h2oGPT Installation Help
The following sections describe how to get a working Python environment on a Linux system.
Install for A100+
E.g. for Ubuntu 20.04, install driver if you haven't already done so:
sudo apt-get update
sudo apt-get -y install nvidia-headless-535-server nvidia-fabricmanager-535 nvidia-utils-535-server
# sudo apt-get -y install nvidia-headless-no-dkms-535-servers
Note that if you run the preceding commands, you don't need to use the NVIDIA developer downloads in the following sections.
Install CUDA Toolkit
If happy with above drivers, then just get run local file for CUDA 11.8:
wget wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run
sudo sh cuda_11.8.0_520.61.05_linux.run
only choose to install toolkit and do not replace existing /usr/local/cuda link if you already have one.
If instead, you want full deb CUDA install cuda coolkit. Pick deb local, e.g. for Ubuntu:
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
Then set the system up to use the freshly installed CUDA location:
echo "export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:/usr/local/cuda/lib64/" >> ~/.bashrc
echo "export CUDA_HOME=/usr/local/cuda" >> ~/.bashrc
echo "export PATH=\$PATH:/usr/local/cuda/bin/" >> ~/.bashrc
source ~/.bashrc
Then reboot the machine, to get everything sync'ed up on restart.
sudo reboot
Compile bitsandbytes
For fast 4-bit and 8-bit training, you need to use bitsandbytes. Note that compiling bitsandbytes is only required if you have a different CUDA version from the ones built into the bitsandbytes PyPI package, which includes CUDA 11.0, 11.1, 11.2, 11.3, 11.4, 11.5, 11.6, 11.7, 11.8, 12.0, and 12.1. In the following example, bitsandbytes is compiled for CUDA 12.1:
git clone http://github.com/TimDettmers/bitsandbytes.git
cd bitsandbytes
git checkout 7c651012fce87881bb4e194a26af25790cadea4f
CUDA_VERSION=121 make cuda12x
CUDA_VERSION=121 python setup.py install
cd ..
Install NVIDIA GPU Manager on systems with multiple A100 or H100 GPUs
To install NVIDIA GPU Manager, run the following:
sudo apt-key del 7fa2af80
distribution=$(. /etc/os-release;echo $ID$VERSION_ID | sed -e 's/\.//g')
wget https://developer.download.nvidia.com/compute/cuda/repos/$distribution/x86_64/cuda-keyring_1.0-1_all.deb
sudo dpkg -i cuda-keyring_1.0-1_all.deb
sudo apt-get update
sudo apt-get install -y datacenter-gpu-manager
# if use 535 drivers, then use 535 below
sudo apt-get install -y libnvidia-nscq-535
sudo systemctl --now enable nvidia-dcgm
dcgmi discovery -l
For more information, see the official GPU Manager user guide.
Install and run NVIDIA Fabric Manager on systems with multiple A100 or H100 GPUs
To install the CUDA drivers for NVIDIA Fabric Manager, run the following:
sudo apt-get install -y cuda-drivers-fabricmanager
Once you've installed Fabric Manager and rebooted your system, run the following to start the NVIDIA Fabric Manager service:
sudo systemctl --now enable nvidia-dcgm
dcgmi discovery -l
sudo systemctl start nvidia-fabricmanager
sudo systemctl status nvidia-fabricmanager
For more information, see the official Fabric Manager user guide.
Optional: Use TensorBoard to inspect training
You can use TensorBoard to inspect the training process. To launch TensorBoard and instruct it to read event files from the runs/ directory, use the following command:
tensorboard --logdir=runs/
For more information, see TensorBoard usage.
Flash Attention
Update: Flash attention specifics are no longer needed. For more information, see https://github.com/h2oai/h2ogpt/issues/128.
To use flash attention with LLaMa, need cuda 11.7 so flash attention module compiles against torch.
E.g. for Ubuntu, one goes to cuda toolkit, then:
wget https://developer.download.nvidia.com/compute/cuda/11.7.0/local_installers/cuda_11.7.0_515.43.04_linux.run
sudo bash ./cuda_11.7.0_515.43.04_linux.run
Then No for symlink change, say continue (not abort), accept license, keep only toolkit selected, select install.
If cuda 11.7 is not your base installation, then when doing pip install -r requirements.txt do instead:
CUDA_HOME=/usr/local/cuda-11.8 pip install -r reqs_optional/requirements_optional_flashattention.txt