"""KLMConfig: Transformers config for KLM 1.5.""" from __future__ import annotations from transformers import PretrainedConfig class KLMConfig(PretrainedConfig): model_type = "klm" def __init__( self, vocab_size: int = 4160, hidden_size: int = 512, num_hidden_layers: int = 24, attention_every: int = 4, gdn_heads: int = 4, gdn_head_dim: int = 128, num_attention_heads: int = 8, num_key_value_heads: int = 4, intermediate_size: int = 1408, max_position_embeddings: int = 2048, rope_theta: float = 10000.0, n_bigrams: int = 125000, n_trigrams: int = 38000, ngram_dim: int = 32, rms_norm_eps: float = 1e-6, gdn_norm_eps: float = 1e-5, decay_floor: float = -20.0, bos_token_id: int = 1, eos_token_id: int = 2, pad_token_id: int = 0, tie_word_embeddings: bool = True, **kwargs, ): self.vocab_size = vocab_size self.hidden_size = hidden_size self.num_hidden_layers = num_hidden_layers self.attention_every = attention_every self.gdn_heads = gdn_heads self.gdn_head_dim = gdn_head_dim self.num_attention_heads = num_attention_heads self.num_key_value_heads = num_key_value_heads self.intermediate_size = intermediate_size self.max_position_embeddings = max_position_embeddings self.rope_theta = rope_theta self.n_bigrams = n_bigrams self.n_trigrams = n_trigrams self.ngram_dim = ngram_dim self.rms_norm_eps = rms_norm_eps self.gdn_norm_eps = gdn_norm_eps self.decay_floor = decay_floor super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, pad_token_id=pad_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs) def is_attention(self, layer: int) -> bool: return layer % self.attention_every == self.attention_every - 1