configuration_neollm.py

# ==================== configuration_neollm.py ====================
from transformers.configuration_utils import PretrainedConfig
from transformers.modeling_rope_utils import rope_config_validation
from transformers.utils import logging

logger = logging.get_logger(__name__)

class NeoLLMConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`NeoLLMModel`]. It is used to instantiate a
    NeoLLM model according to the specified arguments, defining the model architecture.
    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs.
    """
    model_type = "neollm"
    keys_to_ignore_at_inference = []
    
    def __init__(
        self,
        vocab_size=151665,
        hidden_size=512,
        intermediate_size=1536,
        num_hidden_layers=12,
        num_attention_heads=8,
        num_key_value_heads=2,
        hidden_act="xielu",
        max_position_embeddings=32768,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        tie_word_embeddings=True,
        rope_theta=10000.0,
        rope_scaling=None,
        partial_rotary_factor=0.25,
        attention_bias=False,
        attention_dropout=0.1,
        head_dim=64,
        linear_conv_kernel_dim=4,
        linear_key_head_dim=32,
        linear_value_head_dim=32,
        linear_num_key_heads=8,
        linear_num_value_heads=16,
        layer_types=None,
        fan_ratio=0.125,
        fan_ratio_ffn=0.0625,  # NEW: Half of fan_ratio for FFN periodicity modeling
        dropout_rate=0.1,
        **kwargs,
    ):
        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_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.rope_theta = rope_theta
        self.rope_scaling = rope_scaling
        self.partial_rotary_factor = partial_rotary_factor
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.head_dim = head_dim
        
        rope_config_validation(self)
        
        self.layer_types = layer_types
        if self.layer_types is None:
            interval_pattern = kwargs.get("full_attention_interval", 4)
            self.layer_types = [
                "linear_attention" if bool((i + 1) % interval_pattern) else "full_attention"
                for i in range(self.num_hidden_layers)
            ]
        
        # Linear attention parameters
        self.linear_conv_kernel_dim = linear_conv_kernel_dim
        self.linear_key_head_dim = linear_key_head_dim
        self.linear_value_head_dim = linear_value_head_dim
        self.linear_num_key_heads = linear_num_key_heads
        self.linear_num_value_heads = linear_num_value_heads
        
        # FANformer parameters
        self.fan_ratio = fan_ratio  # Used in attention mechanisms
        self.fan_ratio_ffn = fan_ratio_ffn  # NEW: Used in FFN for complementary periodicity
        
        self.dropout_rate = dropout_rate
        
        self.auto_map = {
            "AutoConfig": "configuration_neollm.NeoLLMConfig",
            "AutoModel": "modeling_neollm.NeoLLMModel",
            "AutoModelForCausalLM": "modeling_neollm.NeoLLMForCausalLM"
        }

__all__ = ["NeoLLMConfig"]