MAMBA (2.8B) 🐍 fine-tuned on Synthia-v1.3

Model Card is still WIP!

Base model info

Mamba is a new state space model architecture showing promising performance on information-dense data such as language modeling, where previous subquadratic models fall short of Transformers. It is based on the line of progress on structured state space models, with an efficient hardware-aware design and implementation in the spirit of FlashAttention.

Dataset info

Not Provided

Usage

pip install torch==2.1.0 transformers==4.35.0 causal-conv1d==1.0.0 mamba-ssm==1.0.1

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel

CHAT_TEMPLATE_ID = "HuggingFaceH4/zephyr-7b-beta"

device = "cuda:0" if torch.cuda.is_available() else "cpu"
model_name = "clibrain/mamba-2.8b-ft-synthia-v1.3"

eos_token = "<|endoftext|>"
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.eos_token = eos_token
tokenizer.pad_token = tokenizer.eos_token
tokenizer.chat_template = AutoTokenizer.from_pretrained(CHAT_TEMPLATE_ID).chat_template

model = MambaLMHeadModel.from_pretrained(
        model_name, device=device, dtype=torch.float16)

messages = []
prompt = "Tell me 5 sites to visit in Spain"
messages.append(dict(role="user", content=prompt))

input_ids = tokenizer.apply_chat_template(
            messages, return_tensors="pt", add_generation_prompt=True
).to(device)

out = model.generate(
    input_ids=input_ids,
    max_length=2000,
    temperature=0.9,
    top_p=0.7,
    eos_token_id=tokenizer.eos_token_id,
)

decoded = tokenizer.batch_decode(out)
assistant_message = (
    decoded[0].split("<|assistant|>\n")[-1].replace(eos_token, "")
)

print(assistant_message)

Gradio Demo

git clone https://github.com/mrm8488/mamba-chat.git
cd mamba-chat

pip install -r requirements.txt
pip install -q gradio==4.8.0

python app.py \
--model clibrain/mamba-2.8b-ft-synthia-v1.3 \
--share

Evaluations

Coming soon!

Acknowledgments

Thanks to mamba-chat for heavily inspiring our work