),

('norm_topk_prob', NOT_SUPPORTED_INFO.not_implemented),

('output_router_logits', NOT_SUPPORTED_INFO.not_implemented),

('router_aux_loss_coef', NOT_SUPPORTED_INFO.not_implemented),

("max_window_layers", NOT_SUPPORTED_INFO.not_implemented)],
    "deepseek-v3": [],
})

@dataclass
class MFModelConfig:
    """
    定义MF训练、推理使用的参数，区别于HuggingFace的相关参数集合
    """
    seq_length: int = None
    """Model Seq Length"""

is_dynamic: bool = False
    """Whether model is dynamic shape."""

pad_token_id: int = 0
    """Model pad token id."""

ignore_token_id: int = -100
    """Model ignore token id when training."""

compute_dtype: str = "bfloat16"
    """Linear layer compute dtype."""

layernorm_compute_dtype: str = "float32"
    """LayerNorm compute dtype."""

rotary_dtype: str = "float32"
    """Custom rotary position embedding compute dtype."""

use_eod_reset: bool = False
    """Whether to use eod reset."""

use_flash_attention: bool = True
    """If true, use flash attention for the attention layer."""

class IgnoreAndDeleteParameterDecorator:
    """类装饰器用于拦截并打印__init__中不支持的参数"""

def __init__(self, ignore_type: str = None, extra_ignore_param: List[Tuple[str, str]] = None):
        self.ignore_type = ignore_type
        self.extra_ignore_param = validate_ignore_parameter_format(extra_ignore_param)
        # 添加类级别打印标记
        self.printed_classes = set()

def __call__(self, init_func):
        # 默认忽略的通用参数列表
        ignore_info = IGNORE_COMMON_HF_PARAMETER.copy()

if self.ignore_type:
            model_ignore = IGNORE_HF_MODEL_CONFIG_MAPPING.get(self.ignore_type, [])
            ignore_info.extend(model_ignore)

ignore_info.extend(self.extra_ignore_param)

# 提取所有要忽略的参数名
        ignore_param_names = [item[0] for item in ignore_info]

# 获取原始签名入参
        sig = inspect.signature(init_func)
        all_parameters_kwargs = dict()
        for name, param in sig.parameters.items():
            if param.kind in (inspect.Parameter.VAR_POSITIONAL, inspect.Parameter.VAR_KEYWORD):
                continue
            if name in ('self', 'kwargs'):
                continue
            all_parameters_kwargs.setdefault(name, param.default)

@wraps(init_func)
        def wrapper(self_instance, *args, **kwargs):

# 合并__init__默认签名参数和自定义传入参数

merge_kwargs = {**kwargs, **all_parameters_kwargs}

# 删除传入的忽略参数
            for param_name in ignore_param_names:
                if param_name in merge_kwargs:
                    merge_kwargs.pop(param_name)

# 获取当前类名
            class_name = self_instance.__class__.__name__
            # 检查是否已为该类打印过
            if class_name in self.printed_classes:
                result = init_func(self_instance, *args, **merge_kwargs)
                # 删除需要忽略参数的属性
                for param_name in ignore_param_names:
                    if hasattr(self_instance, param_name):
                        delattr(self_instance, param_name)
                return result

# 标记该类已打印
            self.printed_classes.add(class_name)

# 打印不支持的参数表格
            logger.warning(f"Found unsupported huggingface arguments in {self_instance.__class__.__name__}:")

# 计算列宽
            max_key_len = max(len(str(item[0])) for item in ignore_info) + 2
            max_val_len = max(len(str(item[1])) for item in ignore_info) + 2

# 创建表格边框
            border = f"+{'-' * (max_key_len + 2)}+{'-' * (max_val_len + 2)}+"

# 打印表头
            logger.warning(border)
            logger.warning(f"| {'Argument'.ljust(max_key_len)} | {'Status-Info'.ljust(max_val_len)} |")
            logger.warning(f"|:{'-' * (max_key_len + 1)}|:{'-' * (max_val_len + 1)}|")

# 打印参数行
            for arg, value in ignore_info:
                arg_str = str(arg).ljust(max_key_len)
                val_str = str(value).ljust(max_val_len)
                logger.warning(f"| {arg_str} | {val_str} |")

logger.warning(border)

# 调用原始初始化方法
            result = init_func(self_instance, *args, **merge_kwargs)

# 删除需要忽略参数的属性
            for param_name in ignore_param_names:
                if hasattr(self_instance, param_name):
                    delattr(self_instance, param_name)

return result

return wrapper

def register_mf_parameter(mf_model_kwargs=None):
    """
    装饰器工厂函数，用于自定义 __init__ 方法的 kwargs

参数:
        mf_model_kwargs: 默认的关键字参数，将被添加到或覆盖原始 kwargs

返回:
        装饰器函数
    """
    mf_model_kwargs = asdict(mf_model_kwargs) \
        if mf_model_kwargs is not None is None else asdict(MFModelConfig())

def decorator(init_func):
        def wrapper(self, *args, **kwargs):
            # 合并默认参数和传入参数，传入参数优先级更高
            merged_kwargs = {**mf_model_kwargs, **kwargs}
            return init_func(self, *args, **merged_kwargs)

return wrapper

return decorator
```

Huggingface原始copy模型配置代码，configuration_deepseek.py，做以下增量标准修改：
```
# from transformers.configuration_utils import PretrainedConfig
# 修改1：替换成MindFormers框架定义的PretrainedConfig  
from mindformers.models.configuration_utils import PretrainedConfig
# from transformers.utils import logging
# 修改2：替换成MindFormers框架的logger函数
from mindformers.tools.logger import logger

# logger = logging.get_logger(__name__)  修改3：删除该日志定义逻辑，MF中无需使用

DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
class DeepseekV3Config(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek
    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
    defaults will yield a similar configuration to that of the DeepSeek-V3.

Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

Args:
        vocab_size (`int`, *optional*, defaults to 129280):
            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`DeepseekV3Model`]
        hidden_size (`int`, *optional*, defaults to 4096):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 11008):
            Dimension of the MLP representations.
        moe_intermediate_size (`int`, *optional*, defaults to 1407):
            Dimension of the MoE representations.
        num_hidden_layers (`int`, *optional*, defaults to 32):
            Number of hidden layers in the Transformer decoder.
        num_nextn_predict_layers (`int`, *optional*, defaults to 1):
            Number of nextn predict layers in the DeepSeekV3 Model.
        num_attention_heads (`int`, *optional*, defaults to 32):
            Number of attention heads for each attention layer in the Transformer decoder.
        n_shared_experts (`int`, *optional*, defaults to None):
            Number of shared experts, None means dense model.
        n_routed_experts (`int`, *optional*, defaults to None):
            Number of routed experts, None means dense model.
        routed_scaling_factor (`float`, *optional*, defaults to 1.0):
            Scaling factor or routed experts.
        topk_method (`str`, *optional*, defaults to `gready`):
            Topk method used in routed gate.
        n_group (`int`, *optional*, defaults to None):
            Number of groups for routed experts.
        topk_group (`int`, *optional*, defaults to None):
            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
        num_experts_per_tok (`int`, *optional*, defaults to None):
            Number of selected experts, None means dense model.
        moe_layer_freq (`int`, *optional*, defaults to 1):
            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
        first_k_dense_replace (`int`, *optional*, defaults to 0):
            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
                                                            \--k dense layers--/
        norm_topk_prob (`bool`, *optional*, defaults to False):
            Whether to normalize the weights of the routed experts.
        scoring_func (`str`, *optional*, defaults to 'softmax'):
            Method of computing expert weights.
        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
            Auxiliary loss weight coefficient.
        seq_aux = (`bool`, *optional*, defaults to True):
            Whether to compute the auxiliary loss for each individual sample.
        num_key_value_heads (`int`, *optional*):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
            `num_attention_heads`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to 2048):
            The maximum sequence length that this model might ever be used with.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        pad_token_id (`int`, *optional*):
            Padding token id.
        bos_token_id (`int`, *optional*, defaults to 1):
            Beginning of stream token id.
        eos_token_id (`int`, *optional*, defaults to 2):
            End of stream token id.
        pretraining_tp (`int`, *optional*, defaults to 1):
            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
            issue](https://github.com/pytorch/pytorch/issues/76232).
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether to tie weight embeddings
        rope_theta (`float`, *optional*, defaults to 10000.0):
            The base period of the RoPE embeddings.
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
            `max_position_embeddings` to the expected new maximum.
        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
            Whether to use a bias in the query, key, value and output projection layers during self-attention.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.

```python
    >>> from transformers import DeepseekV3Model, DeepseekV3Config

>>> # Initializing a Deepseek-V3 style configuration
    >>> configuration = DeepseekV3Config()

>>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

model_type = "deepseek_v3"
    keys_to_ignore_at_inference = ["past_key_values"]

# 修改5：使用MF中提供的装饰器，拦截不支持的和注册MF自定义参数
    from .model_config_demo import register_mf_parameter, IgnoreAndDeleteParameterDecorator
    @register_mf_parameter(mf_model_kwargs=MFModelConfig(compute_dtype='bf16', layernorm_compute_dtype="fp32"))
    @IgnoreAndDeleteParameterDecorator(ignore_type='qwen2', extra_ignore_param=[('n_shared_experts', NOT_SUPPORTED_INFO.useless)])
    def __init__(
        self,
        vocab_size=129280,
        hidden_size=7168,
        intermediate_size=18432,
        moe_intermediate_size = 2048,
        num_hidden_layers=61,
        num_nextn_predict_layers=1,
        num_attention_heads=128,
        num_key_value_heads=128,
        n_shared_experts = 1,
        n_routed_experts = 256,
        ep_size = 1,
        routed_scaling_factor = 2.5,
        kv_lora_rank = 512,
        q_lora_rank = 1536,
        qk_rope_head_dim = 64,
        v_head_dim = 128,
        qk_nope_head_dim = 128,
        topk_method = 'noaux_tc',
        n_group = 8,
        topk_group = 4,
        num_experts_per_tok = 8,
        moe_layer_freq = 1,
        first_k_dense_replace = 3,
        norm_topk_prob = True,
        scoring_func = 'sigmoid',
        aux_loss_alpha = 0.001,
        seq_aux = True,
        hidden_act="silu",
        max_position_embeddings=4096,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=0,
        eos_token_id=1,
        pretraining_tp=1,
        tie_word_embeddings=False,
        rope_theta=10000.0,
        rope_scaling=None,
        attention_bias=False,
        attention_dropout=0.0,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.moe_intermediate_size = moe_intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_nextn_predict_layers = num_nextn_predict_layers
        self.num_attention_heads = num_attention_heads
        self.n_shared_experts = n_shared_experts
        self.n_routed_experts = n_routed_experts
        self.ep_size = ep_size
        self.routed_scaling_factor = routed_scaling_factor
        self.kv_lora_rank = kv_lora_rank
        self.q_lora_rank = q_lora_rank
        self.qk_rope_head_dim = qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.qk_nope_head_dim = qk_nope_head_dim
        self.topk_method = topk_method
        self.n_group = n_group
        self.topk_group = topk_group
        self.num_experts_per_tok = num_experts_per_tok
        self.moe_layer_freq = moe_layer_freq
        self.first_k_dense_replace = first_k_dense_replace
        self.norm_topk_prob = norm_topk_prob
        self.scoring_func = scoring_func
        self.aux_loss_alpha = aux_loss_alpha
        self.seq_aux = seq_aux
        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.pretraining_tp = pretraining_tp
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.rope_scaling = rope_scaling
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout

super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )
```

执行效果：
```
config = DeepseekV3Config(use_flash_attention=True)
print(config.to_dict())
config = DeepseekV3Config(use_cache=True)
print(config.use_cache)
```

![输入图片说明](https://foruda.gitee.com/images/1749900404693837363/9679b7c7_9324149.png "屏幕截图")
![
](https://foruda.gitee.com/images/1749903773594439839/035b23ad_9324149.png "屏幕截图")

GVP MindSpore/mindformers
关闭

内容风险标识

评论 (0)

GVPMindSpore/mindformers关闭

内容风险标识