# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/axk2/modular_axk2.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_axk2.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # Copyright 2026 SK Telecom and the HuggingFace Inc. team. All rights reserved. # # 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 RotaryEmbeddingConfigMixin from ...utils import auto_docstring @auto_docstring(checkpoint="skt/A.X-K2") @strict class AXK2Config(PreTrainedConfig, RotaryEmbeddingConfigMixin): r""" n_group (`int`, *optional*): Number of expert groups for grouped routing, used by the larger A.X-K2 releases. `None` (the A.X-K2-Light default) routes over all experts without group restriction. topk_group (`int`, *optional*): Number of expert groups the top-k selection is restricted to when `n_group` is set. mlp_layer_types (`list`, *optional*): MLP type pattern for each layer (`"dense"` or `"sparse"`). Derived from the (legacy) kwargs `first_k_dense_replace` / `moe_layer_freq` when not provided. index_topk (`int`, *optional*, defaults to 2048): Number of top tokens selected by the indexer for sparse attention. index_head_dim (`int`, *optional*, defaults to 128): Head dimension for the indexer projections (DSA). index_n_heads (`int`, *optional*, defaults to 16): Number of heads for the indexer projections (DSA). gated_norm_rank (`int`, *optional*, defaults to 16): Bottleneck rank for the low-rank input-dependent gate used by `AXK2GatedRMSNorm`. The gate wraps `input_layernorm` on every layer and `post_attention_layernorm` on MoE layers. ```python >>> from transformers import AXK2Config, AXK2Model >>> # Initializing an A.X-K2 style configuration >>> configuration = AXK2Config() >>> # Initializing a model from the configuration >>> model = AXK2Model(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "axk2" keys_to_ignore_at_inference = ["past_key_values"] base_model_tp_plan = { "layers.*.self_attn.q_gate_proj": "colwise", "layers.*.self_attn.kv_a_proj_with_mqa": "mla_kv_a_proj", "layers.*.self_attn.kv_b_proj": "colwise", "layers.*.self_attn.o_proj": "rowwise", "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.gate": "ep_router", "layers.*.mlp.experts.gate_up_proj": "grouped_gemm", "layers.*.mlp.experts.down_proj": "grouped_gemm", "layers.*.mlp.experts": "moe_tp_experts", } attribute_map = {"num_local_experts": "n_routed_experts"} vocab_size: int = 163840 hidden_size: int = 2048 intermediate_size: int = 5120 moe_intermediate_size: int = 512 num_hidden_layers: int = 48 num_attention_heads: int = 32 num_key_value_heads: int = 32 n_shared_experts: int = 1 n_routed_experts: int = 128 routed_scaling_factor: float = 2.5 kv_lora_rank: int = 128 q_lora_rank: int = 384 qk_rope_head_dim: int = 32 v_head_dim: int = 64 qk_nope_head_dim: int = 64 # A.X-K2-Light routes without expert groups; the larger A.X-K2 releases set `n_group`/`topk_group` # for DeepSeek-V3-style grouped routing, so both modes are supported (`None` = non-grouped). n_group: int | None = None topk_group: int | None = None num_experts_per_tok: int = 8 norm_topk_prob: bool = True hidden_act: str = "silu" max_position_embeddings: int = 131072 initializer_range: float = 0.02 rms_norm_eps: float = 1e-6 use_cache: bool = True pad_token_id: int | None = None bos_token_id: int | None = 163691 eos_token_id: int | list[int] | None = 163691 tie_word_embeddings: bool = False rope_parameters: dict | None = None mlp_layer_types: list[str] | None = None attention_bias: bool = False attention_dropout: float | int = 0.0 index_topk: int = 2048 index_head_dim: int = 128 index_n_heads: int = 16 head_dim: int = 64 layer_types: list[str] | None = None gated_norm_rank: int = 16 def __post_init__(self, **kwargs): self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim # RoPE applies only to the rope slice, so `head_dim` points at it (the inherited rotary embedding # reads `config.head_dim`). self.head_dim = self.qk_rope_head_dim # Convert from legacy args to mlp layer types if self.mlp_layer_types is None: first_k_dense_replace = kwargs.pop("first_k_dense_replace", 1) moe_layer_freq = kwargs.pop("moe_layer_freq", 1) self.mlp_layer_types = [ "sparse" if i >= first_k_dense_replace and i % moe_layer_freq == 0 else "dense" for i in range(self.num_hidden_layers) ] # Indexer cache needed so DSA to indicate correct cache if self.layer_types is None: self.layer_types = ["deepseek_sparse_attention"] * self.num_hidden_layers super().__post_init__(**kwargs) def validate_architecture(self): super().validate_architecture() if self.q_lora_rank is None or self.q_lora_rank <= 0: raise ValueError( "A.X-K2 requires a positive `q_lora_rank` (the indexer and output gate read the query LoRA " f"bottleneck), got {self.q_lora_rank}." ) if (self.n_group is None) != (self.topk_group is None): raise ValueError( f"`n_group` and `topk_group` must be set together (both `None` for non-grouped routing), got " f"n_group={self.n_group}, topk_group={self.topk_group}." ) if self.n_group is not None: if self.n_routed_experts % self.n_group != 0: raise ValueError( f"`n_routed_experts` ({self.n_routed_experts}) must be divisible by `n_group` ({self.n_group})." ) if self.topk_group > self.n_group: raise ValueError(f"`topk_group` ({self.topk_group}) cannot exceed `n_group` ({self.n_group}).") __all__ = ["AXK2Config"]