File size: 1,727 Bytes
d05246c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
---
license: apache-2.0
language:
- en
tags:
- mamba
- mlx
- cartesia
---

# Model Card for mamba2-1.3b-4bit-mlx

This is an [MLX](https://ml-explore.github.io/mlx)-compatible version of the [mamba2-1.3b](https://huggingface.co/state-spaces/mamba2-1.3b) model, quantized to 4 bits. It uses the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
For more details, see our [blog post](https://cartesia.ai/blog/on-device).

## 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:
```shell 
sudo xcodebuild -license
```

Install the required dependencies: the exact version of `nanobind`, followed by `cartesia-metal`, and finally `cartesia-mlx`, with the following commands:
```shell 
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
```python 
import mlx.core as mx
import cartesia_mlx as cmx

model = cmx.from_pretrained("cartesia-ai/mamba2-1.3b-4bit-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](https://cartesia.ai/), we're building real-time multimodal intelligence for every device.