Adds Doge model implementation with dynamic attention

Introduces a new transformer model architecture featuring dynamic masked attention (DMA) mechanism that adaptively masks tokens based on learned importance scores. The implementation includes support for mixture of experts (MoE) with cross-domain routing, RoPE positional embeddings, and flexible attention backends.

Key features include configurable sliding window attention, dynamic masking for sparse attention patterns, and efficient expert routing for scalable model capacity.

Author: LoserCheems

examples/modeling/configuration_doge.py

@@ -0,0 +1,241 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           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 2025 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
+#
+# The Doge family of small language models is trained by SmallDoge Team.
+#
+# 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
+
+
+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 [SmallDoge/Doge-320M](https://huggingface.co/SmallDoge/Doge-320M).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32768):
+            Vocabulary size of the Doge2 model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`DogeModel`]
+        hidden_size (`int`, *optional*, defaults to 1024):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        hidden_dropout (`float`, *optional*, defaults to 0.0):
+            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`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        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.
+            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.
+                    In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'.
+                    The original max position embeddings used during pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation.
+                    If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<`original_max_position_embeddings`).
+                    Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<`original_max_position_embeddings`).
+                    Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        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*):
+            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).
+            If it is not specified, will default to `num_attention_heads`.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
+        sliding_window (`int`, *optional*):
+            Sliding window attention window size. If not specified, will default to `None`.
+        keep_window_size (`int`, *optional*, defaults to 2048):
+            The window size of tokens that are not dynamically masked, and dynamic masking is only performed when the sequence length exceeds this value.
+        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.
+        num_experts (`int`, *optional*, defaults to 16384):
+            Number of routed experts in the model. This is only used when `is_moe=True`.
+        num_experts_per_tok (`int`, *optional*, defaults to 64):
+            Number of selected experts to route per-token.
+        norm_topk_prob (`bool`, *optional*, defaults to `False`):
+            Whether to normalize the topk probabilities.
+        output_router_logits (`bool`, *optional*, defaults to `False`):
+            Whether or not the router logits should be returned by the model. Enabling this will also
+            allow the model to output the auxiliary loss, including load balancing loss and router z-loss.
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            The aux loss factor for the total loss.
+
+    ```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"]
+    # Default tensor parallel plan for base model `DogeModel`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.dt_proj": "rowwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.input_layernorm.weight": "sequence_parallel",
+        "layers.*.input_residual.weight": "sequence_parallel",
+        "layers.*.post_attention_layernorm.weight": "sequence_parallel",
+        "layers.*.post_attention_residual.weight": "sequence_parallel",
+        "norm.weight": "sequence_parallel",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+        "layers.*.mlp.router_gate": "colwise_rep",
+        "layers.*.mlp.down_embed": "rowwise_rep",
+        "layers.*.mlp.up_embed": "rowwise_rep",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=32768,
+        hidden_size=1024,
+        intermediate_size=2048,
+        num_hidden_layers=32,
+        hidden_dropout=0.0,
+        hidden_act="silu",
+        initializer_range=0.02,
+        rms_norm_eps=1e-06,
+        use_cache=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_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        sliding_window=None,
+        keep_window_size=2048,
+        is_moe=False,
+        num_experts=16384,
+        num_experts_per_tok=64,
+        norm_topk_prob=False,
+        output_router_logits=False,
+        router_aux_loss_coef=0.001,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+
+        self.hidden_dropout = hidden_dropout
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+
+        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_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+        self.sliding_window = sliding_window
+        self.keep_window_size = keep_window_size
+        self.is_moe = is_moe
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.norm_topk_prob = norm_topk_prob
+        self.output_router_logits = output_router_logits
+        self.router_aux_loss_coef = router_aux_loss_coef
+
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, copy it it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            self.num_key_value_heads = num_attention_heads
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["DogeConfig"]