Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

chat_template.jinja

@@ -0,0 +1,86 @@
+[gMASK]<sop>
+{%- if tools -%}
+<|system|>
+# Tools
+
+You may call one or more functions to assist with the user query.
+
+You are provided with function signatures within <tools></tools> XML tags:
+<tools>
+{% for tool in tools %}
+{{ tool | tojson(ensure_ascii=False) }}
+{% endfor %}
+</tools>
+
+For each function call, output the function name and arguments within the following XML format:
+<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call>{%- endif -%}
+{%- macro visible_text(content) -%}
+    {%- if content is string -%}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping -%}
+        {%- for item in content -%}
+            {%- if item is mapping and item.type == 'text' -%}
+                {{- item.text }}
+            {%- elif item is string -%}
+                {{- item }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- else -%}
+        {{- content }}
+    {%- endif -%}
+{%- endmacro -%}
+{%- set ns = namespace(last_user_index=-1) %}
+{%- for m in messages %}
+    {%- if m.role == 'user' %}
+        {% set ns.last_user_index = loop.index0 -%}
+    {%- endif %}
+{%- endfor %}
+{% for m in messages %}
+{%- if m.role == 'user' -%}<|user|>{{ visible_text(m.content) }}
+{%- elif m.role == 'assistant' -%}
+<|assistant|>
+{%- set reasoning_content = '' %}
+{%- set content = visible_text(m.content) %}
+{%- if m.reasoning_content is string %}
+    {%- set reasoning_content = m.reasoning_content %}
+{%- else %}
+    {%- if '</think>' in content %}
+        {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+        {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+    {%- endif %}
+{%- endif %}
+{%- if ((clear_thinking is defined and not clear_thinking) or loop.index0 > ns.last_user_index) and reasoning_content -%}
+{{ '<think>' + reasoning_content.strip() +  '</think>'}}
+{%- else -%}
+{{ '</think>' }}
+{%- endif -%}
+{%- if content.strip() -%}
+{{ content.strip() }}
+{%- endif -%}
+{% if m.tool_calls %}
+{% for tc in m.tool_calls %}
+{%- if tc.function %}
+    {%- set tc = tc.function %}
+{%- endif %}
+{{- '<tool_call>' + tc.name -}}
+{% set _args = tc.arguments %}{% for k, v in _args.items() %}<arg_key>{{ k }}</arg_key><arg_value>{{ v | tojson(ensure_ascii=False) if v is not string else v }}</arg_value>{% endfor %}</tool_call>{% endfor %}
+{% endif %}
+{%- elif m.role == 'tool' -%}
+{%- if m.content is string -%}
+{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+    {{- '<|observation|>' }}
+{%- endif %}
+{{- '<tool_response>' }}
+{{- m.content }}
+{{- '</tool_response>' }}
+{%- else -%}
+<|observation|>{% for tr in m.content %}
+<tool_response>{{ tr.output if tr.output is defined else tr }}</tool_response>{% endfor -%}
+{% endif -%}
+{%- elif m.role == 'system' -%}
+<|system|>{{ visible_text(m.content) }}
+{%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    <|assistant|>{{- '</think>' if (enable_thinking is defined and not enable_thinking) else '<think>' -}}
+{%- endif -%}

config.json

@@ -0,0 +1,110 @@
+{
+  "architectures": [
+    "Glm4MoeLiteSCMForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": 
+    {
+      "AutoConfig": "configuration_glm4_moe_lite_scm.Glm4MoeLiteSCMConfig",
+      "AutoModel": "modeling_glm4_moe_lite_scm.Glm4MoeLiteSCMModel",
+      "AutoModelForCausalLM": "modeling_glm4_moe_lite_scm.Glm4MoeLiteSCMForCausalLM"
+    },
+  "bos_token_id": 0,
+  "dtype": "bfloat16",
+  "eos_token_id": [
+    154820,
+    154827,
+    154829
+  ],
+  "first_k_dense_replace": 1,
+  "head_dim": 64,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 10240,
+  "kv_lora_rank": 512,
+  "max_position_embeddings": 202752,
+  "mlp_layer_types": [
+    "dense",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse",
+    "sparse"
+  ],
+  "model_type": "glm4_moe_lite_scm",
+  "moe_intermediate_size": 1536,
+  "n_group": 1,
+  "n_routed_experts": 64,
+  "n_shared_experts": 1,
+  "norm_topk_prob": true,
+  "num_attention_heads": 20,
+  "num_experts_per_tok": 4,
+  "num_hidden_layers": 47,
+  "num_key_value_heads": 20,
+  "num_nextn_predict_layers": 0,
+  "pad_token_id": 154820,
+  "partial_rotary_factor": 1.0,
+  "pretraining_tp": 1,
+  "q_lora_rank": 768,
+  "qk_head_dim": 256,
+  "qk_nope_head_dim": 192,
+  "qk_rope_head_dim": 64,
+  "rms_norm_eps": 1e-05,
+  "rope_interleave": true,
+  "rope_parameters": {
+    "partial_rotary_factor": 1.0,
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "routed_scaling_factor": 1.8,
+  "tie_word_embeddings": false,
+  "topk_group": 1,
+  "topk_method": "noaux_tc",
+  "transformers_version": "5.0.0",
+  "use_cache": true,
+  "v_head_dim": 256,
+  "vocab_size": 154880
+}

configuration_glm4_moe_lite_for_backconvert.py

@@ -0,0 +1,229 @@
+# 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 transformers.configuration_utils import PreTrainedConfig, layer_type_validation
+from transformers.modeling_rope_utils import RopeParameters
+
+
+class Glm4MoeLiteConfig(PreTrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Glm4MoeLiteModel`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V3.
+    e.g. [bzantium/tiny-deepseek-v3](https://huggingface.co/bzantium/tiny-deepseek-v3)
+    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 154880):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Glm4MoeLiteModel`]
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 10240):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1536):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 47):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 20):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*, defaults to 20):
+            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
+            `num_attention_heads`.
+        n_shared_experts (`int`, *optional*, defaults to 1):
+            Number of shared experts.
+        n_routed_experts (`int`, *optional*, defaults to 64):
+            Number of routed experts.
+        routed_scaling_factor (`float`, *optional*, defaults to 1.8):
+            Scaling factor or routed experts.
+        kv_lora_rank (`int`, *optional*, defaults to 512):
+            Rank of the LoRA matrices for key and value projections.
+        q_lora_rank (`int`, *optional*, defaults to 768):
+            Rank of the LoRA matrices for query projections.
+        qk_rope_head_dim (`int`, *optional*, defaults to 64):
+            Dimension of the query/key heads that use rotary position embeddings.
+        v_head_dim (`int`, *optional*, defaults to 256):
+            Dimension of the value heads.
+        qk_nope_head_dim (`int`, *optional*, defaults to 192):
+            Dimension of the query/key heads that don't use rotary position embeddings.
+        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).
+        num_experts_per_tok (`int`, *optional*, defaults to 4):
+            Number of selected experts, None means dense model.
+        norm_topk_prob (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the weights of the routed experts.
+        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 202752):
+            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`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 0):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 1):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
+            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
+            issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_parameters (`RopeParameters`, *optional*):
+            Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
+            a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
+            with longer `max_position_embeddings`.
+        rope_interleave (`bool`, *optional*, defaults to `True`):
+            Whether to interleave the rotary position embeddings.
+        mlp_layer_types (`list`, *optional*):
+            MLP (Moe vs Dense) pattern for each layer.
+        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.
+
+    ```python
+    >>> from transformers import Glm4MoeLiteModel, Glm4MoeLiteConfig
+
+    >>> # Initializing a Deepseek-V3 style configuration
+    >>> configuration = Glm4MoeLiteConfig()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "glm4_moe_lite"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.experts.gate_up_proj": "local_rowwise",
+        "layers.*.mlp.experts.down_proj": "local_rowwise",
+        "layers.*.mlp.experts": "gather",
+        "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"]),
+    }
+    attribute_map = {
+        "num_local_experts": "n_routed_experts",
+    }
+
+    def __init__(
+        self,
+        vocab_size: int | None = 154880,
+        hidden_size: int | None = 2048,
+        intermediate_size: int | None = 10240,
+        moe_intermediate_size: int | None = 1536,
+        num_hidden_layers: int | None = 47,
+        num_attention_heads: int | None = 20,
+        num_key_value_heads: int | None = 20,
+        n_shared_experts: int | None = 1,
+        n_routed_experts: int | None = 64,
+        routed_scaling_factor: float | None = 1.8,
+        kv_lora_rank: int | None = 512,
+        q_lora_rank: int | None = 768,
+        qk_rope_head_dim: int | None = 64,
+        v_head_dim: int | None = 256,
+        qk_nope_head_dim: int | None = 192,
+        n_group: int | None = 1,
+        topk_group: int | None = 1,
+        num_experts_per_tok: int | None = 4,
+        norm_topk_prob: bool | None = True,
+        hidden_act: str | None = "silu",
+        max_position_embeddings: int | None = 202752,
+        initializer_range: float | None = 0.02,
+        rms_norm_eps: int | None = 1e-5,
+        use_cache: bool | None = True,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = 0,
+        eos_token_id: int | None = 1,
+        pretraining_tp: int | None = 1,
+        tie_word_embeddings: bool | None = False,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        rope_interleave: bool | None = True,
+        mlp_layer_types=None,
+        attention_bias: bool | None = False,
+        attention_dropout: float | None = 0.0,
+        **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
+
+        # Default to MoE from the second layer and on
+        self.mlp_layer_types = mlp_layer_types
+        if self.mlp_layer_types is None:
+            self.mlp_layer_types = ["dense"] + ["sparse"] * (self.num_hidden_layers - 1)
+        layer_type_validation(self.mlp_layer_types, self.num_hidden_layers, attention=False)
+
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.head_dim = qk_rope_head_dim
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.norm_topk_prob = norm_topk_prob
+        self.rope_interleave = rope_interleave
+        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.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.rope_parameters = rope_parameters
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["Glm4MoeLiteConfig"]

configuration_glm4_moe_lite_scm.py

@@ -0,0 +1,229 @@
+# 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 transformers.configuration_utils import PreTrainedConfig, layer_type_validation
+from transformers.modeling_rope_utils import RopeParameters
+
+
+class Glm4MoeLiteSCMConfig(PreTrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Glm4MoeLiteModel`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V3.
+    e.g. [bzantium/tiny-deepseek-v3](https://huggingface.co/bzantium/tiny-deepseek-v3)
+    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 154880):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Glm4MoeLiteModel`]
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 10240):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1536):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 47):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 20):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*, defaults to 20):
+            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
+            `num_attention_heads`.
+        n_shared_experts (`int`, *optional*, defaults to 1):
+            Number of shared experts.
+        n_routed_experts (`int`, *optional*, defaults to 64):
+            Number of routed experts.
+        routed_scaling_factor (`float`, *optional*, defaults to 1.8):
+            Scaling factor or routed experts.
+        kv_lora_rank (`int`, *optional*, defaults to 512):
+            Rank of the LoRA matrices for key and value projections.
+        q_lora_rank (`int`, *optional*, defaults to 768):
+            Rank of the LoRA matrices for query projections.
+        qk_rope_head_dim (`int`, *optional*, defaults to 64):
+            Dimension of the query/key heads that use rotary position embeddings.
+        v_head_dim (`int`, *optional*, defaults to 256):
+            Dimension of the value heads.
+        qk_nope_head_dim (`int`, *optional*, defaults to 192):
+            Dimension of the query/key heads that don't use rotary position embeddings.
+        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).
+        num_experts_per_tok (`int`, *optional*, defaults to 4):
+            Number of selected experts, None means dense model.
+        norm_topk_prob (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the weights of the routed experts.
+        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 202752):
+            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`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 0):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 1):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
+            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
+            issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_parameters (`RopeParameters`, *optional*):
+            Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
+            a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
+            with longer `max_position_embeddings`.
+        rope_interleave (`bool`, *optional*, defaults to `True`):
+            Whether to interleave the rotary position embeddings.
+        mlp_layer_types (`list`, *optional*):
+            MLP (Moe vs Dense) pattern for each layer.
+        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.
+
+    ```python
+    >>> from transformers import Glm4MoeLiteModel, Glm4MoeLiteConfig
+
+    >>> # Initializing a Deepseek-V3 style configuration
+    >>> configuration = Glm4MoeLiteConfig()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "glm4_moe_lite_scm"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.experts.gate_up_proj": "local_rowwise",
+        "layers.*.mlp.experts.down_proj": "local_rowwise",
+        "layers.*.mlp.experts": "gather",
+        "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"]),
+    }
+    attribute_map = {
+        "num_local_experts": "n_routed_experts",
+    }
+
+    def __init__(
+        self,
+        vocab_size: int | None = 154880,
+        hidden_size: int | None = 2048,
+        intermediate_size: int | None = 10240,
+        moe_intermediate_size: int | None = 1536,
+        num_hidden_layers: int | None = 47,
+        num_attention_heads: int | None = 20,
+        num_key_value_heads: int | None = 20,
+        n_shared_experts: int | None = 1,
+        n_routed_experts: int | None = 64,
+        routed_scaling_factor: float | None = 1.8,
+        kv_lora_rank: int | None = 512,
+        q_lora_rank: int | None = 768,
+        qk_rope_head_dim: int | None = 64,
+        v_head_dim: int | None = 256,
+        qk_nope_head_dim: int | None = 192,
+        n_group: int | None = 1,
+        topk_group: int | None = 1,
+        num_experts_per_tok: int | None = 4,
+        norm_topk_prob: bool | None = True,
+        hidden_act: str | None = "silu",
+        max_position_embeddings: int | None = 202752,
+        initializer_range: float | None = 0.02,
+        rms_norm_eps: int | None = 1e-5,
+        use_cache: bool | None = True,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = 0,
+        eos_token_id: int | None = 1,
+        pretraining_tp: int | None = 1,
+        tie_word_embeddings: bool | None = False,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        rope_interleave: bool | None = True,
+        mlp_layer_types=None,
+        attention_bias: bool | None = False,
+        attention_dropout: float | None = 0.0,
+        **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
+
+        # Default to MoE from the second layer and on
+        self.mlp_layer_types = mlp_layer_types
+        if self.mlp_layer_types is None:
+            self.mlp_layer_types = ["dense"] + ["sparse"] * (self.num_hidden_layers - 1)
+        layer_type_validation(self.mlp_layer_types, self.num_hidden_layers, attention=False)
+
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.head_dim = qk_rope_head_dim
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.norm_topk_prob = norm_topk_prob
+        self.rope_interleave = rope_interleave
+        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.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.rope_parameters = rope_parameters
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["Glm4MoeLiteSCMConfig"]

convert_hf_to_scm.py

@@ -0,0 +1,131 @@
+import glob
+import re
+import shutil
+import sys
+
+import accelerate
+import torch
+from configuration_glm4_moe_lite_scm import Glm4MoeLiteSCMConfig
+from modeling_glm4_moe_lite_scm import Glm4MoeLiteSCMForCausalLM
+from transformers.models.glm4_moe_lite.configuration_glm4_moe_lite import Glm4MoeLiteConfig
+from safetensors import safe_open
+
+input_model = sys.argv[1]
+output_model_path = sys.argv[2]
+
+auto_map = {
+    "AutoConfig": "configuration_glm4_moe_lite_scm.Glm4MoeLiteSCMConfig",
+    "AutoModel": "modeling_glm4_moe_lite_scm.Glm4MoeLiteSCMModel",
+    "AutoModelForCausalLM": "modeling_glm4_moe_lite_scm.Glm4MoeLiteSCMForCausalLM"
+},
+
+cfg_standard_moe = Glm4MoeLiteConfig.from_pretrained(input_model)
+cfg_shared_moe = Glm4MoeLiteSCMConfig(
+    auto_map=auto_map,
+    n_group=cfg_standard_moe.n_group,
+    topk_group=cfg_standard_moe.topk_group,
+    n_shared_experts=cfg_standard_moe.n_shared_experts,
+    n_routed_experts=cfg_standard_moe.n_routed_experts,
+    num_experts_per_tok=cfg_standard_moe.num_experts_per_tok,
+    first_k_dense_replace=cfg_standard_moe.first_k_dense_replace,
+    vocab_size=cfg_standard_moe.vocab_size,
+    hidden_size=cfg_standard_moe.hidden_size,
+    intermediate_size=cfg_standard_moe.intermediate_size,
+    num_hidden_layers=cfg_standard_moe.num_hidden_layers,
+    num_attention_heads=cfg_standard_moe.num_attention_heads,
+    num_key_value_heads=cfg_standard_moe.num_key_value_heads,
+    hidden_act=cfg_standard_moe.hidden_act,
+    max_position_embeddings=cfg_standard_moe.max_position_embeddings,
+    initializer_range=cfg_standard_moe.initializer_range,
+    rms_norm_eps=cfg_standard_moe.rms_norm_eps,
+    tie_word_embeddings=cfg_standard_moe.tie_word_embeddings,
+    rope_parameters=cfg_standard_moe.rope_parameters,
+    rope_scaling=cfg_standard_moe.rope_scaling,
+    attention_dropout=cfg_standard_moe.attention_dropout,
+    moe_intermediate_size=cfg_standard_moe.moe_intermediate_size,
+    qk_nope_head_dim=cfg_standard_moe.qk_nope_head_dim,
+    qk_rope_head_dim=cfg_standard_moe.qk_rope_head_dim,
+    v_head_dim=cfg_standard_moe.v_head_dim,
+    partial_rotary_factor=cfg_standard_moe.partial_rotary_factor,
+    num_nextn_predict_layers=0,
+    routed_scaling_factor=cfg_standard_moe.routed_scaling_factor,
+    topk_method=cfg_standard_moe.topk_method,
+    norm_topk_prob=cfg_standard_moe.norm_topk_prob,
+    attention_bias=cfg_standard_moe.attention_bias,
+    q_lora_rank=cfg_standard_moe.q_lora_rank,
+    kv_lora_rank=cfg_standard_moe.kv_lora_rank,
+    eos_token_id=cfg_standard_moe.eos_token_id,
+    pad_token_id=cfg_standard_moe.pad_token_id,
+    torch_dtype=cfg_standard_moe.torch_dtype,
+)
+
+num_experts = cfg_standard_moe.n_routed_experts
+num_hidden_layers = cfg_standard_moe.num_hidden_layers
+
+with accelerate.init_empty_weights():
+    model_shared_moe = Glm4MoeLiteSCMForCausalLM(cfg_shared_moe)
+
+model_shared_moe = model_shared_moe.to(torch.bfloat16)
+new_state_dict = {}
+pattern = f"{input_model}/model-*-of-*.safetensors"
+files = sorted(glob.glob(pattern))
+
+if len(files) == 0:
+    raise FileNotFoundError
+tensors = {}
+
+for file_path in files:
+    print(f"processing {file_path}")
+    with safe_open(file_path, framework="pt", device="cpu") as f:
+        for key in f.keys():
+            tensor = f.get_tensor(key)
+            tensors[key] = tensor
+
+for key in tensors:
+    if f"layers.{num_hidden_layers}" in key:
+        continue
+    if "experts" not in key or "shared_experts" in key:
+        new_state_dict[key] = tensors[key]
+    elif "experts.0" in key:
+        layer_num = int(re.search(r"\d+", key).group())
+        new_state_dict[
+            f"model.layers.{layer_num}.mlp.moe_mlp.output_experts.weight"
+        ] = torch.stack(
+            [
+                tensors[f"model.layers.{layer_num}.mlp.experts.{i}.down_proj.weight"]
+                for i in range(num_experts)
+            ]
+        )
+        new_state_dict[f"model.layers.{layer_num}.mlp.moe_mlp.experts.weight"] = (
+            torch.stack(
+                [
+                    torch.cat(
+                        [
+                            tensors[
+                                f"model.layers.{layer_num}.mlp.experts.{i}.up_proj.weight"
+                            ],
+                            tensors[
+                                f"model.layers.{layer_num}.mlp.experts.{i}.gate_proj.weight"
+                            ],
+                        ],
+                        dim=0,
+                    )
+                    for i in range(num_experts)
+                ]
+            )
+        )
+model_shared_moe.load_state_dict(new_state_dict, strict=True, assign=True)
+model_shared_moe.save_pretrained(output_model_path)
+cfg_shared_moe.save_pretrained(output_model_path)
+
+
+shutil.copy(
+    "modeling_glm4_moe_lite_scm.py",
+    output_model_path + "/" + "modeling_glm4_moe_lite_scm.py",
+)
+shutil.copy(
+    "configuration_glm4_moe_lite_scm.py",
+    output_model_path + "/" + "configuration_glm4_moe_lite_scm.py",
+)
+for i in ["tokenizer_config.json", "tokenizer.json", "chat_template.jinja"]:
+    shutil.copy(input_model + "/" + i, output_model_path + "/" + i)

convert_scm_to_hf.py

@@ -0,0 +1,103 @@
+import glob
+import re
+import shutil
+import sys
+
+import accelerate
+import torch
+from safetensors import safe_open
+from configuration_glm4_moe_lite_scm import Glm4MoeLiteSCMConfig
+
+from modeling_glm4_moe_lite_for_backconvert import Glm4MoeLiteForCausalLM
+from configuration_glm4_moe_lite_for_backconvert import Glm4MoeLiteConfig
+
+input_model = sys.argv[1]
+output_model_path = sys.argv[2]
+
+cfg_shared_moe = Glm4MoeLiteSCMConfig.from_pretrained(input_model)
+cfg_standard_moe = Glm4MoeLiteConfig(
+    n_group=cfg_shared_moe.n_group,
+    topk_group=cfg_shared_moe.topk_group,
+    n_shared_experts=cfg_shared_moe.n_shared_experts,
+    n_routed_experts=cfg_shared_moe.n_routed_experts,
+    num_experts_per_tok=cfg_shared_moe.num_experts_per_tok,
+    first_k_dense_replace=cfg_shared_moe.first_k_dense_replace,
+    vocab_size=cfg_shared_moe.vocab_size,
+    hidden_size=cfg_shared_moe.hidden_size,
+    intermediate_size=cfg_shared_moe.intermediate_size,
+    num_hidden_layers=cfg_shared_moe.num_hidden_layers,
+    num_attention_heads=cfg_shared_moe.num_attention_heads,
+    num_key_value_heads=cfg_shared_moe.num_key_value_heads,
+    hidden_act=cfg_shared_moe.hidden_act,
+    max_position_embeddings=cfg_shared_moe.max_position_embeddings,
+    initializer_range=cfg_shared_moe.initializer_range,
+    rms_norm_eps=cfg_shared_moe.rms_norm_eps,
+    tie_word_embeddings=cfg_shared_moe.tie_word_embeddings,
+    rope_parameters=cfg_shared_moe.rope_parameters,
+    rope_scaling=cfg_shared_moe.rope_scaling,
+    attention_dropout=cfg_shared_moe.attention_dropout,
+    moe_intermediate_size=cfg_shared_moe.moe_intermediate_size,
+    qk_nope_head_dim=cfg_shared_moe.qk_nope_head_dim,
+    qk_rope_head_dim=cfg_shared_moe.qk_rope_head_dim,
+    v_head_dim=cfg_shared_moe.v_head_dim,
+    partial_rotary_factor=cfg_shared_moe.partial_rotary_factor,
+    num_nextn_predict_layers=0,
+    routed_scaling_factor=cfg_shared_moe.routed_scaling_factor,
+    topk_method=cfg_shared_moe.topk_method,
+    norm_topk_prob=cfg_shared_moe.norm_topk_prob,
+    attention_bias=cfg_shared_moe.attention_bias,
+    q_lora_rank=cfg_shared_moe.q_lora_rank,
+    kv_lora_rank=cfg_shared_moe.kv_lora_rank,
+    eos_token_id=cfg_shared_moe.eos_token_id,
+    pad_token_id=cfg_shared_moe.pad_token_id,
+    torch_dtype=cfg_shared_moe.torch_dtype,
+)
+num_experts = cfg_standard_moe.n_shared_experts
+
+with accelerate.init_empty_weights():
+    model_standard_moe = Glm4MoeLiteForCausalLM(cfg_standard_moe)
+
+model_standard_moe = model_standard_moe.to(torch.bfloat16)
+new_state_dict = {}
+pattern = f"{input_model}/model-*-of-*.safetensors"
+files = sorted(glob.glob(pattern))
+
+if len(files) == 0:
+    raise FileNotFoundError
+tensors = {}
+
+for file_path in files:
+    print(f"processing {file_path}")
+    with safe_open(file_path, framework="pt", device="cpu") as f:
+        for key in f.keys():
+            tensor = f.get_tensor(key)
+            tensors[key] = tensor
+
+for key in tensors:
+    if "moe_mlp" not in key:
+        new_state_dict[key] = tensors[key]
+    elif "moe_mlp.output_experts" in key:
+        layer_num = int(re.search(r"\d+", key).group())
+        for i, tensor in enumerate(torch.unbind(tensors[key])):
+            new_state_dict[
+                f"model.layers.{layer_num}.mlp.experts.{i}.down_proj.weight"
+            ] = tensor.contiguous()
+    elif "moe_mlp.experts" in key:
+        layer_num = int(re.search(r"\d+", key).group())
+        for i, tensor in enumerate(torch.unbind(tensors[key])):
+            (
+                new_state_dict[
+                    f"model.layers.{layer_num}.mlp.experts.{i}.up_proj.weight"
+                ],
+                new_state_dict[
+                    f"model.layers.{layer_num}.mlp.experts.{i}.gate_proj.weight"
+                ],
+            ) = torch.chunk(tensor, 2, dim=0)
+
+
+model_standard_moe.load_state_dict(new_state_dict, strict=True, assign=True)
+model_standard_moe.save_pretrained(output_model_path)
+cfg_standard_moe.save_pretrained(output_model_path)
+
+for i in ["tokenizer_config.json", "tokenizer.json", "chat_template.jinja"]:
+    shutil.copy(input_model + "/" + i, output_model_path + "/" + i)

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": [
+    154820,
+    154827,
+    154829
+  ],
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 154820,
+  "transformers_version": "5.0.0",
+  "use_cache": true
+}

model-00001-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:098d0ed1cd2b8179767c229c863f8ec71bfbc568765c24873c45969e165d3cdf
+size 49936318984

model-00002-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a1a35aabc923ea036b25e8026d5278ed305521132273eefea9c1703082153e2
+size 9950566272

model.safetensors.index.json

@@ -0,0 +1,759 @@
+{
+  "metadata": {
+    "total_parameters": 29943390976,
+    "total_size": 59886793728
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00002-of-00002.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.21.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.22.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.22.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.23.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.23.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.24.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.24.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.25.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.25.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.26.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.26.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.27.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.27.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.28.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.28.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.29.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.29.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.30.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.31.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.31.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.32.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.32.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.33.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.33.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.34.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.34.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.35.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.35.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.36.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.36.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.37.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.37.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.38.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.38.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.39.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.39.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.39.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.39.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.39.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.39.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.39.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.39.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.39.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.40.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.40.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.mlp.moe_mlp.experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.self_attn.kv_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.self_attn.kv_a_proj_with_mqa.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.self_attn.kv_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.self_attn.q_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.self_attn.q_a_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.40.self_attn.q_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.41.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.mlp.moe_mlp.experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.self_attn.kv_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.self_attn.kv_a_proj_with_mqa.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.self_attn.kv_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.self_attn.q_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.self_attn.q_a_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.41.self_attn.q_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.42.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.mlp.moe_mlp.experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.self_attn.kv_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.self_attn.kv_a_proj_with_mqa.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.self_attn.kv_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.self_attn.q_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.self_attn.q_a_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.42.self_attn.q_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.43.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.mlp.moe_mlp.experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.self_attn.kv_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.self_attn.kv_a_proj_with_mqa.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.self_attn.kv_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.self_attn.q_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.self_attn.q_a_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.43.self_attn.q_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.44.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.mlp.moe_mlp.experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.self_attn.kv_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.self_attn.kv_a_proj_with_mqa.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.self_attn.kv_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.self_attn.q_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.self_attn.q_a_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.44.self_attn.q_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.45.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.mlp.moe_mlp.experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.self_attn.kv_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.self_attn.kv_a_proj_with_mqa.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.self_attn.kv_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.self_attn.q_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.self_attn.q_a_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.45.self_attn.q_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.mlp.gate.e_score_correction_bias": "model-00002-of-00002.safetensors",
+    "model.layers.46.mlp.gate.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.mlp.moe_mlp.experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.mlp.moe_mlp.output_experts.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.mlp.shared_experts.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.mlp.shared_experts.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.mlp.shared_experts.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.self_attn.kv_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.self_attn.kv_a_proj_with_mqa.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.self_attn.kv_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.self_attn.q_a_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.self_attn.q_a_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.46.self_attn.q_b_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.gate.e_score_correction_bias": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.gate.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.moe_mlp.experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.moe_mlp.output_experts.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.shared_experts.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.shared_experts.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.shared_experts.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.kv_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.kv_a_proj_with_mqa.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.kv_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.q_a_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.q_a_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.q_b_proj.weight": "model-00001-of-00002.safetensors",
+    "model.norm.weight": "model-00002-of-00002.safetensors"
+  }
+}

modeling_glm4_moe_lite_for_backconvert.py

@@ -0,0 +1,743 @@
+# 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.
+
+
+import math
+from collections.abc import Callable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from transformers import initialization as init
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernel_func_from_hub
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, is_grouped_mm_available
+from transformers.utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
+
+try:
+    from .configuration_glm4_moe_lite_for_backconvert import Glm4MoeLiteConfig
+except:
+    from configuration_glm4_moe_lite_for_backconvert import Glm4MoeLiteConfig
+
+
+class Glm4MoeLiteRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Glm4MoeLiteConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: Glm4MoeLiteConfig | None = None,
+        device: Optional["torch.device"] = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        dim = int(head_dim * partial_rotary_factor)
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @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 maybe_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)
+
+
+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)
+
+
+@use_kernel_func_from_hub("rotary_pos_emb")
+def apply_rotary_pos_emb(q, k, cos, sin, 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.
+        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)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+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: torch.Tensor | None,
+    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 apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    r"""
+    TODO let's just use the original freqcis computation to not have the view
+    transpose + reshape! This is not optimized!
+    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`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        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)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class Glm4MoeLiteAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Glm4MoeLiteConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+        self.num_heads = config.num_attention_heads
+
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.qk_head_dim = config.qk_head_dim
+
+        self.is_causal = True
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
+        else:
+            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
+            self.q_a_layernorm = Glm4MoeLiteRMSNorm(config.q_lora_rank)
+            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            config.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = Glm4MoeLiteRMSNorm(self.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+        if self.config.rope_parameters.get("rope_type", "default") != "default":
+            mscale_all_dim = self.config.rope_parameters.get("mscale_all_dim", 0)
+            scaling_factor = self.config.rope_parameters["factor"]
+            if mscale_all_dim:
+                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
+                self.scaling = self.scaling * mscale * mscale
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
+        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)
+
+        if self.q_lora_rank is None:
+            q_states = self.q_proj(hidden_states)
+        else:
+            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+
+        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        cos, sin = position_embeddings
+        if self.config.rope_interleave:  # support using interleaved weights for efficiency
+            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
+        else:
+            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
+
+        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,
+        )
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Glm4MoeLiteMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.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 Glm4MoeLiteTopkRouter(nn.Module):
+    def __init__(self, config: Glm4MoeLiteConfig):
+        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))
+
+    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))
+        return router_logits
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class Glm4MoeLiteRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Glm4MoeLiteRMSNorm 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}"
+
+
+#@use_experts_implementation
+# class Glm4MoeLiteNaiveMoe(nn.Module):
+#     """Collection of expert weights stored as 3D tensors."""
+
+#     def __init__(self, config):
+#         super().__init__()
+#         self.num_experts = config.num_local_experts
+#         self.hidden_dim = config.hidden_size
+#         self.intermediate_dim = config.moe_intermediate_size
+#         self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
+#         self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
+#         self.act_fn = ACT2FN[config.hidden_act]
+
+#     def forward(
+#         self,
+#         hidden_states: torch.Tensor,
+#         top_k_index: torch.Tensor,
+#         top_k_weights: torch.Tensor,
+#     ) -> torch.Tensor:
+#         final_hidden_states = torch.zeros_like(hidden_states)
+#         with torch.no_grad():
+#             expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
+#             expert_mask = expert_mask.permute(2, 1, 0)
+#             expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+
+#         for expert_idx in expert_hit:
+#             expert_idx = expert_idx[0]
+#             if expert_idx == self.num_experts:
+#                 continue
+#             top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+#             current_state = hidden_states[token_idx]
+#             gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
+#             current_hidden_states = self.act_fn(gate) * up
+#             current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
+#             current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
+#             final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+
+#         return final_hidden_states
+
+
+class Glm4MoeLiteMoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        #self.experts = Glm4MoeLiteNaiveMoe(config)
+        self.experts = nn.ModuleList(
+            [Glm4MoeLiteMLP(
+                config, intermediate_size=config.moe_intermediate_size
+            ) for _ in range(config.n_routed_experts)]
+        )
+        self.gate = Glm4MoeLiteTopkRouter(config)
+        self.shared_experts = Glm4MoeLiteMLP(
+            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
+        )
+        self.n_routed_experts = config.n_routed_experts
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.top_k = config.num_experts_per_tok
+
+    def route_tokens_to_experts(self, router_logits):
+        router_logits = router_logits.sigmoid()
+        router_logits_for_choice = router_logits + self.gate.e_score_correction_bias
+        group_scores = (
+            router_logits_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 = router_logits_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]
+        topk_weights = router_logits.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
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        router_logits = self.gate(hidden_states)
+        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+class Glm4MoeLiteDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: Glm4MoeLiteConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Glm4MoeLiteAttention(config, layer_idx)
+
+        if config.mlp_layer_types[layer_idx] == "sparse":
+            self.mlp = Glm4MoeLiteMoE(config)
+        else:
+            self.mlp = Glm4MoeLiteMLP(config)
+
+        self.input_layernorm = Glm4MoeLiteRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_layernorm = Glm4MoeLiteRMSNorm(config.hidden_size, config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> 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_values=past_key_values,
+            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 Glm4MoeLitePreTrainedModel(PreTrainedModel):
+    config: Glm4MoeLiteConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Glm4MoeLiteDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": Glm4MoeLiteDecoderLayer,
+        "attentions": Glm4MoeLiteAttention,
+    }
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, Glm4MoeLiteTopkRouter):
+            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            init.zeros_(module.e_score_correction_bias)
+        # elif isinstance(module, Glm4MoeLiteNaiveMoe):
+        #     init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
+        #     init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
+
+
+@auto_docstring
+class Glm4MoeLiteModel(Glm4MoeLitePreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.47.*"]
+
+    def __init__(self, config: Glm4MoeLiteConfig):
+        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(
+            [Glm4MoeLiteDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Glm4MoeLiteRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Glm4MoeLiteRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        cache_position: torch.LongTensor | None = None,
+        use_cache: bool | None = 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(config=self.config)
+
+        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(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            )
+
+        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=position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_embeddings=position_embeddings,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class Glm4MoeLiteForCausalLM(Glm4MoeLitePreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Glm4MoeLiteModel(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()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Glm4MoeLiteForCausalLM
+
+        >>> model = Glm4MoeLiteForCausalLM.from_pretrained("meta-glm4_moe_lite/Glm4MoeLite-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-glm4_moe_lite/Glm4MoeLite-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__ = ["Glm4MoeLitePreTrainedModel", "Glm4MoeLiteModel", "Glm4MoeLiteForCausalLM"]

modeling_glm4_moe_lite_scm.py

@@ -0,0 +1,745 @@
+# 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.
+
+
+import math
+from collections.abc import Callable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from transformers import initialization as init
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernel_func_from_hub
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, is_grouped_mm_available
+from transformers.utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
+
+import scattermoe
+
+try:
+    from .configuration_glm4_moe_lite_scm import Glm4MoeLiteSCMConfig
+except:
+    from configuration_glm4_moe_lite_scm import Glm4MoeLiteSCMConfig
+
+
+class Glm4MoeLiteSCMRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Glm4MoeLiteSCMConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: Glm4MoeLiteSCMConfig | None = None,
+        device: Optional["torch.device"] = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        dim = int(head_dim * partial_rotary_factor)
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @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 maybe_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)
+
+
+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)
+
+
+@use_kernel_func_from_hub("rotary_pos_emb")
+def apply_rotary_pos_emb(q, k, cos, sin, 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.
+        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)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+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: torch.Tensor | None,
+    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 apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    r"""
+    TODO let's just use the original freqcis computation to not have the view
+    transpose + reshape! This is not optimized!
+    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`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        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)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class Glm4MoeLiteSCMAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Glm4MoeLiteSCMConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+        self.num_heads = config.num_attention_heads
+
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.qk_head_dim = config.qk_head_dim
+
+        self.is_causal = True
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
+        else:
+            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
+            self.q_a_layernorm = Glm4MoeLiteSCMRMSNorm(config.q_lora_rank)
+            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            config.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = Glm4MoeLiteSCMRMSNorm(self.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+        if self.config.rope_parameters.get("rope_type", "default") != "default":
+            mscale_all_dim = self.config.rope_parameters.get("mscale_all_dim", 0)
+            scaling_factor = self.config.rope_parameters["factor"]
+            if mscale_all_dim:
+                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
+                self.scaling = self.scaling * mscale * mscale
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
+        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)
+
+        if self.q_lora_rank is None:
+            q_states = self.q_proj(hidden_states)
+        else:
+            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+
+        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        cos, sin = position_embeddings
+        if self.config.rope_interleave:  # support using interleaved weights for efficiency
+            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
+        else:
+            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
+
+        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,
+        )
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Glm4MoeLiteSCMMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.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 Glm4MoeLiteSCMTopkRouter(nn.Module):
+    def __init__(self, config: Glm4MoeLiteSCMConfig):
+        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))
+
+    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))
+        return router_logits
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class Glm4MoeLiteSCMRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Glm4MoeLiteSCMRMSNorm 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}"
+
+
+#@use_experts_implementation
+# class _Glm4MoeLiteSCMNaiveMoe(nn.Module):
+#     """Collection of expert weights stored as 3D tensors."""
+
+#     def __init__(self, config):
+#         super().__init__()
+#         self.num_experts = config.num_local_experts
+#         self.hidden_dim = config.hidden_size
+#         self.intermediate_dim = config.moe_intermediate_size
+#         self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
+#         self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
+#         self.act_fn = ACT2FN[config.hidden_act]
+
+#     def forward(
+#         self,
+#         hidden_states: torch.Tensor,
+#         top_k_index: torch.Tensor,
+#         top_k_weights: torch.Tensor,
+#     ) -> torch.Tensor:
+#         final_hidden_states = torch.zeros_like(hidden_states)
+#         with torch.no_grad():
+#             expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
+#             expert_mask = expert_mask.permute(2, 1, 0)
+#             expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+
+#         for expert_idx in expert_hit:
+#             expert_idx = expert_idx[0]
+#             if expert_idx == self.num_experts:
+#                 continue
+#             top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+#             current_state = hidden_states[token_idx]
+#             gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
+#             current_hidden_states = self.act_fn(gate) * up
+#             current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
+#             current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
+#             final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+
+#         return final_hidden_states
+
+
+class Glm4MoeLiteSCMMoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.moe_mlp = scattermoe.mlp.GLUMLP(
+            input_size=self.config.hidden_size,
+            hidden_size=self.config.moe_intermediate_size,
+            num_experts=self.config.n_routed_experts,
+            top_k=self.config.num_experts_per_tok,
+            activation=ACT2FN[config.hidden_act],
+        )
+        self.gate = Glm4MoeLiteSCMTopkRouter(config)
+        self.shared_experts = Glm4MoeLiteSCMMLP(
+            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
+        )
+        self.n_routed_experts = config.n_routed_experts
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.top_k = config.num_experts_per_tok
+
+    def route_tokens_to_experts(self, router_logits):
+        router_logits = router_logits.sigmoid()
+        router_logits_for_choice = router_logits + self.gate.e_score_correction_bias
+        group_scores = (
+            router_logits_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 = router_logits_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]
+        topk_weights = router_logits.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
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        router_logits = self.gate(hidden_states)
+        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+
+        hidden_states = self.moe_mlp(hidden_states, topk_weights.to(torch.bfloat16), topk_indices).view(*orig_shape)
+
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+class Glm4MoeLiteSCMDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: Glm4MoeLiteSCMConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Glm4MoeLiteSCMAttention(config, layer_idx)
+
+        if config.mlp_layer_types[layer_idx] == "sparse":
+            self.mlp = Glm4MoeLiteSCMMoE(config)
+        else:
+            self.mlp = Glm4MoeLiteSCMMLP(config)
+
+        self.input_layernorm = Glm4MoeLiteSCMRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_layernorm = Glm4MoeLiteSCMRMSNorm(config.hidden_size, config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> 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_values=past_key_values,
+            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 Glm4MoeLiteSCMPreTrainedModel(PreTrainedModel):
+    config: Glm4MoeLiteSCMConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Glm4MoeLiteSCMDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": Glm4MoeLiteSCMDecoderLayer,
+        "attentions": Glm4MoeLiteSCMAttention,
+    }
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, Glm4MoeLiteSCMTopkRouter):
+            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            init.zeros_(module.e_score_correction_bias)
+
+
+@auto_docstring
+class Glm4MoeLiteSCMModel(Glm4MoeLiteSCMPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.47.*"]
+
+    def __init__(self, config: Glm4MoeLiteSCMConfig):
+        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(
+            [Glm4MoeLiteSCMDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Glm4MoeLiteSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Glm4MoeLiteSCMRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        cache_position: torch.LongTensor | None = None,
+        use_cache: bool | None = 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(config=self.config)
+
+        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(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            )
+
+        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=position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_embeddings=position_embeddings,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class Glm4MoeLiteSCMForCausalLM(Glm4MoeLiteSCMPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Glm4MoeLiteSCMModel(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()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Glm4MoeLiteSCMForCausalLM
+
+        >>> model = Glm4MoeLiteSCMForCausalLM.from_pretrained("meta-glm4_moe_lite/Glm4MoeLiteSCM-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-glm4_moe_lite/Glm4MoeLiteSCM-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__ = ["Glm4MoeLiteSCMPreTrainedModel", "Glm4MoeLiteSCMModel", "Glm4MoeLiteSCMForCausalLM"]

modeling_glm4_moe_lite_scm_liger.py

@@ -0,0 +1,724 @@
+# 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.
+
+
+import math
+from collections.abc import Callable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from transformers import initialization as init
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernel_func_from_hub
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, is_grouped_mm_available
+from transformers.utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
+
+import scattermoe
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+try:
+    from .configuration_glm4_moe_lite_scm import Glm4MoeLiteSCMConfig
+except:
+    from configuration_glm4_moe_lite_scm import Glm4MoeLiteSCMConfig
+
+
+class Glm4MoeLiteSCMRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Glm4MoeLiteSCMConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: Glm4MoeLiteSCMConfig | None = None,
+        device: Optional["torch.device"] = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        dim = int(head_dim * partial_rotary_factor)
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @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 maybe_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)
+
+
+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)
+
+
+@use_kernel_func_from_hub("rotary_pos_emb")
+def apply_rotary_pos_emb(q, k, cos, sin, 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.
+        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)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+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: torch.Tensor | None,
+    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 apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    r"""
+    TODO let's just use the original freqcis computation to not have the view
+    transpose + reshape! This is not optimized!
+    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`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        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)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class Glm4MoeLiteSCMAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Glm4MoeLiteSCMConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+        self.num_heads = config.num_attention_heads
+
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.qk_head_dim = config.qk_head_dim
+
+        self.is_causal = True
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
+        else:
+            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
+            self.q_a_layernorm = Glm4MoeLiteSCMRMSNorm(config.q_lora_rank)
+            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            config.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = Glm4MoeLiteSCMRMSNorm(self.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+        if self.config.rope_parameters.get("rope_type", "default") != "default":
+            mscale_all_dim = self.config.rope_parameters.get("mscale_all_dim", 0)
+            scaling_factor = self.config.rope_parameters["factor"]
+            if mscale_all_dim:
+                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
+                self.scaling = self.scaling * mscale * mscale
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
+        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)
+
+        if self.q_lora_rank is None:
+            q_states = self.q_proj(hidden_states)
+        else:
+            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+
+        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        cos, sin = position_embeddings
+        if self.config.rope_interleave:  # support using interleaved weights for efficiency
+            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
+        else:
+            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
+
+        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,
+        )
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Glm4MoeLiteSCMMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.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 Glm4MoeLiteSCMTopkRouter(nn.Module):
+    def __init__(self, config: Glm4MoeLiteSCMConfig):
+        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))
+
+    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))
+        return router_logits
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class Glm4MoeLiteSCMRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Glm4MoeLiteSCMRMSNorm 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 Glm4MoeLiteSCMMoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.moe_mlp = scattermoe.mlp.GLUMLP(
+            input_size=self.config.hidden_size,
+            hidden_size=self.config.moe_intermediate_size,
+            num_experts=self.config.n_routed_experts,
+            top_k=self.config.num_experts_per_tok,
+            activation=ACT2FN[config.hidden_act],
+        )
+        self.gate = Glm4MoeLiteSCMTopkRouter(config)
+        self.shared_experts = Glm4MoeLiteSCMMLP(
+            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
+        )
+        self.n_routed_experts = config.n_routed_experts
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.top_k = config.num_experts_per_tok
+
+    def route_tokens_to_experts(self, router_logits):
+        router_logits = router_logits.sigmoid()
+        router_logits_for_choice = router_logits + self.gate.e_score_correction_bias
+        group_scores = (
+            router_logits_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 = router_logits_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]
+        topk_weights = router_logits.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
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        router_logits = self.gate(hidden_states)
+        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+
+        hidden_states = self.moe_mlp(hidden_states, topk_weights.to(torch.bfloat16), topk_indices).view(*orig_shape)
+
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+class Glm4MoeLiteSCMDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: Glm4MoeLiteSCMConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Glm4MoeLiteSCMAttention(config, layer_idx)
+
+        if config.mlp_layer_types[layer_idx] == "sparse":
+            self.mlp = Glm4MoeLiteSCMMoE(config)
+        else:
+            self.mlp = Glm4MoeLiteSCMMLP(config)
+
+        self.input_layernorm = Glm4MoeLiteSCMRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_layernorm = Glm4MoeLiteSCMRMSNorm(config.hidden_size, config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> 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_values=past_key_values,
+            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 Glm4MoeLiteSCMPreTrainedModel(PreTrainedModel):
+    config: Glm4MoeLiteSCMConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Glm4MoeLiteSCMDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": Glm4MoeLiteSCMDecoderLayer,
+        "attentions": Glm4MoeLiteSCMAttention,
+    }
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, Glm4MoeLiteSCMTopkRouter):
+            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            init.zeros_(module.e_score_correction_bias)
+
+
+@auto_docstring
+class Glm4MoeLiteSCMModel(Glm4MoeLiteSCMPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.47.*"]
+
+    def __init__(self, config: Glm4MoeLiteSCMConfig):
+        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(
+            [Glm4MoeLiteSCMDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Glm4MoeLiteSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Glm4MoeLiteSCMRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        cache_position: torch.LongTensor | None = None,
+        use_cache: bool | None = 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(config=self.config)
+
+        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(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            )
+
+        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=position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_embeddings=position_embeddings,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class Glm4MoeLiteSCMForCausalLM(Glm4MoeLiteSCMPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Glm4MoeLiteSCMModel(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()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Glm4MoeLiteSCMForCausalLM
+
+        >>> model = Glm4MoeLiteSCMForCausalLM.from_pretrained("meta-glm4_moe_lite/Glm4MoeLiteSCM-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-glm4_moe_lite/Glm4MoeLiteSCM-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
+        kept_hidden_states = hidden_states[:, slice_indices, :]
+        shift_labels = kwargs.pop("shift_labels", None)
+        logits = None
+
+        loss = None
+
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+        if skip_logits:
+            loss = LigerForCausalLMLoss(
+                hidden_states=kept_hidden_states,
+                lm_head_weight=self.lm_head.weight,
+                labels=labels,
+                shift_labels=shift_labels,
+                hidden_size=self.config.hidden_size,
+                **kwargs,
+            )
+        else:
+            logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+            if labels is not None:
+                loss = self.loss_function(logits, labels, self.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__ = ["Glm4MoeLiteSCMPreTrainedModel", "Glm4MoeLiteSCMModel", "Glm4MoeLiteSCMForCausalLM"]

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d
+size 20217442

tokenizer_config.json

@@ -0,0 +1,321 @@
+{
+  "added_tokens_decoder": {
+    "154820": {
+      "content": "<|endoftext|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154821": {
+      "content": "[MASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154822": {
+      "content": "[gMASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154823": {
+      "content": "[sMASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154824": {
+      "content": "<sop>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154825": {
+      "content": "<eop>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154826": {
+      "content": "<|system|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154827": {
+      "content": "<|user|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154828": {
+      "content": "<|assistant|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154829": {
+      "content": "<|observation|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154830": {
+      "content": "<|begin_of_image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154831": {
+      "content": "<|end_of_image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154832": {
+      "content": "<|begin_of_video|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154833": {
+      "content": "<|end_of_video|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154834": {
+      "content": "<|begin_of_audio|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154835": {
+      "content": "<|end_of_audio|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154836": {
+      "content": "<|begin_of_transcription|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154837": {
+      "content": "<|end_of_transcription|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154838": {
+      "content": "<|code_prefix|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154839": {
+      "content": "<|code_middle|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154840": {
+      "content": "<|code_suffix|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154841": {
+      "content": "<think>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154842": {
+      "content": "</think>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154843": {
+      "content": "<tool_call>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154844": {
+      "content": "</tool_call>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154845": {
+      "content": "<tool_response>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154846": {
+      "content": "</tool_response>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154847": {
+      "content": "<arg_key>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154848": {
+      "content": "</arg_key>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154849": {
+      "content": "<arg_value>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154850": {
+      "content": "</arg_value>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154851": {
+      "content": "/nothink",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154852": {
+      "content": "<|begin_of_box|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154853": {
+      "content": "<|end_of_box|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154854": {
+      "content": "<|image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154855": {
+      "content": "<|video|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|endoftext|>",
+    "[MASK]",
+    "[gMASK]",
+    "[sMASK]",
+    "<sop>",
+    "<eop>",
+    "<|system|>",
+    "<|user|>",
+    "<|assistant|>",
+    "<|observation|>",
+    "<|begin_of_image|>",
+    "<|end_of_image|>",
+    "<|begin_of_video|>",
+    "<|end_of_video|>",
+    "<|begin_of_audio|>",
+    "<|end_of_audio|>",
+    "<|begin_of_transcription|>",
+    "<|end_of_transcription|>"
+  ],
+  "clean_up_tokenization_spaces": false,
+  "do_lower_case": false,
+  "eos_token": "<|endoftext|>",
+  "extra_special_tokens": {},
+  "model_max_length": 128000,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "left",
+  "remove_space": false,
+  "tokenizer_class": "PreTrainedTokenizer"
+}