GLM-4 Update (#39393)

* one commit with full

* Create glm4_moe.md

* Update check_config_docstrings.py

* Update __init__.py

* update

* argue

* argue: router problem

* 1

* Update test_modeling_glm4_moe.py

* Update test_modeling_glm4_moe.py

* Update test_modeling_glm4_moe.py

* Update modular_glm4_moe.py

* update

* use dsv3 pretrainmodel in modular

* update for test

* upodate new modular

* use LlamaAttention and avoid use  CohereAttention cause repeat norm

* update the modular

* update attn modular

* update

* Update modular_glm4_moe.py

* MTP layer is need to ignore

* fix gradient error using with dots_1 method

* Update test_modeling_glm4_moe.py

* Update test_modeling_glm4_moe.py

* Update test_modeling_glm4_moe.py

---------

Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co>

docs/source/en/_toctree.yml

@@ -475,6 +475,8 @@
         title: GLM
       - local: model_doc/glm4
         title: glm4
+      - local: model_doc/glm4_moe
+        title: glm4_moe
       - local: model_doc/openai-gpt
         title: GPT
       - local: model_doc/gpt_neo

docs/source/en/model_doc/glm4_moe.md

@@ -0,0 +1,35 @@
+<!--Copyright 2025 The ZhipuAI Inc. 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Glm4Moe
+
+## Overview
+
+This will update After model release.
+
+## Glm4MoeConfig
+
+[[autodoc]] Glm4MoeConfig
+
+## Glm4MoeModel
+
+[[autodoc]] Glm4MoeModel
+    - forward
+
+## Glm4MoeForCausalLM
+
+[[autodoc]] Glm4MoeForCausalLM
+    - forward

src/transformers/models/auto/configuration_auto.py

@@ -153,6 +153,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("git", "GitConfig"),
         ("glm", "GlmConfig"),
         ("glm4", "Glm4Config"),
+        ("glm4_moe", "Glm4MoeConfig"),
         ("glm4v", "Glm4vConfig"),
         ("glm4v_text", "Glm4vTextConfig"),
         ("glpn", "GLPNConfig"),
@@ -547,6 +548,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("git", "GIT"),
         ("glm", "GLM"),
         ("glm4", "GLM4"),
+        ("glm4_moe", "Glm4MoE"),
         ("glm4v", "GLM4V"),
         ("glm4v_text", "GLM4V"),
         ("glpn", "GLPN"),

src/transformers/models/auto/modeling_auto.py

@@ -144,6 +144,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("git", "GitModel"),
         ("glm", "GlmModel"),
         ("glm4", "Glm4Model"),
+        ("glm4_moe", "Glm4MoeModel"),
         ("glm4v", "Glm4vModel"),
         ("glm4v_text", "Glm4vTextModel"),
         ("glpn", "GLPNModel"),
@@ -606,6 +607,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("git", "GitForCausalLM"),
         ("glm", "GlmForCausalLM"),
         ("glm4", "Glm4ForCausalLM"),
+        ("glm4_moe", "Glm4MoeForCausalLM"),
         ("got_ocr2", "GotOcr2ForConditionalGeneration"),
         ("gpt-sw3", "GPT2LMHeadModel"),
         ("gpt2", "GPT2LMHeadModel"),

src/transformers/models/auto/tokenization_auto.py

@@ -269,6 +269,7 @@ TOKENIZER_MAPPING_NAMES = OrderedDict[str, tuple[Optional[str], Optional[str]]](
         ("git", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
         ("glm", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
         ("glm4", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
+        ("glm4_moe", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
         ("glm4v", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
         ("gpt-sw3", ("GPTSw3Tokenizer" if is_sentencepiece_available() else None, None)),
         ("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),

src/transformers/models/glm4_moe/__init__.py

@@ -0,0 +1,27 @@
+# Copyright 2025 The HuggingFace 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 typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_glm4_moe import *
+    from .modeling_glm4_moe import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/glm4_moe/configuration_glm4_moe.py

@@ -0,0 +1,242 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/glm4_moe/modular_glm4_moe.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_glm4_moe.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# 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.
+
+from ...configuration_utils import PretrainedConfig
+from ...modeling_rope_utils import rope_config_validation
+
+
+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,
+        )
+
+
+__all__ = ["Glm4MoeConfig"]

src/transformers/models/glm4_moe/modeling_glm4_moe.py

@@ -0,0 +1,649 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/glm4_moe/modular_glm4_moe.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_glm4_moe.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# 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.
+
+from typing import Callable, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache
+from ...generation import GenerationMixin
+from ...integrations import use_kernel_forward_from_hub
+from ...masking_utils import create_causal_mask
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
+from ...utils.generic import check_model_inputs
+from .configuration_glm4_moe import Glm4MoeConfig
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    # Keep half or full tensor for later concatenation
+    rotary_dim = cos.shape[-1]
+    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
+    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
+
+    # Apply rotary embeddings on the first half or full tensor
+    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
+    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
+
+    # Concatenate back to full shape
+    q_embed = torch.cat([q_embed, q_pass], dim=-1)
+    k_embed = torch.cat([k_embed, k_pass], dim=-1)
+    return q_embed, k_embed
+
+
+class Glm4MoeAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Glm4MoeConfig, layer_idx: Optional[int] = None):
+        super().__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)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape)
+        key_states = self.k_proj(hidden_states).view(hidden_shape)
+        value_states = self.v_proj(hidden_states).view(hidden_shape)
+
+        if self.use_qk_norm:  # main diff from Llama
+            query_states = self.q_norm(query_states)
+            key_states = self.k_norm(key_states)
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Glm4MoeMLP(nn.Module):
+    def __init__(self, config, hidden_size=None, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class Glm4MoeTopkRouter(nn.Module):
+    def __init__(self, config: Glm4MoeConfig):
+        super().__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))
+
+    @torch.no_grad()
+    def get_topk_indices(self, scores):
+        scores_for_choice = scores.view(-1, self.n_routed_experts) + self.e_score_correction_bias.unsqueeze(0)
+        group_scores = (
+            scores_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .reshape(-1, self.n_routed_experts)
+        )
+        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
+        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
+        return topk_indices
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.view(-1, self.config.hidden_size)
+        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+        scores = router_logits.sigmoid()
+        topk_indices = self.get_topk_indices(scores)
+        topk_weights = scores.gather(1, topk_indices)
+        if self.norm_topk_prob:
+            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weights /= denominator
+        topk_weights = topk_weights * self.routed_scaling_factor
+        return topk_indices, topk_weights
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class Glm4MoeRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Glm4MoeRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class Glm4MoeMoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.experts = nn.ModuleList(
+            [
+                Glm4MoeMLP(config, intermediate_size=config.moe_intermediate_size)
+                for _ in range(config.n_routed_experts)
+            ]
+        )
+        self.gate = Glm4MoeTopkRouter(config)
+        self.shared_experts = Glm4MoeMLP(
+            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
+        )
+
+    def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor):
+        r"""
+        CALL FOR CONTRIBUTION! I don't have time to optimise this right now, but expert weights need to be fused
+        to not have to do a loop here (deepseek has 256 experts soooo yeah).
+        """
+        final_hidden_states = torch.zeros_like(hidden_states, dtype=topk_weights.dtype)
+        expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=len(self.experts))
+        expert_mask = expert_mask.permute(2, 0, 1)
+
+        for expert_idx in range(len(self.experts)):
+            expert = self.experts[expert_idx]
+            mask = expert_mask[expert_idx]
+            token_indices, weight_indices = torch.where(mask)
+
+            if token_indices.numel() > 0:
+                expert_weights = topk_weights[token_indices, weight_indices]
+                expert_input = hidden_states[token_indices]
+                expert_output = expert(expert_input)
+                weighted_output = expert_output * expert_weights.unsqueeze(-1)
+                final_hidden_states.index_add_(0, token_indices, weighted_output)
+
+        # in original deepseek, the output of the experts are gathered once we leave this module
+        # thus the moe module is itelsf an IsolatedParallel module
+        # and all expert are "local" meaning we shard but we don't gather
+        return final_hidden_states.type(hidden_states.dtype)
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        topk_indices, topk_weights = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.moe(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+class Glm4MoeDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: Glm4MoeConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = Glm4MoeAttention(config=config, layer_idx=layer_idx)
+
+        if layer_idx >= config.first_k_dense_replace:
+            self.mlp = Glm4MoeMoE(config)
+        else:
+            self.mlp = Glm4MoeMLP(config)
+
+        self.input_layernorm = Glm4MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Glm4MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+@auto_docstring
+class Glm4MoePreTrainedModel(PreTrainedModel):
+    config: Glm4MoeConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Glm4MoeDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_static_cache = False
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": Glm4MoeDecoderLayer,
+        "attentions": Glm4MoeAttention,
+    }
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, Glm4MoeRMSNorm):
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, Glm4MoeTopkRouter):
+            module.weight.data.normal_(mean=0.0, std=std)
+
+
+class Glm4MoeRotaryEmbedding(nn.Module):
+    def __init__(self, config: Glm4MoeConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+@auto_docstring
+class Glm4MoeModel(Glm4MoePreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.92.*", r"model\.layers\.46.*"]
+
+    def __init__(self, config: Glm4MoeConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Glm4MoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Glm4MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Glm4MoeRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position: torch.Tensor = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            input_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class Glm4MoeForCausalLM(Glm4MoePreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Glm4MoeModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Glm4MoeForCausalLM
+
+        >>> model = Glm4MoeForCausalLM.from_pretrained("meta-glm4_moe/Glm4Moe-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-glm4_moe/Glm4Moe-2-7b-hf")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = ["Glm4MoePreTrainedModel", "Glm4MoeModel", "Glm4MoeForCausalLM"]

src/transformers/models/glm4_moe/modular_glm4_moe.py

@@ -0,0 +1,329 @@
+# 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):
+    _supports_static_cache = 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",
+]

tests/models/glm4_moe/test_modeling_glm4_moe.py

@@ -0,0 +1,135 @@
+# 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.
+"""Testing suite for the PyTorch GLM-4-MoE model."""
+
+import unittest
+
+import torch
+from packaging import version
+
+from transformers import is_torch_available
+from transformers.testing_utils import (
+    cleanup,
+    require_read_token,
+    require_torch,
+    require_torch_accelerator,
+    slow,
+    torch_device,
+)
+
+from ...causal_lm_tester import CausalLMModelTest, CausalLMModelTester
+
+
+if is_torch_available():
+    from transformers import AutoTokenizer, Glm4MoeConfig, Glm4MoeForCausalLM, Glm4MoeModel
+
+
+class Glm4MoeModelTester(CausalLMModelTester):
+    if is_torch_available():
+        config_class = Glm4MoeConfig
+        base_model_class = Glm4MoeModel
+        causal_lm_class = Glm4MoeForCausalLM
+
+    def __init__(
+        self,
+        parent,
+        n_routed_experts=8,
+        n_shared_experts=1,
+        n_group=1,
+        topk_group=1,
+        num_experts_per_tok=8,
+    ):
+        super().__init__(parent=parent, num_experts_per_tok=num_experts_per_tok)
+        self.n_routed_experts = n_routed_experts
+        self.n_shared_experts = n_shared_experts
+        self.n_group = n_group
+        self.topk_group = topk_group
+
+
+@require_torch
+class Glm4MoeModelTest(CausalLMModelTest, unittest.TestCase):
+    all_model_classes = (
+        (
+            Glm4MoeModel,
+            Glm4MoeForCausalLM,
+        )
+        if is_torch_available()
+        else ()
+    )
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": Glm4MoeModel,
+            "text-generation": Glm4MoeForCausalLM,
+        }
+        if is_torch_available()
+        else {}
+    )
+    test_headmasking = False
+    test_pruning = False
+    fx_compatible = False
+    model_tester_class = Glm4MoeModelTester
+    # used in `test_torch_compile_for_training`
+    _torch_compile_train_cls = Glm4MoeForCausalLM if is_torch_available() else None
+
+
+@require_torch_accelerator
+@require_read_token
+@slow
+class Glm4MoeIntegrationTest(unittest.TestCase):
+    def tearDown(self):
+        # See LlamaIntegrationTest.tearDown(). Can be removed once LlamaIntegrationTest.tearDown() is removed.
+        cleanup(torch_device, gc_collect=False)
+
+    @slow
+    @require_torch_accelerator
+    @require_read_token
+    def test_compile_static_cache(self):
+        # `torch==2.2` will throw an error on this test (as in other compilation tests), but torch==2.1.2 and torch>2.2
+        # work as intended. See https://github.com/pytorch/pytorch/issues/121943
+        if version.parse(torch.__version__) < version.parse("2.3.0"):
+            self.skipTest(reason="This test requires torch >= 2.3 to run.")
+
+        NUM_TOKENS_TO_GENERATE = 40
+        EXPECTED_TEXT_COMPLETION = [
+            'hello, world!\'\'\')\nprint(\'hello, world!\')\nprint("hello, world!")\nprint("hello, world!")\nprint("hello, world!")\nprint("hello, world!")\nprint("hello, world!")\n',
+            "tell me the story of the first Thanksgiving. commonly known as the Pilgrims, arrived in the autumn of 1620. They were seeking religious freedom and a new life in the Plymouth Colony. Their first",
+        ]
+
+        prompts = ["[gMASK]<sop>hello", "[gMASK]<sop>tell me"]
+        tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4.5")
+        model = Glm4MoeForCausalLM.from_pretrained(
+            "THUDM/GLM-4.5", device_map=torch_device, torch_dtype=torch.bfloat16
+        )
+        inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(model.device)
+
+        # Dynamic Cache
+        generated_ids = model.generate(**inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False)
+        dynamic_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT_COMPLETION, dynamic_text)
+
+        # Static Cache
+        generated_ids = model.generate(
+            **inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False, cache_implementation="static"
+        )
+        static_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT_COMPLETION, static_text)
+
+        # Static Cache + compile
+        model._cache = None  # clear cache object, initialized when we pass `cache_implementation="static"`
+        model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
+        generated_ids = model.generate(
+            **inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False, cache_implementation="static"
+        )
+        static_compiled_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT_COMPLETION, static_compiled_text)