Arctic is a dense-MoE Hybrid transformer architecture pre-trained from scratch by the Snowflake AI
Research Team. We are releasing model checkpoints for both the base and instruct-tuned versions of
Arctic under an Apache-2.0 license. This means you can use them freely in your own research,
prototypes, and products. Please see our blog
Snowflake Arctic: The Best LLM for Enterprise AI — Efficiently Intelligent, Truly Open
for more information on Arctic and links to other relevant resources such as our series of cookbooks
covering topics around training your own custom MoE models, how to produce high-quality training data,
and much more.
Arctic combines a 10B dense transformer model with a residual 128x3.66B MoE MLP resulting in 480B
total and 17B active parameters chosen using a top-2 gating. For more details about Arctic's model
Architecture, training process, data, etc. see our series of cookbooks.
Usage
Arctic is currently supported with transformers by leveraging the
custom code feature,
to use this you simply need to add trust_remote_code=True to your AutoTokenizer and AutoModelForCausalLM calls.
However, we recommend that you use a transformers version at or above 4.39:
pip install transformers>=4.39.0
Arctic leverages several features from DeepSpeed, you will need to
install the DeepSpeed 0.14.2 or higher to get all of these required features:
pip install deepspeed>=0.14.2
Inference examples
Due to the model size we recommend using a single 8xH100 instance from your
favorite cloud provider such as: AWS p5.48xlarge,
Azure ND96isr_H100_v5, etc.
In this example we are using FP8 quantization provided by DeepSpeed in the backend, we can also use FP6
quantization by specifying q_bits=6 in the QuantizationConfig config. The "150GiB" setting
for max_memory is required until we can get DeepSpeed's FP quantization supported natively as a
HFQuantizer which we
are actively working on.
