| |
| |
| |
|
|
| from transformers.configuration_utils import PretrainedConfig |
|
|
|
|
| class ScatterbrainConfig(PretrainedConfig): |
| r""" |
| Configuration class for Scatterbrain - a model that loops a single MoE layer multiple times. |
| |
| Args: |
| vocab_size (`int`, *optional*, defaults to 151936): |
| Vocabulary size of the model. |
| hidden_size (`int`, *optional*, defaults to 2048): |
| Dimension of the hidden representations. |
| intermediate_size (`int`, *optional*, defaults to 6144): |
| Dimension of the MLP intermediate representations (for dense layers). |
| moe_intermediate_size (`int`, *optional*, defaults to 768): |
| Dimension of the MLP intermediate representations for each expert. |
| num_hidden_layers (`int`, *optional*, defaults to 1): |
| Number of physical decoder layers. For Scatterbrain this is always 1. |
| num_loop_iterations (`int`, *optional*, defaults to 30): |
| Number of times to loop through the single layer. |
| num_attention_heads (`int`, *optional*, defaults to 32): |
| Number of attention heads for each attention layer. |
| num_key_value_heads (`int`, *optional*, defaults to 4): |
| Number of key_value heads for Grouped Query Attention. |
| head_dim (`int`, *optional*, defaults to 128): |
| Dimension of each attention head. |
| hidden_act (`str`, *optional*, defaults to `"silu"`): |
| Activation function in the MLP. |
| max_position_embeddings (`int`, *optional*, defaults to 262144): |
| Maximum sequence length. |
| initializer_range (`float`, *optional*, defaults to 0.02): |
| Standard deviation for weight initialization. |
| rms_norm_eps (`float`, *optional*, defaults to 1e-6): |
| Epsilon for RMS normalization. |
| use_cache (`bool`, *optional*, defaults to `True`): |
| Whether to use KV cache. |
| tie_word_embeddings (`bool`, *optional*, defaults to `False`): |
| Whether to tie input and output embeddings. |
| rope_theta (`float`, *optional*, defaults to 10000000.0): |
| Base period for rotary embeddings. |
| attention_dropout (`float`, *optional*, defaults to 0.0): |
| Dropout ratio for attention weights. |
| num_experts (`int`, *optional*, defaults to 128): |
| Number of experts in the MoE layer. |
| num_experts_per_tok (`int`, *optional*, defaults to 8): |
| Number of experts activated per token. |
| norm_topk_prob (`bool`, *optional*, defaults to `True`): |
| Whether to normalize top-k routing probabilities. |
| router_aux_loss_coef (`float`, *optional*, defaults to 0.001): |
| Coefficient for router auxiliary loss. |
| """ |
|
|
| model_type = "scatterbrain" |
| keys_to_ignore_at_inference = ["past_key_values"] |
|
|
| def __init__( |
| self, |
| vocab_size=151936, |
| hidden_size=2048, |
| intermediate_size=6144, |
| moe_intermediate_size=768, |
| num_hidden_layers=1, |
| num_loop_iterations=30, |
| num_attention_heads=32, |
| num_key_value_heads=4, |
| head_dim=128, |
| hidden_act="silu", |
| max_position_embeddings=262144, |
| initializer_range=0.02, |
| rms_norm_eps=1e-6, |
| use_cache=True, |
| tie_word_embeddings=False, |
| rope_theta=10000000.0, |
| rope_scaling=None, |
| attention_dropout=0.0, |
| attention_bias=False, |
| num_experts=128, |
| num_experts_per_tok=8, |
| norm_topk_prob=True, |
| output_router_logits=False, |
| router_aux_loss_coef=0.001, |
| decoder_sparse_step=1, |
| mlp_only_layers=None, |
| sliding_window=None, |
| use_sliding_window=False, |
| max_window_layers=48, |
| pad_token_id=None, |
| bos_token_id=151643, |
| eos_token_id=151645, |
| **kwargs, |
| ): |
| self.vocab_size = vocab_size |
| self.hidden_size = hidden_size |
| self.intermediate_size = intermediate_size |
| self.moe_intermediate_size = moe_intermediate_size |
| self.num_hidden_layers = num_hidden_layers |
| self.num_loop_iterations = num_loop_iterations |
| self.num_attention_heads = num_attention_heads |
| self.num_key_value_heads = num_key_value_heads |
| self.head_dim = head_dim |
| self.hidden_act = hidden_act |
| self.max_position_embeddings = max_position_embeddings |
| self.initializer_range = initializer_range |
| self.rms_norm_eps = rms_norm_eps |
| self.use_cache = use_cache |
| self.rope_theta = rope_theta |
| self.rope_scaling = rope_scaling |
| self.attention_dropout = attention_dropout |
| self.attention_bias = attention_bias |
| self.num_experts = num_experts |
| self.num_experts_per_tok = num_experts_per_tok |
| self.norm_topk_prob = norm_topk_prob |
| self.output_router_logits = output_router_logits |
| self.router_aux_loss_coef = router_aux_loss_coef |
| self.decoder_sparse_step = decoder_sparse_step |
| self.mlp_only_layers = mlp_only_layers if mlp_only_layers is not None else [] |
| self.sliding_window = sliding_window |
| self.use_sliding_window = use_sliding_window |
| self.max_window_layers = max_window_layers |
|
|
| 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, |
| ) |
|
|
