| |
| |
|
|
| import typing |
|
|
| from transformers.configuration_utils import PreTrainedConfig |
| from transformers.modeling_rope_utils import RopeParameters |
|
|
|
|
| class DomynEdgeConfig(PreTrainedConfig): |
| """Configuration of a DomynEdge model, with per-layer position encoding, sliding window and head size.""" |
|
|
| model_type: typing.ClassVar[str] = "domynedge" |
| default_theta: float = 10_000 |
|
|
| def __init__( |
| self, |
| vocab_size: int = 115269, |
| hidden_size: int = 2560, |
| intermediate_size: int = 9728, |
| num_hidden_layers: int = 36, |
| num_attention_heads: int = 32, |
| num_key_value_heads: int = 8, |
| max_position_embeddings: int = 8192, |
| initializer_range: float = 0.02, |
| rms_norm_eps: float = 1e-6, |
| *, |
| use_cache: bool = True, |
| tie_word_embeddings: bool = True, |
| rope_parameters: RopeParameters | None = None, |
| max_window_layers: int | None = None, |
| attention_bias: bool = False, |
| attention_dropout: float = 0.0, |
| position_layers: list[str] = ["nope", "nope"] + ["rope", "rope", "rope", "nope"] * 8 + ["nope", "nope"], |
| layer_types: list[str] = ["full_attention", "full_attention"] |
| + ["sliding_attention", "sliding_attention", "sliding_attention", "full_attention"] * 8 |
| + ["full_attention", "full_attention"], |
| sliding_window: int | None = 1025, |
| head_sizes: list[int] = [256, 256] + [64, 64, 64, 256] * 8 + [256, 256], |
| pad_token_id: int | None = None, |
| bos_token_id: int | None = None, |
| eos_token_id: int | None = None, |
| **kwargs, |
| ): |
| """Initialize and validate the per-layer settings, which must all have `num_hidden_layers` entries.""" |
| self.vocab_size = vocab_size |
| self.max_position_embeddings = max_position_embeddings |
| 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.max_window_layers = num_hidden_layers if max_window_layers is None else max_window_layers |
|
|
| |
| if num_key_value_heads is None: |
| num_key_value_heads = num_attention_heads |
|
|
| self.num_key_value_heads = num_key_value_heads |
| self.initializer_range = initializer_range |
| self.rms_norm_eps = rms_norm_eps |
| self.use_cache = use_cache |
| self.attention_bias = attention_bias |
| self.attention_dropout = attention_dropout |
|
|
| self.position_layers = position_layers |
| self.layer_types = layer_types |
| self.sliding_window = sliding_window |
| self.head_sizes = head_sizes |
|
|
| |
| if len(self.position_layers) != self.num_hidden_layers: |
| msg = ( |
| f"Length of position_layers ({len(self.position_layers)}) " |
| f"must be equal to num_hidden_layers ({self.num_hidden_layers})" |
| ) |
| raise ValueError(msg) |
| if any(pl not in ["rope", "nope"] for pl in self.position_layers): |
| msg = f"position_layers must be one of 'rope', 'nope'. Got {self.position_layers}" |
| raise ValueError(msg) |
| |
| if len(self.layer_types) != self.num_hidden_layers: |
| msg = ( |
| f"Length of layer_types ({len(self.layer_types)}) " |
| f"must be equal to num_hidden_layers ({self.num_hidden_layers})" |
| ) |
| raise ValueError(msg) |
| if "sliding_attention" in self.layer_types and self.sliding_window is None: |
| msg = "sliding_window must be set when layer_types contains 'sliding_attention'" |
| raise ValueError(msg) |
| |
| if len(head_sizes) != self.num_hidden_layers: |
| msg = ( |
| f"Length of head_sizes ({len(head_sizes)}) " |
| f"must be equal to num_hidden_layers ({self.num_hidden_layers})" |
| ) |
| raise ValueError(msg) |
| if any(not isinstance(hs, int) for hs in head_sizes): |
| msg = f"head_sizes must be a list of integers. Got {head_sizes}" |
| raise ValueError(msg) |
| |
| rope_dims = [head_size for rope, head_size in zip(position_layers, head_sizes, strict=True) if rope == "rope"] |
| if len(set(rope_dims)) > 1: |
| msg = f"All rope layers must have the same head size. Got {rope_dims}" |
| raise ValueError(msg) |
|
|
| kwargs.pop("rope_dim", None) |
| kwargs.pop("window_sizes", None) |
| self.rope_dim = rope_dims[0] |
|
|
| |
| |
| |
| self.rope_parameters = rope_parameters |
|
|
| self.pad_token_id = pad_token_id |
| self.bos_token_id = bos_token_id |
| self.eos_token_id = eos_token_id |
| self.tie_word_embeddings = tie_word_embeddings |
|
|
| if "per_layer_config" not in kwargs: |
| kwargs["per_layer_config"] = { |
| layer_idx: {"head_dim": head_size} for layer_idx, head_size in enumerate(head_sizes) |
| } |
|
|
| super().__init__(**kwargs) |
|
|
| @property |
| def window_sizes(self) -> list[tuple[int, int]]: |
| """The torchtitan (left, right) attention window of each layer, (-1, 0) for full attention. |
| |
| Legacy implementation to be compatible with vLLM v0.33, to be removed when updated from vLLM side. |
| """ |
| return [ |
| (self.sliding_window - 1, 0) if layer_type == "sliding_attention" and self.sliding_window else (-1, 0) |
| for layer_type in self.layer_types |
| ] |
|
|
|
|
| __all__ = ["DomynEdgeConfig"] |
|
|
