| """ Bailing MoE model configuration """ |
|
|
| from transformers.configuration_utils import PretrainedConfig |
|
|
|
|
| class BailingMoeLinearConfig(PretrainedConfig): |
| model_type = "bailing_moe_linear" |
|
|
| 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, |
| use_bias=True, |
| rms_norm_eps=1e-05, |
| norm_head=False, |
| tie_word_embeddings=False, |
| 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, |
| norm_topk_prob=True, |
| moe_intermediate_size=None, |
| first_k_dense_replace=0, |
| head_dim=None, |
| output_router_logits=False, |
| layer_group_size=1, |
| use_linear_silu=False, |
| linear_rope=True, |
| use_linear_gqa=False, |
| use_low_rank=False, |
| rotary_type='full-1d', |
| linear_mode='chunk', |
| **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 |
|
|
| |
| self.num_experts = num_experts |
| self.num_shared_experts = num_shared_experts |
| self.num_experts_per_tok = num_experts_per_tok |
| self.norm_topk_prob = norm_topk_prob |
| self.moe_intermediate_size = moe_intermediate_size |
| self.first_k_dense_replace = first_k_dense_replace |
| self.output_router_logits = output_router_logits |
|
|
| |
| self.layer_group_size = layer_group_size |
| self.use_linear_silu = use_linear_silu |
| self.linear_rope = linear_rope |
| self.use_linear_gqa = use_linear_gqa |
| self.use_low_rank = use_low_rank |
| self.rotary_type = rotary_type |
| self.linear_mode = linear_mode |
|
|
| super().__init__(pad_token_id=pad_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs) |
|
|
