|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
"""Boost model configuration""" |
|
|
|
from transformers.configuration_utils import PretrainedConfig |
|
from transformers.utils import logging |
|
|
|
|
|
logger = logging.get_logger(__name__) |
|
|
|
|
|
class BoostConfig(PretrainedConfig): |
|
|
|
model_type = "boost" |
|
keys_to_ignore_at_inference = ["past_key_values"] |
|
|
|
def __init__( |
|
self, |
|
vocab_size=151936, |
|
hidden_size=4096, |
|
intermediate_size=22016, |
|
num_hidden_layers=32, |
|
num_attention_heads=32, |
|
num_key_value_heads=32, |
|
hidden_act="silu", |
|
max_position_embeddings=32768, |
|
initializer_range=0.02, |
|
rms_norm_eps=1e-6, |
|
use_cache=True, |
|
tie_word_embeddings=False, |
|
rope_theta=10000.0, |
|
use_sliding_window=False, |
|
sliding_window=4096, |
|
max_window_layers=28, |
|
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.num_hidden_layers = num_hidden_layers |
|
self.num_attention_heads = num_attention_heads |
|
self.use_sliding_window = use_sliding_window |
|
self.sliding_window = sliding_window |
|
self.max_window_layers = max_window_layers |
|
|
|
|
|
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.use_cache = use_cache |
|
self.rope_theta = rope_theta |
|
self.attention_dropout = attention_dropout |
|
|
|
super().__init__( |
|
tie_word_embeddings=tie_word_embeddings, |
|
**kwargs, |
|
) |