| from transformers import PretrainedConfig |
|
|
| class PebbleConfig(PretrainedConfig): |
| model_type = "pebble_10m" |
|
|
| def __init__( |
| self, |
| vocab_size=2048, |
| hidden_size=384, |
| intermediate_size=1536, |
| num_hidden_layers=8, |
| num_attention_heads=6, |
| block_pattern="mmma|mmma", |
| hybrid_ratio="3:1 mamba2:attention", |
| max_position_embeddings=512, |
| rms_norm_eps=1e-6, |
| tie_word_embeddings=True, |
| mamba2=None, |
| attention=None, |
| **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.block_pattern = block_pattern |
| self.hybrid_ratio = hybrid_ratio |
| self.max_position_embeddings = max_position_embeddings |
| self.rms_norm_eps = rms_norm_eps |
| self.tie_word_embeddings = tie_word_embeddings |
| |
| |
| self.mamba2 = mamba2 or { |
| "d_state": 128, "d_conv": 4, "expand": 2, |
| "headdim": 96, "use_mem_eff_path": True |
| } |
| self.attention = attention or { |
| "rope_theta": 10000.0, "is_causal": True |
| } |
| super().__init__(**kwargs) |
