Installation

This is a fork of VLLM for AMD MI25/50/60. I assume you already have rocm 6.2.2 installed with GPU drivers. If not, use these commands (assuming you have Ubuntu 22.04):

sudo apt update
sudo apt install "linux-headers-$(uname -r)" "linux-modules-extra-$(uname -r)"
sudo usermod -a -G render,video $LOGNAME # Add the current user to the render and video groups
wget https://repo.radeon.com/amdgpu-install/6.2.2/ubuntu/jammy/amdgpu-install_6.2.60202-1_all.deb
sudo apt install ./amdgpu-install_6.2.60202-1_all.deb
sudo apt update
sudo apt install amdgpu-dkms rocm
# linker
sudo tee --append /etc/ld.so.conf.d/rocm.conf <<EOF
/opt/rocm/lib
/opt/rocm/lib64
EOF
sudo ldconfig

# path
export PATH=$PATH:/opt/rocm-6.2.2/bin

# verify drivers
dkms status

# You need to close all windows and reboot to make changes effective.
reboot

Now you have rocm 6.2.2. If you have a single AMD GPU, you should be fine with just following commands below. If you have multiple AMD GPUs (e.g. 2xAMD50/60), then install my fork of triton 3.1.0 - https://github.com/Said-Akbar/triton-gcn5.git.

Install VLLM.

# create venv
python3 -m venv vllmenv
source vllmenv/bin/activate
git clone https://github.com/Said-Akbar/vllm-rocm.git
cd vllm-rocm
pip3 install -r requirements-rocm.txt
# uninstall torch that comes with vllm. You need torch 2.6.0 or greater.
pip3 uninstall torch torchaudio torchvision
pip3 install --no-cache-dir --pre torch>=2.6 torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.2
pip3 install amdsmi
# there might be setuptools and setuptools_scm version conflict. In that case, uninstall both first.
# pip3 uninstall setuptools setuptools_scm
pip3 install setuptools setuptools_scm
pip3 install --upgrade cmake
# build vllm
python3 setup.py develop

If everything went well, you should be able to use vllm! It supports GGUF, GPTQ, FP16 in AMD GPUs. Since we did not build/install triton, we need to use VLLM_USE_TRITON_FLASH_ATTN=0 when serving a model. Example command:

VLLM_USE_TRITON_FLASH_ATTN=0 vllm serve /home/ai-llm/Downloads/models/Meta-Llama-3.1-8B-Instruct-Q4_K_M/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf --max_model_len 1024

Note that if you need tensor parallelism with 2 or 4 AMD GPUs, you need to install my fork of triton 3.1.0 (why version 3.1.0? Because we need to match the version with pytorch-rocm-triton which was 3.1.0 at the time of writing).

End of custom description.

Easy, fast, and cheap LLM serving for everyone

| Documentation | Blog | Paper | Discord | Twitter/X | Developer Slack |

---

Latest News 🔥

• [2024/12] vLLM joins pytorch ecosystem! Easy, Fast, and Cheap LLM Serving for Everyone!
• [2024/11] We hosted the seventh vLLM meetup with Snowflake! Please find the meetup slides from vLLM team here, and Snowflake team here.
• [2024/10] We have just created a developer slack (slack.vllm.ai) focusing on coordinating contributions and discussing features. Please feel free to join us there!
• [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team here. Learn more from the talks from other vLLM contributors and users!
• [2024/09] We hosted the sixth vLLM meetup with NVIDIA! Please find the meetup slides here.
• [2024/07] We hosted the fifth vLLM meetup with AWS! Please find the meetup slides here.
• [2024/07] In partnership with Meta, vLLM officially supports Llama 3.1 with FP8 quantization and pipeline parallelism! Please check out our blog post here.
• [2024/06] We hosted the fourth vLLM meetup with Cloudflare and BentoML! Please find the meetup slides here.
• [2024/04] We hosted the third vLLM meetup with Roblox! Please find the meetup slides here.
• [2024/01] We hosted the second vLLM meetup with IBM! Please find the meetup slides here.
• [2023/10] We hosted the first vLLM meetup with a16z! Please find the meetup slides here.
• [2023/08] We would like to express our sincere gratitude to Andreessen Horowitz (a16z) for providing a generous grant to support the open-source development and research of vLLM.
• [2023/06] We officially released vLLM! FastChat-vLLM integration has powered LMSYS Vicuna and Chatbot Arena since mid-April. Check out our blog post.

---

About

vLLM is a fast and easy-to-use library for LLM inference and serving.

vLLM is fast with:

• State-of-the-art serving throughput
• Efficient management of attention key and value memory with PagedAttention
• Continuous batching of incoming requests
• Fast model execution with CUDA/HIP graph
• Quantizations: GPTQ, AWQ, INT4, INT8, and FP8.
• Optimized CUDA kernels, including integration with FlashAttention and FlashInfer.
• Speculative decoding
• Chunked prefill

Performance benchmark: We include a performance benchmark at the end of our blog post. It compares the performance of vLLM against other LLM serving engines (TensorRT-LLM, SGLang and LMDeploy). The implementation is under nightly-benchmarks folder and you can reproduce this benchmark using our one-click runnable script.

vLLM is flexible and easy to use with:

• Seamless integration with popular Hugging Face models
• High-throughput serving with various decoding algorithms, including parallel sampling, beam search, and more
• Tensor parallelism and pipeline parallelism support for distributed inference
• Streaming outputs
• OpenAI-compatible API server
• Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron.
• Prefix caching support
• Multi-lora support

vLLM seamlessly supports most popular open-source models on HuggingFace, including:

• Transformer-like LLMs (e.g., Llama)
• Mixture-of-Expert LLMs (e.g., Mixtral)
• Embedding Models (e.g. E5-Mistral)
• Multi-modal LLMs (e.g., LLaVA)

Find the full list of supported models here.

Getting Started

Install vLLM with pip or from source:

pip install vllm

Visit our documentation to learn more.

• Installation
• Quickstart
• Supported Models

Contributing

We welcome and value any contributions and collaborations. Please check out CONTRIBUTING.md for how to get involved.

Sponsors

vLLM is a community project. Our compute resources for development and testing are supported by the following organizations. Thank you for your support!

• a16z
• AMD
• Anyscale
• AWS
• Crusoe Cloud
• Databricks
• DeepInfra
• Dropbox
• Google Cloud
• Lambda Lab
• Nebius
• NVIDIA
• Replicate
• Roblox
• RunPod
• Sequoia Capital
• Skywork AI
• Trainy
• UC Berkeley
• UC San Diego
• ZhenFund

We also have an official fundraising venue through OpenCollective. We plan to use the fund to support the development, maintenance, and adoption of vLLM.

Citation

If you use vLLM for your research, please cite our paper:

@inproceedings{kwon2023efficient,
  title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
  author={Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica},
  booktitle={Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles},
  year={2023}
}

Contact Us

• For technical questions and feature requests, please use Github issues or discussions.
• For discussing with fellow users, please use Discord.
• For coordinating contributions and development, please use Slack.
• For security disclosures, please use Github's security advisory feature.
• For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.

Media Kit

• If you wish to use vLLM's logo, please refer to our media kit repo.