|
|
|
|
|
|
|
from transformers import BertConfig as TransformersBertConfig |
|
|
|
|
|
class JBertConfig(TransformersBertConfig): |
|
def __init__( |
|
self, |
|
model_max_length: int = 8192, |
|
attention_probs_dropout_prob: float = 0.0, |
|
**kwargs, |
|
): |
|
"""Configuration class for MosaicBert. |
|
|
|
Args: |
|
model_max_length (int): Use `model_max_length` to determine how large of an alibi tensor to |
|
create when initializing the model. You should be able to ignore this parameter in most cases. |
|
Defaults to 8192. |
|
attention_probs_dropout_prob (float): By default, turn off attention dropout in Mosaic BERT |
|
(otherwise, Flash Attention will be off by default). Defaults to 0.0. |
|
""" |
|
super().__init__( |
|
attention_probs_dropout_prob=attention_probs_dropout_prob, **kwargs |
|
) |
|
self.model_max_length = model_max_length |
|
|