| from transformers.configuration_utils import PretrainedConfig |
|
|
|
|
| class Grok1Config(PretrainedConfig): |
| model_type = "grok-1" |
| keys_to_ignore_at_inference = ["past_key_values"] |
|
|
| def __init__( |
| self, |
| vocab_size=32000, |
| hidden_size=4096, |
| intermediate_size=32768, |
| num_hidden_layers=32, |
| num_attention_heads=32, |
| num_key_value_heads=32, |
| attn_output_multiplier=1.0, |
| max_attn_value=1.0, |
| max_position_embeddings=4096, |
| embedding_multiplier_scale: float = 1.0, |
| output_multiplier_scale: float = 1.0, |
| rms_norm_eps=1e-5, |
| use_cache=True, |
| pad_token_id=None, |
| bos_token_id=1, |
| eos_token_id=2, |
| tie_word_embeddings=True, |
| num_experts_per_tok=2, |
| num_experts=8, |
| output_router_logits=False, |
| router_aux_loss_coef=0.001, |
| **kwargs |
| ): |
| self.vocab_size = vocab_size |
| self.attn_output_multiplier = attn_output_multiplier |
| self.max_attn_value = max_attn_value |
| self.max_position_embeddings = max_position_embeddings |
| self.embedding_multiplier_scale = embedding_multiplier_scale |
| self.output_multiplier_scale = output_multiplier_scale |
| self.hidden_size = hidden_size |
| self.intermediate_size = intermediate_size |
| self.num_hidden_layers = num_hidden_layers |
| self.num_attention_heads = num_attention_heads |
|
|
| |
| if num_key_value_heads is None: |
| num_key_value_heads = num_attention_heads |
|
|
| self.num_key_value_heads = num_key_value_heads |
| self.rms_norm_eps = rms_norm_eps |
| self.use_cache = use_cache |
|
|
| self.num_experts_per_tok = num_experts_per_tok |
| self.num_experts = num_experts |
| self.output_router_logits = output_router_logits |
| self.router_aux_loss_coef = router_aux_loss_coef |
| super().__init__( |
| pad_token_id=pad_token_id, |
| bos_token_id=bos_token_id, |
| eos_token_id=eos_token_id, |
| tie_word_embeddings=tie_word_embeddings, |
| **kwargs, |
| ) |
|
|
