# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/kimi_linear/modular_kimi_linear.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_kimi_linear.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 ...utils import auto_docstring @auto_docstring(checkpoint="moonshotai/Kimi-Linear-48B-A3B-Instruct") @strict class KimiLinearConfig(PreTrainedConfig): r""" n_group (`int`, *optional*, defaults to 8): Number of groups for routed experts. mlp_layer_types (`list[str]`, *optional*): List of layer types for the MLP or MoE layers. Defaults to None. linear_head_dim (`int`, *optional*): Dimension of each head in linear attention layers. Defaults to 128. linear_num_heads (`int`, *optional*): Number of heads for the linear attention layers. Defaults to 32. linear_conv_kernel_dim (`int`, *optional*, defaults to 4): Kernel size for the short convolution applied to queries, keys, and values in linear attention layers. """ model_type = "kimi_linear" 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", } vocab_size: int = 163840 hidden_size: int = 2304 intermediate_size: int = 9216 moe_intermediate_size: int = 1024 num_hidden_layers: int = 27 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 = 2.446 kv_lora_rank: int = 512 q_lora_rank: int | None = None qk_rope_head_dim: int = 64 v_head_dim: int | None = 128 qk_nope_head_dim: int = 128 n_group: int = 1 topk_group: int | None = 1 num_experts_per_tok: int | None = 8 norm_topk_prob: bool = True hidden_act: str = "silu" max_position_embeddings: int = 1048576 initializer_range: float = 0.02 rms_norm_eps: float = 1e-5 use_cache: bool = True pad_token_id: int | None = 163839 bos_token_id: int | None = 163584 eos_token_id: int | list[int] | None = 163586 pretraining_tp: int | None = 1 tie_word_embeddings: bool = False attention_bias: bool = False attention_dropout: float | int | None = 0.0 num_local_experts: int = 256 mlp_layer_types: list[str] | None = None layer_types: list[str] | None = None linear_head_dim: int = 128 linear_num_heads: int = 32 linear_conv_kernel_dim: int = 4 def __post_init__(self, **kwargs): 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) # Checkpoint stores linear attention attributes in a config sub-dict: if it's there, extract them linear_attn_config = kwargs.get("linear_attn_config", {}) self.linear_head_dim = linear_attn_config.get("head_dim", self.linear_head_dim) self.linear_num_heads = linear_attn_config.get("num_heads", self.linear_num_heads) self.linear_conv_kernel_dim = linear_attn_config.get("short_conv_kernel_size", self.linear_conv_kernel_dim) # For layer types, the precedence is: checkpoint config > layer types > default if self.layer_types is None: if "full_attn_layers" in linear_attn_config and "kda_layers" in linear_attn_config: layer_types = [None] * self.num_hidden_layers for layer in linear_attn_config["full_attn_layers"]: layer_types[layer - 1] = "full_attention" # types are 1-indexed in the checkpoint for layer in linear_attn_config["kda_layers"]: layer_types[layer - 1] = "linear_attention" self.layer_types = layer_types else: self.layer_types = [ "full_attention" if i and i % 4 == 0 else "linear_attention" for i in range(self.num_hidden_layers) ] # Same for MLP layer types, which indicate MLP or MoE if self.mlp_layer_types is None: first_k_dense_replace = kwargs.get("first_k_dense_replace", 1) self.mlp_layer_types = [ "dense" if i < first_k_dense_replace else "sparse" for i in range(self.num_hidden_layers) ] __all__ = ["KimiLinearConfig"]