# 맘바[[mamba]]

## 개요[[overview]]

맘바(Mamba) 모델은 Albert Gu, Tri Dao가 제안한 [맘바: 선택적 상태 공간을 이용한 선형 시간 시퀀스 모델링](https://huggingface.co/papers/2312.00752)라는 논문에서 소개 되었습니다.

이 모델은 `state-space-models`을 기반으로 한 새로운 패러다임 아키텍처입니다. 직관적인 이해를 얻고 싶다면 [이곳](https://srush.github.io/annotated-s4/)을 참고 하세요.

해당 논문의 초록입니다:

*현재 딥러닝에서 흥미로운 응용 프로그램을 구동하는 대부분의 기초 모델들은 거의 보편적으로 트랜스포머 아키텍처와 그 핵심 어텐션 모듈을 기반으로 합니다. 선형 어텐션, 게이트된 컨볼루션과 순환 모델, 구조화된 상태 공간 모델(SSM) 등 많은 준이차시간(subquadratic-time) 아키텍처가 긴 시퀀스에 대한 트랜스포머의 계산 비효율성을 해결하기 위해 개발되었지만, 언어와 같은 중요한 양식에서는 어텐션만큼 성능을 내지 못했습니다. 우리는 이러한 모델의 주요 약점이 내용 기반 추론을 수행하지 못한다는 점임을 알고 몇 가지를 개선했습니다. 첫째, SSM 매개변수를 입력의 함수로 만드는 것만으로도 이산 모달리티(discrete modalities)의 약점을 해결할 수 있어, 현재 토큰에 따라 시퀀스 길이 차원을 따라 정보를 선택적으로 전파하거나 잊을 수 있게 합니다. 둘째, 이러한 변경으로 효율적인 컨볼루션을 사용할 수 없게 되었지만, 우리는 순환 모드에서 하드웨어를 인식하는 병렬 알고리즘을 설계했습니다. 우리는 이러한 선택적 SSM을 어텐션이나 MLP 블록도 없는 단순화된 종단간 신경망 아키텍처인 맘바에 통합시켰습니다. 맘바는 빠른 추론(트랜스포머보다 5배 높은 처리량)과 시퀀스 길이에 대한 선형 확장성을 누리며, 백만 길이 시퀀스까지 실제 데이터에서 성능이 향상됩니다. 일반적인 시퀀스 모델 백본으로서 맘바는 언어, 오디오, 유전체학과 같은 여러 양식에서 최첨단 성능을 달성합니다. 언어 모델링에서 우리의 맘바-3B 모델은 같은 크기의 트랜스포머를 능가하고 두 배 크기의 트랜스포머와 맞먹는 성능을 보이며, 사전 훈련과 다운스트림 평가 모두에서 성능을 나타납니다.*

팁:

- 맘바는 고전적인 트랜스포머와 견줄 만한 새로운 `상태 공간 모델` 아키텍처입니다. 이는 구조화된 상태 공간 모델의 발전 선상에 있으며, [플래시어텐션](https://github.com/Dao-AILab/flash-attention)의 정신을 따르는 효율적인 하드웨어 인식 설계와 구현을 특징으로 합니다.
- 맘바는 `어텐션` 레이어와 동등한 `믹서(mixer)` 레이어를 쌓습니다. `맘바`의 핵심 로직은 `MambaMixer` 클래스에 있습니다.
- 두 가지 구현이 공존합니다: 하나는 최적화되어 빠른 cuda커널을 사용하고, 다른 하나는 단순하지만 모든 장치에서 실행할 수 있습니다!
- 현재 구현은 원본 cuda커널을 활용합니다: 맘바를 위한 플래시 어텐션의 역할을 하는 것은 [`mamba-ssm`](https://github.com/state-spaces/mamba)와 [`causal_conv1d`](https://github.com/Dao-AILab/causal-conv1d) 저장소에 호스팅되어 있습니다. 하드웨어가 지원한다면 반드시 설치하세요!
- cuda 커널을 최적화하는 방향 보다는, 단순하지만 모든 장치에서 실행가능하도록하는 방향인 '단순구현'의 성능을 빠르게 향상시키는 기여를 더 환영하고 있습니다. 🤗

이 모델은 [ArthurZ](https://huggingface.co/ArthurZ)에 의해 기여되었습니다.
원본 코드는 [이곳](https://github.com/state-spaces/mamba)에서 확인할 수 있습니다.

# 사용

### 간단한 생성 예제
 
```python 
from transformers import MambaConfig, MambaForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("state-spaces/mamba-130m-hf")
model = MambaForCausalLM.from_pretrained("state-spaces/mamba-130m-hf")
input_ids = tokenizer("Hey how are you doing?", return_tensors= "pt")["input_ids"]

out = model.generate(input_ids, max_new_tokens=10)
print(tokenizer.batch_decode(out))
```

### Peft 파인튜닝
느린 버전은 학습에서 아주 안정적이진 않습니다. 빠른 버전은 `float32`가 필요합니다!

```python 
from datasets import load_dataset
from trl import SFTConfig, SFTTrainer
from peft import LoraConfig

model_id = "state-spaces/mamba-130m-hf"
dataset = load_dataset("Abirate/english_quotes", split="train")
training_args = SFTConfig(dataset_text_field="quote")
lora_config =  LoraConfig(target_modules=["x_proj", "embeddings", "in_proj", "out_proj"])
trainer = SFTTrainer(
    model=model_id,
    args=training_args,
    train_dataset=dataset,
    peft_config=lora_config,
)
trainer.train()
```

## MambaConfig[[transformers.MambaConfig]]

#### transformers.MambaConfig[[transformers.MambaConfig]]

[Source](https://github.com/huggingface/transformers/blob/v5.8.0/src/transformers/models/mamba/configuration_mamba.py#L26)

This is the configuration class to store the configuration of a MambaModel. It is used to instantiate a Mamba
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the [state-spaces/mamba-2.8b](https://huggingface.co/state-spaces/mamba-2.8b)

Configuration objects inherit from [PreTrainedConfig](/docs/transformers/v5.8.0/ko/main_classes/configuration#transformers.PreTrainedConfig) and can be used to control the model outputs. Read the
documentation from [PreTrainedConfig](/docs/transformers/v5.8.0/ko/main_classes/configuration#transformers.PreTrainedConfig) for more information.

Example:

```python
>>> from transformers import MambaConfig, MambaModel

>>> # Initializing a Mamba configuration
>>> configuration = MambaConfig()

>>> # Initializing a model (with random weights) from the configuration
>>> model = MambaModel(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config
```

**Parameters:**

vocab_size (`int`, *optional*, defaults to `50280`) : 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 `768`) : Dimension of the hidden representations.

state_size (`int`, *optional*, defaults to `16`) : Size of the SSM state (latent state dimension) in the Mamba layers.

num_hidden_layers (`int`, *optional*, defaults to `32`) : Number of hidden layers in the Transformer decoder.

layer_norm_epsilon (`float`, *optional*, defaults to 1e-05) : The epsilon to use in the layer normalization layers.

pad_token_id (`int`, *optional*, defaults to `0`) : Token id used for padding in the vocabulary.

bos_token_id (`int`, *optional*, defaults to `0`) : Token id used for beginning-of-stream in the vocabulary.

eos_token_id (`Union[int, list[int]]`, *optional*, defaults to `0`) : Token id used for end-of-stream in the vocabulary.

expand (`int`, *optional*, defaults to 2) : Expanding factor used to determine the intermediate size.

conv_kernel (`int`, *optional*, defaults to `4`) : The size of the convolutional kernel.

use_bias (`bool`, *optional*, defaults to `False`) : Whether or not to use bias in ["in_proj", "out_proj"] of the mixer block

use_conv_bias (`bool`, *optional*, defaults to `True`) : Whether or not to use bias in the convolution layer of the mixer block.

hidden_act (`str`, *optional*, defaults to `silu`) : The non-linear activation function (function or string) in the decoder. For example, `"gelu"`, `"relu"`, `"silu"`, etc.

initializer_range (`float`, *optional*, defaults to `0.1`) : The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

residual_in_fp32 (`bool`, *optional*, defaults to `True`) : Whether or not residuals should be in `float32`. If set to `False` residuals will keep the same `dtype` as the rest of the model

time_step_rank (`Union[str, int]`, *optional*, defaults to `auto`) : Rank of the delta (time step) projection. Can be `"auto"` to set it automatically.

time_step_scale (`float`, *optional*, defaults to `1.0`) : Scale applied to the time step delta before discretization.

time_step_min (`float`, *optional*, defaults to `0.001`) : Minimum `time_step` used to bound `dt_proj.bias`.

time_step_max (`float`, *optional*, defaults to `0.1`) : Maximum `time_step` used to bound `dt_proj.bias`.

time_step_init_scheme (`str`, *optional*, defaults to `random`) : Initialization scheme for the time step delta. Can be `"random"` or `"uniform"`.

time_step_floor (`float`, *optional*, defaults to `0.0001`) : Minimum allowed value for the discrete time step delta after softplus activation.

rescale_prenorm_residual (`bool`, *optional*, defaults to `False`) : Whether or not to rescale `out_proj` weights when initializing.

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.

use_mambapy (`bool`, *optional*, defaults to `False`) : Determines the fallback strategy during training if the CUDA-based official implementation of Mamba is not available. If `True`, the mamba.py implementation is used. If `False`, the naive and slower implementation is used. Consider switching to the naive version if memory is limited.

use_associative_scan (`bool`, *optional*, defaults to `True`) : Whether to use PyTorch's `torch._higher_order_ops.associative_scan` for the parallel scan instead of the naive sequential implementation. The associative scan is only active during `torch.compile` tracing and requires torch >= 2.9.0. Both paths are tested to produce numerically identical results (see `test_associative_scan_matches_sequential`). Set to `False` to fall back to the sequential loop.

tie_word_embeddings (`bool`, *optional*, defaults to `True`) : Whether to tie weight embeddings according to model's `tied_weights_keys` mapping.

## MambaModel[[transformers.MambaModel]]

#### transformers.MambaModel[[transformers.MambaModel]]

[Source](https://github.com/huggingface/transformers/blob/v5.8.0/src/transformers/models/mamba/modeling_mamba.py#L519)

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

This model inherits from [PreTrainedModel](/docs/transformers/v5.8.0/ko/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.

forwardtransformers.MambaModel.forwardhttps://github.com/huggingface/transformers/blob/v5.8.0/src/transformers/models/mamba/modeling_mamba.py#L544[{"name": "input_ids", "val": ": torch.LongTensor | None = None"}, {"name": "inputs_embeds", "val": ": torch.LongTensor | None = None"}, {"name": "cache_params", "val": ": transformers.cache_utils.Cache | None = None"}, {"name": "use_cache", "val": ": bool | None = None"}, {"name": "output_hidden_states", "val": ": bool | None = None"}, {"name": "return_dict", "val": ": bool | None = None"}, {"name": "attention_mask", "val": ": torch.LongTensor | None = None"}, {"name": "**kwargs", "val": ""}]- **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/v5.8.0/ko/model_doc/auto#transformers.AutoTokenizer). See [PreTrainedTokenizer.encode()](/docs/transformers/v5.8.0/ko/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode) and
  [PreTrainedTokenizer.__call__()](/docs/transformers/v5.8.0/ko/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__) for details.

  [What are input IDs?](../glossary#input-ids)
- **inputs_embeds** (`torch.LongTensor` 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.
- **cache_params** (`Cache`, *optional*) --
  If passed along, the model uses the previous state in all the blocks (which will give the output for the
  `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
- **use_cache** (`bool`, *optional*) --
  If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
- **output_hidden_states** (`bool`, *optional*) --
  Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
  more detail.
- **return_dict** (`bool`, *optional*) --
  Whether or not to return a [ModelOutput](/docs/transformers/v5.8.0/ko/main_classes/output#transformers.utils.ModelOutput) instead of a plain tuple.
- **attention_mask** (`torch.LongTensor` 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)0`MambaOutput` or `tuple(torch.FloatTensor)`A `MambaOutput` 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 ([MambaConfig](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaConfig)) and inputs.
The [MambaModel](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaModel) 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)`, *optional*, defaults to `None`) -- Sequence of hidden-states at the output of the last layer of the model.
- **cache_params** (`~cache_utils.Cache`, *optional*, defaults to `None`) -- The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
  avoid providing the old `input_ids`.

  Includes both the State space model state matrices after the selective scan, and the Convolutional states
- **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.

**Parameters:**

config ([MambaModel](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaModel)) : 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/v5.8.0/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained) method to load the model weights.

**Returns:**

``MambaOutput` or `tuple(torch.FloatTensor)``

A `MambaOutput` 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 ([MambaConfig](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaConfig)) and inputs.

## MambaLMHeadModel[[transformers.MambaForCausalLM]]

#### transformers.MambaForCausalLM[[transformers.MambaForCausalLM]]

[Source](https://github.com/huggingface/transformers/blob/v5.8.0/src/transformers/models/mamba/modeling_mamba.py#L614)

The MAMBA Model transformer with a language modeling head on top (linear layer with weights tied to the input
embeddings).

This model inherits from [PreTrainedModel](/docs/transformers/v5.8.0/ko/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.

forwardtransformers.MambaForCausalLM.forwardhttps://github.com/huggingface/transformers/blob/v5.8.0/src/transformers/models/mamba/modeling_mamba.py#L655[{"name": "input_ids", "val": ": torch.LongTensor | None = None"}, {"name": "attention_mask", "val": ": torch.LongTensor | None = None"}, {"name": "inputs_embeds", "val": ": torch.FloatTensor | None = None"}, {"name": "cache_params", "val": ": transformers.cache_utils.Cache | None = None"}, {"name": "labels", "val": ": torch.LongTensor | None = None"}, {"name": "output_hidden_states", "val": ": bool | None = None"}, {"name": "return_dict", "val": ": bool | None = None"}, {"name": "use_cache", "val": ": bool | None = None"}, {"name": "logits_to_keep", "val": ": int | torch.Tensor = 0"}, {"name": "**kwargs", "val": ""}]- **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/v5.8.0/ko/model_doc/auto#transformers.AutoTokenizer). See [PreTrainedTokenizer.encode()](/docs/transformers/v5.8.0/ko/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode) and
  [PreTrainedTokenizer.__call__()](/docs/transformers/v5.8.0/ko/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__) for details.

  [What are input IDs?](../glossary#input-ids)
- **attention_mask** (`torch.LongTensor` 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.
- **cache_params** (`Cache`, *optional*) --
  If passed along, the model uses the previous state in all the blocks (which will give the output for the
  `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
- **labels** (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*) --
  Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
  `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
  are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
- **output_hidden_states** (`bool`, *optional*) --
  Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
  more detail.
- **return_dict** (`bool`, *optional*) --
  Whether or not to return a [ModelOutput](/docs/transformers/v5.8.0/ko/main_classes/output#transformers.utils.ModelOutput) instead of a plain tuple.
- **use_cache** (`bool`, *optional*) --
  If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
- **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).0`MambaCausalLMOutput` or `tuple(torch.FloatTensor)`A `MambaCausalLMOutput` 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 ([MambaConfig](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaConfig)) and inputs.
The [MambaForCausalLM](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaForCausalLM) 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).
- **cache_params** (`~cache_utils.Cache`, *optional*, defaults to `None`) -- The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
  avoid providing the old `input_ids`.

  Includes both the State space model state matrices after the selective scan, and the Convolutional states
- **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.

Example:

```python
```

**Parameters:**

config ([MambaForCausalLM](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaForCausalLM)) : 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/v5.8.0/ko/main_classes/model#transformers.PreTrainedModel.from_pretrained) method to load the model weights.

**Returns:**

``MambaCausalLMOutput` or `tuple(torch.FloatTensor)``

A `MambaCausalLMOutput` 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 ([MambaConfig](/docs/transformers/v5.8.0/ko/model_doc/mamba#transformers.MambaConfig)) and inputs.