| from transformers import PretrainedConfig | |
| class PersadianNanoV4Config(PretrainedConfig): | |
| model_type = "persadian_nano_v4" | |
| def __init__( | |
| self, | |
| hidden_size=512, | |
| intermediate_size=1024, | |
| num_hidden_layers=12, | |
| num_attention_heads=8, | |
| num_key_value_heads=4, | |
| num_experts=4, | |
| num_experts_per_tok=2, | |
| progressive_experts=True, | |
| min_experts=1, | |
| max_experts=2, | |
| use_hyper_connection=True, | |
| use_compressed_attention=True, | |
| compress_ratio=4, | |
| vocab_size=50257, | |
| max_position_embeddings=2048, | |
| bos_token_id=50256, | |
| eos_token_id=50256, | |
| rope_theta=10000.0, | |
| layer_norm_eps=1e-5, | |
| dropout=0.1, | |
| attention_dropout=0.0, | |
| use_flash_attention=True, | |
| **kwargs | |
| ): | |
| super().__init__(**kwargs) | |
| self.hidden_size = hidden_size | |
| self.intermediate_size = intermediate_size | |
| self.num_hidden_layers = num_hidden_layers | |
| self.num_attention_heads = num_attention_heads | |
| self.num_key_value_heads = num_key_value_heads | |
| self.num_experts = num_experts | |
| self.num_experts_per_tok = num_experts_per_tok | |
| self.progressive_experts = progressive_experts | |
| self.min_experts = min_experts | |
| self.max_experts = max_experts | |
| self.use_hyper_connection = use_hyper_connection | |
| self.use_compressed_attention = use_compressed_attention | |
| self.compress_ratio = compress_ratio | |
| self.vocab_size = vocab_size | |
| self.max_position_embeddings = max_position_embeddings | |
| self.bos_token_id = bos_token_id | |
| self.eos_token_id = eos_token_id | |
| self.rope_theta = rope_theta | |
| self.layer_norm_eps = layer_norm_eps | |
| self.dropout = dropout | |
| self.attention_dropout = attention_dropout | |
| self.use_flash_attention = use_flash_attention | |