Skip to content

Latest commit

 

History

History
99 lines (88 loc) · 3.88 KB

README_WHEEL.md

File metadata and controls

99 lines (88 loc) · 3.88 KB
Raw

Python Wheel

Building wheel for your platform

git clone https://github.com/h2oai/h2ogpt.git
cd h2ogpt
python setup.py bdist_wheel

Note that Coqui TTS is not installed due to issues with librosa. Use one-click, docker, or manual install scripts to get Coqui TTS. Also, AMD ROC and others are supported, but need manual edits to the reqs_optional/requirements_optional_llamacpp_gpt4all.txt file to select it and comment out others.

Install in fresh env, avoiding being inside h2ogpt directory or a directory where it is a sub directory. For CUDA GPU do:

export CUDA_HOME=/usr/local/cuda-12.1
export PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cu121 https://huggingface.github.io/autogptq-index/whl/cu121"
set CMAKE_ARGS=-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all
set LLAMA_CUBLAS=1
set FORCE_CMAKE=1

for the cmake args, choose e llama_cpp_python ARGS for your system according to llama_cpp_python backend documentation. Note for some reason things will fail with llama_cpp_python if don't add all cuda arches, and building with all those arches does take some time. Then pip install:

pip install <h2ogpt_path>/dist/h2ogpt-0.1.0-py3-none-any.whl[cuda]
pip install flash-attn==2.4.2

and pick your CUDA version, where <h2ogpt_path> is the relative path to the h2ogpt repo where the wheel was built. Replace 0.1.0 with actual version built if more than one.

For non CUDA cases, e.g. CPU, Metal M1/M2 do:

pip install <h2ogpt_path>/dist/h2ogpt-0.1.0-py3-none-any.whl[cpu]

A wheel online is provided for this and can be installed as follows: First, if using conda, DocTR can be enabled using above installation if first doing:

conda install weasyprint pygobject -c conda-forge -y

second run:

export CMAKE_ARGS="-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all"
export CUDA_HOME=/usr/local/cuda-12.1
export PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cu121 https://huggingface.github.io/autogptq-index/whl/cu121"
pip install h2ogpt==0.2.0[cuda] --index-url https://downloads.h2ogpt.h2o.ai --extra-index-url https://pypi.org/simple --no-cache
pip install flash-attn==2.4.2

for CUDA support. If conda and those packages weren't installed, this would exclude some DocTR support that is provided otherwise also by docker, one-click installer for windows and mac, or manual windows/linux installers.

Checks

Once the wheel is built, if you do:

python -m pip check

and you should see:

No broken requirements found.

PyPI

For PyPI, we use a more limited set of packages built like:

PYPI=1 python setup.py bdist_wheel

which can be installed with basic CUDA support like:

# For other GPUs etc. see: https://github.com/abetlen/llama-cpp-python?tab=readme-ov-file#supported-backends
# required for PyPi wheels that do not allow URLs, so uses generic llama_cpp_python package:
export CMAKE_ARGS="-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all"
export CUDA_HOME=/usr/local/cuda-12.1
export PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cu121 https://huggingface.github.io/autogptq-index/whl/cu121"
# below [cuda] assumes CUDA 12.1 for some packages like AutoAWQ etc.
pip install h2ogpt[cuda]
pip install flash-attn==2.4.2

Run

To run h2oGPT, do, e.g.

CUDA_VISIBLE_DEVICES=0 python -m h2ogpt.generate --base_model=llama

or inside python:

from h2ogpt.generate import main
main(base_model='llama')

See src/gen.py for all documented options one can pass to main(). E.g. to start LLaMa7B:

from h2ogpt.generate import main
main(base_model='meta-llama/Llama-2-7b-chat-hf',
          prompt_type='llama2',
          save_dir='save_gpt7',
          score_model=None,
          max_max_new_tokens=2048,
          max_new_tokens=1024,
          num_async=10,
          top_k_docs=-1)