configuration_xmodel.py

# Copyright (c) 2023 XiaoDuo AI. All rights reserved.

from transformers.configuration_utils import PretrainedConfig
from transformers.modeling_rope_utils import rope_config_validation
from transformers.utils import logging
from typing_extensions import Self

logger = logging.get_logger(__name__)


class XmodelConfig(PretrainedConfig):
    model_type = "xmodel"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
            self,
            vocab_size=32000,
            hidden_size=1536,
            intermediate_size=4096,
            num_hidden_layers=48,
            num_attention_heads=24,
            num_key_value_heads=8,
            hidden_act="silu",
            max_position_embeddings=131072,
            initializer_range=0.1,
            rms_norm_eps=1e-5,
            use_cache=True,
            pad_token_id=None,
            bos_token_id=1,
            eos_token_id=2,
            pretraining_tp=1,
            tie_word_embeddings=True,
            rope_theta=500000.0,
            rope_scaling=None,
            attention_bias=False,
            attention_dropout=0.0,
            mlp_bias=False,
            hidden_act_param=0.03,
            scale_emb=12,
            dim_model_base=256,
            scale_depth=1.4,
            **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        # self.intermediate_size = intermediate_size
        if intermediate_size is None:
            self.intermediate_size = find_multiple(int(8 * hidden_size / 3), 256)
        else:
            self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads

        # 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.hidden_act_param = hidden_act_param
        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
        self.mlp_bias = mlp_bias
        self.scale_emb = scale_emb
        self.dim_model_base = dim_model_base
        self.scale_depth = scale_depth

        self.auto_map = {
            "AutoConfig": "configuration_xmodel.XmodelConfig",
            "AutoModelForCausalLM": "modeling_xmodel.XmodelForCausalLM"
        }

        # Validate the correctness of rotary position embeddings parameters
        # BC: if there is a 'type' field, move it to 'rope_type'.
        if self.rope_scaling is not None and "type" in self.rope_scaling:
            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
        rope_config_validation(self)

        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,
        )

    @classmethod
    def from_name(cls, name: str) -> Self:
        return cls(**xmodel_configs[name])


xmodel_configs = {
    "nano": dict(num_hidden_layers=8,
                 num_attention_heads=4,
                 num_key_value_heads=1,
                 hidden_size=256,
                 tie_word_embeddings=True,
                 intermediate_size=640),

    "nano_old": dict(num_hidden_layers=6,
                     num_attention_heads=6,
                     num_key_value_heads=1,
                     hidden_size=192,
                     tie_word_embeddings=False),

    "micro": dict(num_hidden_layers=12,
                  num_attention_heads=6,
                  num_key_value_heads=1,
                  hidden_size=384,
                  tie_word_embeddings=True,
                  intermediate_size=960),

    "micro_old": dict(num_hidden_layers=6,
                      num_attention_heads=6,
                      num_key_value_heads=1,
                      hidden_size=384,
                      tie_word_embeddings=False),

    "tiny": dict(num_hidden_layers=18,
                 num_attention_heads=8,
                 num_key_value_heads=4,
                 hidden_size=512,
                 tie_word_embeddings=True,
                 intermediate_size=1280),

    "tiny_old": dict(num_hidden_layers=8,
                     num_attention_heads=8,
                     num_key_value_heads=2,
                     hidden_size=512,
                     tie_word_embeddings=False),

    # GPT-1 & Bert-Base
    "small": dict(num_hidden_layers=30,
                  num_attention_heads=9,
                  num_key_value_heads=3,
                  hidden_size=576,
                  tie_word_embeddings=True,
                  intermediate_size=1440),

    "small_old": dict(num_hidden_layers=12,
                      num_attention_heads=12,
                      num_key_value_heads=3,
                      hidden_size=768,
                      tie_word_embeddings=False),

    # Bert-Large
    "medium": dict(num_hidden_layers=32,
                   num_attention_heads=15,
                   num_key_value_heads=5,
                   hidden_size=960,
                   tie_word_embeddings=True,
                   intermediate_size=2400),

    "medium_old": dict(num_hidden_layers=24,
                       num_attention_heads=16,
                       num_key_value_heads=4,
                       hidden_size=1024,
                       tie_word_embeddings=False),

    # GPT-2
    "xl": dict(num_hidden_layers=48,
               num_attention_heads=24,
               num_key_value_heads=8,
               hidden_size=1536,
               tie_word_embeddings=True,
               intermediate_size=3840),  # GPT-2

    "xl_old": dict(num_hidden_layers=24,
                   num_attention_heads=32,
                   num_key_value_heads=4,
                   hidden_size=2048,
                   tie_word_embeddings=False),

}


def find_multiple(n: int, k: int) -> int:
    if n % k == 0:
        return n
    return n + k - (n % k)