transformers/pr_40546/en/model_doc/afmoe.md

# 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](/docs/transformers/pr_40546/en/main_classes/pipelines#transformers.Pipeline) or the [AutoModel](/docs/transformers/pr_40546/en/model_doc/auto#transformers.AutoModel).

```python
from transformers import pipeline

pipeline = pipeline(
    task="text-generation",
    model="arcee-ai/Trinity-Mini",
    device=0
)

output = pipeline("The key innovation in mixture of experts is")
print(output[0]["generated_text"])
```

```python
import torch

from transformers import AfmoeForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("arcee-ai/Trinity-Mini")
model = AfmoeForCausalLM.from_pretrained(
    "arcee-ai/Trinity-Mini",
    device_map="auto"
)

inputs = tokenizer("The key innovation in mixture of experts is", return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=50)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## 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[[transformers.AfmoeConfig]]

- **vocab_size** (`int`, *optional*, defaults to `200192`) --
  Vocabulary size of the model. Defines the number of different tokens that can be represented by the `input_ids`.
- **hidden_size** (`int`, *optional*, defaults to `2048`) --
  Dimension of the hidden representations.
- **intermediate_size** (`int`, *optional*, defaults to `6144`) --
  Dimension of the 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 attention head dimension. If None, it will default to hidden_size // num_attention_heads
- **hidden_act** (`str`, *optional*, defaults to `silu`) --
  The non-linear activation function (function or string) in the decoder. For example, `"gelu"`,
  `"relu"`, `"silu"`, etc.
- **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` or when the model is a decoder-only generative model.
- **tie_word_embeddings** (`bool`, *optional*, defaults to `False`) --
  Whether to tie weight embeddings according to model's `tied_weights_keys` mapping.
- **rope_parameters** (`Union[~modeling_rope_utils.RopeParameters, dict]`, *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 or routed experts.
- **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.
- **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 attention window size. If `None`, no sliding window is applied.
- **layer_types** (`list[str]`, *optional*) --
  A list that explicitly maps each layer index with its layer type. If not provided, it will be automatically
  generated based on config values.
- **attention_dropout** (`Union[float, int]`, *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)`.
- **eos_token_id** (`Union[int, list[int]]`, *optional*) --
  Token id used for end-of-stream in the vocabulary.
- **pad_token_id** (`int`, *optional*) --
  Token id used for padding in the vocabulary.
- **bos_token_id** (`int`, *optional*) --
  Token id used for beginning-of-stream in the vocabulary.
- **attention_bias** (`bool`, *optional*, defaults to `False`) --
  Whether to use a bias in the query, key, value and output projection layers during self-attention.

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.

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
```

## AfmoeModel[[transformers.AfmoeModel]]

- **config** ([AfmoeConfig](/docs/transformers/pr_40546/en/model_doc/afmoe#transformers.AfmoeConfig)) --
  Model configuration class with all the parameters of the model. Initializing with a config file does not
  load the weights associated with the model, only the configuration. Check out the
  [from_pretrained()](/docs/transformers/pr_40546/en/main_classes/model#transformers.PreTrainedModel.from_pretrained) method to load the model weights.

The bare Afmoe Model outputting raw hidden-states without any specific head on top.

This model inherits from [PreTrainedModel](/docs/transformers/pr_40546/en/main_classes/model#transformers.PreTrainedModel). Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)

This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.

- **input_ids** (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.

  Indices can be obtained using [AutoTokenizer](/docs/transformers/pr_40546/en/model_doc/auto#transformers.AutoTokenizer). See [PreTrainedTokenizer.encode()](/docs/transformers/pr_40546/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode) and
  [PreTrainedTokenizer.__call__()](/docs/transformers/pr_40546/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__) for details.

  [What are input IDs?](../glossary#input-ids)
- **attention_mask** (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

  - 1 for tokens that are **not masked**,
  - 0 for tokens that are **masked**.

  [What are attention masks?](../glossary#attention-mask)
- **inputs_embeds** (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) --
  Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
  is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
  model's internal embedding lookup matrix.
- **position_ids** (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`.

  [What are position IDs?](../glossary#position-ids)
- **past_key_values** (`~cache_utils.Cache`, *optional*) --
  Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
  blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
  returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

  Only [Cache](/docs/transformers/pr_40546/en/internal/generation_utils#transformers.Cache) instance is allowed as input, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
  If no `past_key_values` are passed, [DynamicCache](/docs/transformers/pr_40546/en/internal/generation_utils#transformers.DynamicCache) will be initialized by default.

  The model will output the same cache format that is fed as input.

  If `past_key_values` are used, the user is expected to input only unprocessed `input_ids` (those that don't
  have their past key value states given to this model) of shape `(batch_size, unprocessed_length)` instead of all `input_ids`
  of shape `(batch_size, sequence_length)`.
- **use_cache** (`bool`, *optional*) --
  If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
  `past_key_values`).`MoeModelOutputWithPast` or `tuple(torch.FloatTensor)`A `MoeModelOutputWithPast` or a tuple of
`torch.FloatTensor` (if `return_dict=False` is passed or when `config.return_dict=False`) comprising various
elements depending on the configuration (`None`) and inputs.
The [AfmoeModel](/docs/transformers/pr_40546/en/model_doc/afmoe#transformers.AfmoeModel) forward method, overrides the `__call__` special method.

Although the recipe for forward pass needs to be defined within this function, one should call the `Module`
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.

- **last_hidden_state** (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`) -- Sequence of hidden-states at the output of the last layer of the model.
- **past_key_values** (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`) -- It is a [Cache](/docs/transformers/pr_40546/en/internal/generation_utils#transformers.Cache) instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

  Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
  `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
  input) to speed up sequential decoding.
- **hidden_states** (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`) -- Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
  one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

  Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
- **attentions** (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`) -- Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
  sequence_length)`.

  Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
  heads.
- **router_logits** (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_probs=True` and `config.add_router_probs=True` is passed or when `config.output_router_probs=True`) -- Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.

  Raw router logtis (post-softmax) that are computed by MoE routers, these terms are used to compute the auxiliary
  loss for Mixture of Experts models.

## AfmoeForCausalLM[[transformers.AfmoeForCausalLM]]

- **config** ([AfmoeForCausalLM](/docs/transformers/pr_40546/en/model_doc/afmoe#transformers.AfmoeForCausalLM)) --
  Model configuration class with all the parameters of the model. Initializing with a config file does not
  load the weights associated with the model, only the configuration. Check out the
  [from_pretrained()](/docs/transformers/pr_40546/en/main_classes/model#transformers.PreTrainedModel.from_pretrained) method to load the model weights.

The Afmoe Model for causal language modeling.

This model inherits from [PreTrainedModel](/docs/transformers/pr_40546/en/main_classes/model#transformers.PreTrainedModel). Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)

This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.

- **input_ids** (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.

  Indices can be obtained using [AutoTokenizer](/docs/transformers/pr_40546/en/model_doc/auto#transformers.AutoTokenizer). See [PreTrainedTokenizer.encode()](/docs/transformers/pr_40546/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode) and
  [PreTrainedTokenizer.__call__()](/docs/transformers/pr_40546/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__) for details.

  [What are input IDs?](../glossary#input-ids)
- **attention_mask** (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

  - 1 for tokens that are **not masked**,
  - 0 for tokens that are **masked**.

  [What are attention masks?](../glossary#attention-mask)
- **position_ids** (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`.

  [What are position IDs?](../glossary#position-ids)
- **past_key_values** (`~cache_utils.Cache`, *optional*) --
  Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
  blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
  returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

  Only [Cache](/docs/transformers/pr_40546/en/internal/generation_utils#transformers.Cache) instance is allowed as input, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
  If no `past_key_values` are passed, [DynamicCache](/docs/transformers/pr_40546/en/internal/generation_utils#transformers.DynamicCache) will be initialized by default.

  The model will output the same cache format that is fed as input.

  If `past_key_values` are used, the user is expected to input only unprocessed `input_ids` (those that don't
  have their past key value states given to this model) of shape `(batch_size, unprocessed_length)` instead of all `input_ids`
  of shape `(batch_size, sequence_length)`.
- **inputs_embeds** (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) --
  Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
  is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
  model's internal embedding lookup matrix.
- **labels** (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
  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]`.
- **use_cache** (`bool`, *optional*) --
  If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
  `past_key_values`).
- **output_router_logits** (`bool`, *optional*) --
  Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
  should not be returned during inference.
- **logits_to_keep** (`Union[int, torch.Tensor]`, *optional*, defaults to `0`) --
  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).`MoeCausalLMOutputWithPast` or `tuple(torch.FloatTensor)`A `MoeCausalLMOutputWithPast` or a tuple of
`torch.FloatTensor` (if `return_dict=False` is passed or when `config.return_dict=False`) comprising various
elements depending on the configuration ([AfmoeConfig](/docs/transformers/pr_40546/en/model_doc/afmoe#transformers.AfmoeConfig)) and inputs.
The [AfmoeForCausalLM](/docs/transformers/pr_40546/en/model_doc/afmoe#transformers.AfmoeForCausalLM) forward method, overrides the `__call__` special method.

Although the recipe for forward pass needs to be defined within this function, one should call the `Module`
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.

- **loss** (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided) -- Language modeling loss (for next-token prediction).
- **logits** (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`) -- Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
- **aux_loss** (`torch.FloatTensor`, *optional*, returned when `labels` is provided) -- aux_loss for the sparse modules.
- **router_logits** (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_probs=True` and `config.add_router_probs=True` is passed or when `config.output_router_probs=True`) -- Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.

  Raw router logtis (post-softmax) that are computed by MoE routers, these terms are used to compute the auxiliary
  loss for Mixture of Experts models.
- **past_key_values** (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`) -- It is a [Cache](/docs/transformers/pr_40546/en/internal/generation_utils#transformers.Cache) instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

  Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
  `past_key_values` input) to speed up sequential decoding.
- **hidden_states** (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`) -- Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
  one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

  Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
- **attentions** (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`) -- Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
  sequence_length)`.

  Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
  heads.

Example:

```python
>>> from transformers import AutoTokenizer, AfmoeForCausalLM

>>> model = AfmoeForCausalLM.from_pretrained("meta-afmoe/Afmoe-2-7b-hf")
>>> tokenizer = AutoTokenizer.from_pretrained("meta-afmoe/Afmoe-2-7b-hf")

>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")

>>> # Generate
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
```