|
--- |
|
license: llama3.3 |
|
tags: |
|
- distributed-inference |
|
- text-generation |
|
--- |
|
|
|
This is converted **Llama 3.3 70B Instruct** model to [Distributed Llama](https://github.com/b4rtaz/distributed-llama) format. The model is quantized to Q40 format. Due to Huggingface limitations, the model is split into 11 parts. Before use, you need to combine the parts together. |
|
|
|
To run this model, you need approximately 42 GB of RAM on a single device, or approximately 42 GB of RAM distributed across 2, 4, 8, or 16 devices connected in a cluster (more informations how to do it you can find [here](https://github.com/b4rtaz/distributed-llama)). |
|
|
|
## 🚀 How to Run? |
|
|
|
1. ⏬ Download the model. You have two options: |
|
* Download this repository and combine all parts together by using the `cat` command. |
|
* Download the model by using the `launch.py` script from Distributed Llama repository: `python launch.py llama3_3_70b_instruct_q40` |
|
4. ⏬ Download [Distributed Llama](https://github.com/b4rtaz/distributed-llama) repository. |
|
5. 🔨 Build Distributed Llama: |
|
``` |
|
make dllama |
|
``` |
|
4. 🚀 Run Distributed Llama: |
|
``` |
|
./dllama chat --model llama3_3_70b_instruct_q40.m --tokenizer dllama_tokenizer_llama_3_3.t --buffer-float-type q80 --max-seq-len 2048 --nthreads 64 |
|
``` |
|
|
|
## 🎩 License |
|
|
|
You need to accept the Llama 3.3 license before downloading this model. |
|
|
|
[Llama 3.3 Community License](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/LICENSE) |
