Mamba 370m Hf

🚀 Mamba

This repository houses the mamba-2.8b model compatible with the transformers library. The checkpoints remain unaltered, while the complete config.json and tokenizer have been uploaded to this repository.

🚀 Quick Start

To use this model, you need to perform the following installations.

📦 Installation

First, you need to install the transformers library from the main branch until transformers=4.39.0 is officially released:

pip install git+https://github.com/huggingface/transformers@main

We also suggest installing both causal_conv_1d and mamba-ssm using the following commands:

pip install causal-conv1d>=1.2.0
pip install mamba-ssm

If either of these two libraries is not installed, the "eager" implementation will be used. Otherwise, the more optimized cuda kernels will be employed.

💻 Usage Examples

Basic Usage

You can utilize the classic generate API for text generation:

>>> from transformers import MambaConfig, MambaForCausalLM, AutoTokenizer
>>> import torch

>>> tokenizer = AutoTokenizer.from_pretrained("state-spaces/mamba-370m-hf")
>>> model = MambaForCausalLM.from_pretrained("state-spaces/mamba-370m-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))
["Hey how are you doing?\n\nI'm doing great.\n\nI"]

Advanced Usage

Here is an example of fine - tuning the model using the peft library. Note that we recommend keeping the model in float32 during fine - tuning:

from datasets import load_dataset
from trl import SFTTrainer
from peft import LoraConfig
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
tokenizer = AutoTokenizer.from_pretrained("state-spaces/mamba-370m-hf")
model = AutoModelForCausalLM.from_pretrained("state-spaces/mamba-370m-hf")
dataset = load_dataset("Abirate/english_quotes", split="train")
training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    logging_dir='./logs',
    logging_steps=10,
    learning_rate=2e-3
)
lora_config =  LoraConfig(
        r=8,
        target_modules=["x_proj", "embeddings", "in_proj", "out_proj"],
        task_type="CAUSAL_LM",
        bias="none"
)
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=training_args,
    peft_config=lora_config,
    train_dataset=dataset,
    dataset_text_field="quote",
)
trainer.train()