cartesia-pytorch

Models

This repository contains the PyTorch implementation of the Rene and Llamba models, which are large-scale language models trained by Cartesia.

Rene

Rene is a 1.3 billion-parameter language model, which is the first model in a series of models trained by Cartesia. Rene has a hybrid architecture based on Mamba-2, with feedforward and sliding window attention layers interspersed. It uses the allenai/OLMo-1B-hf tokenizer. Rene was pretrained on 1.5 trillion tokens of the Dolma-1.7 dataset. For more details, see our blog post.

Llamba

The Llamba model series is a family of highly efficient recurrent language models distilled from meta-llama/Llama-3.x into the Mamba-2 architecture, developed in collaboration between Cartesia and CMU’s Goomba Lab. The series includes Llamba-1B, Llamba-3B, and Llamba-8B, delivering high inference throughput while maintaining competitive benchmark performance. Llamba models scale linearly with input length and were trained on 8B, 10B, and 12B tokens, respectively, demonstrating the effectiveness of distillation in large language models.

Usage

This is the PyTorch version of the package, and it's intended to run on CUDA devices. For use on Mac computers, please install the native MLX version instead.

Installation

The Rene model depends on the cartesia-pytorch package, which can be installed with pip as follows:

pip install --no-binary :all: cartesia-pytorch

Generation example

python -m evals.generation \
--model Rene \
--prompt "Rene Descartes was" \
--promptlen 100 \
--genlen 100 \
--dtype bfloat16 \
--temperature 1.0 \
--top_k 1 \
--top_p 0.99 \
--min_p 0.0 \
--repetition_penalty 1.0

To generate using another model, simply replace Rene with the desired model name, e.g. Llamba-8B.

python -m evals.generation \
--model Llamba-8B \
--prompt "My favorite book is" \
--promptlen 100 \
--genlen 100 \
--dtype bfloat16 \
--temperature 1.0 \
--top_k 1 \
--top_p 0.99 \
--min_p 0.0 \
--repetition_penalty 1.0298

Evaluation example

You can use our cartesia_lm_eval wrapper around the Language Model Evaluation Harness to evaluate our model on standard text benchmarks. Example command (clone this repo and run the below from within the cartesia-pytorch directory):

python -m evals.cartesia_lm_eval --model rene_ssm --model_args pretrained=cartesia-ai/Rene-v0.1-1.3b-pytorch,trust_remote_code=True --trust_remote_code --tasks copa,hellaswag,piqa,arc_easy,arc_challenge,winogrande,openbookqa --cache_requests true --batch_size auto:4 --output_path outputs/rene_evals/

python -m evals.cartesia_lm_eval --model llamba_ssm --model_args pretrained=cartesia-ai/Llamba-8B,trust_remote_code=True --trust_remote_code --tasks hellaswag,piqa,arc_easy,arc_challenge,winogrande,mmlu --cache_requests true --batch_size auto:4 --output_path outputs/llamba_evals/

About Cartesia

At Cartesia, we're building real-time multimodal intelligence for every device.