# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/hy_v4/modular_hy_v4.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_hy_v4.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # Copyright 2026 Tencent HunYuan Team 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 RopeParameters from ...utils import auto_docstring @auto_docstring(checkpoint="tencent/Hy4-preview") @strict class HYV4Config(PreTrainedConfig): r""" n_group (`int`, *optional*, defaults to 1): Number of expert groups for routing. HYV4 selects experts globally, so this is 1 (one group holding every expert). topk_group (`int`, *optional*, defaults to 1): Number of expert groups kept during routing. With `n_group=1` this reuses `Glm4MoeLiteTopkRouter` as a plain global top-k. mlp_layer_types (`list[str]`, *optional*): Per-layer MLP type, either `"dense"` or `"sparse"`. Defaults to one dense layer followed by sparse MoE layers. index_topk (`int`, *optional*, defaults to 2048): Maximum number of key positions selected by each DSA query. index_head_dim (`int`, *optional*, defaults to 128): Hidden dimension of each DSA indexer head. index_n_heads (`int`, *optional*, defaults to 16): Number of DSA indexer heads. indexer_types (`list[str]`, *optional*): Per-layer DSA indexer type, either `"full"` or `"shared"`. A shared layer reuses the most recent full indexer in the same forward request. hc_mult (`int`, *optional*, defaults to 4): Number of hidden-state channels maintained by iHC. hc_magnitude (`float`, *optional*, defaults to 2.0): Scale applied to the iHC post-gating branch. hc_eps (`float`, *optional*, defaults to 1e-6): Numerical epsilon added to iHC sigmoid gates. learnable_sink_init (`float`, *optional*, defaults to 0.0): Initial value of each learned attention-sink logit. swiglu_limit (`float`, *optional*, defaults to 10.0): Magnitude of the routed-expert SwiGLU clamp. Values at or below zero disable the clamp. """ model_type = "hy_v4" keys_to_ignore_at_inference = ["past_key_values"] attribute_map = { "num_local_experts": "n_routed_experts", } base_model_tp_plan = { "layers.*.self_attn.q_b_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.*.self_attn.gate_proj": "colwise", "layers.*.self_attn.sinks": "colwise", "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", } vocab_size: int = 120832 hidden_size: int = 2816 intermediate_size: int = 6912 moe_intermediate_size: int = 768 num_hidden_layers: int = 34 num_attention_heads: int = 32 num_key_value_heads: int = 32 head_dim: int = 256 hidden_act: str = "silu" max_position_embeddings: int = 262144 initializer_range: float = 0.006 rms_norm_eps: float = 1e-5 use_cache: bool = True pad_token_id: int | None = 120002 bos_token_id: int | None = 120000 eos_token_id: int | list[int] | None = 120025 tie_word_embeddings: bool = False attention_bias: bool = False attention_dropout: float = 0.0 n_routed_experts: int = 256 output_router_logits: bool = False n_shared_experts: int = 1 num_experts_per_tok: int = 8 routed_scaling_factor: float = 2.827 norm_topk_prob: bool = True n_group: int = 1 topk_group: int = 1 q_lora_rank: int = 1536 kv_lora_rank: int = 512 qk_nope_head_dim: int = 192 qk_rope_head_dim: int = 64 v_head_dim: int = 256 mlp_layer_types: list[str] | None = None layer_types: list[str] | None = None index_topk: int = 2048 index_head_dim: int = 128 index_n_heads: int = 16 indexer_types: list[str] | None = None hc_mult: int = 4 hc_magnitude: float = 2.0 hc_eps: float = 1e-6 learnable_sink_init: float = 0.0 swiglu_limit: float = 10.0 rope_parameters: RopeParameters | dict | None = None def __post_init__(self, **kwargs): # MLA expands the latent to one key/value per query head, so keys are never grouped. self.num_key_value_heads = self.num_attention_heads 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. self.head_dim = self.qk_rope_head_dim if self.mlp_layer_types is None: self.mlp_layer_types = ["dense"] * min(1, self.num_hidden_layers) + ["sparse"] * max( self.num_hidden_layers - 1, 0 ) if self.layer_types is None: self.layer_types = ["indexed_attention"] * self.num_hidden_layers if self.indexer_types is None: self.indexer_types = [ "full" if layer_idx == 0 or (layer_idx - 1) % 4 == 0 else "shared" for layer_idx in range(self.num_hidden_layers) ] super().__post_init__(**kwargs) __all__ = ["HYV4Config"]