recastmlp_llama/configuration_recastmlp_llama.py

from transformers import PretrainedConfig


class RECASTMLP_llama(PretrainedConfig):
    model_type = "recastmlp_llama"
    attribute_map = {
        "hidden_size": "hidden_size",
        "num_attention_heads": "num_attention_heads",
    }

    def __init__(
        self,
        vocab_size=128256,
        hidden_size=4096,
        intermediate_size=14336,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=8,
        hidden_act="silu",
        max_position_embeddings=131072,
        initializer_range=0.02,
        rms_norm_eps=1e-5,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=128000,
        eos_token_id=128001,
        pretraining_tp=1,
        tie_word_embeddings=False,
        rope_theta=500000.0,
        rope_scaling={
            "factor": 8.0,
            "low_freq_factor": 1.0,
            "high_freq_factor": 4.0,
            "original_max_position_embeddings": 8192,
            "rope_type": "llama3",
        },
        attention_bias=False,
        attention_dropout=0.0,
        mlp_bias=False,
        # Template-specific configs
        num_templates=4,
        num_groups=8,
        num_cf=1,
        torch_dtype="bfloat16",
        **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.pretraining_tp = pretraining_tp
        self.use_cache = use_cache
        self.mlp_bias = mlp_bias
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.rope_theta = rope_theta
        self.rope_scaling = rope_scaling
        self.torch_dtype = torch_dtype

        # Template-specific configs
        self.num_templates = num_templates
        self.num_groups = num_groups
        self.num_cf = num_cf

        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
        )