Add afmoe model (#42168)

* Add AFMoE model support

* Address review feedback for AFMoE implementation

* Add flex attention support to AFMoE model

* Fix expert_bias routing in AFMoE

* Remove test-results directory

* Address PR review feedback for AFMoE model

* fix(afmoe): ensure RMSNorm output dtype matches input dtype)

* properly return attn weights

* fix most tests

* cleanup
Remove shared expert if else as defaults to 2
Remove `route_norm` as it default to `True`.

Make test smaller faster

* fix input embeds api

* update rope API, smaller test and should be good to go

* oups wront place to skip unittest

* quality

* update

* rope parameter docstring fill

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Arthur <arthur.zucker@gmail.com>

Author: alyosha-swamy

.gitignore

@@ -175,3 +175,4 @@ tags
 
 # Cursor IDE files
 .cursor/
+test-results/

docs/source/en/_toctree.yml

@@ -384,6 +384,8 @@
     title: Main Classes
   - sections:
     - sections:
+      - local: model_doc/afmoe
+        title: AFMoE
       - local: model_doc/albert
         title: ALBERT
       - local: model_doc/apertus

docs/source/en/model_doc/afmoe.md

@@ -0,0 +1,130 @@
+<!--Copyright 2025 Arcee AI and The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+*This model was released on {release_date} and added to Hugging Face Transformers on 2025-11-18.*
+
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
+</div>
+
+# AFMoE
+
+AFMoE (Arcee Foundational Mixture of Experts) is a decoder-only transformer model that extends the Llama architecture with a sparse Mixture of Experts (MoE) approach. The model combines token-choice routing with shared experts and employs several architectural innovations for efficient inference and improved performance.
+
+## Key Architecture Features
+
+AFMoE introduces several key modifications to the standard transformer architecture:
+
+- **Mixture of Experts with Shared Experts**: Combines routed experts (activated per-token via learned routing) with always-active shared experts for stable base computation
+- **Token-Choice Routing**: Uses sigmoid or softmax-based routing with normalization and scaling for expert selection
+- **Q/K Normalization and Gating**: Applies RMSNorm to query and key projections and uses sigmoid gating on attention outputs for improved stability
+- **Hybrid Attention Patterns**: Alternates between sliding window attention and full attention across layers for efficiency with long contexts
+- **Dual Normalization**: Uses pre- and post-normalization around both attention and MLP blocks for training stability
+- **Configurable Dense Layers**: Allows initial layers to use dense MLPs before transitioning to sparse MoE layers
+
+The model supports extended context lengths with RoPE embeddings and includes all standard Transformers features including Flash Attention 2, SDPA, gradient checkpointing, and quantization support.
+
+> [!TIP]
+> AFMoE is particularly well-suited for scenarios requiring efficient scaling through sparsity while maintaining strong performance. The shared experts provide a stable computation baseline while routed experts enable model capacity scaling.
+
+The example below demonstrates how to generate text with AFMoE using [`Pipeline`] or the [`AutoModel`].
+
+<hfoptions id="usage">
+<hfoption id="Pipeline">
+
+```py
+import torch
+from transformers import pipeline
+
+pipeline = pipeline(
+    task="text-generation",
+    model="arcee-ai/Trinity-Mini",
+    torch_dtype=torch.bfloat16,
+    device=0
+)
+
+output = pipeline("The key innovation in mixture of experts is")
+print(output[0]["generated_text"])
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```py
+import torch
+from transformers import AutoTokenizer, AfmoeForCausalLM
+
+tokenizer = AutoTokenizer.from_pretrained("arcee-ai/Trinity-Mini")
+model = AfmoeForCausalLM.from_pretrained(
+    "arcee-ai/Trinity-Mini",
+    torch_dtype=torch.bfloat16,
+    device_map="auto"
+)
+
+inputs = tokenizer("The key innovation in mixture of experts is", return_tensors="pt")
+with torch.no_grad():
+    outputs = model.generate(**inputs, max_new_tokens=50)
+
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+</hfoption>
+</hfoptions>
+
+## Model Architecture Details
+
+### Expert Routing
+
+AFMoE uses token-choice routing where each token independently selects top-k experts based on router logits. The routing mechanism includes:
+
+- Configurable scoring function (sigmoid or softmax)
+- Optional route normalization for balanced expert utilization
+- Route scaling to control expert contribution strength
+- Bias correction for expert selection
+
+### Shared Experts
+
+Unlike standard MoE models, AFMoE includes shared experts that are always activated for every token, providing:
+
+- A stable computation baseline across all tokens
+- Reduced variance in model outputs
+- Better handling of out-of-distribution inputs
+
+### Attention Mechanism
+
+The hybrid attention pattern alternates between:
+
+- **Sliding Window Attention**: For efficiency on long sequences, with configurable window size
+- **Full Attention**: Applied every N layers (configurable via `global_attn_every_n_layers`) for global context
+
+All attention layers include Q/K normalization and output gating for improved training dynamics.
+
+## AfmoeConfig
+
+[[autodoc]] AfmoeConfig
+
+## AfmoeModel
+
+[[autodoc]] AfmoeModel
+    - forward
+
+## AfmoeForCausalLM
+
+[[autodoc]] AfmoeForCausalLM
+    - forward

splitted_tests.txt

@@ -0,0 +1 @@
+tests/models/afmoe/test_modeling_afmoe.py

src/transformers/models/afmoe/__init__.py

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

src/transformers/models/afmoe/configuration_afmoe.py

@@ -0,0 +1,210 @@
+# coding=utf-8
+# Copyright 2025 Arcee AI 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.
+"""AFMoE model configuration"""
+
+from typing import Optional
+
+from ...configuration_utils import PreTrainedConfig, layer_type_validation
+from ...modeling_rope_utils import RopeParameters
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class AfmoeConfig(PreTrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AfmoeModel`]. It is used to instantiate an
+    AFMoE model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of [arcee-ai/Trinity-Mini](https://huggingface.co/arcee-ai/Trinity-Mini).
+
+    AFMoE is an Adaptive Feedforward MoE (Mixture of Experts) model with token-choice routing, shared experts, and a
+    hybrid attention mechanism combining sliding window and full attention patterns.
+
+    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 200192):
+            Vocabulary size of the AFMoE model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`AfmoeModel`].
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 6144):
+            Dimension of the dense MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1408):
+            Intermediate size of the routed expert MLPs.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_dense_layers (`int`, *optional*, defaults to 1):
+            Number of initial dense layers before MoE layers begin. Layers with index < num_dense_layers will use
+            standard dense MLPs instead of MoE.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            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, check out [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to
+            `num_attention_heads`.
+        head_dim (`int`, *optional*, defaults to 128):
+            The dimension of each attention head.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the MLP blocks.
+        max_position_embeddings (`int`, *optional*, defaults to 16384):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the RMS normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_parameters (`RopeParameters`, *optional*):
+            Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
+            a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
+            with longer `max_position_embeddings`.
+        num_experts (`int`, *optional*, defaults to 64):
+            Number of routed experts in MoE layers.
+        num_experts_per_tok (`int`, *optional*, defaults to 6):
+            Number of experts to route each token to. This is the top-k value for the token-choice routing.
+        num_shared_experts (`int`, *optional*, defaults to 2):
+            Number of shared experts that are always activated for all tokens.
+        route_scale (`float`, *optional*, defaults to 1.0):
+            Scaling factor applied to routing weights.
+        global_attn_every_n_layers (`int`, *optional*, defaults to 4):
+            The frequency of full attention layers. Every Nth layer will use full attention, while others use sliding
+            window attention.
+        sliding_window (`int`, *optional*, defaults to 1024):
+            Sliding window size for local attention layers.
+        layer_types (`list[str]`, *optional*):
+            A list that explicitly maps each layer index with its attention type. Each element should be either
+            "sliding_attention" or "full_attention". If not provided, it will be automatically generated based on
+            `global_attn_every_n_layers`.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mup_enabled (`bool`, *optional*, defaults to `False`):
+            Whether to enable muP (Maximal Update Parametrization) input scaling. When enabled, input embeddings
+            are scaled by `sqrt(hidden_size)`.
+
+    Example:
+    ```python
+    >>> from transformers import AfmoeModel, AfmoeConfig
+
+    >>> # Initializing an AFMoE configuration
+    >>> configuration = AfmoeConfig()
+
+    >>> # Initializing a model from the afmoe-small-sft-v1 style configuration
+    >>> model = AfmoeModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "afmoe"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default pipeline parallel plan for base model
+    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: Optional[int] = 200192,
+        hidden_size: Optional[int] = 2048,
+        intermediate_size: Optional[int] = 6144,
+        moe_intermediate_size: Optional[int] = 1408,
+        num_hidden_layers: Optional[int] = 32,
+        num_dense_layers: Optional[int] = 1,
+        num_attention_heads: Optional[int] = 16,
+        num_key_value_heads: Optional[int] = None,
+        head_dim: Optional[int] = 128,
+        hidden_act: Optional[str] = "silu",
+        max_position_embeddings: Optional[int] = 16384,
+        initializer_range: Optional[float] = 0.02,
+        rms_norm_eps: Optional[float] = 1e-5,
+        use_cache: Optional[bool] = True,
+        tie_word_embeddings: Optional[bool] = False,
+        rope_theta: Optional[float] = 10000.0,
+        rope_parameters: Optional[RopeParameters | dict[str, RopeParameters]] = None,
+        num_experts: Optional[int] = 64,
+        num_experts_per_tok: Optional[int] = 6,
+        num_shared_experts: Optional[int] = 2,
+        route_scale: Optional[float] = 1.0,
+        global_attn_every_n_layers: Optional[int] = 4,
+        sliding_window: Optional[int] = 1024,
+        layer_types: Optional[list] = None,
+        attention_dropout: Optional[float] = 0.0,
+        mup_enabled: Optional[bool] = False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_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_parameters = rope_parameters
+
+        # MoE specific
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_experts_per_tok = num_experts_per_tok
+        self.num_experts = num_experts
+        self.num_shared_experts = num_shared_experts
+        self.route_scale = route_scale
+        self.attention_bias = False
+
+        # Attention specific
+        self.attention_dropout = attention_dropout
+        self.global_attn_every_n_layers = global_attn_every_n_layers
+        self.sliding_window = sliding_window
+        self.mup_enabled = mup_enabled
+        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)
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["AfmoeConfig"]