Mambaoutai is the result of all the experiments and training runs described in the following blog post, where all details about the model series is shared. Mambaoutai is series of small mamba checkpoints released for the community to explore, trained on French, English and code. We run two different decay phases with the WSD-scheduler, and release model checkpoints pretrained both with and without instruction data.

Usage

You need to install transformers from main until transformers=4.39.0 is released.

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

We also recommend you to install both causal-conv1d and mamba-ssm using:

bash

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

If any of these two is not installed, the "eager" implementation will be used(not recommended). Otherwise the more optimised CUDA kernels will be used.

Generation

Use this snippet of code to generate text from the model:

python

from transformers import MambaConfig, MambaForCausalLM, AutoTokenizer
import torch

if model_has_instruct_data:
	# use chat tokens
	prompt = ”<start_user>Tell me something about Paris.<end_message><start_assistant>”
else:
	# prompt the non-instructed tuned model gently
	prompt = ”This is a text about Paris. Paris is”

tokenizer = AutoTokenizer.from_pretrained("lightonai/mambaoutai")
model = MambaForCausalLM.from_pretrained("lightonai/mambaoutai")
input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"]

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

Training checkpoints

You can find some of the training checkpoints in the repo branch. On branch corresponding to the model at some point in time during training.

You can do inference with these training checkpoints by adding the revision parameter to the from_pretrained method. For example, to load the model checkpoint after 30000 steps of pretraining, you can use the following code:

python

from transformers import MambaConfig, MambaForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("lightonai/mambaoutai", revision="pre-30000")
model = MambaForCausalLM.from_pretrained("lightonai/mambaoutai", revision="pre-30000")
input_ids = tokenizer("What is a mamba?", return_tensors="pt")["input_ids"]

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

On-device Inference

Since Mambaoutai is only 1.6B parameters, it can be run on a CPU with reasonable speed.

Here is an example of how to run it on llama.cpp:

bash

# Clone llama.cpp repository and compile it from source
git clone https://github.com/ggerganov/llama.cpp\
cd llama.cpp
make

# Create a venv and install dependencies
conda create -n mamba-cpp python=3.10
conda activate mamba-cpp
pip install -r requirements/requirements-convert-hf-to-gguf.txt

# Download the weights, tokenizer, config, tokenizer_config and special_tokens_map from this repo and
# put them in a directory 'Mambaoutai/' 
mkdir Mambaoutai

# Convert the weights to GGUF format
python convert-hf-to-gguf.py Mambaoutai

# Run inference with a prompt
./main -m Mambaoutai/ggml-model-f16.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 1

Training Hardware

The model checkpoints with no instruction data have been fully trained on an NVIDIA DGX H100 provided by OVH Cloud, whereas the decay phases with instruction data have been carried out on an HPE Cray with 8xH100 on Orange Cloud Avenue. The ablation experiments were conducted on 16 nodes(4xA100-40GB) on MeluXina.

Model hyperparameters

More details about the model hyperparameters are given in the table below :

Parameter	Value
d_model	2688
n_layer	28
vocab_size	65024
context_len	4096
rms_norm	true
residual_in_fp32	true
fused_add_norm	true
conv_kernel	4
d_inner	5376
state_size	16
dtype	bfloat16
tie_word_embeddings	false
non embeddings params	1.27B