| """HuggingFace configuration for the Rose X1 architecture.""" |
| from transformers import PretrainedConfig |
|
|
|
|
| class RoseX1Config(PretrainedConfig): |
| model_type = "rose_x1" |
|
|
| def __init__( |
| self, |
| vocab_size=32768, |
| hidden_size=512, |
| intermediate_size=1720, |
| num_hidden_layers=24, |
| num_attention_heads=8, |
| num_key_value_heads=2, |
| head_dim=None, |
| max_position_embeddings=2048, |
| hidden_act="silu", |
| rms_norm_eps=1e-5, |
| attention_bias=False, |
| mlp_bias=False, |
| attention_dropout=0.0, |
| tie_word_embeddings=True, |
| rope_theta=100000.0, |
| rope_scaling=None, |
| initializer_range=0.02, |
| use_cache=True, |
| |
| use_qk_norm=True, |
| refresh_gate_enabled=True, |
| refresh_gate_inject_layers=None, |
| refresh_gate_kernel_size=9, |
| **kwargs, |
| ): |
| self.vocab_size = vocab_size |
| 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.head_dim = head_dim if head_dim is not None else hidden_size // num_attention_heads |
| self.max_position_embeddings = max_position_embeddings |
| self.hidden_act = hidden_act |
| self.rms_norm_eps = rms_norm_eps |
| self.attention_bias = attention_bias |
| self.mlp_bias = mlp_bias |
| self.attention_dropout = attention_dropout |
| self.tie_word_embeddings = tie_word_embeddings |
| self.rope_theta = rope_theta |
| self.rope_scaling = rope_scaling |
| self.initializer_range = initializer_range |
| self.use_cache = use_cache |
| self.use_qk_norm = use_qk_norm |
| self.refresh_gate_enabled = refresh_gate_enabled |
| self.refresh_gate_inject_layers = ( |
| list(refresh_gate_inject_layers) if refresh_gate_inject_layers else [] |
| ) |
| self.refresh_gate_kernel_size = refresh_gate_kernel_size |
| super().__init__(**kwargs) |
