Parrot can be installed as a Python library from the github repository.
- OS: Linux, Ubuntu 20.04
- GPU: cc >= 7.0 (Ours: NVIDIA A100, NVIDIA A6000)
- CUDA version: >= 12.1
- DL Framework: PyTorch >= 2.1.0 with CUDA 12.1.
pip install torch==2.1.0 --upgrade --index-url https://download.pytorch.org/whl/cu121
git clone --recursive https://github.com/microsoft/ParrotServe.git
source .env
- Step 1: Install basic requirements.
pip install -r requirements.txt
- Step 2: Install necessary dependencies listed in
3rdparty
folder.
cd 3rdparty/vllm
pip install -e .
- Step 3 (Optional): Install Optional dependencies.
(Optional) FastChat and Langchain are used only in our benchmark.
cd 3rdparty/FastChat
pip install -e ".[model_worker,webui]"
cd 3rdparty/langchain/libs/langchain
pip install -e .
! Important Notes:
Triton 2.0.0 has some bugs in Kernel memory issues. So we enforce the version to be 2.1.0 here. You will see some dependencies warnings, but it will not affect the common usages. (The similar error also happens in LightLLM kernels.)
pip install triton==2.1.0
(In the root folder of Parrot)
pip install -e .
When installing is done, check your installation by
import parrot