Skip to content

LLAMA.md

Latest commit

 

History

History
70 lines (63 loc) · 2.75 KB

LLAMA.md

File metadata and controls

70 lines (63 loc) · 2.75 KB

How to Run Llama

How to Run Llama 2

  1. Download Llama 2 weights from Meta. This project supports 7B, 7B-chat, 13B, 13B-chat, 70B and 70B-chat models.
  2. Open the llama-2-7b/params.json file:
  • replace "vocab_size": -1 to "vocab_size": 32000,
  • add a new property: "max_seq_len": 2048.
  1. Install dependencies of the converter:
cd converter && pip install -r requirements.txt
  1. Convert weights to Distributed Llama format. This will take a bit of time. The script requires Python 3.
python convert-llama.py /path/to/meta/llama-2-7b q40
  1. Download the tokenizer for Llama 2:
wget https://huggingface.co/b4rtaz/llama-2-distributed-llama/resolve/main/dllama-llama2-tokenizer.t
  1. Build the project:
make dllama
  1. Run:
./dllama inference --model dllama_llama-2-7b_q40.bin --tokenizer dllama-llama2-tokenizer.t --weights-float-type q40 --buffer-float-type q80 --prompt "Hello world" --steps 16 --nthreads 4

In the table below, you can find the expected size of the converted weights with different floating-point types.

Model Original size Float32 Float16 Q40
Llama 2 7B 13.48 GB 25.10GB 3.95 GB
Llama 2 13B 26.03 GB 7.35 GB
Llama 2 70B 137.97 GB 36.98 GB

How to Run Llama 3

You can skip the 1 - 9 points by downloading the converted Llama 3 8B Q40 model from here.

  1. Get an access to the model on Llama 3 website.
  2. Clone the https://github.com/meta-llama/llama3 repository.
  3. Run the download.sh script to download the model.
  4. For Llama 3 8B model you should have the following files:
    • Meta-Llama-3-8B/consolidated.00.pth
    • Meta-Llama-3-8B/params.json
    • Meta-Llama-3-8B/tokenizer.model
  5. Open params.json and add a new property: "max_seq_len": 8192.
  6. Clone the https://github.com/b4rtaz/distributed-llama.git repository.
  7. Install dependencies of the converter:
cd converter && pip install -r requirements.txt
  1. Convert the model to the Distributed Llama format:
python converter/convert-llama.py path/to/Meta-Llama-3-8B q40
  1. Convert the tokenizer to the Distributed Llama format:
python converter/convert-tokenizer-llama3.py path/to/tokenizer.model
  1. Build the project:
make dllama
  1. Run the Distributed Llama:
./dllama inference --weights-float-type q40 --buffer-float-type q80 --prompt "My name is" --steps 128 --nthreads 8 --model dllama_meta-llama-3-8b_q40.bin --tokenizer llama3-tokenizer.t