# coding=utf-8 # Copyright 2025 The ZhipuAI Inc. team and 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. """PyTorch GLM-4-MOE model.""" from typing import Optional import torch import torch.utils.checkpoint from torch import nn from ...configuration_utils import PretrainedConfig from ...modeling_rope_utils import rope_config_validation from ...utils import logging from ..cohere.modeling_cohere import CohereAttention from ..deepseek_v3.modeling_deepseek_v3 import ( DeepseekV3DecoderLayer, DeepseekV3ForCausalLM, DeepseekV3MLP, DeepseekV3Model, DeepseekV3PreTrainedModel, DeepseekV3RMSNorm, DeepseekV3TopkRouter, ) from ..gpt_neox.modeling_gpt_neox import apply_rotary_pos_emb # noqa logger = logging.get_logger(__name__) class Glm4MoeConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Glm4MoeModel`]. It is used to instantiate a Glm4Moe model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of [THUDM/GLM-4-100B-A10B](https://huggingface.co/THUDM/GLM-4-100B-A10B). Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 151552): Vocabulary size of the Glm4Moe model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Glm4MoeModel`] hidden_size (`int`, *optional*, defaults to 4096): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 10944): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 46): Number of hidden layers in the Transformer encoder. num_attention_heads (`int`, *optional*, defaults to 96): Number of attention heads for each attention layer in the Transformer encoder. partial_rotary_factor (`float`, *optional*, defaults to 0.5): The factor of the partial rotary position. num_key_value_heads (`int`, *optional*, defaults to 8): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the decoder. max_position_embeddings (`int`, *optional*, defaults to 131072): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rms_norm_eps (`float`, *optional*, defaults to 1e-05): The epsilon used by the rms normalization layers. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether the model's input and output word embeddings should be tied. rope_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. rope_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value accordingly. Expected contents: `rope_type` (`str`): The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', 'llama3'], with 'default' being the original RoPE implementation. `factor` (`float`, *optional*): Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length. `original_max_position_embeddings` (`int`, *optional*): Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during pretraining. `attention_factor` (`float`, *optional*): Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention computation. If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value. `beta_fast` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear ramp function. If unspecified, it defaults to 32. `beta_slow` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear ramp function. If unspecified, it defaults to 1. `short_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to short contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `long_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to long contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `low_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE `high_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. moe_intermediate_size (`int`, *optional*, defaults to 1408): Intermediate size of the routed expert. num_experts_per_tok (`int`, *optional*, defaults to 8): number of experts per token. n_shared_experts (`int`, *optional*, defaults to 1): Number of shared experts. n_routed_experts (`int`, *optional*, defaults to 128): Number of routed experts. routed_scaling_factor (`float`, *optional*, defaults to 1.0): Scaling factor or routed experts. n_group (`int`, *optional*, defaults to 1): Number of groups for routed experts. topk_group (`int`, *optional*, defaults to 1): Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups). first_k_dense_replace (`int`, *optional*, defaults to 1): Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head). \--k dense layers--/ norm_topk_prob (`bool`, *optional*, defaults to `True`): Whether to normalize the topk probabilities. use_qk_norm (`bool`, *optional*, defaults to `False`): Whether to use query-key normalization in the attention ```python >>> from transformers import Glm4MoeModel, Glm4MoeConfig >>> # Initializing a Glm4Moe style configuration >>> configuration = Glm4MoeConfig() >>> # Initializing a model from the GLM-4-MOE-100B-A10B style configuration >>> model = Glm4MoeModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "glm4_moe" keys_to_ignore_at_inference = ["past_key_values"] # Default tensor parallel plan for base model `Glm4Moe` base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise", "layers.*.self_attn.k_proj": "colwise", "layers.*.self_attn.v_proj": "colwise", "layers.*.self_attn.o_proj": "rowwise", "layers.*.mlp.experts.*.gate_proj": "colwise", "layers.*.mlp.experts.*.up_proj": "colwise", "layers.*.mlp.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"]), } def __init__( self, vocab_size=151552, hidden_size=4096, intermediate_size=10944, num_hidden_layers=46, num_attention_heads=96, partial_rotary_factor=0.5, num_key_value_heads=8, hidden_act="silu", max_position_embeddings=131072, initializer_range=0.02, rms_norm_eps=1e-5, use_cache=True, tie_word_embeddings=False, rope_theta=10000.0, rope_scaling=None, attention_bias=False, attention_dropout=0.0, moe_intermediate_size=1408, num_experts_per_tok=8, n_shared_experts=1, n_routed_experts=128, routed_scaling_factor=1.0, n_group=1, topk_group=1, first_k_dense_replace=1, norm_topk_prob=True, use_qk_norm=False, **kwargs, ): 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.partial_rotary_factor = partial_rotary_factor self.num_key_value_heads = num_key_value_heads self.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = rope_scaling self.attention_bias = attention_bias self.attention_dropout = attention_dropout # Validate the correctness of rotary position embeddings parameters # BC: if there is a 'type' field, move it to 'rope_type'. if self.rope_scaling is not None and "type" in self.rope_scaling: self.rope_scaling["rope_type"] = self.rope_scaling["type"] rope_config_validation(self) # MoE arguments self.moe_intermediate_size = moe_intermediate_size self.num_experts_per_tok = num_experts_per_tok self.n_group = n_group self.topk_group = topk_group self.n_shared_experts = n_shared_experts self.n_routed_experts = n_routed_experts self.routed_scaling_factor = routed_scaling_factor self.first_k_dense_replace = first_k_dense_replace self.norm_topk_prob = norm_topk_prob self.use_qk_norm = use_qk_norm super().__init__( tie_word_embeddings=tie_word_embeddings, **kwargs, ) class Glm4MoeAttention(CohereAttention, nn.Module): def __init__(self, config: Glm4MoeConfig, layer_idx: Optional[int] = None): nn.Module.__init__() self.config = config self.layer_idx = layer_idx self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads self.scaling = self.head_dim**-0.5 self.attention_dropout = config.attention_dropout self.is_causal = True self.q_proj = nn.Linear( config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias ) self.k_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.v_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False) self.use_qk_norm = config.use_qk_norm if self.use_qk_norm: self.q_norm = Glm4MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps) self.k_norm = Glm4MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps) class Glm4MoeMLP(DeepseekV3MLP): pass class Glm4MoeTopkRouter(DeepseekV3TopkRouter, nn.Module): def __init__(self, config: Glm4MoeConfig): nn.Module.__init__() self.config = config self.top_k = config.num_experts_per_tok self.n_routed_experts = config.n_routed_experts self.routed_scaling_factor = config.routed_scaling_factor self.n_group = config.n_group self.topk_group = config.topk_group self.norm_topk_prob = config.norm_topk_prob self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size))) self.register_buffer("e_score_correction_bias", torch.zeros((self.n_routed_experts), dtype=torch.float32)) class Glm4MoeRMSNorm(DeepseekV3RMSNorm): pass class Glm4MoeDecoderLayer(DeepseekV3DecoderLayer): pass class Glm4MoePreTrainedModel(DeepseekV3PreTrainedModel): _can_compile_fullgraph = False class Glm4MoeModel(DeepseekV3Model): _keys_to_ignore_on_load_unexpected = [r"model\.layers\.92.*", r"model\.layers\.46.*"] class Glm4MoeForCausalLM(DeepseekV3ForCausalLM): pass __all__ = [ "Glm4MoeConfig", "Glm4MoePreTrainedModel", "Glm4MoeModel", "Glm4MoeForCausalLM", ]