vllm-xft 0.5.5.4

A high-throughput and memory-efficient inference and serving engine for LLMs

This is a fork of vLLM to support xfastertransformer backend. This version is based on official vllm v0.4.2.

Notice

🎉🎉🎉Continuous batching and distributed is supported. 🎇🎇🎇

• BeamSearch is not support yet.(WIP)
• LORA is not support yet.(WIP)

Install

From PyPI

pip install vllm-xft

From Source

python3 setup.py bdist_wheel --verbose

Usage

Python offline

python examples/offline_inference_xfastertransformer.py

Serving(OpenAI Compatible Server)

python -m vllm.entrypoints.openai.api_server \
        --model /data/llama-2-7b-chat-cpu \
        --tokenizer /data/llama-2-7b-chat-hf \
        --dtype fp16 \
        --kv-cache-dtype fp16 \
        --served-model-name xft \
        --port 8000 \
        --trust-remote-code \

• --max-num-batched-tokens: max batched token, default value is max(MAX_SEQ_LEN_OF_MODEL, 2048).
• --max-num-seqs: max seqs batch, default is 256.

More Arguments please refer to vllm official docs

Query example

  curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
  "model": "xft",
  "prompt": "San Francisco is a",
  "max_tokens": 512,
  "temperature": 0
  }'

Distributed(Multi-rank)

Use oneCCL's mpirun to run the workload. The master (rank 0) is the same as the single-rank above, and the slaves (rank > 0) should use the following command:

python -m vllm.entrypoints.slave --dtype fp16 --model ${MODEL_PATH} --kv-cache-dtype fp16

Please keep params of slaves align with master.

Serving(OpenAI Compatible Server)

Here is a example on 2Socket platform, 48 cores pre socket.

OMP_NUM_THREADS=48 mpirun \
        -n 1 numactl --all -C 0-47 -m 0 \
          python -m vllm.entrypoints.openai.api_server \
            --model ${MODEL_PATH} \
            --tokenizer ${TOKEN_PATH} \
            --dtype bf16 \
            --kv-cache-dtype fp16 \
            --served-model-name xft \
            --port 8000 \
            --trust-remote-code \
        : -n 1 numactl --all -C 48-95 -m 1 \
          python -m vllm.entrypoints.slave \
            --dtype bf16 \
            --model ${MODEL_PATH} \
            --kv-cache-dtype fp16

Easy, fast, and cheap LLM serving for everyone

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

---

vLLM & NVIDIA Triton User Meetup (Monday, September 9, 5pm-9pm PT) at Fort Mason, San Francisco

We are excited to announce our sixth vLLM Meetup, in collaboration with NVIDIA Triton Team. Join us to hear the vLLM's recent update about performance. Register now here and be part of the event!

---

Latest News 🔥

• [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 that compares the performance of vLLM against other LLM serving engines (TensorRT-LLM, text-generation-inference and lmdeploy).

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
• 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}
}