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 +1,68 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- if '</think>' in content %}
+            {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+        {%- endif %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {% generation %}
+        {{- content }}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>' }}
+        {% endgeneration %}
+        {{- '\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n<think>\n' }}
+{%- endif %}

config.json

@@ -0,0 +1,79 @@
+{
+  "architectures": [
+    "AfmoeSCMForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_afmoe_scm.AfmoeSCMConfig",
+    "AutoModel": "modeling_afmoe_scm.AfmoeSCMModel",
+    "AutoModelForCausalLM": "modeling_afmoe_scm.AfmoeSCMForCausalLM"
+  },
+  "dtype": "bfloat16",
+  "global_attn_every_n_layers": 4,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "load_balance_coeff": 0.001,
+  "max_position_embeddings": 131072,
+  "model_type": "afmoe",
+  "moe_intermediate_size": 1024,
+  "mup_enabled": true,
+  "n_group": 1,
+  "num_attention_heads": 32,
+  "num_dense_layers": 2,
+  "num_expert_groups": 1,
+  "num_experts": 128,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 4,
+  "num_limited_groups": 1,
+  "num_shared_experts": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000,
+  "route_norm": true,
+  "route_scale": 2.826,
+  "score_func": "sigmoid",
+  "sliding_window": 2048,
+  "tie_word_embeddings": false,
+  "topk_group": 1,
+  "transformers_version": "4.57.3",
+  "use_cache": true,
+  "vocab_size": 200192
+}

configuration_afmoe_scm.py

@@ -0,0 +1,133 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.configuration_utils import layer_type_validation
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+class AfmoeSCMConfig(PretrainedConfig):
+    """
+    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).
+    """
+    model_type = "afmoe"
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        num_hidden_layers: int = 32,
+        vocab_size: int = 200192,
+        hidden_size: int = 2048,
+        intermediate_size: int = 6144,
+        moe_intermediate_size=1408,
+        num_dense_layers=1,
+        num_attention_heads=16,
+        num_key_value_heads=None,
+        head_dim=128,
+        hidden_act="silu",
+        max_position_embeddings=16384,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        num_experts=64,
+        num_experts_per_tok=6,
+        num_shared_experts=2,
+        num_expert_groups=1,
+        num_limited_groups=1,
+        score_func="sigmoid",
+        route_norm=True,
+        route_scale=1.0,
+        global_attn_every_n_layers=4,
+        sliding_window=1024,
+        mup_enabled=False,
+        layer_types=None,
+        attention_dropout: float = 0.0,
+        n_group: int = 1,
+        topk_group: int = 1,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_dense_layers = num_dense_layers
+        self.num_attention_heads = num_attention_heads
+        self.head_dim = head_dim
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        
+        
+        # MoE specific
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_experts_per_tok = num_experts_per_tok
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts = num_experts
+        self.num_shared_experts = num_shared_experts
+        self.num_expert_groups = num_expert_groups
+        self.num_limited_groups = num_limited_groups
+        self.score_func = score_func
+        self.route_norm = route_norm
+        self.route_scale = route_scale
+
+
+        # Attention specific
+        self.attention_dropout = attention_dropout
+        self.global_attn_every_n_layers = global_attn_every_n_layers
+        self.sliding_window = sliding_window
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention" if bool((i + 1) % global_attn_every_n_layers) else "full_attention" for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+        # muP specific
+        self.mup_enabled = mup_enabled
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+
+
+        # Validate rope configs
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["AfmoeSCMConfig"]

convert_hf_to_scm.py

@@ -0,0 +1,128 @@
+import glob
+import re
+import shutil
+import sys
+
+import accelerate
+import torch
+from configuration_afmoe_scm import AfmoeSCMConfig
+from modeling_afmoe_scm import AfmoeSCMForCausalLM
+from configuration_afmoe import AfmoeConfig
+from safetensors import safe_open
+
+input_model = sys.argv[1]
+output_model_path = sys.argv[2]
+
+auto_map = {
+    "AutoConfig": "configuration_afmoe_scm.AfmoeSCMConfig",
+    "AutoModel": "modeling_afmoe_scm.AfmoeSCMModel",
+    "AutoModelForCausalLM": "modeling_afmoe_scm.AfmoeSCMForCausalLM"
+}
+
+cfg_standard_moe = AfmoeConfig.from_pretrained(input_model)
+cfg_shared_moe = AfmoeSCMConfig(
+    auto_map=auto_map,
+    layer_types=cfg_standard_moe.layer_types,
+    global_attn_every_n_layers=cfg_standard_moe.global_attn_every_n_layers,
+    load_balance_coeff=cfg_standard_moe.load_balance_coeff,
+    mup_enabled=cfg_standard_moe.mup_enabled,
+    num_dense_layers=cfg_standard_moe.num_dense_layers,
+    num_expert_groups=cfg_standard_moe.num_expert_groups,
+    num_limited_groups=cfg_standard_moe.num_limited_groups,
+    route_norm=cfg_standard_moe.route_norm,
+    route_scale=cfg_standard_moe.route_scale,
+    score_func=cfg_standard_moe.score_func,
+    topk_group=cfg_standard_moe.topk_group,
+    num_shared_experts=cfg_standard_moe.num_shared_experts,
+    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,
+    use_cache=cfg_standard_moe.use_cache,
+    tie_word_embeddings=cfg_standard_moe.tie_word_embeddings,
+    rope_theta=cfg_standard_moe.rope_theta,
+    rope_scaling=cfg_standard_moe.rope_scaling,
+    sliding_window=cfg_standard_moe.sliding_window,
+    attention_dropout=cfg_standard_moe.attention_dropout,
+    moe_intermediate_size=cfg_standard_moe.moe_intermediate_size,
+    num_experts_per_tok=cfg_standard_moe.num_experts_per_tok,
+    num_experts=cfg_standard_moe.num_experts,
+    head_dim=cfg_standard_moe.head_dim,
+    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.num_experts
+
+with accelerate.init_empty_weights():
+    model_shared_moe = AfmoeSCMForCausalLM(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 "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_afmoe_scm.py",
+    output_model_path + "/" + "modeling_afmoe_scm.py",
+)
+shutil.copy(
+    "configuration_afmoe_scm.py",
+    output_model_path + "/" + "configuration_afmoe_scm.py",
+)
+for i in ["special_tokens_map.json", "tokenizer_config.json", "tokenizer.json", "chat_template.jinja"]:
+    shutil.copy(input_model + "/" + i, output_model_path + "/" + i)

convert_scm_to_hf.py

@@ -0,0 +1,117 @@
+import glob
+import re
+import shutil
+import sys
+
+import accelerate
+import torch
+from safetensors import safe_open
+from configuration_afmoe_scm import AfmoeSCMConfig
+from modeling_afmoe import AfmoeForCausalLM
+from configuration_afmoe import AfmoeConfig
+
+input_model = sys.argv[1]
+output_model_path = sys.argv[2]
+
+auto_map = {
+    "AutoConfig": "configuration_afmoe.AfmoeConfig",
+    "AutoModel": "modeling_afmoe.AfmoeModel",
+    "AutoModelForCausalLM": "modeling_afmoe.AfmoeForCausalLM"
+}
+cfg_shared_moe = AfmoeSCMConfig.from_pretrained(input_model)
+cfg_standard_moe = AfmoeConfig(
+    auto_map=auto_map,
+    layer_types=cfg_shared_moe.layer_types,
+    global_attn_every_n_layers=cfg_shared_moe.global_attn_every_n_layers,
+    load_balance_coeff=cfg_shared_moe.load_balance_coeff,
+    mup_enabled=cfg_shared_moe.mup_enabled,
+    num_dense_layers=cfg_shared_moe.num_dense_layers,
+    num_expert_groups=cfg_shared_moe.num_expert_groups,
+    num_limited_groups=cfg_shared_moe.num_limited_groups,
+    route_norm=cfg_shared_moe.route_norm,
+    route_scale=cfg_shared_moe.route_scale,
+    score_func=cfg_shared_moe.score_func,
+    topk_group=cfg_shared_moe.topk_group,
+    num_shared_experts=cfg_shared_moe.num_shared_experts,
+    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,
+    use_cache=cfg_shared_moe.use_cache,
+    tie_word_embeddings=cfg_shared_moe.tie_word_embeddings,
+    rope_theta=cfg_shared_moe.rope_theta,
+    rope_scaling=cfg_shared_moe.rope_scaling,
+    sliding_window=cfg_shared_moe.sliding_window,
+    attention_dropout=cfg_shared_moe.attention_dropout,
+    moe_intermediate_size=cfg_shared_moe.moe_intermediate_size,
+    num_experts_per_tok=cfg_shared_moe.num_experts_per_tok,
+    num_experts=cfg_shared_moe.num_experts,
+    head_dim=cfg_shared_moe.head_dim,
+    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_shared_moe.num_experts
+
+with accelerate.init_empty_weights():
+    model_standard_moe = AfmoeForCausalLM(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)
+
+shutil.copy(
+    "modeling_afmoe.py",
+    output_model_path + "/" + "modeling_afmoe.py",
+)
+shutil.copy(
+    "configuration_afmoe.py",
+    output_model_path + "/" + "configuration_afmoe.py",
+)
+
+for i in ["special_tokens_map.json", "tokenizer_config.json", "tokenizer.json", "chat_template.jinja"]:
+    shutil.copy(input_model + "/" + i, output_model_path + "/" + i)
+

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "4.57.3"
+}

model-00001-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6e76bc7093ed7b73fca644c785e14eca4d7f6297018b8fab5004c36bd512182
+size 4504252888

model-00002-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ea06032723cbc8909ce6eb47f26852316f288570933262287292410bd029007
+size 4988115224

model-00003-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:940e643ea5a42d83fcea1eccd458908d675862f92d8c56ebdb3a687c54cbbf47
+size 4476952592

model-00004-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5ba37cbf9942002c7a0ac95237771d8bf549f9eb555c9e95800fdbeb3ee9d39
+size 3961053608

model-00005-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:702192634903333bf135612a0a11ac04d916e2b4bec44b26fbdf0169184583b9
+size 4988115264

model-00006-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b8dba99959db39cb05970a5aec1e0c849f9847a8f01c136096a5a101d313ecc
+size 4476952640

model-00007-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a79061db20e5ee5b92f1aa79fb94e7065aef1d663e18abe2209375940fd42993
+size 3961053624

model-00008-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:539a15a7c3e16efe21767557dd55de0ef848e2dbc46dc27c33505a1746240e5f
+size 4988115264

model-00009-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79cc9fcefee2a2d72cdb5b2fc7e46c65483883a9a41bf8c063caa8bcdda5bd10
+size 4476952640

model-00010-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:654643950d9f4373bd130d163e7ed2ae19651f4154a2d901b8409b0b507f38e0
+size 3961053624

model-00011-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:906da265ef65fc9087d9e1bb192df384a9827170b600019f2b1ec2c3dcd6eae3
+size 4988115264

model-00012-of-00012.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:776c52063a0936746fa9bf67d82edbb4395399699e8fadbaee88fbae0004c572
+size 2477284056

model.safetensors.index.json

@@ -0,0 +1,579 @@
+{
+  "metadata": {
+    "total_parameters": 26123974400,
+    "total_size": 52247948800
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00012-of-00012.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.post_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.pre_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.self_attn.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.self_attn.k_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.self_attn.q_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.post_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.pre_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.self_attn.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.self_attn.k_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.self_attn.q_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.mlp.expert_bias": "model-00004-of-00012.safetensors",
+    "model.layers.10.mlp.moe_mlp.experts.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.mlp.moe_mlp.output_experts.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.mlp.router.gate.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.mlp.shared_experts.down_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.mlp.shared_experts.gate_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.mlp.shared_experts.up_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.post_mlp_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.pre_mlp_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.self_attn.gate_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.self_attn.k_norm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.self_attn.q_norm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.mlp.expert_bias": "model-00004-of-00012.safetensors",
+    "model.layers.11.mlp.moe_mlp.experts.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.mlp.moe_mlp.output_experts.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.mlp.router.gate.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.mlp.shared_experts.down_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.mlp.shared_experts.gate_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.mlp.shared_experts.up_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.post_mlp_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.pre_mlp_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.self_attn.gate_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.self_attn.k_norm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.self_attn.q_norm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.mlp.expert_bias": "model-00004-of-00012.safetensors",
+    "model.layers.12.mlp.moe_mlp.experts.weight": "model-00005-of-00012.safetensors",
+    "model.layers.12.mlp.moe_mlp.output_experts.weight": "model-00005-of-00012.safetensors",
+    "model.layers.12.mlp.router.gate.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.mlp.shared_experts.down_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.mlp.shared_experts.gate_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.mlp.shared_experts.up_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.post_mlp_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.pre_mlp_layernorm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.self_attn.gate_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.self_attn.k_norm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.self_attn.q_norm.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00004-of-00012.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.mlp.expert_bias": "model-00005-of-00012.safetensors",
+    "model.layers.13.mlp.moe_mlp.experts.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.mlp.moe_mlp.output_experts.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.mlp.router.gate.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.mlp.shared_experts.down_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.mlp.shared_experts.gate_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.mlp.shared_experts.up_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.post_mlp_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.pre_mlp_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.self_attn.gate_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.self_attn.k_norm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.self_attn.q_norm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.mlp.expert_bias": "model-00005-of-00012.safetensors",
+    "model.layers.14.mlp.moe_mlp.experts.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.mlp.moe_mlp.output_experts.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.mlp.router.gate.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.mlp.shared_experts.down_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.mlp.shared_experts.gate_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.mlp.shared_experts.up_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.post_mlp_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.pre_mlp_layernorm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.self_attn.gate_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.self_attn.k_norm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.self_attn.q_norm.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.mlp.expert_bias": "model-00006-of-00012.safetensors",
+    "model.layers.15.mlp.moe_mlp.experts.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.mlp.moe_mlp.output_experts.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.mlp.router.gate.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.mlp.shared_experts.down_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.mlp.shared_experts.gate_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.mlp.shared_experts.up_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.post_mlp_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.pre_mlp_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.self_attn.gate_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.self_attn.k_norm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.self_attn.q_norm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00005-of-00012.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.mlp.expert_bias": "model-00006-of-00012.safetensors",
+    "model.layers.16.mlp.moe_mlp.experts.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.mlp.moe_mlp.output_experts.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.mlp.router.gate.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.mlp.shared_experts.down_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.mlp.shared_experts.gate_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.mlp.shared_experts.up_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.post_mlp_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.pre_mlp_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.self_attn.gate_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.self_attn.k_norm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.self_attn.q_norm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.mlp.expert_bias": "model-00006-of-00012.safetensors",
+    "model.layers.17.mlp.moe_mlp.experts.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.mlp.moe_mlp.output_experts.weight": "model-00007-of-00012.safetensors",
+    "model.layers.17.mlp.router.gate.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.mlp.shared_experts.down_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.mlp.shared_experts.gate_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.mlp.shared_experts.up_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.post_mlp_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.pre_mlp_layernorm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.self_attn.gate_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.self_attn.k_norm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.self_attn.q_norm.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00006-of-00012.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.mlp.expert_bias": "model-00007-of-00012.safetensors",
+    "model.layers.18.mlp.moe_mlp.experts.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.mlp.moe_mlp.output_experts.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.mlp.router.gate.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.mlp.shared_experts.down_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.mlp.shared_experts.gate_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.mlp.shared_experts.up_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.post_mlp_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.pre_mlp_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.self_attn.gate_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.self_attn.k_norm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.self_attn.q_norm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.mlp.expert_bias": "model-00007-of-00012.safetensors",
+    "model.layers.19.mlp.moe_mlp.experts.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.mlp.moe_mlp.output_experts.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.mlp.router.gate.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.mlp.shared_experts.down_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.mlp.shared_experts.gate_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.mlp.shared_experts.up_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.post_mlp_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.pre_mlp_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.self_attn.gate_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.self_attn.k_norm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.self_attn.q_norm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.mlp.expert_bias": "model-00001-of-00012.safetensors",
+    "model.layers.2.mlp.moe_mlp.experts.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.mlp.moe_mlp.output_experts.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.mlp.router.gate.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.mlp.shared_experts.down_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.mlp.shared_experts.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.mlp.shared_experts.up_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.post_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.pre_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.self_attn.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.self_attn.k_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.self_attn.q_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.mlp.expert_bias": "model-00007-of-00012.safetensors",
+    "model.layers.20.mlp.moe_mlp.experts.weight": "model-00008-of-00012.safetensors",
+    "model.layers.20.mlp.moe_mlp.output_experts.weight": "model-00008-of-00012.safetensors",
+    "model.layers.20.mlp.router.gate.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.mlp.shared_experts.down_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.mlp.shared_experts.gate_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.mlp.shared_experts.up_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.post_mlp_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.pre_mlp_layernorm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.self_attn.gate_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.self_attn.k_norm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.self_attn.q_norm.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00007-of-00012.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.mlp.expert_bias": "model-00008-of-00012.safetensors",
+    "model.layers.21.mlp.moe_mlp.experts.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.mlp.moe_mlp.output_experts.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.mlp.router.gate.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.mlp.shared_experts.down_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.mlp.shared_experts.gate_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.mlp.shared_experts.up_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.post_mlp_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.pre_mlp_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.self_attn.gate_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.self_attn.k_norm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.self_attn.q_norm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.mlp.expert_bias": "model-00008-of-00012.safetensors",
+    "model.layers.22.mlp.moe_mlp.experts.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.mlp.moe_mlp.output_experts.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.mlp.router.gate.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.mlp.shared_experts.down_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.mlp.shared_experts.gate_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.mlp.shared_experts.up_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.post_mlp_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.pre_mlp_layernorm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.self_attn.gate_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.self_attn.k_norm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.self_attn.q_norm.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.mlp.expert_bias": "model-00009-of-00012.safetensors",
+    "model.layers.23.mlp.moe_mlp.experts.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.mlp.moe_mlp.output_experts.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.mlp.router.gate.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.mlp.shared_experts.down_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.mlp.shared_experts.gate_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.mlp.shared_experts.up_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.post_mlp_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.pre_mlp_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.self_attn.gate_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.self_attn.k_norm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.self_attn.q_norm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00008-of-00012.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.mlp.expert_bias": "model-00009-of-00012.safetensors",
+    "model.layers.24.mlp.moe_mlp.experts.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.mlp.moe_mlp.output_experts.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.mlp.router.gate.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.mlp.shared_experts.down_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.mlp.shared_experts.gate_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.mlp.shared_experts.up_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.post_mlp_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.pre_mlp_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.self_attn.gate_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.self_attn.k_norm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.self_attn.q_norm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.mlp.expert_bias": "model-00009-of-00012.safetensors",
+    "model.layers.25.mlp.moe_mlp.experts.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.mlp.moe_mlp.output_experts.weight": "model-00010-of-00012.safetensors",
+    "model.layers.25.mlp.router.gate.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.mlp.shared_experts.down_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.mlp.shared_experts.gate_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.mlp.shared_experts.up_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.post_mlp_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.pre_mlp_layernorm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.self_attn.gate_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.self_attn.k_norm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.self_attn.q_norm.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00009-of-00012.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.mlp.expert_bias": "model-00010-of-00012.safetensors",
+    "model.layers.26.mlp.moe_mlp.experts.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.mlp.moe_mlp.output_experts.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.mlp.router.gate.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.mlp.shared_experts.down_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.mlp.shared_experts.gate_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.mlp.shared_experts.up_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.post_mlp_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.pre_mlp_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.self_attn.gate_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.self_attn.k_norm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.self_attn.q_norm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.mlp.expert_bias": "model-00010-of-00012.safetensors",
+    "model.layers.27.mlp.moe_mlp.experts.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.mlp.moe_mlp.output_experts.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.mlp.router.gate.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.mlp.shared_experts.down_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.mlp.shared_experts.gate_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.mlp.shared_experts.up_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.post_mlp_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.pre_mlp_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.self_attn.gate_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.self_attn.k_norm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.self_attn.q_norm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.mlp.expert_bias": "model-00010-of-00012.safetensors",
+    "model.layers.28.mlp.moe_mlp.experts.weight": "model-00011-of-00012.safetensors",
+    "model.layers.28.mlp.moe_mlp.output_experts.weight": "model-00011-of-00012.safetensors",
+    "model.layers.28.mlp.router.gate.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.mlp.shared_experts.down_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.mlp.shared_experts.gate_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.mlp.shared_experts.up_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.post_mlp_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.pre_mlp_layernorm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.self_attn.gate_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.self_attn.k_norm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.self_attn.k_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.self_attn.q_norm.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.self_attn.q_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.28.self_attn.v_proj.weight": "model-00010-of-00012.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.mlp.expert_bias": "model-00011-of-00012.safetensors",
+    "model.layers.29.mlp.moe_mlp.experts.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.mlp.moe_mlp.output_experts.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.mlp.router.gate.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.mlp.shared_experts.down_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.mlp.shared_experts.gate_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.mlp.shared_experts.up_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.post_mlp_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.pre_mlp_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.self_attn.gate_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.self_attn.k_norm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.self_attn.k_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.self_attn.q_norm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.self_attn.q_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.29.self_attn.v_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.mlp.expert_bias": "model-00001-of-00012.safetensors",
+    "model.layers.3.mlp.moe_mlp.experts.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.mlp.moe_mlp.output_experts.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.mlp.router.gate.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.mlp.shared_experts.down_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.mlp.shared_experts.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.mlp.shared_experts.up_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.post_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.pre_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.self_attn.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.self_attn.k_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.self_attn.q_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.mlp.expert_bias": "model-00011-of-00012.safetensors",
+    "model.layers.30.mlp.moe_mlp.experts.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.mlp.moe_mlp.output_experts.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.mlp.router.gate.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.mlp.shared_experts.down_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.mlp.shared_experts.gate_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.mlp.shared_experts.up_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.post_mlp_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.pre_mlp_layernorm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.self_attn.gate_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.self_attn.k_norm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.self_attn.k_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.self_attn.q_norm.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.self_attn.q_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.30.self_attn.v_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.mlp.expert_bias": "model-00012-of-00012.safetensors",
+    "model.layers.31.mlp.moe_mlp.experts.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.mlp.moe_mlp.output_experts.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.mlp.router.gate.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.mlp.shared_experts.down_proj.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.mlp.shared_experts.gate_proj.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.mlp.shared_experts.up_proj.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.post_mlp_layernorm.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.pre_mlp_layernorm.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.self_attn.gate_proj.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.self_attn.k_norm.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.self_attn.k_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.self_attn.q_norm.weight": "model-00012-of-00012.safetensors",
+    "model.layers.31.self_attn.q_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.31.self_attn.v_proj.weight": "model-00011-of-00012.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.mlp.expert_bias": "model-00001-of-00012.safetensors",
+    "model.layers.4.mlp.moe_mlp.experts.weight": "model-00002-of-00012.safetensors",
+    "model.layers.4.mlp.moe_mlp.output_experts.weight": "model-00002-of-00012.safetensors",
+    "model.layers.4.mlp.router.gate.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.mlp.shared_experts.down_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.mlp.shared_experts.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.mlp.shared_experts.up_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.post_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.pre_mlp_layernorm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.self_attn.gate_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.self_attn.k_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.self_attn.q_norm.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00012.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.mlp.expert_bias": "model-00002-of-00012.safetensors",
+    "model.layers.5.mlp.moe_mlp.experts.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.mlp.moe_mlp.output_experts.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.mlp.router.gate.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.mlp.shared_experts.down_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.mlp.shared_experts.gate_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.mlp.shared_experts.up_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.post_mlp_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.pre_mlp_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.self_attn.gate_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.self_attn.k_norm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.self_attn.q_norm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.mlp.expert_bias": "model-00002-of-00012.safetensors",
+    "model.layers.6.mlp.moe_mlp.experts.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.mlp.moe_mlp.output_experts.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.mlp.router.gate.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.mlp.shared_experts.down_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.mlp.shared_experts.gate_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.mlp.shared_experts.up_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.post_mlp_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.pre_mlp_layernorm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.self_attn.gate_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.self_attn.k_norm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.self_attn.q_norm.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.mlp.expert_bias": "model-00003-of-00012.safetensors",
+    "model.layers.7.mlp.moe_mlp.experts.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.mlp.moe_mlp.output_experts.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.mlp.router.gate.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.mlp.shared_experts.down_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.mlp.shared_experts.gate_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.mlp.shared_experts.up_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.post_mlp_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.pre_mlp_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.self_attn.gate_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.self_attn.k_norm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.self_attn.q_norm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00002-of-00012.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.mlp.expert_bias": "model-00003-of-00012.safetensors",
+    "model.layers.8.mlp.moe_mlp.experts.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.mlp.moe_mlp.output_experts.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.mlp.router.gate.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.mlp.shared_experts.down_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.mlp.shared_experts.gate_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.mlp.shared_experts.up_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.post_mlp_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.pre_mlp_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.self_attn.gate_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.self_attn.k_norm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.self_attn.q_norm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.mlp.expert_bias": "model-00003-of-00012.safetensors",
+    "model.layers.9.mlp.moe_mlp.experts.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.mlp.moe_mlp.output_experts.weight": "model-00004-of-00012.safetensors",
+    "model.layers.9.mlp.router.gate.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.mlp.shared_experts.down_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.mlp.shared_experts.gate_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.mlp.shared_experts.up_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.post_mlp_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.pre_mlp_layernorm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.self_attn.gate_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.self_attn.k_norm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.self_attn.q_norm.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00003-of-00012.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00003-of-00012.safetensors",
+    "model.norm.weight": "model-00012-of-00012.safetensors"
+  }
+}

modeling_afmoe_scm.py

@@ -0,0 +1,661 @@
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import (
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel, ALL_ATTENTION_FUNCTIONS
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.masking_utils import (
+    create_causal_mask,
+    create_sliding_window_causal_mask,
+)
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.integrations import use_kernel_forward_from_hub
+
+import scattermoe
+
+try:
+    from .configuration_afmoe_scm import AfmoeSCMConfig
+except:
+    from configuration_afmoe_scm import AfmoeSCMConfig
+
+class AfmoeSCMRotaryEmbedding(nn.Module):
+
+    def __init__(self, config: AfmoeSCMConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float().to(x.device) @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = 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)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    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)
+
+@use_kernel_forward_from_hub("RMSNorm")
+class AfmoeSCMRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float):
+        """
+        AfmoeSCMRMSNorm 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}"
+
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    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
+
+
+class AfmoeSCMMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size or config.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):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class AfmoeSCMTokenChoiceRouter(nn.Module):
+    """Token-choice top-K router for MoE routing."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+        self.score_func = config.score_func
+        self.route_norm = config.route_norm
+        self.route_scale = config.route_scale
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)     
+
+    def forward(self, hidden_states, expert_bias: torch.Tensor | None):
+        _, _, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+
+        scores = self.gate(hidden_states)
+
+        # Apply scoring function in float32 for stability
+        if self.score_func == "sigmoid":
+            scores = torch.sigmoid(scores.to(torch.float32))
+        else:
+            scores = F.softmax(scores.to(torch.float32), dim=-1)
+
+        if expert_bias is not None:
+            _, selected_experts = torch.topk(scores + expert_bias, k=self.top_k, dim=1)
+            top_scores = scores.gather(dim=1, index=selected_experts)
+        else:
+            top_scores, selected_experts = torch.topk(scores, k=self.top_k, dim=1)
+
+        # Normalize weights if using sigmoid
+        if self.score_func == "sigmoid" and self.route_norm:
+            denominator = top_scores.sum(dim=-1, keepdim=True) + 1e-20
+            top_scores = top_scores / denominator
+
+        top_scores = top_scores * self.route_scale
+        return top_scores, selected_experts
+
+class AfmoeSCMMoE(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.router = AfmoeSCMTokenChoiceRouter(config)
+
+        self.shared_experts = None
+        if config.num_shared_experts > 0:
+            self.shared_experts = AfmoeSCMMLP(
+                config, config.moe_intermediate_size * config.num_shared_experts
+            )
+        self.moe_mlp = scattermoe.mlp.GLUMLP(
+            input_size=self.config.hidden_size,
+            hidden_size=self.config.moe_intermediate_size,
+            num_experts=self.config.num_experts,
+            top_k=self.config.num_experts_per_tok,
+            activation=ACT2FN[config.hidden_act],
+        )
+        
+        #self.experts = nn.ModuleList(
+        #    [AfmoeSCMMLP(
+        #        config, intermediate_size=config.moe_intermediate_size
+        #    ) for _ in range(config.num_experts)]
+        #)
+        self.expert_bias = nn.Parameter(torch.zeros(config.num_experts, dtype=torch.float32), requires_grad=False)
+        
+
+    def forward(self, hidden_states):
+        batch_size, seq_len, hidden_dim = hidden_states.shape
+        hidden_states_flat = hidden_states.view(-1, hidden_dim)
+
+        # Get routing decisions
+        top_scores, selected_experts = self.router(hidden_states, self.expert_bias)
+
+        # Process through shared experts
+        if self.shared_experts is not None:
+            shared_output = self.shared_experts(hidden_states_flat)
+        else:
+            shared_output = torch.zeros_like(hidden_states_flat)
+
+        hs_0 = self.moe_mlp(hidden_states_flat, top_scores.to(torch.bfloat16), selected_experts)
+        output = hs_0 + shared_output
+
+        return output.view(batch_size, seq_len, hidden_dim)
+
+
+class AfmoeSCMAttention(nn.Module):
+    """Multi-headed attention with local/global pattern and gating."""
+
+    def __init__(self, config: AfmoeSCMConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_local_attention = config.layer_types[layer_idx] == "sliding_attention"
+        self.sliding_window = config.sliding_window if self.is_local_attention else None
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, config.hidden_size, bias=False
+        )
+
+        self.q_norm = AfmoeSCMRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = AfmoeSCMRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+
+        self.gate_proj = nn.Linear(
+            config.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:   
+
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape)
+        key_states = self.k_proj(hidden_states).view(hidden_shape)
+        value_states = self.v_proj(hidden_states).view(hidden_shape)
+        gate_states = self.gate_proj(hidden_states)
+
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+        
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        if self.is_local_attention:
+            cos, sin = position_embeddings
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            cache_kwargs = {"cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[
+                self.config._attn_implementation
+            ]
+
+        output, _ = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask=attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=self.sliding_window,
+            **kwargs,
+        )
+
+        output = output.view(*input_shape, -1).contiguous()
+        output = output * F.sigmoid(gate_states)
+        return self.o_proj(output)
+
+
+class AfmoeSCMDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: AfmoeSCMConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+
+        self.self_attn = AfmoeSCMAttention(config=config, layer_idx=layer_idx)
+        self.attention_type = config.layer_types[layer_idx]
+
+        # Dual normalization for attention
+        self.input_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        # Dual normalization for FFN
+        self.pre_mlp_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_mlp_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        # MoE or dense FFN
+        self.moe_enabled = layer_idx >= config.num_dense_layers
+        if self.moe_enabled:
+            self.mlp = AfmoeSCMMoE(config)
+        else:
+            self.mlp = AfmoeSCMMLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        # Self Attention with dual normalization
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = residual + hidden_states
+
+        # FFN with dual normalization
+        residual = hidden_states
+        hidden_states = self.pre_mlp_layernorm(hidden_states)
+
+        if self.moe_enabled:
+            hidden_states = self.mlp(hidden_states)
+        else:
+            hidden_states = self.mlp(hidden_states)
+
+        hidden_states = self.post_mlp_layernorm(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class AfmoeSCMPreTrainedModel(PreTrainedModel):
+    config_class = AfmoeSCMConfig
+    base_model_prefix = "model"
+    _no_split_modules = ["AfmoeSCMDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _keep_in_fp32_modules = [
+        "input_layernorm",
+        "post_attention_layernorm",
+        "pre_mlp_layernorm",
+        "post_mlp_layernorm",
+        "q_norm",
+        "k_norm",
+        "norm",
+    ]
+    _supports_sdpa = True
+    _supports_attention_backend = True
+    supports_gradient_checkpointing = True
+
+
+class AfmoeSCMModel(AfmoeSCMPreTrainedModel):
+    _no_split_modules = ["AfmoeSCMDecoderLayer"]
+
+    def __init__(self, config: AfmoeSCMConfig):
+        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(
+            [
+                AfmoeSCMDecoderLayer(config, layer_idx)
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = AfmoeSCMRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[list[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> MoeModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        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.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+            }
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+                "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs),
+            }
+
+        hidden_states = inputs_embeds
+
+        # Apply muP input scaling if enabled
+        if self.config.mup_enabled:
+            hidden_states = hidden_states * (self.config.hidden_size**0.5)
+
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+class AfmoeSCMForCausalLM(AfmoeSCMPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = AfmoeSCMModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        token_type_ids: Optional[torch.Tensor] = None,  # will be ignored
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        outputs: MoeModelOutputWithPast = 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
+        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, labels, self.vocab_size, **kwargs)
+
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+
+__all__ = [
+    "AfmoeSCMForCausalLM",
+    "AfmoeSCMModel",
+    "AfmoeSCMPreTrainedModel",
+]

modeling_afmoe_scm_liger.py

@@ -0,0 +1,675 @@
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import (
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel, ALL_ATTENTION_FUNCTIONS
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.masking_utils import (
+    create_causal_mask,
+    create_sliding_window_causal_mask,
+)
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.integrations import use_kernel_forward_from_hub
+
+import scattermoe
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.rms_norm import LigerRMSNorm as AfmoeSCMRMSNorm
+
+try:
+    from .configuration_afmoe_scm import AfmoeSCMConfig
+except:
+    from configuration_afmoe_scm import AfmoeSCMConfig
+
+class AfmoeSCMRotaryEmbedding(nn.Module):
+
+    def __init__(self, config: AfmoeSCMConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float().to(x.device) @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = 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)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    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)
+
+#@use_kernel_forward_from_hub("RMSNorm")
+class _AfmoeSCMRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float):
+        """
+        AfmoeSCMRMSNorm 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}"
+
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    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
+
+
+class AfmoeSCMMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size or config.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):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class AfmoeSCMTokenChoiceRouter(nn.Module):
+    """Token-choice top-K router for MoE routing."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+        self.score_func = config.score_func
+        self.route_norm = config.route_norm
+        self.route_scale = config.route_scale
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)     
+
+    def forward(self, hidden_states, expert_bias: torch.Tensor | None):
+        _, _, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+
+        scores = self.gate(hidden_states)
+
+        # Apply scoring function in float32 for stability
+        if self.score_func == "sigmoid":
+            scores = torch.sigmoid(scores.to(torch.float32))
+        else:
+            scores = F.softmax(scores.to(torch.float32), dim=-1)
+
+        if expert_bias is not None:
+            _, selected_experts = torch.topk(scores + expert_bias, k=self.top_k, dim=1)
+            top_scores = scores.gather(dim=1, index=selected_experts)
+        else:
+            top_scores, selected_experts = torch.topk(scores, k=self.top_k, dim=1)
+
+        # Normalize weights if using sigmoid
+        if self.score_func == "sigmoid" and self.route_norm:
+            denominator = top_scores.sum(dim=-1, keepdim=True) + 1e-20
+            top_scores = top_scores / denominator
+
+        top_scores = top_scores * self.route_scale
+        return top_scores, selected_experts
+
+class AfmoeSCMMoE(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.router = AfmoeSCMTokenChoiceRouter(config)
+
+        self.shared_experts = None
+        if config.num_shared_experts > 0:
+            self.shared_experts = AfmoeSCMMLP(
+                config, config.moe_intermediate_size * config.num_shared_experts
+            )
+        self.moe_mlp = scattermoe.mlp.GLUMLP(
+            input_size=self.config.hidden_size,
+            hidden_size=self.config.moe_intermediate_size,
+            num_experts=self.config.num_experts,
+            top_k=self.config.num_experts_per_tok,
+            activation=ACT2FN[config.hidden_act],
+        )
+        
+        self.expert_bias = nn.Parameter(torch.zeros(config.num_experts, dtype=torch.float32), requires_grad=False)
+        
+
+    def forward(self, hidden_states):
+        batch_size, seq_len, hidden_dim = hidden_states.shape
+        hidden_states_flat = hidden_states.view(-1, hidden_dim)
+
+        # Get routing decisions
+        top_scores, selected_experts = self.router(hidden_states, self.expert_bias)
+
+        # Process through shared experts
+        if self.shared_experts is not None:
+            shared_output = self.shared_experts(hidden_states_flat)
+        else:
+            shared_output = torch.zeros_like(hidden_states_flat)
+
+        hs_0 = self.moe_mlp(hidden_states_flat, top_scores.to(torch.bfloat16), selected_experts)
+        output = hs_0 + shared_output
+
+        return output.view(batch_size, seq_len, hidden_dim)
+
+
+class AfmoeSCMAttention(nn.Module):
+    """Multi-headed attention with local/global pattern and gating."""
+
+    def __init__(self, config: AfmoeSCMConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_local_attention = config.layer_types[layer_idx] == "sliding_attention"
+        self.sliding_window = config.sliding_window if self.is_local_attention else None
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, config.hidden_size, bias=False
+        )
+
+        self.q_norm = AfmoeSCMRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = AfmoeSCMRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+
+        self.gate_proj = nn.Linear(
+            config.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:   
+
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape)
+        key_states = self.k_proj(hidden_states).view(hidden_shape)
+        value_states = self.v_proj(hidden_states).view(hidden_shape)
+        gate_states = self.gate_proj(hidden_states)
+
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+        
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        if self.is_local_attention:
+            cos, sin = position_embeddings
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            cache_kwargs = {"cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[
+                self.config._attn_implementation
+            ]
+
+        output, _ = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask=attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=self.sliding_window,
+            **kwargs,
+        )
+
+        output = output.view(*input_shape, -1).contiguous()
+        output = output * F.sigmoid(gate_states)
+        return self.o_proj(output)
+
+
+class AfmoeSCMDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: AfmoeSCMConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+
+        self.self_attn = AfmoeSCMAttention(config=config, layer_idx=layer_idx)
+        self.attention_type = config.layer_types[layer_idx]
+
+        # Dual normalization for attention
+        self.input_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        # Dual normalization for FFN
+        self.pre_mlp_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_mlp_layernorm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        # MoE or dense FFN
+        self.moe_enabled = layer_idx >= config.num_dense_layers
+        if self.moe_enabled:
+            self.mlp = AfmoeSCMMoE(config)
+        else:
+            self.mlp = AfmoeSCMMLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        # Self Attention with dual normalization
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = residual + hidden_states
+
+        # FFN with dual normalization
+        residual = hidden_states
+        hidden_states = self.pre_mlp_layernorm(hidden_states)
+
+        if self.moe_enabled:
+            hidden_states = self.mlp(hidden_states)
+        else:
+            hidden_states = self.mlp(hidden_states)
+
+        hidden_states = self.post_mlp_layernorm(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class AfmoeSCMPreTrainedModel(PreTrainedModel):
+    config_class = AfmoeSCMConfig
+    base_model_prefix = "model"
+    _no_split_modules = ["AfmoeSCMDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _keep_in_fp32_modules = [
+        "input_layernorm",
+        "post_attention_layernorm",
+        "pre_mlp_layernorm",
+        "post_mlp_layernorm",
+        "q_norm",
+        "k_norm",
+        "norm",
+    ]
+    _supports_sdpa = True
+    _supports_attention_backend = True
+    supports_gradient_checkpointing = True
+
+
+class AfmoeSCMModel(AfmoeSCMPreTrainedModel):
+    _no_split_modules = ["AfmoeSCMDecoderLayer"]
+
+    def __init__(self, config: AfmoeSCMConfig):
+        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(
+            [
+                AfmoeSCMDecoderLayer(config, layer_idx)
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = AfmoeSCMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = AfmoeSCMRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[list[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> MoeModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        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.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+            }
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+                "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs),
+            }
+
+        hidden_states = inputs_embeds
+
+        # Apply muP input scaling if enabled
+        if self.config.mup_enabled:
+            hidden_states = hidden_states * (self.config.hidden_size**0.5)
+
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+class AfmoeSCMForCausalLM(AfmoeSCMPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = AfmoeSCMModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        token_type_ids: Optional[torch.Tensor] = None,  # will be ignored
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        outputs: MoeModelOutputWithPast = 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
+        loss = None
+        # Only compute necessary logits
+        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 MoeCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+
+__all__ = [
+    "AfmoeSCMForCausalLM",
+    "AfmoeSCMModel",
+    "AfmoeSCMPreTrainedModel",
+]

special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|pad|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0058acf26a6e7228298b0c9fed2a87fcb3f6cb5f84752cfde101b9e68b380918
+size 14614841

tokenizer_config.json

@@ -0,0 +1,271 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "<|start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "<|channel|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "7": {
+      "content": "<|message|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "8": {
+      "content": "<|end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "9": {
+      "content": "<|fitm_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "10": {
+      "content": "<|fitm_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "11": {
+      "content": "<|fitm_hole|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "12": {
+      "content": "<|pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "13": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "14": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "15": {
+      "content": "<|reserved_special_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "16": {
+      "content": "<|reserved_special_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "17": {
+      "content": "<|reserved_special_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "18": {
+      "content": "<|reserved_special_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "19": {
+      "content": "<|reserved_special_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "20": {
+      "content": "<|reserved_special_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "21": {
+      "content": "<|reserved_special_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "22": {
+      "content": "<|reserved_special_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "23": {
+      "content": "<|reserved_special_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "24": {
+      "content": "<|reserved_special_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "25": {
+      "content": "<|reserved_special_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "26": {
+      "content": "<|reserved_special_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "27": {
+      "content": "<|reserved_special_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "28": {
+      "content": "<|reserved_special_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "29": {
+      "content": "<|reserved_special_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "30": {
+      "content": "<|reserved_special_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "31": {
+      "content": "<|reserved_special_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "extra_special_tokens": {},
+  "model_max_length": 65536,
+  "pad_token": "<|pad|>",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "use_default_system_prompt": false
+}