@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
md 82 lines 4.4 KB
Rendered Raw
1 ---
2 hide:
3 - navigation
4 - toc
5 ---
6
7 # Welcome to vLLM
8
9 <figure markdown="span">
10 ![](./assets/logos/vllm-logo-text-light.png){ align="center" alt="vLLM Light" class="logo-light" width="60%" }
11 ![](./assets/logos/vllm-logo-text-dark.png){ align="center" alt="vLLM Dark" class="logo-dark" width="60%" }
12 </figure>
13
14 <p style="text-align:center">
15 <strong>Easy, fast, and cheap LLM serving for everyone
16 </strong>
17 </p>
18
19 <p style="text-align:center">
20 <script async defer src="https://buttons.github.io/buttons.js"></script>
21 <a class="github-button" href="https://github.com/vllm-project/vllm" data-show-count="true" data-size="large" aria-label="Star">Star</a>
22 <a class="github-button" href="https://github.com/vllm-project/vllm/subscription" data-show-count="true" data-icon="octicon-eye" data-size="large" aria-label="Watch">Watch</a>
23 <a class="github-button" href="https://github.com/vllm-project/vllm/fork" data-show-count="true" data-icon="octicon-repo-forked" data-size="large" aria-label="Fork">Fork</a>
24 </p>
25
26 vLLM is a fast and easy-to-use library for LLM inference and serving.
27
28 Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has grown into one of the most active open-source AI projects built and maintained by a diverse community of many dozens of academic institutions and companies from over 2000 contributors.
29
30 Where to get started with vLLM depends on the type of user. If you are looking to:
31
32 - Run open-source models on vLLM, we recommend starting with the [Quickstart Guide](./getting_started/quickstart.md)
33 - Build applications with vLLM, we recommend starting with the [User Guide](./usage/README.md)
34 - Build vLLM, we recommend starting with [Developer Guide](./contributing/README.md)
35
36 For information about the development of vLLM, see:
37
38 - [Roadmap](https://roadmap.vllm.ai)
39 - [Releases](https://github.com/vllm-project/vllm/releases)
40
41 vLLM is fast with:
42
43 - State-of-the-art serving throughput
44 - Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
45 - Continuous batching of incoming requests, chunked prefill, prefix caching
46 - Fast and flexible model execution with piecewise and full CUDA/HIP graphs
47 - Quantization: FP8, MXFP8/MXFP4, NVFP4, INT8, INT4, GPTQ/AWQ, GGUF, compressed-tensors, ModelOpt, TorchAO, and [more](https://docs.vllm.ai/en/latest/features/quantization/index.html)
48 - Optimized attention kernels including FlashAttention, FlashInfer, TRTLLM-GEN, FlashMLA, and Triton
49 - Optimized GEMM/MoE kernels for various precisions using CUTLASS, TRTLLM-GEN, CuTeDSL
50 - Speculative decoding including n-gram, suffix, EAGLE, DFlash
51 - Automatic kernel generation and graph-level transformations using torch.compile
52 - Disaggregated prefill, decode, and encode
53
54 vLLM is flexible and easy to use with:
55
56 - Seamless integration with popular Hugging Face models
57 - High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
58 - Tensor, pipeline, data, expert, and context parallelism for distributed inference
59 - Streaming outputs
60 - Generation of structured outputs using xgrammar or guidance
61 - Tool calling and reasoning parsers
62 - OpenAI-compatible API server, plus Anthropic Messages API and gRPC support
63 - Efficient multi-LoRA support for dense and MoE layers
64 - Support for NVIDIA GPUs, AMD GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon, MetaX GPU, and more.
65
66 vLLM seamlessly supports 200+ model architectures on HuggingFace, including:
67
68 - Decoder-only LLMs (e.g., Llama, Qwen, Gemma)
69 - Mixture-of-Expert LLMs (e.g., Mixtral, DeepSeek-V3, Qwen-MoE, GPT-OSS)
70 - Hybrid attention and state-space models (e.g., Mamba, Qwen3.5)
71 - Multi-modal models (e.g., LLaVA, Qwen-VL, Pixtral)
72 - Embedding and retrieval models (e.g., E5-Mistral, GTE, ColBERT)
73 - Reward and classification models (e.g., Qwen-Math)
74
75 Find the full list of supported models [here](./models/supported_models.md).
76
77 For more information, check out the following:
78
79 - [vLLM announcing blog post](https://blog.vllm.ai/2023/06/20/vllm.html) (intro to PagedAttention)
80 - [vLLM paper](https://arxiv.org/abs/2309.06180) (SOSP 2023)
81 - [How continuous batching enables 23x throughput in LLM inference while reducing p50 latency](https://www.anyscale.com/blog/continuous-batching-llm-inference) by Cade Daniel et al.
82 - [vLLM Meetups](community/meetups.md)