File size: 3,545 Bytes
9fda39c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | """LLaDA2 MoE model configuration"""
from transformers.configuration_utils import PretrainedConfig
class LLaDA2MoeConfig(PretrainedConfig):
model_type = "llada2_moe"
def __init__(
self,
vocab_size=30592,
hidden_size=1024,
intermediate_size=None,
num_hidden_layers=24,
num_attention_heads=16,
num_key_value_heads=0,
hidden_act="silu",
use_qkv_bias=False, # llada2 only
use_qk_norm=False,
use_bias=True, # llada2 only
rms_norm_eps=1e-05,
norm_head=False, # llada2 only
tie_word_embeddings=False, # PretrainedConfig key, here change default value.
embedding_dropout=0.1,
attention_dropout=0.1,
output_dropout=0.1,
initializer_range=0.02,
max_position_embeddings=16384,
rope_theta=10000.0,
use_cache=True,
use_sliding_window=False,
sliding_window=4096,
max_window_layers=28,
rope_scaling=None,
pad_token_id=126081,
num_experts=16,
num_shared_experts=0,
num_experts_per_tok=2,
n_group=8,
topk_group=4,
routed_scaling_factor=2.5,
moe_intermediate_size=None,
first_k_dense_replace=0,
head_dim=None,
output_router_logits=False,
partial_rotary_factor=0.5,
# @sicheng: add more parameters for coreset selection
mode='aggregate', # aggregate
core_size=18, # coreset upper bound
core_threshold=0.6, # coreset selection threshold
**kwargs,
):
self.num_hidden_layers = num_hidden_layers
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_attention_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.use_qkv_bias = use_qkv_bias
self.use_bias = use_bias
self.norm_head = norm_head
self.rms_norm_eps = rms_norm_eps
self.embedding_dropout = embedding_dropout
self.attention_dropout = attention_dropout
self.output_dropout = output_dropout
self.initializer_range = initializer_range
self.max_position_embeddings = max_position_embeddings
self.rope_theta = rope_theta
self.use_cache = use_cache
self.use_sliding_window = use_sliding_window
self.sliding_window = sliding_window
self.max_window_layers = max_window_layers
self.head_dim = head_dim or self.hidden_size // self.num_attention_heads
self.rope_scaling = rope_scaling
# MoE configs
self.num_experts = num_experts
self.num_shared_experts = num_shared_experts
self.num_experts_per_tok = num_experts_per_tok
self.n_group = n_group
self.topk_group = topk_group
self.moe_intermediate_size = moe_intermediate_size
self.first_k_dense_replace = first_k_dense_replace
self.output_router_logits = output_router_logits
self.routed_scaling_factor = routed_scaling_factor
self.partial_rotary_factor = partial_rotary_factor
# @sicheng: add more tokens for coreset selection implementation
self.mode = mode # aggregate
self.core_size = core_size # coreset upper bound
self.core_threshold = core_threshold # coreset selection threshold
super().__init__(pad_token_id=pad_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs)
|