| from transformers import PretrainedConfig |
|
|
|
|
| class SlidingWindowConfig(PretrainedConfig): |
| model_type = "sliding_window" |
| |
| def __init__( |
| self, |
| vocab_size=50304, |
| hidden_size=768, |
| intermediate_size=None, |
| hidden_ratio=4, |
| num_hidden_layers=12, |
| num_heads=12, |
| num_kv_heads=None, |
| hidden_act="swish", |
| max_position_embeddings=2048, |
| initializer_range=0.02, |
| norm_eps=1e-6, |
| use_cache=True, |
| pad_token_id=None, |
| bos_token_id=1, |
| eos_token_id=2, |
| tie_word_embeddings=False, |
| attention_bias=False, |
| fuse_norm=True, |
| fuse_cross_entropy=True, |
| use_rope=False, |
| |
| window_size=2, |
| qk_norm=False, |
| qk_norm_share_param_across_head=False, |
| use_k_shift=False, |
| use_v_shift=False, |
| elementwise_affine=True, |
| **kwargs, |
| ): |
| self.vocab_size = vocab_size |
| self.hidden_size = hidden_size |
| self.intermediate_size = intermediate_size or hidden_ratio * hidden_size |
| self.hidden_ratio = hidden_ratio |
| self.num_hidden_layers = num_hidden_layers |
| self.num_heads = num_heads |
| self.num_kv_heads = num_kv_heads |
| self.hidden_act = hidden_act |
| self.max_position_embeddings = max_position_embeddings |
| self.initializer_range = initializer_range |
| self.norm_eps = norm_eps |
| self.use_cache = use_cache |
| self.tie_word_embeddings = tie_word_embeddings |
| self.attention_bias = attention_bias |
| self.fuse_norm = fuse_norm |
| self.fuse_cross_entropy = fuse_cross_entropy |
| self.use_rope = use_rope |
| |
| |
| self.window_size = window_size |
| |
| self.qk_norm = qk_norm |
| self.qk_norm_share_param_across_head = qk_norm_share_param_across_head |
| self.use_k_shift = use_k_shift |
| self.use_v_shift = use_v_shift |
| self.elementwise_affine = elementwise_affine |
|
|
| super().__init__( |
| pad_token_id=pad_token_id, |
| bos_token_id=bos_token_id, |
| eos_token_id=eos_token_id, |
| **kwargs, |
| ) |
