# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/bailing_hybrid/modular_bailing_hybrid.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_bailing_hybrid.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # Copyright 2026 The HuggingFace Inc. team # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from huggingface_hub.dataclasses import strict from ...configuration_utils import PreTrainedConfig from ...modeling_rope_utils import RopeParameters from ...utils import auto_docstring @auto_docstring(checkpoint="inclusionAI/Ling-3.0-Flash") @strict class BailingHybridConfig(PreTrainedConfig): r""" n_group (`int`, *optional*, defaults to 8): Number of expert groups used by the group-limited router. first_k_dense_replace (`int`, *optional*, defaults to 2): Number of initial decoder layers that use a dense MLP instead of routed experts. num_mtp_layers (`int`, *optional*, defaults to 1): Number of Multi-Token Prediction (MTP) modules available to append after the base transformer model. When `0`, the model behaves as a standard decoder. When `>0`, each extra module can predict one additional future token at inference time (speculative decoding via `generate(..., use_mtp=True)`). rope_theta (`float`, *optional*, defaults to 6000000.0): Base period of the rotary position embeddings used by MLA layers. short_conv_kernel_size (`int`, *optional*, defaults to 4): Kernel size of the short depthwise convolutions in KDA layers. layer_group_size (`int`, *optional*, defaults to 6): Number of layers in each hybrid attention group. The last layer in every group uses full MLA and the other layers use KDA. kda_safe_gate (`bool`, *optional*, defaults to `True`): Whether to use the bounded KDA forget gate. kda_lower_bound (`float`, *optional*, defaults to -5.0): Lower bound of the KDA forget gate in log space. mlp_layer_types (`list[str]`, *optional*): Per-layer MLP types. Derived from `first_k_dense_replace` when omitted. mtp_loss_scaling_factor (`float`, *optional*, defaults to 0.0): Scaling factor used for the auxiliary multi-token-prediction loss during pretraining. """ model_type = "bailing_hybrid" keys_to_ignore_at_inference = ["past_key_values"] base_model_tp_plan = { "layers.*.mlp.experts.gate_up_proj": "packed_colwise", "layers.*.mlp.experts.down_proj": "rowwise", "layers.*.mlp.experts": "moe_tp_experts", "layers.*.mlp.shared_experts.gate_proj": "colwise", "layers.*.mlp.shared_experts.up_proj": "colwise", "layers.*.mlp.shared_experts.down_proj": "rowwise", "layers.*.mlp.gate_proj": "colwise", "layers.*.mlp.up_proj": "colwise", "layers.*.mlp.down_proj": "rowwise", } base_model_pp_plan = { "embed_tokens": (["input_ids"], ["inputs_embeds"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } base_model_ep_plan = { "layers.*.mlp.experts.gate_up_proj": "grouped_gemm", "layers.*.mlp.experts.down_proj": "grouped_gemm", "layers.*.mlp.experts": "ep_dispatch_experts", } attribute_map = { "max_position_embeddings": "model_max_length", "norm_topk_prob": "moe_renormalize", "n_group": "num_expert_group", "num_local_experts": "num_experts", "num_experts_per_tok": "num_experts_per_token", "n_shared_experts": "num_shared_experts", "num_mtp_layers": "num_nextn_predict_layers", } vocab_size: int = 157184 hidden_size: int = 2560 intermediate_size: int = 6144 moe_intermediate_size: int = 768 num_hidden_layers: int = 42 num_attention_heads: int = 32 num_key_value_heads: int | None = 32 n_shared_experts: int = 1 output_router_logits: bool = False routed_scaling_factor: float = 2.5 kv_lora_rank: int = 512 qk_rope_head_dim: int = 64 v_head_dim: int = 128 qk_nope_head_dim: int = 128 n_group: int = 8 topk_group: int = 4 num_experts_per_tok: int = 8 first_k_dense_replace: int = 2 norm_topk_prob: bool = True hidden_act: str = "silu" max_position_embeddings: int = 262144 initializer_range: float = 0.02 rms_norm_eps: float = 1e-6 use_cache: bool = True pad_token_id: int | None = 156892 bos_token_id: int | None = None eos_token_id: int | list[int] | None = 156895 pretraining_tp: int | None = 1 tie_word_embeddings: bool = False rope_parameters: RopeParameters | dict | None = None attention_bias: bool = False attention_dropout: float = 0.0 num_mtp_layers: int = 1 num_local_experts: int = 512 rope_theta: float | int = 6_000_000.0 head_dim: int = 128 short_conv_kernel_size: int = 4 layer_group_size: int = 6 kda_safe_gate: bool = True kda_lower_bound: float = -5.0 layer_types: list[str] | None = None mlp_layer_types: list[str] | None = None mtp_loss_scaling_factor: float | int = 0.0 def __post_init__(self, **kwargs): self.linear_head_dim = self.head_dim self.linear_num_heads = self.num_attention_heads self.linear_conv_kernel_dim = self.short_conv_kernel_size if self.layer_types is None: self.layer_types = [ "full_attention" if (layer_idx + 1) % self.layer_group_size == 0 else "linear_attention" for layer_idx in range(self.num_hidden_layers) ] if self.mlp_layer_types is None: self.mlp_layer_types = [ "dense" if layer_idx < self.first_k_dense_replace else "sparse" for layer_idx in range(self.num_hidden_layers) ] if self.rope_parameters is None: self.rope_parameters = {"rope_type": "default", "rope_theta": self.rope_theta} if self.num_key_value_heads is None: self.num_key_value_heads = self.num_attention_heads self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim self.head_dim = self.qk_rope_head_dim super().__post_init__(**kwargs) __all__ = ["BailingHybridConfig"]