58 lines
1.6 KiB
Python
58 lines
1.6 KiB
Python
|
from transformers import PretrainedConfig
|
||
|
|
|
||
|
|
|
||
|
|
class VeyraConfig(PretrainedConfig):
|
||
|
|
model_type = "veyra"
|
||
|
|
|
||
|
|
def __init__(
|
||
|
|
self,
|
||
|
|
vocab_size=8192,
|
||
|
|
d_model=512,
|
||
|
|
hidden_size=512,
|
||
|
|
n_q_heads=8,
|
||
|
|
num_attention_heads=8,
|
||
|
|
n_kv_heads=2,
|
||
|
|
num_key_value_heads=2,
|
||
|
|
intermediate_size=2048,
|
||
|
|
max_seq_len=1024,
|
||
|
|
max_position_embeddings=1024,
|
||
|
|
layer_pattern=("A", "M", "A", "M", "A", "M", "A", "M"),
|
||
|
|
num_hidden_layers=8,
|
||
|
|
rms_norm_eps=1e-6,
|
||
|
|
rope_theta=10000.0,
|
||
|
|
tie_word_embeddings=True,
|
||
|
|
bos_token_id=0,
|
||
|
|
eos_token_id=1,
|
||
|
|
pad_token_id=2,
|
||
|
|
unk_token_id=3,
|
||
|
|
**kwargs,
|
||
|
|
):
|
||
|
|
self.vocab_size = vocab_size
|
||
|
|
self.d_model = d_model
|
||
|
|
self.hidden_size = hidden_size
|
||
|
|
|
||
|
|
self.n_q_heads = n_q_heads
|
||
|
|
self.num_attention_heads = num_attention_heads
|
||
|
|
self.n_kv_heads = n_kv_heads
|
||
|
|
self.num_key_value_heads = num_key_value_heads
|
||
|
|
|
||
|
|
self.intermediate_size = intermediate_size
|
||
|
|
self.max_seq_len = max_seq_len
|
||
|
|
self.max_position_embeddings = max_position_embeddings
|
||
|
|
|
||
|
|
self.layer_pattern = list(layer_pattern)
|
||
|
|
self.num_hidden_layers = num_hidden_layers
|
||
|
|
|
||
|
|
self.rms_norm_eps = rms_norm_eps
|
||
|
|
self.rope_theta = rope_theta
|
||
|
|
self.tie_word_embeddings = tie_word_embeddings
|
||
|
|
|
||
|
|
super().__init__(
|
||
|
|
bos_token_id=bos_token_id,
|
||
|
|
eos_token_id=eos_token_id,
|
||
|
|
pad_token_id=pad_token_id,
|
||
|
|
unk_token_id=unk_token_id,
|
||
|
|
tie_word_embeddings=tie_word_embeddings,
|
||
|
|
**kwargs,
|
||
|
|
)
|