|
from transformers.configuration_utils import PretrainedConfig |
|
|
|
|
|
class Grok1Config(PretrainedConfig): |
|
model_type = "grok-1" |
|
keys_to_ignore_at_inference = ["past_key_values"] |
|
|
|
def __init__( |
|
self, |
|
vocab_size=32000, |
|
hidden_size=4096, |
|
intermediate_size=32768, |
|
num_hidden_layers=32, |
|
num_attention_heads=32, |
|
num_key_value_heads=32, |
|
attn_output_multiplier=1.0, |
|
max_attn_value=1.0, |
|
max_position_embeddings=4096, |
|
embedding_multiplier_scale: float = 1.0, |
|
output_multiplier_scale: float = 1.0, |
|
rms_norm_eps=1e-5, |
|
use_cache=True, |
|
pad_token_id=None, |
|
bos_token_id=1, |
|
eos_token_id=2, |
|
tie_word_embeddings=True, |
|
num_experts_per_tok=2, |
|
num_experts=8, |
|
output_router_logits=False, |
|
router_aux_loss_coef=0.001, |
|
**kwargs |
|
): |
|
self.vocab_size = vocab_size |
|
self.attn_output_multiplier = attn_output_multiplier |
|
self.max_attn_value = max_attn_value |
|
self.max_position_embeddings = max_position_embeddings |
|
self.embedding_multiplier_scale = embedding_multiplier_scale |
|
self.output_multiplier_scale = output_multiplier_scale |
|
self.hidden_size = hidden_size |
|
self.intermediate_size = intermediate_size |
|
self.num_hidden_layers = num_hidden_layers |
|
self.num_attention_heads = num_attention_heads |
|
|
|
|
|
if num_key_value_heads is None: |
|
num_key_value_heads = num_attention_heads |
|
|
|
self.num_key_value_heads = num_key_value_heads |
|
self.rms_norm_eps = rms_norm_eps |
|
self.use_cache = use_cache |
|
|
|
self.num_experts_per_tok = num_experts_per_tok |
|
self.num_experts = num_experts |
|
self.output_router_logits = output_router_logits |
|
self.router_aux_loss_coef = router_aux_loss_coef |
|
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, |
|
) |
|
|
