mamba2-780m-8bit-mlx

Model Card for mamba2-780m-8bit-mlx

This is an MLX-compatible version of the mamba2-780m model, quantized to 8 bits. It uses the EleutherAI/gpt-neox-20b tokenizer. For more details, see our blog post.

Usage

Installation

This model requires the cartesia-metal and cartesia-mlx packages.

Installation requires Xcode, which can be downloaded from https://developer.apple.com/xcode/. Accept the license agreement with:

sudo xcodebuild -license

Install the required dependencies: the exact version of nanobind, followed by cartesia-metal, and finally cartesia-mlx, with the following commands:

pip install nanobind@git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
pip install git+https://github.com/cartesia-ai/edge.git#subdirectory=cartesia-metal
pip install cartesia-mlx

Note: This package has been tested on macOS Sonoma 14.1 with the M3 chip.

Generation example

import mlx.core as mx
import cartesia_mlx as cmx

model = cmx.from_pretrained("cartesia-ai/mamba2-780m-8bit-mlx")
model.set_dtype(mx.float32)   

prompt = "Rene Descartes was"

print(prompt, end="", flush=True)
for text in model.generate(
    prompt,
    max_tokens=500,
    eval_every_n=5,
    verbose=True,
    top_p=0.99,
    temperature=0.85,
):
    print(text, end="", flush=True)

About Cartesia

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