This repository showcases my comprehensive guide to deploying the Llama2-7B model on Google Cloud VM, using NVIDIA GPUs. As an open-source alternative to commercial LLMs such as OpenAI's GPT and Google's Palm. By setting up your own private LLM instance with this guide, you can benefit from its capabilities while prioritizing data confidentiality.
Google Cloud VM:
- 24 vCPU
- 96 Gb RAM
- 2 x NVIDAN L4 (24Gb VRAM x 2)
- 250 Gb SSD
The Llama models, developed by MetaAI, are a series of breakthroughs in open-source AI. The Llama 2 model is a standout in the AI world, primarily because it's open-source. This means anyone can access and utilize its capabilities freely, fostering innovation and broader application. More crucially, its open-source nature addresses pressing data privacy concerns. Users can run Llama 2 locally, ensuring their data remains in their control and sidestepping the privacy issues tied to many commercial models. While many are familiar with renowned models like GPT-3.5 and Google's Palm, the Llama2-70B stands out not just for its competitive performance - verified through research paper and human evaluations - but also for its open-source nature.
| Benchmark (shots) | GPT-3.5 | GPT-4 | PaLM | PaLM-2-L | Llama 2 |
|---|---|---|---|---|---|
| MMLU (5-shot) | 70.0 | 86.4 | 69.3 | 78.3 | 68.9 |
| TriviaQA (1-shot) | – | – | 81.4 | 86.1 | 85.0 |
| Natural Questions (1-shot) | – | – | 29.3 | 37.5 | 33.0 |
| GSM8K (8-shot) | 57.1 | 92.0 | 56.5 | 80.7 | 56.8 |
| HumanEval (0-shot) | 48.1 | 67.0 | 26.2 | – | 29.9 |
| BIG-Bench Hard (3-shot) | – | – | 52.3 | 65.7 | 51.2 |
Llama 2 models are a state-of-the-art collection of pretrained and fine-tuned generative text models.
🚨 Important: Before accessing the models, visit the Meta website (insert link here) to accept the license terms and acceptable use policy. After doing so, you can request access to a specific model, and your request will be processed within 1-2 business days.
Llama 2 consists of models ranging from 7 billion to 70 billion parameters. The Llama-2-Chat variants are specifically optimized for dialogue use cases and they demonstrate significant performance improvements over other open-source chat models. Based on various benchmarks and human evaluations, Llama-2-Chat models offer comparable performance to popular closed-source models like ChatGPT and PaLM.
- meta-llama/Llama-2-13b-chat-hf - Last updated: 8 days ago
- meta-llama/Llama-2-13b-hf - Last updated: 8 days ago
- meta-llama/Llama-2-70b-hf - Last updated: 8 days ago
- meta-llama/Llama-2-70b-chat-hf - Last updated: 8 days ago
- meta-llama/Llama-2-7b-chat-hf - Last updated: 8 days ago
- meta-llama/Llama-2-7b-hf - Last updated: 8 days ago
- meta-llama/Llama-2-7b - Last updated: 20 days ago
- meta-llama/Llama-2-13b - Last updated: 20 days ago
- meta-llama/Llama-2-13b-chat - Last updated: 20 days ago
- meta-llama/Llama-2-70b - Last updated: 20 days ago
- meta-llama/Llama-2-70b-chat - Last updated: 20 days ago
- meta-llama/Llama-2-7b-chat - Last updated: 20 days ago
- Context Size: Doubled from 2048 to 4096 tokens, enabling more in-depth content processing.
- Training: Improved with 40% more tokens, leading to better performance on benchmarks.
- Commercial Use: Llama 2 is available for commercial applications, broadening its accessibility.
- Llama2 Chat: A fine-tuned chat version, assessed to be more helpful in responses than existing models like ChatGPT.
Graphs of the model's performance indicate that larger models, especially the 70 billion parameter variant, could still improve with more training and tokens. Even the smaller 7 billion parameter model shows strong performance, though closer to its peak.
Llama2 Chat has been evaluated to provide more helpful answers compared to ChatGPT. It's a free model that can be utilized for various purposes, including commercial ones.
In a nutshell, Llama 2 is an innovative AI model that offers advancements in context understanding, training efficiency, and practical applications. It marks a significant step forward in the field and offers exciting possibilities for developers and businesses.
- Commercial Use: The Llama 2 model is available for commercial purposes unless you're a business with over 700 million monthly active users. In which case, there might be restrictions or licensing requirements.
- Comparison with Industry Leaders: In benchmarks, Llama 2 has shown performance that, while not quite on par with proprietary models like GPT-4 and Palm 2L, is competitive especially considering its open-source nature.
- In the MMLU benchmark, Llama 2 scored 68, which is between the scores of GPT-4 (86) and GPT-3. For context, the MMLU benchmark assesses performance on a wide range of tasks.
- In the MMLU benchmark, Llama 2 scored 68, which is between the scores of GPT-4 (86) and GPT-3. For context, the MMLU benchmark assesses performance on a wide range of tasks.
- Position among Open-Source Models: When compared to other open-source models, Llama 2 stands out as a leading performer, especially on the MMLU benchmark. This suggests that, among freely available models, Llama 2 offers the most advanced capabilities currently available.
- Coding Capability: One exception to Llama 2's top-tier performance is in coding tasks, where the MPT model, another open-source contender, outperforms it in the 7 billion parameter category.
The release of the Llama 2 model democratizes access to powerful AI. Companies and individual developers alike can harness advanced machine learning capabilities without incurring the high costs associated with using proprietary models from big players in the AI industry.
This guide will help you to set up Llama-7B on your Google Cloud VM equipped with NVIDIA GPUs.
- Google Cloud VM: 24vCPU, 96 RAM, and 2*NVIDAN L4s (24Gb VRAM x 2)
sudo su -
passwd
passwd tharindu_sankalpa
usermod -a -G sudo,adm tharindu_sankalpa
vi /etc/ssh/sshd_config
systemctl restart sshd
After making changes to the SSH configuration, connect to the server:
ssh tharindu_sankalpa@YOUR_VM_IP_ADDRESS
sudo apt-get update
sudo apt-get upgrade
Check current NVIDIA status:
nvidia-smi
Output should be something like this.
tharindu_sankalpa@llama2-model-endpoint2:~$ nvidia-smi
Tue Aug 8 16:45:14 2023
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.125.06 Driver Version: 525.125.06 CUDA Version: 12.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA L4 Off | 00000000:00:03.0 Off | 0 |
| N/A 61C P8 19W / 72W | 70MiB / 23034MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 NVIDIA L4 Off | 00000000:00:04.0 Off | 0 |
| N/A 61C P8 19W / 72W | 4MiB / 23034MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 1531 G /usr/lib/xorg/Xorg 59MiB |
| 0 N/A N/A 1590 G /usr/bin/gnome-shell 10MiB |
| 1 N/A N/A 1531 G /usr/lib/xorg/Xorg 4MiB |
+-----------------------------------------------------------------------------+
tharindu_sankalpa@llama2-model-endpoint2:~$
tharindu_sankalpa@llama2-model-endpoint2:~$
Install the NVIDIA driver:
sudo apt-get install nvidia-driver-525
sudo reboot now
Once the system is rebooted, check your VM's availability:
ping YOUR_VM_IP_ADDRESS
ssh tharindu_sankalpa@YOUR_VM_IP_ADDRESS
Fetch and configure CUDA:
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/12.1.0/local_installers/cuda-repo-ubuntu2004-12-1-local_12.1.0-530.30.02-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2004-12-1-local_12.1.0-530.30.02-1_amd64.deb
sudo cp /var/cuda-repo-ubuntu2004-12-1-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install cuda
Reboot the VM:
sudo reboot now
wget https://repo.anaconda.com/miniconda/Miniconda3-py39_23.5.2-0-Linux-x86_64.sh
chmod +x Miniconda3-py39_23.5.2-0-Linux-x86_64.sh
./Miniconda3-py39_23.5.2-0-Linux-x86_64.sh
source .bashrc
conda create --name tf_gpu_env python=3.9.13
conda activate tf_gpu_env
pip install tensorflow
Verify CUDA compiler version:
nvcc --version
sudo apt install nvidia-cuda-toolkit
wget https://storage.googleapis.com/windows-server-imaage-bucket/cudnn-11.3-linux-x64-v8.2.1.32.tgz
tar -xzvf cudnn-11.3-linux-x64-v8.2.1.32.tgz
sudo cp -P cuda/include/cudnn*.h /usr/local/cuda/include/
sudo cp -P cuda/lib64/libcudnn* /usr/local/cuda/lib64/
sudo chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*
Then, adjust the environment and libraries:
conda install -c conda-forge cudatoolkit=11.8.0
pip install nvidia-cudnn-cu11==8.6.0.163
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
echo 'CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)"))' >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
echo 'export LD_LIBRARY_PATH=$CONDA_PREFIX/lib/:$CUDNN_PATH/lib:$LD_LIBRARY_PATH' >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
Test TensorFlow and GPU:
python3 -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"
python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
conda create -n llama2 python=3.9
conda activate llama2
git clone https://github.com/thisserand/llama2_local.git
cd llama2_local/
pip install -r requirements.txt
huggingface-cli login
# Follow on-screen prompts to enter your Hugging Face tokenpython llama.py --model_name="meta-llama/Llama-2-7b-chat-hf"(llama2) tharindu_sankalpa@llama2-model-endpoint2:~/llama2_local$ python llama.py --model_name="meta-llama/Llama-2-7b-chat-hf"
Downloading (…)lve/main/config.json: 100%|███████████████████████| 614/614 [00:00<00:00, 93.6kB/s]
Downloading (…)fetensors.index.json: 100%|███████████████████| 26.8k/26.8k [00:00<00:00, 4.50MB/s]
Downloading (…)of-00002.safetensors: 100%|████████████████████| 9.98G/9.98G [00:29<00:00, 344MB/s]
Downloading (…)of-00002.safetensors: 100%|████████████████████| 3.50G/3.50G [00:08<00:00, 436MB/s]
Downloading shards: 100%|███████████████████████████████████████████| 2/2 [00:37<00:00, 18.78s/it]
Loading checkpoint shards: 100%|████████████████████████████████████| 2/2 [00:09<00:00, 4.60s/it]
Downloading (…)neration_config.json: 100%|███████████████████████| 167/167 [00:00<00:00, 28.2kB/s]
Downloading (…)okenizer_config.json: 100%|████████████████████████| 776/776 [00:00<00:00, 491kB/s]
Downloading tokenizer.model: 100%|██████████████████████████████| 500k/500k [00:00<00:00, 230MB/s]
Downloading (…)/main/tokenizer.json: 100%|███████████████████| 1.84M/1.84M [00:00<00:00, 8.66MB/s]
Downloading (…)cial_tokens_map.json: 100%|████████████████████████| 414/414 [00:00<00:00, 280kB/s]
Running on local URL: http://127.0.0.1:7860
Running on public URL: https://ca14ace05f9cd9f845.gradio.live
This share link expires in 72 hours. For free permanent hosting and GPU upgrades, run `gradio deploy` from Terminal to deploy to Spaces (https://huggingface.co/spaces)
The evaluation of three prominent language models: GPT-4, GPT-3.5, and Llama2-70B. We use a simple set of prompts to assess the performance, clarity, and quality of answers each model produces.
Below are the prompts used for evaluation:
- What are the core takeaways from the "7 Habits of Highly Effective People" book?
- In the context of the book, it often differentiates between personality ethics and character ethics. Can you elaborate on their differences and signify which one holds more importance?
- Could you provide examples to illustrate the difference between personality ethics and character ethics?
- What steps can I take to develop my character ethics in order to achieve enduring success?
- How can I introspectively identify and understand my personal values?
- Could you provide a structured questionnaire that aids in the identification and reflection of one's core values?
https://chat.openai.com/share/6c9deabe-e086-4095-b5f7-687b6c6350a1
https://chat.openai.com/share/54e3e52a-4a97-4438-bda8-d33bc2d73ffd
