Upload DogeForCausalLM

config.json

@@ -1,46 +1,46 @@
-{
-  "_name_or_path": "./results/Doge-160M/checkpoint-21600",
-  "architectures": [
-    "DogeForCausalLM"
-  ],
-  "attention_dropout": 0.0,
-  "auto_map": {
-    "AutoConfig": "configuration_doge.DogeConfig",
-    "AutoModelForCausalLM": "modeling_doge.DogeForCausalLM"
-  },
-  "bos_token_id": 0,
-  "dynamic_mask_ratio": 0.0,
-  "eos_token_id": 1,
-  "expert_retrieval_size": 64,
-  "hidden_act": "silu",
-  "hidden_bias": false,
-  "hidden_dropout": 0.0,
-  "hidden_size": 768,
-  "initializer_range": 0.02,
-  "intermediate_size": 1536,
-  "is_causal": false,
-  "is_moe": false,
-  "max_position_embeddings": 2048,
-  "model_type": "doge",
-  "num_attention_heads": 6,
-  "num_cdmoe_experts": 16348,
-  "num_cdmoe_experts_per_head": 8,
-  "num_cdmoe_heads": 4,
-  "num_channels": 3,
-  "num_hidden_layers": 24,
-  "num_key_value_heads": 3,
-  "pad_token_id": 2,
-  "patch_size": 16,
-  "rms_norm_eps": 1e-06,
-  "rope_scaling": {
-    "factor": 4.0,
-    "original_max_position_embeddings": 2048,
-    "rope_type": "dynamic"
-  },
-  "rope_theta": 10000.0,
-  "tie_word_embeddings": true,
-  "torch_dtype": "float32",
-  "transformers_version": "4.48.1",
-  "use_cache": true,
-  "vocab_size": 32768
-}
+{
+  "_name_or_path": "SmallDoge/Doge-160M-checkpoint",
+  "architectures": [
+    "DogeForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_doge.DogeConfig",
+    "AutoModelForCausalLM": "modeling_doge.DogeForCausalLM"
+  },
+  "bos_token_id": 0,
+  "dynamic_mask_ratio": 0.0,
+  "eos_token_id": 1,
+  "expert_retrieval_size": 64,
+  "hidden_act": "silu",
+  "hidden_bias": false,
+  "hidden_dropout": 0.0,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 1536,
+  "is_causal": false,
+  "is_moe": false,
+  "max_position_embeddings": 2048,
+  "model_type": "doge",
+  "num_attention_heads": 6,
+  "num_cdmoe_experts": 16348,
+  "num_cdmoe_experts_per_head": 8,
+  "num_cdmoe_heads": 4,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 3,
+  "pad_token_id": 2,
+  "patch_size": 16,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": {
+    "factor": 4.0,
+    "original_max_position_embeddings": 2048,
+    "rope_type": "dynamic"
+  },
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.48.3",
+  "use_cache": true,
+  "vocab_size": 32768
+}

configuration_doge.py

@@ -1,9 +1,14 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/doge/modular_doge.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_doge.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
 # coding=utf-8
 # Copyright 2024 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
 #
 # This code is based on the Wonderful Matrices paper implementation.
-#
-#     https://arxiv.org/abs/2412.11834
+# The Doge family of small language models is trained by Jingze Shi.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -16,8 +21,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""PyTorch Doge model configuration"""
-
 from transformers.configuration_utils import PretrainedConfig
 from transformers.modeling_rope_utils import rope_config_validation
 
@@ -25,7 +28,7 @@ from transformers.modeling_rope_utils import rope_config_validation
 class DogeConfig(PretrainedConfig):
     r"""
     This is the configuration class to store the configuration of a [`DogeModel`]. It is used to instantiate an Doge
-    model according to the specified arguments, defining the model architecture like [JingzeShi/Doge-20M](https://huggingface.co/JingzeShi/Doge-20M).
+    model according to the specified arguments, defining the model architecture like [SmallDoge/Doge-20M](https://huggingface.co/SmallDoge/Doge-20M).
 
     Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
     documentation from [`PretrainedConfig`] for more information.
@@ -33,10 +36,6 @@ class DogeConfig(PretrainedConfig):
     Args:
         vocab_size (`int`, *optional*, defaults to 32768):
             Vocabulary size of the Doge model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`DogeModel`]
-        num_channels (`int`, *optional*, defaults to 3):
-            Number of channels in the input image.
-        patch_size (`int`, *optional*, defaults to 16):
-            Patch size of Vision Transformer Embeddings.
         hidden_size (`int`, *optional*, defaults to 1024):
             Dimension of the hidden representations.
         intermediate_size (`int`, *optional*, defaults to 2048):
@@ -49,25 +48,41 @@ class DogeConfig(PretrainedConfig):
             Dropout probability for each sequence transformation and state transformation module.
         hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
             The non-linear activation function (function or string) in the decoder.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        bos_token_id (`int`, *optional*, defaults to 0):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 1):
+            End of stream token id.
+        pad_token_id (`int`, *optional*, defaults to 2):
+            Padding token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
         max_position_embeddings (`int`, *optional*, defaults to 2048):
             The maximum sequence length that this model might ever be used with.
         rope_theta (`float`, *optional*, defaults to 10000.0):
             The base period of the RoPE embeddings.
         rope_scaling (`Dict`, *optional*):
-            Dictionary containing the scaling configuration for the RoPE embeddings. 
+            Dictionary containing the scaling configuration for the RoPE embeddings.
             NOTE: if you apply new rope type and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value accordingly.
+            Doge family of small models use `{ 'rope_type': 'dynamic', 'factor': 4.0, 'original_max_position_embeddings': 2048 }` as the default value.
             Expected contents:
                 `rope_type` (`str`):
                     The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', 'llama3'], with 'default' being the original RoPE implementation.
                 `factor` (`float`, *optional*):
-                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. 
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings.
                     In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length.
                 `original_max_position_embeddings` (`int`, *optional*):
-                    Used with 'dynamic', 'longrope' and 'llama3'. 
+                    Used with 'dynamic', 'longrope' and 'llama3'.
                     The original max position embeddings used during pretraining.
                 `attention_factor` (`float`, *optional*):
                     Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
-                    computation. 
+                    computation.
                     If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value.
                 `beta_fast` (`float`, *optional*):
                     Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
@@ -76,54 +91,51 @@ class DogeConfig(PretrainedConfig):
                     Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
                     ramp function. If unspecified, it defaults to 1.
                 `short_factor` (`List[float]`, *optional*):
-                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<`original_max_position_embeddings`). 
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<`original_max_position_embeddings`).
                     Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2
                 `long_factor` (`List[float]`, *optional*):
-                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<`original_max_position_embeddings`). 
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<`original_max_position_embeddings`).
                     Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2
                 `low_freq_factor` (`float`, *optional*):
                     Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
                 `high_freq_factor` (`float`, *optional*):
                     Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
-        initializer_range (`float`, *optional*, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the rms normalization layers.
-        use_cache (`bool`, *optional*, defaults to `True`):
-            Whether or not the model should return the last key/values attentions (not used by all models). Only
-            relevant if `config.is_decoder=True`.
-        pad_token_id (`int`, *optional*, defaults to 0):
-            Padding token id.
-        bos_token_id (`int`, *optional*, defaults to 1):
-            Beginning of stream token id.
-        eos_token_id (`int`, *optional*, defaults to 2):
-            End of stream token id.
-        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
-            Whether to tie weight embeddings
         num_attention_heads (`int`, *optional*, defaults to 8):
             Number of attention heads for each attention layer in the Transformer decoder.
-        num_key_value_heads (`int`, *optional*, defaults to `None`):
-            This is the number of key_value heads that should be used to implement Grouped Query Attention. 
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention.
             If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
-            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. 
-            When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. 
-            For more details checkout [this paper](https://arxiv.org/pdf/2305.13245.pdf). 
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used.
+            When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group.
+            For more details checkout [this paper](https://arxiv.org/pdf/2305.13245.pdf).
             If it is not specified, will default to `num_attention_heads`.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
-        dynamic_mask_ratio (`float`, *optional*, defaults to 0.0, range [0, 1]):
-            The ratio to control the proportion of the dynamic mask filled with the minimum value.
+        dynamic_mask_ratio (`float`, *optional*, defaults to 0.0):
+            The ratio to control the proportion of the dynamic mask filled with the minimum value. For more details checkout [this paper](https://arxiv.org/pdf/2412.11834).
         is_moe (`bool`, *optional*, defaults to `False`):
-            Whether to use the Cross Domain Mixture of Experts, if `True`, the MoE will inherit the MLP to initialize
+            Whether to use the Cross Domain Mixture of Experts, if `True`, the MoE will inherit the MLP to initialize. For more details checkout [this paper](https://arxiv.org/pdf/2412.11834).
         num_cdmoe_experts (`int`, *optional*, defaults to 16348):
-            Number of Private Experts for the Cross Domain Mixture of Experts. calculation formula: :math:`\text{num_cdmoe_experts} = (32 \times \text{num_cdmoe_heads})^2`
+            Number of Experts for the Cross Domain Mixture of Experts.
         num_cdmoe_heads (`int`, *optional*, defaults to 4):
-            Number of heads of Private Experts for the Cross Domain Mixture of Experts.
+            Number of retrieval heads, used to mix multi-head experts.
         num_cdmoe_experts_per_head (`int`, *optional*, defaults to 8):
-            Number of Private Experts per head for the Cross Domain Mixture of Experts.
+            Number of Experts per retrieval head, used to mix multi-head experts.
         expert_retrieval_size (`int`, *optional*, defaults to 64):
-            Dimension of the Expert retrieval states for the Cross Domain Mixture of Experts.
-    """
+            Dimension of the Expert retrieval states for calculating the dot product of query and key to determine the expert index.
+
+    ```python
+    >>> from transformers import DogeConfig, DogeModel
+
+    >>> # Initializing a Doge-320M style configuration
+    >>> configuration = DogeConfig()
+
+    >>> # Initializing a model from the Doge-320M style configuration
+    >>> model = DogeModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
 
     model_type = "doge"
     keys_to_ignore_at_inference = ["past_key_values"]
@@ -132,7 +144,7 @@ class DogeConfig(PretrainedConfig):
         "layers.*.self_attn.q_proj": "colwise",
         "layers.*.self_attn.k_proj": "colwise",
         "layers.*.self_attn.v_proj": "colwise",
-        "layers.*.self_attn.dt_proj": "colwise",
+        "layers.*.self_attn.dt_proj": "rowwise",
         "layers.*.self_attn.o_proj": "rowwise",
         "layers.*.mlp.gate_proj": "colwise",
         "layers.*.mlp.up_proj": "colwise",
@@ -142,33 +154,26 @@ class DogeConfig(PretrainedConfig):
     def __init__(
         self,
         vocab_size=32768,
-        num_channels=3,
-        patch_size=16,
         hidden_size=1024,
         intermediate_size=2048,
         num_hidden_layers=32,
         hidden_bias=False,
         hidden_dropout=0.0,
         hidden_act="silu",
-        max_position_embeddings=2048,
-        rope_theta=10000.0,
-        rope_scaling={
-            "rope_type": "dynamic",
-            "factor": 4.0,
-            "original_max_position_embeddings": 2048,
-        },
         initializer_range=0.02,
         rms_norm_eps=1e-06,
         use_cache=True,
         bos_token_id=0,
         eos_token_id=1,
         pad_token_id=2,
-        tie_word_embeddings=True,
+        tie_word_embeddings=False,
+        max_position_embeddings=2048,
+        rope_theta=10000.0,
+        rope_scaling=None,
         num_attention_heads=8,
         num_key_value_heads=None,
         attention_dropout=0.0,
         dynamic_mask_ratio=0.0,
-        is_causal=False,
         is_moe=False,
         num_cdmoe_experts=16348,
         num_cdmoe_heads=4,
@@ -177,29 +182,24 @@ class DogeConfig(PretrainedConfig):
         **kwargs,
     ):
         self.vocab_size = vocab_size
-        self.num_channels = num_channels
-        self.patch_size = patch_size
         self.hidden_size = hidden_size
         self.intermediate_size = intermediate_size
         self.num_hidden_layers = num_hidden_layers
+
         self.hidden_bias = hidden_bias
         self.hidden_dropout = hidden_dropout
         self.hidden_act = hidden_act
-        self.max_position_embeddings = max_position_embeddings
-        self.rope_theta = rope_theta
-        self.rope_scaling = rope_scaling
         self.initializer_range = initializer_range
         self.rms_norm_eps = rms_norm_eps
         self.use_cache = use_cache
-        self.bos_token_id = bos_token_id
-        self.eos_token_id = eos_token_id
-        self.pad_token_id = pad_token_id
-        self.tie_word_embeddings = tie_word_embeddings
+
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
         self.num_attention_heads = num_attention_heads
         self.num_key_value_heads = num_key_value_heads
         self.attention_dropout = attention_dropout
         self.dynamic_mask_ratio = dynamic_mask_ratio
-        self.is_causal = is_causal
         self.is_moe = is_moe
         self.num_cdmoe_experts = num_cdmoe_experts
         self.num_cdmoe_heads = num_cdmoe_heads

generation_config.json

@@ -3,5 +3,5 @@
   "bos_token_id": 0,
   "eos_token_id": 1,
   "pad_token_id": 2,
-  "transformers_version": "4.48.1"
+  "transformers_version": "4.48.3"
 }

modeling_doge.py

@@ -1,9 +1,14 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/doge/modular_doge.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_doge.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
 # coding=utf-8
 # Copyright 2024 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
 #
 # This code is based on the Wonderful Matrices paper implementation.
-#
-#     https://arxiv.org/abs/2412.11834
+# The Doge family of small language models is trained by Jingze Shi.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -16,24 +21,19 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""PyTorch Doge model."""
 
 import math
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
 import torch.nn.functional as F
-import torch.utils.checkpoint
 from torch import nn
 
 from transformers.activations import ACT2FN
 from transformers.cache_utils import Cache, DynamicCache, StaticCache
 from transformers.generation import GenerationMixin
-from transformers.modeling_outputs import (
-    BaseModelOutputWithPast,
-    CausalLMOutputWithPast,
-    SequenceClassifierOutputWithPast,
-)
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
 from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
 from transformers.modeling_utils import PreTrainedModel
 from transformers.processing_utils import Unpack
@@ -41,30 +41,24 @@ from transformers.utils import (
     LossKwargs,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
-    is_torch_greater_or_equal,
+    is_torch_flex_attn_available,
     logging,
     replace_return_docstrings,
 )
 from .configuration_doge import DogeConfig
 
-try:
-    from einx import add as einx_add
-except ImportError:
-    einx_add = None
-
-if is_torch_greater_or_equal("2.5"):
+if is_torch_flex_attn_available():
     from torch.nn.attention.flex_attention import flex_attention
 
-
 logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = "DogeConfig"
 
 
-class RMSNorm(nn.Module):
+class DogeRMSNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
         """
-        RMSNorm is equivalent to T5LayerNorm
+        DogeRMSNorm is equivalent to T5LayerNorm
         """
         super().__init__()
         self.weight = nn.Parameter(torch.ones(hidden_size))
@@ -81,7 +75,7 @@ class RMSNorm(nn.Module):
         return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
 
 
-class Residual(nn.Module):
+class DogeResidual(nn.Module):
     def __init__(self, hidden_size):
         super().__init__()
         self.weight = nn.Parameter(torch.ones(hidden_size))
@@ -93,23 +87,21 @@ class Residual(nn.Module):
         return f"{tuple(self.weight.shape)}"
 
 
-class RotaryEmbedding(nn.Module):
-    def __init__(self, config: Optional[DogeConfig] = None):
+class DogeRotaryEmbedding(nn.Module):
+    def __init__(self, config: DogeConfig, device=None):
         super().__init__()
-        self.rope_kwargs = {}
-
-        if config.rope_scaling is not None:
+        # 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.base = config.rope_theta
 
         self.config = config
         self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
-        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, **self.rope_kwargs)
+        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
 
@@ -121,13 +113,14 @@ class RotaryEmbedding(nn.Module):
         """
         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.rope_kwargs
-            )
+            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
 
@@ -136,7 +129,7 @@ class RotaryEmbedding(nn.Module):
         if "dynamic" in self.rope_type:
             self._dynamic_frequency_update(position_ids, device=x.device)
 
-        # core RoPE block
+        # 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)
@@ -156,15 +149,13 @@ class RotaryEmbedding(nn.Module):
 
 
 def rotate_half(x):
-    """
-    Rotates half the hidden dims of the input.
-    """
+    """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_QK_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_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:
@@ -176,10 +167,11 @@ def apply_QK_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
             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.
+            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.
     """
@@ -192,8 +184,8 @@ def apply_QK_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
 
 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)
+    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:
@@ -202,6 +194,148 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
 
 
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    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(-1, -2)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = F.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def sdpa_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    dropout: float = 0.0,
+    scaling: Optional[float] = None,
+    is_causal: Optional[bool] = None,
+    **kwargs,
+) -> Tuple[torch.Tensor, None]:
+    key = repeat_kv(key, module.num_key_value_groups)
+    value = repeat_kv(value, module.num_key_value_groups)
+
+    causal_mask = attention_mask
+    if attention_mask is not None:
+        causal_mask = causal_mask[:, :, :, : key.shape[-2]]
+
+    # SDPA with memory-efficient backend is bugged with non-contiguous inputs and custom attn_mask for some torch versions
+    # Reference: https://github.com/pytorch/pytorch/issues/112577.
+    query = query.contiguous()
+    key = key.contiguous()
+    value = value.contiguous()
+
+    # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+    # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+    if is_causal is None:
+        is_causal = causal_mask is None and query.shape[2] > 1
+
+    # Shapes (e.g. query.shape[2]) are tensors during jit tracing, resulting in `is_causal` being a tensor.
+    # We convert it to a bool for the SDPA kernel that only accepts bools.
+    if torch.jit.is_tracing() and isinstance(is_causal, torch.Tensor):
+        is_causal = is_causal.item()
+
+    # NOTE: As of pytorch 2.5.1, SDPA backward pass of cuDNN is still incorrect, so we disable cuDNN SDPA (see https://github.com/pytorch/pytorch/issues/138581)
+    torch.backends.cuda.enable_cudnn_sdp(False)
+    attn_output = F.scaled_dot_product_attention(
+        query=query,
+        key=key,
+        value=value,
+        attn_mask=causal_mask,
+        dropout_p=dropout,
+        scale=scaling,
+        is_causal=is_causal,
+    )
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, None
+
+
+def flex_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: Optional[float] = None,
+    is_causal: Optional[bool] = None,
+    softcap: Optional[float] = None,
+    head_mask: Optional[torch.Tensor] = None,
+    **kwargs,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    causal_mask = attention_mask
+    if attention_mask is not None:
+        causal_mask = causal_mask[:, :, :, : key.shape[-2]]
+
+    if is_causal is None:
+        is_causal = causal_mask is None and query.shape[2] > 1
+
+    def causal_mod(score, batch, head, q_idx, kv_idx):
+        if softcap is not None:
+            score = softcap * torch.tanh(score / softcap)
+        if causal_mask is not None:
+            score = score + causal_mask[batch][0][q_idx][kv_idx]
+        if head_mask is not None:
+            score = score + head_mask[batch][head][0][0]
+        return score
+
+    def dynamic_mod(score, batch, head, q_idx, kv_idx):
+        if softcap is not None:
+            score = softcap * torch.tanh(score / softcap)
+        if causal_mask is not None:
+            score = score + causal_mask[batch][head][q_idx][kv_idx]
+        if head_mask is not None:
+            score = score + head_mask[batch][head][0][0]
+        return score
+
+    # TODO: flex_attention: As of pytorch 2.5.1, captured buffers that require grad are not yet supported.
+    # NOTE: So we only use flex_attention in inference mode.
+    mask_mod = causal_mod if is_causal or module.training else dynamic_mod
+
+    attn_output, attention_weights = flex_attention(
+        query=query,
+        key=key,
+        value=value,
+        score_mod=mask_mod,
+        enable_gqa=True,
+        scale=scaling,
+        # Last time checked on PyTorch == 2.5.1: Flex Attention always computes the lse regardless.
+        # For simplification, we thus always return it as no additional computations are introduced.
+        return_lse=True,
+    )
+    # lse is returned in float32
+    attention_weights = attention_weights.to(value.dtype)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attention_weights
+
+
+ALL_ATTENTION_FUNCTIONS = {
+    "eager": eager_attention_forward,
+    "sdpa": sdpa_attention_forward,
+    "flex_attention": flex_attention_forward,
+}
+
+
 class DogeDynamicMaskAttention(nn.Module):
     """Dynamic Mask Attention from 'Wonderful Matrices' paper."""
 
@@ -209,48 +343,28 @@ class DogeDynamicMaskAttention(nn.Module):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
-        self.head_dim = config.hidden_size // config.num_attention_heads
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
         self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
-        self.scaling = self.head_dim ** -0.5
+        self.scaling = self.head_dim**-0.5
         self.attention_dropout = config.attention_dropout
         self.dynamic_mask_ratio = config.dynamic_mask_ratio
-        self.is_causal = config.is_causal
 
-        self.ALL_ATTENTION_FUNCTIONS = {
-            "eager": self.eager_attention_forward,
-            "flex_attention": self.flex_attention_forward,
-            "sdpa": self.sdpa_attention_forward,
-        }
-
-        # Q K V O projections
         self.q_proj = nn.Linear(
-            config.hidden_size,
-            config.num_attention_heads * self.head_dim,
-            bias=config.hidden_bias
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.hidden_bias
         )
         self.k_proj = nn.Linear(
-            config.hidden_size,
-            config.num_key_value_heads * self.head_dim,
-            bias=config.hidden_bias
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.hidden_bias
         )
         self.v_proj = nn.Linear(
-            config.hidden_size,
-            config.num_key_value_heads * self.head_dim,
-            bias=config.hidden_bias
-        )
-        # dynamic mask for the QK^T attention score matrix
-        self.A = nn.Parameter(
-            torch.zeros(config.num_attention_heads)
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.hidden_bias
         )
+        # dynamic mask for the QK^T attention weights matrix
+        self.A = nn.Parameter(torch.zeros(config.num_attention_heads))
         self.dt_proj = nn.Linear(
-            config.num_key_value_heads * self.head_dim,
-            config.num_attention_heads,
-            bias=config.hidden_bias
+            config.num_key_value_heads * self.head_dim, config.num_attention_heads, bias=config.hidden_bias
         )
         self.o_proj = nn.Linear(
-            config.num_attention_heads * self.head_dim,
-            config.hidden_size,
-            bias=config.hidden_bias
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.hidden_bias
         )
 
     def forward(
@@ -261,7 +375,7 @@ class DogeDynamicMaskAttention(nn.Module):
         past_key_value: Optional[Cache] = None,
         cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
-    ) -> Tuple[torch.Tensor, Optional[Cache]]:
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -270,21 +384,18 @@ class DogeDynamicMaskAttention(nn.Module):
         value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
 
         cos, sin = position_embeddings
-        query_states, key_states = apply_QK_rotary_pos_emb(query_states, key_states, cos, sin)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
         if past_key_value is not None:
             # sin and cos are specific to RoPE models; cache_position needed for the static cache
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        
-        dynamic_mask = None
-        if self.is_causal is False:
-            # calculate dynamic mask from value_states
-            # NOTE: If these weights are not trained in causal mode, a mask of all ones will be returned, which will not affect the training results of causal mode
-            # TODO: The main reason for setting causal mode is that the Flex Attention kernel does not yet support score_mod functions with learnable parameters. However, we can continue training from the causal checkpoint later.
-            dt_states = self.dt_proj(value_states.transpose(1, 2).reshape(value_states.shape[0], value_states.shape[-2], -1))
-            dynamic_mask = torch.exp(self.A * F.softplus(dt_states)).transpose(-1, -2)
+        # calculate dynamic mask from value_states
+        dt_states = self.dt_proj(
+            value_states.transpose(1, 2).reshape(value_states.shape[0], value_states.shape[-2], -1)
+        )
+        dynamic_mask = torch.exp(self.A * F.softplus(dt_states)).transpose(-1, -2)
         attn_mask = self.prepare_dynamic_mask(
             hidden_states=hidden_states,
             dynamic_mask=dynamic_mask,
@@ -292,11 +403,18 @@ class DogeDynamicMaskAttention(nn.Module):
             attention_mask=attention_mask,
         )
 
-        attention_interface: Callable = self.eager_attention_forward
+        attention_interface: Callable = eager_attention_forward
         if self.config._attn_implementation != "eager":
-            attention_interface = self.ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
-        
-        attn_output = attention_interface(
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
             query_states,
             key_states,
             value_states,
@@ -308,7 +426,7 @@ class DogeDynamicMaskAttention(nn.Module):
 
         attn_output = attn_output.reshape(*input_shape, -1).contiguous()
         attn_output = self.o_proj(attn_output)
-        return attn_output
+        return attn_output, attn_weights
 
     def prepare_dynamic_mask(
         self,
@@ -341,113 +459,9 @@ class DogeDynamicMaskAttention(nn.Module):
             attn_mask = attention_mask
 
         return attn_mask
-    
-    def eager_attention_forward(
-        self,
-        query: torch.Tensor,
-        key: torch.Tensor,
-        value: torch.Tensor,
-        attention_mask: Optional[torch.Tensor],
-        scaling: float,
-        dropout: float = 0.0,
-        **kwargs,
-    ) -> torch.Tensor:
-        key_states = repeat_kv(key, self.num_key_value_groups)
-        value_states = repeat_kv(value, self.num_key_value_groups)
-
-        # compute attention scores matrix
-        attn_weights = torch.matmul(query, key_states.transpose(-1, -2)) * scaling
-        if attention_mask is not None:
-            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
-            attn_weights = attn_weights + causal_mask
-        
-        # upcast attention scores to fp32
-        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
-        attn_weights = F.dropout(attn_weights, p=dropout, training=self.training)
-
-        # apply attention scores to value states
-        attn_output = torch.matmul(attn_weights, value_states)
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        return attn_output
-    
-    def sdpa_attention_forward(
-        self,
-        query: torch.Tensor,
-        key: torch.Tensor,
-        value: torch.Tensor,
-        attention_mask: Optional[torch.Tensor],
-        scaling: float,
-        dropout: float = 0.0,
-        **kwargs,
-    ) -> torch.Tensor:
-        key = repeat_kv(key, self.num_key_value_groups)
-        value = repeat_kv(value, self.num_key_value_groups)
-
-        causal_mask = attention_mask
-        if attention_mask is not None:
-            causal_mask = causal_mask[:, :, :, : key.shape[-2]]
-
-        # SDPA with memory-efficient backend is bugged with non-contiguous inputs and custom attn_mask for some torch versions
-        # Reference: https://github.com/pytorch/pytorch/issues/112577.
-        query = query.contiguous()
-        key = key.contiguous()
-        value = value.contiguous()
-
-        # NOTE: As of pytorch 2.5.1, cuDNN's SDPA backward pass is still incorrect, so we disable cuDNN SDPA (see https://github.com/pytorch/pytorch/issues/138581)
-        torch.backends.cuda.enable_cudnn_sdp(False)
-        attn_output = F.scaled_dot_product_attention(
-            query,
-            key,
-            value,
-            attn_mask=causal_mask,
-            dropout_p=dropout,
-            scale=scaling,
-            # enable_gqa=True,
-        )
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        return attn_output
-    
-    def flex_attention_forward(
-        self,
-        query: torch.Tensor,
-        key: torch.Tensor,
-        value: torch.Tensor,
-        attention_mask: Optional[torch.Tensor],
-        scaling: float,
-        dropout: float = 0.0,
-        **kwargs,
-    ) -> torch.Tensor:
-        causal_mask = attention_mask
-        if attention_mask is not None:
-            causal_mask = causal_mask[:, :, :, : key.shape[-2]]
-
-        # TODO: flex_attention: As of pytorch 2.5.1, captured buffers that require grad are not yet supported.
-        # NOTE: So we only use flex_attention in inference mode.
-
-        def causal_mod(score, batch, head, q_idx, kv_idx):
-            score = score + causal_mask[batch][0][q_idx][kv_idx]
-            return score
-        
-        def dynamic_mod(score, batch, head, q_idx, kv_idx):
-            score = score + causal_mask[batch][head][q_idx][kv_idx]
-            return score
-        
-        mask_mod = causal_mod if self.is_causal else dynamic_mod
-        
-        attn_output = flex_attention(
-            query,
-            key,
-            value,
-            score_mod=mask_mod,
-            scale=scaling,
-            enable_gqa=True,
-        )
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        return attn_output
 
 
 class DogeMLP(nn.Module):
-
     def __init__(self, config: DogeConfig):
         super().__init__()
         self.hidden_dim = config.hidden_size
@@ -482,11 +496,11 @@ class DogeCDMoE(DogeMLP):
         self.num_keys = int(math.sqrt(self.num_cdmoe_experts))
 
         # queries and keys for retrieval experts
-        self.queries = nn.Linear(self.hidden_dim, self.num_cdmoe_heads * self.expert_retrieval_dim, bias=False)
-        self.keys = nn.Parameter(torch.zeros(self.num_cdmoe_heads, self.num_keys, 2, self.expert_retrieval_dim // 2))
+        self.queries_proj = nn.Linear(self.hidden_dim, self.num_cdmoe_heads * self.expert_retrieval_dim, bias=False)
+        self.keys = nn.Parameter(torch.zeros(self.num_cdmoe_heads, self.expert_retrieval_dim, self.num_keys))
 
         # experts
-        self.down_embed  = nn.Embedding(self.num_cdmoe_experts, self.hidden_dim)
+        self.down_embed = nn.Embedding(self.num_cdmoe_experts, self.hidden_dim)
         self.up_embed = nn.Embedding(self.num_cdmoe_experts, self.hidden_dim)
 
     def forward(
@@ -496,30 +510,28 @@ class DogeCDMoE(DogeMLP):
     ) -> torch.Tensor:
         bsz, seq_len, _ = hidden_states.shape
 
-        # get similarity with queries and keys
-        queries = self.queries(hidden_states)
-        queries = queries.view(bsz, seq_len, 2, self.num_cdmoe_heads, -1).permute(2, 0, 1, 3, 4)
-        sim = torch.einsum("p b t h n, h k p n -> p b t h k", queries, self.keys)
-
-        # get experts with the highest similarity
-        (scores_x, scores_y), (indices_x, indices_y) = sim.topk(self.num_cdmoe_experts_per_head, dim=-1)
-        if einx_add is not None:
-            all_scores = einx_add("... i, ... j -> ... (i j)", scores_x, scores_y)
-            all_indices = einx_add("... i, ... j -> ... (i j)", indices_x * self.num_keys, indices_y)
-        else:
-            all_scores = scores_x.unsqueeze(-1) + scores_y.unsqueeze(-2)
-            all_scores = all_scores.view(*scores_x.shape[:-1], -1)
-            all_indices = (indices_x.unsqueeze(-1) * self.num_keys) + indices_y.unsqueeze(-2)
-            all_indices = all_indices.view(*indices_x.shape[:-1], -1)
+        # get routing weights with queries and keys
+        queries = self.queries_proj(hidden_states)
+        queries = queries.view(2, self.num_cdmoe_heads, bsz * seq_len, -1)
+        keys = self.keys.view(2, self.num_cdmoe_heads, -1, self.num_keys)
+        routing_weights = torch.matmul(queries, keys)
+        routing_weights = routing_weights.transpose(-2, -3).view(2, bsz, seq_len, self.num_cdmoe_heads, self.num_keys)
+
+        # get experts with the highest routing weights
+        (scores_x, scores_y), (indices_x, indices_y) = routing_weights.topk(self.num_cdmoe_experts_per_head, dim=-1)
+        all_scores = scores_x.unsqueeze(-1) + scores_y.unsqueeze(-2)
+        all_scores = all_scores.view(*scores_x.shape[:-1], -1)
+        all_indices = (indices_x.unsqueeze(-1) * self.num_keys) + indices_y.unsqueeze(-2)
+        all_indices = all_indices.view(*indices_x.shape[:-1], -1)
         scores, pk_indices = all_scores.topk(self.num_cdmoe_experts_per_head, dim=-1)
         indices = all_indices.gather(-1, pk_indices)
         down_embed = self.down_embed(indices)
         up_embed = self.up_embed(indices)
 
         # mix experts states with cross domain states
-        experts_weights = torch.einsum("b t d, b t h k d -> b t h k", hidden_states, down_embed)
+        experts_weights = torch.sum(hidden_states[:, :, None, None, :] * down_embed, dim=-1)
         experts_weights = self.act_fn(experts_weights) * scores.softmax(dim=-1)
-        experts_states = torch.einsum("b t h k, b t h k d -> b t d", experts_weights, up_embed)
+        experts_states = torch.sum(experts_weights[:, :, :, :, None] * up_embed, dim=(-2, -3))
         hidden_states = self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
         hidden_states = hidden_states + experts_states
         return hidden_states
@@ -530,13 +542,13 @@ class DogeDecoderLayer(nn.Module):
         super().__init__()
         self.hidden_dropout = config.hidden_dropout
 
-        self.pre_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.pre_layernorm = DogeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.self_attn = DogeDynamicMaskAttention(config=config, layer_idx=layer_idx)
-        self.pre_residual = Residual(config.hidden_size)
+        self.pre_residual = DogeResidual(config.hidden_size)
 
-        self.post_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.feed_forward = DogeMLP(config) if config.is_moe == False else DogeCDMoE(config)
-        self.post_residual = Residual(config.hidden_size)
+        self.post_layernorm = DogeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.feed_forward = DogeMLP(config) if not config.is_moe else DogeCDMoE(config)
+        self.post_residual = DogeResidual(config.hidden_size)
 
     def forward(
         self,
@@ -550,15 +562,16 @@ class DogeDecoderLayer(nn.Module):
         position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
         **kwargs,
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
-
         # sequence transformation
         residual = hidden_states
         hidden_states = self.pre_layernorm(hidden_states)
-        hidden_states = self.self_attn(
+        hidden_states, self_attn_weights = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
             position_ids=position_ids,
             past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
             cache_position=cache_position,
             position_embeddings=position_embeddings,
             **kwargs,
@@ -596,6 +609,8 @@ DOGE_START_DOCSTRING = r"""
             load the weights associated with the model, only the configuration. Check out the
             [`~PreTrainedModel.from_pretrained`] method to load the model weights.
 """
+
+
 @add_start_docstrings(
     "The bare Doge Model outputting raw hidden-states without any specific head on top.",
     DOGE_START_DOCSTRING,
@@ -607,7 +622,7 @@ class DogePreTrainedModel(PreTrainedModel):
     _no_split_modules = ["DogeDecoderLayer"]
     _skip_keys_device_placement = ["past_key_values"]
     _supports_sdpa = True
-    _supports_flex_attn = True
+    # _supports_flex_attn = True # TODO: enable this when flex_attention is fully supported
     _supports_cache_class = True
     _supports_quantized_cache = True
     _supports_static_cache = True
@@ -718,11 +733,11 @@ class DogeModel(DogePreTrainedModel):
         self.vocab_size = config.vocab_size
 
         self.word_embed = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
-        self.rotary_emb = RotaryEmbedding(config)
+        self.rotary_emb = DogeRotaryEmbedding(config)
         self.layers = nn.ModuleList(
             [DogeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
         )
-        self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.final_layernorm = DogeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.gradient_checkpointing = False
 
         # Initialize weights and apply final processing
@@ -849,9 +864,27 @@ class DogeModel(DogePreTrainedModel):
         past_key_values: Cache,
         output_attentions: bool,
     ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and (attention_mask == 0.0).any():
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
 
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
         dtype, device = input_tensor.dtype, input_tensor.device
         sequence_length = input_tensor.shape[1]
         if using_static_cache:
@@ -863,9 +896,9 @@ class DogeModel(DogePreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        # in case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
         causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
-            attention_mask=attention_mask,
+            attention_mask,
             sequence_length=sequence_length,
             target_length=target_length,
             dtype=dtype,
@@ -874,17 +907,29 @@ class DogeModel(DogePreTrainedModel):
             batch_size=input_tensor.shape[0],
         )
 
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type in ["cuda", "xpu"]
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
         return causal_mask
-    
+
     @staticmethod
     def _prepare_4d_causal_attention_mask_with_cache_position(
-        attention_mask: torch.Tensor = None,
-        sequence_length: int = None,
-        target_length: int = None,
-        dtype: torch.dtype = None,
-        device: torch.device = None,
-        cache_position: torch.Tensor = None,
-        batch_size: int = None,
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
         **kwargs,
     ):
         """
@@ -915,8 +960,7 @@ class DogeModel(DogePreTrainedModel):
         else:
             min_dtype = torch.finfo(dtype).min
             causal_mask = torch.full(
-                (sequence_length, target_length),
-                fill_value=min_dtype, dtype=dtype, device=device,
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
             )
             if sequence_length != 1:
                 causal_mask = torch.triu(causal_mask, diagonal=1)
@@ -934,9 +978,6 @@ class DogeModel(DogePreTrainedModel):
         return causal_mask
 
 
-class KwargsForCausalLM(LossKwargs): ...
-
-
 class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
     _tied_weights_keys = ["lm_head.weight"]
     _tp_plan = {"lm_head": "colwise_rep"}
@@ -962,7 +1003,7 @@ class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
 
     def set_output_embeddings(self, new_embeddings):
         self.lm_head = new_embeddings
-    
+
     def get_decoder(self):
         return self.model
 
@@ -984,8 +1025,8 @@ class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        num_logits_to_keep: int = 0,
-        **kwargs: Unpack[KwargsForCausalLM],
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[LossKwargs],
     ) -> Union[Tuple, CausalLMOutputWithPast]:
         r"""
         Args:
@@ -994,10 +1035,12 @@ class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
                 config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                 (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
 
-            num_logits_to_keep (`int`, *optional*):
-                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+            logits_to_keep (`int`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
                 `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                 token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
 
         Returns:
 
@@ -1006,8 +1049,8 @@ class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
         ```python
          >>> from transformers import AutoTokenizer, AutoModelForCausalLM
 
-        >>> model = AutoModelForCausalLM.from_pretrained("JingzeShi/Doge-20M-Instruct")
-        >>> tokenizer = AutoTokenizer.from_pretrained("JingzeShi/Doge-20M-Instruct")
+        >>> model = AutoModelForCausalLM.from_pretrained("SmallDoge/Doge-20M")
+        >>> tokenizer = AutoTokenizer.from_pretrained("SmallDoge/Doge-20M")
 
         >>> prompt = "Hey, are you conscious? Can you talk to me?"
         >>> inputs = tokenizer(prompt, return_tensors="pt")
@@ -1039,9 +1082,9 @@ class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
         )
 
         hidden_states = outputs[0]
-
         # only compute necessary logits, and do not upcast them to float if we are not computing the loss
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        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:
@@ -1060,111 +1103,32 @@ class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
         )
 
 
-class DogePatchEmbedding(nn.Module):
-    """
-    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial `hidden_states` of shape `(batch_size, seq_len, hidden_size)` to be consumed by a Transformer.
-    """
-
-    def __init__(self, config: DogeConfig):
-        super().__init__()
-
-        self.num_channels = config.num_channels
-        self.patch_size = config.patch_size
-        self.hidden_dim = config.hidden_size
-
-        self.sequence_proj = nn.Conv2d(self.num_channels, self.hidden_dim, kernel_size=self.patch_size, stride=self.patch_size)
-        self.state_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=config.hidden_bias)
-
-    def forward(
-        self,
-        pixel_values: torch.Tensor,
-    ) -> torch.Tensor:
-        image_embedding = self.sequence_proj(pixel_values).flatten(2).transpose(1, 2)
-        image_embedding = self.state_proj(image_embedding)
-        return image_embedding
-
-
-class DogeForCausalVLM(DogeForCausalLM):
-    _tied_weights_keys = ["lm_head.weight"]
-
-    def __init__(self, config: DogeConfig):
-        super().__init__(config)
-        self.config = config
-        self.pixel_embed = DogePatchEmbedding(config)
-
-        # Initialize weights and apply final processing
-        self.post_init()
-    
-    def forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        pixel_values: torch.FloatTensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[torch.Tensor] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        labels: Optional[torch.LongTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        return_dict: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        num_logits_to_keep: int = 0,
-        **loss_kwargs,
-    ) -> Union[Tuple, CausalLMOutputWithPast]:
-        # TODO: @wubingheng111: refer to Llava for implementating the forward method
-        ...
-    
-    def prepare_inputs_for_generation(
-        self,
-        input_ids=None,
-        pixel_values=None,
-        past_key_values=None,
-        input_embeds=None,
-        attention_mask=None,
-        cache_position=None,
-        num_logits_to_keep=None,
-        **kwargs,
-    ):
-        model_inputs = self.model.prepare_inputs_for_generation(
-            input_ids,
-            past_key_values=past_key_values,
-            inputs_embeds=input_embeds,
-            attention_mask=attention_mask,
-            cache_position=cache_position,
-            num_logits_to_keep=num_logits_to_keep,
-            **kwargs,
-        )
-
-        if cache_position[0] == 0:
-            model_inputs["pixel_values"] = pixel_values
-
-        return model_inputs
-
-
 @add_start_docstrings(
     """
     The Doge Model transformer with a sequence classification head on top (linear layer).
 
-    [`DogeForSequenceClassification`] uses the last token in order to do the classification, as other causal models (e.g. GPT-2) do.
+    [`DogeForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
 
-    Since it does classification on the last token, it requires to know the position of the last token. 
-    If a `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. 
-    If no `pad_token_id` is defined, it simply takes the last value in each row of the batch. 
-    Since it cannot guess the padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in each row of the batch).
-    """
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    DOGE_START_DOCSTRING,
 )
 class DogeForSequenceClassification(DogePreTrainedModel):
     def __init__(self, config: DogeConfig):
         super().__init__(config)
-        self.config = config
         self.num_labels = config.num_labels
 
         self.model = DogeModel(config)
-        self.classifier = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+        self.config = config
 
         # Initialize weights and apply final processing
-        self.init_weights()
+        self.post_init()
 
     def get_input_embeddings(self):
         return self.model.word_embed
@@ -1188,14 +1152,14 @@ class DogeForSequenceClassification(DogePreTrainedModel):
     ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
-            Labels for computing the sequence classification/regression loss. 
-            Indices should be in `[0, ..., config.num_labels - 1]`. 
-            If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
         """
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
-        outputs = self.model(
-            input_ids=input_ids,
+        transformer_outputs = self.model(
+            input_ids,
             attention_mask=attention_mask,
             position_ids=position_ids,
             past_key_values=past_key_values,
@@ -1205,8 +1169,8 @@ class DogeForSequenceClassification(DogePreTrainedModel):
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
         )
-        hidden_states = outputs[0]
-        logits = self.classifier(hidden_states)
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
 
         if input_ids is not None:
             batch_size = input_ids.shape[0]
@@ -1216,35 +1180,36 @@ class DogeForSequenceClassification(DogePreTrainedModel):
         if self.config.pad_token_id is None and batch_size != 1:
             raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
         if self.config.pad_token_id is None:
-            sequence_lengths = -1
+            last_non_pad_token = -1
+        elif input_ids is not None:
+            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
+            non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
+            token_indices = torch.arange(input_ids.shape[-1], device=logits.device)
+            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
         else:
-            if input_ids is not None:
-                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
-                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
-                sequence_lengths = sequence_lengths % input_ids.shape[-1]
-                sequence_lengths = sequence_lengths.to(logits.device)
-            else:
-                sequence_lengths = -1
+            last_non_pad_token = -1
+            logger.warning_once(
+                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+            )
 
-        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token]
 
         loss = None
         if labels is not None:
-            loss = self.loss_function(
-                logits=logits,
-                labels=labels,
-                pooled_logits=pooled_logits,
-                config=self.config,
-            )
+            loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
 
         if not return_dict:
-            output = (pooled_logits,) + outputs[1:]
+            output = (pooled_logits,) + transformer_outputs[1:]
             return ((loss,) + output) if loss is not None else output
 
         return SequenceClassifierOutputWithPast(
             loss=loss,
             logits=pooled_logits,
-            past_key_values=outputs.past_key_values,
-            hidden_states=outputs.hidden_states,
-            attentions=outputs.attentions,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
         )
+
+
+__all__ = ["DogeForCausalLM", "DogeModel", "DogePreTrainedModel", "DogeForSequenceClassification"]