HuggingFace_transformer/docs/source/en/model_doc/mamba.md

Mamba

Mamba is a selective structured state space model (SSMs) designed to work around Transformers computational inefficiency when dealing with long sequences. It is a completely attention-free architecture, and comprised of a combination of H3 and gated MLP blocks (Mamba block). Mamba's "content-based reasoning" allows it to focus on specific parts of an input depending on the current token. Mamba also uses a new hardware-aware parallel algorithm to compensate for the lack of convolutional operations. As a result, Mamba has fast inference and can scale to very long sequences.

You can find all the original Mamba checkpoints under the State Space Models organization.

Tip

This model was contributed by Molbap and AntonV. Click on the Mamba models in the right sidebar for more examples of how to apply Mamba to different language tasks.

The example below demonstrates how to generate text with [Pipeline], [AutoModel], and from the command line.

import torch
from transformers import pipeline

pipeline = pipeline(
    task="text-generation",
    model="state-spaces/mamba-130m-hf",
    torch_dtype=torch.float16,
    device=0
)
pipeline("Plants create energy through a process known as")

import torch  
from transformers import AutoModelForCausalLM, AutoTokenizer  

tokenizer = AutoTokenizer.from_pretrained("state-spaces/mamba-130m-hf")
model = AutoModelForCausalLM.from_pretrained("state-spaces/mamba-130m-hf", torch_dtype=torch.float16, device_map="auto",)  
input_ids = tokenizer("Plants create energy through a process known as", return_tensors="pt").to("cuda")  

output = model.generate(**input_ids)  
print(tokenizer.decode(output[0], skip_special_tokens=True)

echo -e "Plants create energy through a process known as" | transformers run --task text-generation --model state-spaces/mamba-130m-hf --device 0

Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the Quantization overview for more available quantization backends.

The example below uses torchao to only quantize the weights to 4-bit integers.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig
from torchao.quantization import Int4WeightOnlyConfig

quantization_config = Int4WeightOnlyConfig(group_size=128)
quantization_config = TorchAoConfig(quant_type=quant_config)
tokenizer = AutoTokenizer.from_pretrained("state-spaces/mamba-2.8b-hf")
model = AutoModelForCausalLM.from_pretrained("state-spaces/mamba-2.8b-hf", torch_dtype=torch.bfloat16, quantization_config=quantization_config, device_map="auto",)
input_ids = tokenizer("Plants create energy through a process known as", return_tensors="pt").to("cuda")

output = model.generate(**input_ids)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Notes

• The current implementation uses the original CUDA kernels. The FlashAttention equivalent implementation is hosted in the mamba-ssm and causal_conv1d repositories. Make sure to install them if your hardware supports it!

• Mamba stacks mixer layers which are equivalent to Attention layers. You can find the main logic of Mamba in the MambaMixer class.

• The example below demonstrates how to fine-tune Mamba with PEFT.

  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()
  

MambaCache

autodoc MambaCache - update_conv_state - update_ssm_state - reset

MambaConfig

autodoc MambaConfig

MambaModel

autodoc MambaModel - forward

MambaLMHeadModel

autodoc MambaForCausalLM - forward