# Technical Documentation: OpenBMB/vllm > ℹ️ **Provenance:** Multi-Tier Ingestion · [CodeWiki Reference](https://codewiki.google/github.com/OpenBMB/vllm) ### Overview (`README.md`)

vLLM

Easy, fast, and cheap LLM serving for everyone

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

--- The first vLLM meetup in 2025 is happening on January 22nd, Wednesday, with Google Cloud in San Francisco! We will talk about vLLM's performant V1 architecture, Q1 roadmap, Google Cloud's innovation around vLLM: networking, Cloud Run, Vertex, and TPU! [Register Now](https://lu.ma/zep56hui) --- *Latest News* 🔥 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone! - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing). - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://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](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users! - [2024/09] We hosted [the sixth vLLM meetup](https://lu.ma/87q3nvnh) with NVIDIA! Please find the meetup slides [here](https://docs.google.com/presentation/d/1wrLGwytQfaOTd5wCGSPNhoaW3nq0E-9wqyP7ny93xRs/edit?usp=sharing). - [2024/07] We hosted [the fifth vLLM meetup](https://lu.ma/lp0gyjqr) with AWS! Please find the meetup slides [here](https://docs.google.com/presentation/d/1RgUD8aCfcHocghoP3zmXzck9vX3RCI9yfUAB2Bbcl4Y/edit?usp=sharing). - [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](https://blog.vllm.ai/2024/07/23/llama31.html). - [2024/06] We hosted [the fourth vLLM meetup](https://lu.ma/agivllm) with Cloudflare and BentoML! Please find the meetup slides [here](https://docs.google.com/presentation/d/1iJ8o7V2bQEi0BFEljLTwc5G1S10_Rhv3beed5oB0NJ4/edit?usp=sharing). - [2024/04] We hosted [the third vLLM meetup](https://robloxandvllmmeetup2024.splashthat.com/) with Roblox! Please find the meetup slides [here](https://docs.google.com/presentation/d/1A--47JAK4BJ39t954HyTkvtfwn0fkqtsL8NGFuslReM/edit?usp=sharing). - [2024/01] We hosted [the second vLLM meetup](https://lu.ma/ygxbpzhl) with IBM! Please find the meetup slides [here](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing). - [2023/10] We hosted [the first vLLM meetup](https://lu.ma/first-vllm-meetup) with a16z! Please find the meetup slides [here](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing). - [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (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](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai). --- ## About vLLM is a fast and easy-to-use library for LLM inference and serving. Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evloved into a community-driven project with contributions from both academia and industry. vLLM is fast with: - State-of-the-art serving throughput - Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html) - Continuous batching of incoming requests - Fast model execution with CUDA/HIP graph - Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), 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](https://blog.vllm.ai/2024/09/05/perf-update.html). It compares the performance of vLLM against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [SGLang](https://github.com/sgl-project/sglang) and [LMDeploy](https://github.com/InternLM/lmdeploy)). The implementation is under [nightly-benchmarks folder](.buildkite/nightly-benchmarks/) and you can [reproduce](https://github.com/vllm-project/vllm/issues/8176) 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, Deepseek-V2 and V3) - Embedding Models (e.g. E5-Mistral) - Multi-modal LLMs (e.g., LLaVA) Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html). ## Getting Started Install vLLM with `pip` or [from source](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/index.html#build-wheel-from-source): ```bash pip install vllm ``` Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more. - [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html) - [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html) - [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html) ## Contributing We welcome and value any contributions and collaborations. Please check out [CONTRIBUTING.md](./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! Cash Donations: - a16z - Dropbox - Sequoia Capital - Skywork AI - ZhenFund Compute Resources: - AMD - Anyscale - AWS - Crusoe Cloud - Databricks - DeepInfra - Google Cloud - Lambda Lab - Nebius - Novita AI - NVIDIA - Replicate - Roblox - RunPod - Trainy - UC Berkeley - UC San Diego Slack Sponsor: Anyscale We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). 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](https://arxiv.org/abs/2309.06180): ```bibtex @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](https://github.com/vllm-project/media-kit). --- ### Overview (`benchmarks/README.md`) # Benchmarking vLLM ## Downloading the ShareGPT dataset You can download the dataset by running: ```bash wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json ``` ## Downloading the ShareGPT4V dataset The json file refers to several image datasets (coco, llava, etc.). The benchmark scripts will ignore a datapoint if the referred image is missing. ```bash wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json mkdir coco -p wget http://images.cocodataset.org/zips/train2017.zip -O coco/train2017.zip unzip coco/train2017.zip -d coco/ ``` --- ### Overview (`docs/README.md`) # vLLM documents ## Build the docs ```bash # Install dependencies. pip install -r requirements-docs.txt # Build the docs. make clean make html ``` ## Open the docs with your browser ```bash python -m http.server -d build/html/ ``` Launch your browser and open localhost:8000. --- ### Overview (`.buildkite/nightly-benchmarks/README.md`) # vLLM benchmark suite ## Introduction This directory contains two sets of benchmark for vllm. - Performance benchmark: benchmark vllm's performance under various workload, for **developers** to gain clarity on whether their PR improves/degrades vllm's performance - Nightly benchmark: compare vllm's performance against alternatives (tgi, trt-llm and lmdeploy), for **the public** to know when to choose vllm. See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results. ## Performance benchmark quick overview **Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!), with different models. **Benchmarking Duration**: about 1hr. **For benchmarking developers**: please try your best to constraint the duration of benchmarking to about 1 hr so that it won't take forever to run. ## Nightly benchmark quick overview **Benchmarking Coverage**: Fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) on Llama-3 8B, 70B and Mixtral 8x7B. **Benchmarking engines**: vllm, TGI, trt-llm and lmdeploy. **Benchmarking Duration**: about 3.5hrs. ## Trigger the benchmark Performance benchmark will be triggered when: - A PR being merged into vllm. - Every commit for those PRs with `perf-benchmarks` label AND `ready` label. Nightly benchmark will be triggered when: - Every commit for those PRs with `perf-benchmarks` label and `nightly-benchmarks` label. ## Performance benchmark details See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases. #### Latency test Here is an example of one test inside `latency-tests.json`: ```json [ { "test_name": "latency_llama8B_tp1", "parameters": { "model": "meta-llama/Meta-Llama-3-8B", "tensor_parallel_size": 1, "load_format": "dummy", "num_iters_warmup": 5, "num_iters": 15 } }, ] ``` In this example: - The `test_name` attributes is a unique identifier for the test. In `latency-tests.json`, it must start with `latency_`. - The `parameters` attribute control the command line arguments to be used for `benchmark_latency.py`. Note that please use underline `_` instead of the dash `-` when specifying the command line arguments, and `run-performance-benchmarks.sh` will convert the underline to dash when feeding the arguments to `benchmark_latency.py`. For example, the corresponding command line arguments for `benchmark_latency.py` will be `--model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --load-format dummy --num-iters-warmup 5 --num-iters 15` Note that the performance numbers are highly sensitive to the value of the parameters. Please make sure the parameters are set correctly. WARNING: The benchmarking script will save json results by itself, so please do not configure `--output-json` parameter in the json file. #### Throughput test The tests are specified in `throughput-tests.json`. The syntax is similar to `latency-tests.json`, except for that the parameters will be fed forward to `benchmark_throughput.py`. The number of this test is also stable -- a slight change on the value of this number might vary the performance numbers by a lot. #### Serving test We test the throughput by using `benchmark_serving.py` with request rate = inf to cover the online serving overhead. The corresponding parameters are in `serving-tests.json`, and here is an example: ``` [ { "test_name": "serving_llama8B_tp1_sharegpt", "qps_list": [1, 4, 16, "inf"], "server_parameters": { "model": "meta-llama/Meta-Llama-3-8B", "tensor_parallel_size": 1, "swap_space": 16, "disable_log_stats": "", "disable_log_requests": "", "load_format": "dummy" }, "client_parameters": { "model": "meta-llama/Meta-Llama-3-8B", "backend": "vllm", "dataset_name": "sharegpt", "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json", "num_prompts": 200 } }, ] ``` Inside this example: - The `test_name` attribute is also a unique identifier for the test. It must start with `serving_`. - The `server-parameters` includes the command line arguments for vLLM server. - The `client-parameters` includes the command line arguments for `benchmark_serving.py`. - The `qps_list` controls the list of qps for test. It will be used to configure the `--request-rate` parameter in `benchmark_serving.py` The number of this test is less stable compared to the delay and latency benchmarks (due to randomized sharegpt dataset sampling inside `benchmark_serving.py`), but a large change on this number (e.g. 5% change) still vary the output greatly. WARNING: The benchmarking script will save json results by itself, so please do not configure `--save-results` or other results-saving-related parameters in `serving-tests.json`. #### Visualizing the results The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](tests/descriptions.md) with real benchmarking results. You can find the result presented as a table inside the `buildkite/performance-benchmark` job page. If you do not see the table, please wait till the benchmark finish running. The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file. The raw benchmarking results (in the format of json files) are in the `Artifacts` tab of the benchmarking. ## Nightly test details See [nightly-descriptions.md](nightly-descriptions.md) for the detailed description on test workload, models and docker containers of benchmarking other llm engines. #### Workflow - The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines. - Inside each container, we run [run-nightly-suite.sh](run-nightly-suite.sh), which will probe the serving engine of the current container. - The `run-nightly-suite.sh` will redirect the request to `tests/run-[llm serving engine name]-nightly.sh`, which parses the workload described in [nightly-tests.json](tests/nightly-tests.json) and performs the benchmark. - At last, we run [scripts/plot-nightly-results.py](scripts/plot-nightly-results.py) to collect and plot the final benchmarking results, and update the results to buildkite. #### Nightly tests In [nightly-tests.json](tests/nightly-tests.json), we include the command line arguments for benchmarking commands, together with the benchmarking test cases. The format is highly similar to performance benchmark. #### Docker containers The docker containers for benchmarking are specified in `nightly-pipeline.yaml`. WARNING: the docker versions are HARD-CODED and SHOULD BE ALIGNED WITH `nightly-descriptions.md`. The docker versions need to be hard-coded as there are several version-specific bug fixes inside `tests/run-[llm serving engine name]-nightly.sh`. WARNING: populating `trt-llm` to latest version is not easy, as it requires updating several protobuf files in [tensorrt-demo](https://github.com/neuralmagic/tensorrt-demo.git). --- ### Overview (`csrc/quantization/machete/Readme.md`) # Machete (Mixed Precision Cutlass-Based GEMM) Machete is a spiritual successor to the Marlin kernel but optimized for Hopper architectures and based on Cutlass. Being based on Cutlass, new type pairs and epilogues are easier to add compared to Marlin. ## Overview Machete effectively performs ``` scale_type = w_s.dtype compute_type = a.dtype out = (w_q.to(scale_type) * w_s - w_z.to(scale_type)) @ a ``` Where `w_q` is a quantized weight matrix, `w_s` is the quantization scales, and `w_z` is the quantization zeropoints. > **_NOTE:_** `w_z` is added after the scales so we can use FMA operations, but this means they must have the scales pre-applied if the supplied zeropoints assume that they will be subtracted before the scales are applied. ## API The main optimization within Machete is prepacking the weight matrix to more closely match the tensor core layouts, allowing for wider shared memory loads when loading the weight matrix. This means that the weight matrix must be prepacked before calling `machete_gemm`. The flow looks something like: ``` from vllm import _custom_ops as ops ... W_q_packed = ops.machete_prepack_B(w_q, wtype) output = ops.machete_gemm( a, b_q=W_q_packed, b_type=wtype, b_scales=w_s, b_group_size=group_size ) ``` ## Code Generation Since Machete is based on Cutlass, we can generate multiple type pairs and different tile shapes using the same kernel template. We generate multiple instantiations of this template using `generate.py`. New type pairs (`TypeConfig`s) can be appended to `impl_configs` (in `generate()`), and these will get automatically generated (assuming they can be supported without issues). For each `TypeConfig`, you must also provide an `ImplConfig`, which bundles a `TypeConfig` with a list of `ScheduleConfig`s, `Specialization`s, and a default heuristic. The `ScheduleConfig`s (which contain info on tile shapes, tile scheduler, etc.) can perform differently for different problem shapes, and there is almost never one `ScheduleConfig` that works well for all problem shapes, so it is generally beneficial to generate different `ScheduleConfig`s for different potential problem shapes. This is where the heuristic comes in. For each `TypeConfig`, a default heuristic should be provided. This maps different problem shapes to different `ScheduleConfig`s and is used when the user does not provide the `schedule` parameter to `machete_gemm`. The `Specialization`s define what feature combinations to generate, i.e., `with_zeropoints`, `with_scales`, etc. We can reduce compile times and the final binary size by limiting the set of feature combinations we generate. --- ### Overview (`examples/online_serving/chart-helm/README.md`) # Helm Charts This directory contains a Helm chart for deploying the vllm application. The chart includes configurations for deployment, autoscaling, resource management, and more. ## Files - Chart.yaml: Defines the chart metadata including name, version, and maintainers. - ct.yaml: Configuration for chart testing. - lintconf.yaml: Linting rules for YAML files. - values.schema.json: JSON schema for validating values.yaml. - values.yaml: Default values for the Helm chart. - templates/_helpers.tpl: Helper templates for defining common configurations. - templates/configmap.yaml: Template for creating ConfigMaps. - templates/custom-objects.yaml: Template for custom Kubernetes objects. - templates/deployment.yaml: Template for creating Deployments. - templates/hpa.yaml: Template for Horizontal Pod Autoscaler. - templates/job.yaml: Template for Kubernetes Jobs. - templates/poddisruptionbudget.yaml: Template for Pod Disruption Budget. - templates/pvc.yaml: Template for Persistent Volume Claims. - templates/secrets.yaml: Template for Kubernetes Secrets. - templates/service.yaml: Template for creating Services. --- ### Overview (`examples/online_serving/prometheus_grafana/README.md`) # Prometheus and Grafana This is a simple example that shows you how to connect vLLM metric logging to the Prometheus/Grafana stack. For this example, we launch Prometheus and Grafana via Docker. You can checkout other methods through [Prometheus](https://prometheus.io/) and [Grafana](https://grafana.com/) websites. Install: - [`docker`](https://docs.docker.com/engine/install/) - [`docker compose`](https://docs.docker.com/compose/install/linux/#install-using-the-repository) ## Launch Prometheus metric logging is enabled by default in the OpenAI-compatible server. Launch via the entrypoint: ```bash vllm serve mistralai/Mistral-7B-v0.1 \ --max-model-len 2048 \ --disable-log-requests ``` Launch Prometheus and Grafana servers with `docker compose`: ```bash docker compose up ``` Submit some sample requests to the server: ```bash wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json python3 ../../benchmarks/benchmark_serving.py \ --model mistralai/Mistral-7B-v0.1 \ --tokenizer mistralai/Mistral-7B-v0.1 \ --endpoint /v1/completions \ --dataset-name sharegpt \ --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json \ --request-rate 3.0 ``` Navigating to [`http://localhost:8000/metrics`](http://localhost:8000/metrics) will show the raw Prometheus metrics being exposed by vLLM. ## Grafana Dashboard Navigate to [`http://localhost:3000`](http://localhost:3000). Log in with the default username (`admin`) and password (`admin`). ### Add Prometheus Data Source Navigate to [`http://localhost:3000/connections/datasources/new`](http://localhost:3000/connections/datasources/new) and select Prometheus. On Prometheus configuration page, we need to add the `Prometheus Server URL` in `Connection`. For this setup, Grafana and Prometheus are running in separate containers, but Docker creates DNS name for each containers. You can just use `http://prometheus:9090`. Click `Save & Test`. You should get a green check saying "Successfully queried the Prometheus API.". ### Import Dashboard Navigate to [`http://localhost:3000/dashboard/import`](http://localhost:3000/dashboard/import), upload `grafana.json`, and select the `prometheus` datasource. You should see a screen that looks like the following: --- ### Overview (`examples/other/fp8/README.md`) # FP8 KV Cache This utility extracts the KV cache scaling factors from a quantized HF (Hugging Face) model. The extracted scaling factors are saved to a JSON file, which can later be used by vLLM (variable-length language model) during runtime. This tool is particularly useful when the KV cache data type is FP8 and is intended for use on ROCm (AMD GPU) platforms. ## Prerequisites - Python 3.x - PyTorch - NumPy - Hugging Face Transformers - Hugging Face Hub - AMMO Before incorporating the FP8 datatype for inference workloads, you must adhere to the following steps: 1. Install all necessary prerequisites and dependencies. 2. Convert HF model into a quantized HF model. 3. Extract KV Cache Scaling Factors from quantized HF model. 4. Load KV Cache Scaling Factors into VLLM. ### 2. Convert HF model into a quantized HF model. Note: The following steps are adapted from the [TensorRT-LLM repository](https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/quantization/README.md). `quantize.py` (examples/other/fp8/quantizer/quantize.py) uses the quantization toolkit (AMMO) to calibrate the PyTorch models and export TensorRT-LLM checkpoints. Each TensorRT-LLM checkpoint contains a config file (in .json format) and one or several rank weight files (in .safetensors format). The detailed quantization toolkit (AMMO) conversion guide for FP8 can be found at `examples/other/fp8/quantizer/README.md`. ### 3. Extract KV Cache Scaling Factors from quantized HF model. `extract_scales.py` (examples/other/fp8/extract_scales.py) can be utilized to extract the KV cache scaling factors from your quantized HF model, however at the moment, this tool exclusively supports Llama 2 models. It is also important to note the following: 1. **File Structure**: The utility operates under the assumption that all parameters, including KV cache scaling factors, corresponding to a particular Tensor Parallelism (TP) rank are stored in a single file. These files must adhere to a specific naming convention where the TP rank is immediately identified after a specific keyword (e.g., "rank") in the filename. 2. **TP Decomposition**: The utility assumes consistency between the TP decomposition employed by the quantizer tool and that used by vLLM. 3. **AMMO Compatibility**: Currently, the generated KV cache scaling factors for AMMO remain uniform across all TP ranks. ```python # prerequisites: # - Quantized HF LLaMa 2 model python3 examples/other/fp8/extract_scales.py --help Usage: extract_scales.py [-h] --quantized_model QUANTIZED_MODEL [--load_format {auto,safetensors,npz,pt}] [--output_dir OUTPUT_DIR] [--output_name OUTPUT_NAME] [--tp_size TP_SIZE] KV Scale Extraction Example optional arguments: --quantized_model: Specify either the local path to, or name of, a quantized HF model. It is expected that the quantization format is FP8_E4M3, for use on ROCm (AMD GPU). Optional arguments: --cache_dir: Specify a cache directory to use in the event of a HF model download. (Default: None) --load_format: Specify the format of the model's tensor files containing the KV cache scaling factors. (Choices: auto, safetensors, npz, pt; Default: auto) --revision: Specify the model's revision number. (Default: None) --output_dir: Specify the output directory. By default the KV cache scaling factors will be saved in the model directory. (Default: None) --output_name: Specify the output filename. (Default: kv_cache_scales.json) --tp_size: Specify the tensor-parallel (TP) size that the quantized model should correspond to. If specified, during KV cache scaling factor extraction the observed TP size will be checked against this and an error will be raised if there is a mismatch. (Default: None) ``` ```python Example: python3 examples/other/fp8/extract_scales.py --quantized_model --tp_size --output_dir ``` ### 4. Load KV Cache Scaling Factors into VLLM. This script evaluates the inference throughput of language models using various backends such as vLLM. It measures the time taken to process a given number of prompts and generate sequences for each prompt. The recently generated KV cache scaling factors are now integrated into the benchmarking process and allow for KV cache scaling factors to be utilized for FP8. ``` /* Detailed source-code truncated for AI context efficiency. */ ```for common inference criteria. --quantization-param-path QUANT_PARAM_JSON Path to the JSON file containing the KV cache scaling factors. This should generally be supplied, when KV cache dtype is FP8. Otherwise, KV cache scaling factors default to 1.0, which may cause accuracy issues. FP8_E5M2 (without scaling) is only supported on cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead supported for common inference criteria. ``` Example: ```console python3 benchmarks/benchmark_throughput.py --input-len --output-len -tp --kv-cache-dtype fp8 --quantization-param-path --model ``` --- ### Overview (`vllm/distributed/kv_transfer/README.md`) # Distributed KV cache transfer This folder implements distributed KV cache transfer across vLLM instances. Currently the main usecase is for disaggregated prefilling. ## Abstractions The KV cache transfer contains three layer of abstractions: - KV pipe: a FIFO pipe for torch.tensor transmission. Key APIs: `send_tensor` and `recv_tensor`. - KV lookup buffer: a lookup buffer for KV caches. Key: the tokens, value: the KV caches (and/or hidden states). Key APIs: `insert` and `drop_select` (similar to SQL semantics). - KV connector: a connector that connects the KV pipe and KV lookup buffer to vLLM. Key APIs: `send_kv_caches_and_hidden_states` and `recv_kv_caches_and_hidden_states`. Why we need KV lookup buffer: FIFO pipe itself is not enough as prefill vLLM worker may process requests in a different order compared to decode vLLM worker. Say the QPS is really high, prefill worker may handle requests in order A -> B -> C, but the decode worker may process request C first. This is not the case that can be naturally handled by FIFO pipe, so we provide KV lookup buffer to help translate a FIFO pipe to a lookup buffer. NOTE: KV pipe layer is bypassible: you can skip this layer if your distributed communication service already supports key-value-based lookup (like redis or RDMA database). NOTE: If you want to not only transfer KV caches, but adjust the model execution flow of vLLM as well (for example, allow vLLM to receive KV caches on some tokens and do prefill on the remaining tokens), you can bypass both KV pipe layer and KV lookup buffer layer, and directly implement on KV connector layer. Bear in mind that as vLLM's model input is constantly changing, this implementation will likely be broken when vLLM has new updates. ## Disaggregated prefilling The example usage is in [this file](../../../examples/online_serving/disaggregated_prefill.sh). Here is the diagram of how we run disaggretgated prefilling. --- ### Overview (`examples/other/fp8/quantizer/README.md`) ### Quantizer Utilities `quantize.py`: NVIDIA Quantization utilities using TensorRT-Model-Optimizer, ported from TensorRT-LLM: [`examples/quantization/quantize.py`](https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/quantization/quantize.py) ### Prerequisite #### AMMO (AlgorithMic Model Optimization) Installation: nvidia-ammo 0.7.1 or later `pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com nvidia-ammo` #### AMMO Download (code and docs) `https://developer.nvidia.com/downloads/assets/cuda/files/nvidia-ammo/nvidia_ammo-0.5.0.tar.gz` `https://developer.nvidia.com/downloads/assets/cuda/files/nvidia-ammo/nvidia_ammo-0.7.1.tar.gz` ### Usage #### Run on H100 system for speed if FP8; number of GPUs depends on the model size #### Example: quantize Llama2-7b model from HF to FP8 with FP8 KV Cache: `python quantize.py --model-dir ./ll2-7b --dtype float16 --qformat fp8 --kv-cache-dtype fp8 --output-dir ./ll2_7b_fp8 --calib-size 512 --tp-size 1` Outputs: model structure, quantized model & parameters (with scaling factors) are in JSON and Safetensors (npz is generated only for the reference) ``` # ll ./ll2_7b_fp8/ total 19998244 drwxr-xr-x 2 root root 4096 Feb 7 01:08 ./ drwxrwxr-x 8 1060 1061 4096 Feb 7 01:08 ../ -rw-r--r-- 1 root root 176411 Feb 7 01:08 llama_tp1.json -rw-r--r-- 1 root root 13477087480 Feb 7 01:09 llama_tp1_rank0.npz -rw-r--r-- 1 root root 7000893272 Feb 7 01:08 rank0.safetensors # ``` --- ### Source Index (`docs/source/index.md`) # Welcome to vLLM ```{figure} ./assets/logos/vllm-logo-text-light.png :align: center :alt: vLLM :class: no-scaled-link :width: 60% ``` ```{raw} html

Easy, fast, and cheap LLM serving for everyone

Star Watch Fork

``` vLLM is a fast and easy-to-use library for LLM inference and serving. Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evloved into a community-driven project with contributions from both academia and industry. vLLM is fast with: - State-of-the-art serving throughput - Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html) - Continuous batching of incoming requests - Fast model execution with CUDA/HIP graph - Quantization: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), INT4, INT8, and FP8 - Optimized CUDA kernels, including integration with FlashAttention and FlashInfer. - Speculative decoding - Chunked prefill vLLM is flexible and easy to use with: - Seamless integration with popular HuggingFace 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, Gaudi® accelerators and GPUs, PowerPC CPUs, TPU, and AWS Trainium and Inferentia Accelerators. - Prefix caching support - Multi-lora support For more information, check out the following: - [vLLM announcing blog post](https://vllm.ai) (intro to PagedAttention) - [vLLM paper](https://arxiv.org/abs/2309.06180) (SOSP 2023) - [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. - [vLLM Meetups](#meetups) ## Documentation % How to start using vLLM? ```{toctree} :caption: Getting Started :maxdepth: 1 getting_started/installation/index getting_started/quickstart getting_started/examples/examples_index getting_started/troubleshooting getting_started/faq ``` % What does vLLM support? ```{toctree} :caption: Models :maxdepth: 1 models/generative_models models/pooling_models models/supported_models models/extensions/index ``` % Additional capabilities ```{toctree} :caption: Features :maxdepth: 1 features/quantization/index features/lora features/tool_calling features/structured_outputs features/automatic_prefix_caching features/disagg_prefill features/spec_decode features/compatibility_matrix ``` % Details about running vLLM ```{toctree} :caption: Inference and Serving :maxdepth: 1 serving/offline_inference serving/openai_compatible_server serving/multimodal_inputs serving/distributed_serving serving/metrics serving/engine_args serving/env_vars serving/usage_stats serving/integrations/index ``` % Scaling up vLLM for production ```{toctree} :caption: Deployment :maxdepth: 1 deployment/docker deployment/k8s deployment/nginx deployment/frameworks/index deployment/integrations/index ``` % Making the most out of vLLM ```{toctree} :caption: Performance :maxdepth: 1 performance/optimization performance/benchmarks ``` % Explanation of vLLM internals ```{toctree} :caption: Design Documents :maxdepth: 2 design/arch_overview design/huggingface_integration design/plugin_system design/kernel/paged_attention design/mm_processing design/automatic_prefix_caching design/multiprocessing ``` % How to contribute to the vLLM project ```{toctree} :caption: Developer Guide :maxdepth: 2 contributing/overview contributing/profiling/profiling_index contributing/dockerfile/dockerfile contributing/model/index contributing/vulnerability_management ``` % Technical API specifications ```{toctree} :caption: API Reference :maxdepth: 2 api/offline_inference/index api/engine/index api/inference_params api/multimodal/index api/model/index ``` % Latest news and acknowledgements ```{toctree} :caption: Community :maxdepth: 1 community/meetups community/sponsors ``` ## Indices and tables - {ref}`genindex` - {ref}`modindex` --- ### Source Contributing Overview (`docs/source/contributing/overview.md`) # Contributing to vLLM Thank you for your interest in contributing to vLLM! Our community is open to everyone and welcomes all kinds of contributions, no matter how small or large. There are several ways you can contribute to the project: - Identify and report any issues or bugs. - Request or add support for a new model. - Suggest or implement new features. - Improve documentation or contribute a how-to guide. We also believe in the power of community support; thus, answering queries, offering PR reviews, and assisting others are also highly regarded and beneficial contributions. Finally, one of the most impactful ways to support us is by raising awareness about vLLM. Talk about it in your blog posts and highlight how it's driving your incredible projects. Express your support on social media if you're using vLLM, or simply offer your appreciation by starring our repository! ## License See . ## Developing Depending on the kind of development you'd like to do (e.g. Python, CUDA), you can choose to build vLLM with or without compilation. Check out the [building from source](#build-from-source) documentation for details. ## Testing ```bash pip install -r requirements-dev.txt # Linting, formatting and static type checking pre-commit install # You can manually run pre-commit with pre-commit run --all-files # Unit tests pytest tests/ ``` ```{note} Currently, the repository is not fully checked by `mypy`. ``` ## Issues If you encounter a bug or have a feature request, please [search existing issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue) first to see if it has already been reported. If not, please [file a new issue](https://github.com/vllm-project/vllm/issues/new/choose), providing as much relevant information as possible. ```{important} If you discover a security vulnerability, please follow the instructions [here](gh-file:SECURITY.md#reporting-a-vulnerability). ``` ## Pull Requests & Code Reviews Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process. ### DCO and Signed-off-by When contributing changes to this project, you must agree to the . Commits must include a `Signed-off-by:` header which certifies agreement with the terms of the DCO. Using `-s` with `git commit` will automatically add this header. ### PR Title and Classification Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following: - `[Bugfix]` for bug fixes. - `[CI/Build]` for build or continuous integration improvements. - `[Doc]` for documentation fixes and improvements. - `[Model]` for adding a new model or improving an existing model. Model name should appear in the title. - `[Frontend]` For changes on the vLLM frontend (e.g., OpenAI API server, `LLM` class, etc.) - `[Kernel]` for changes affecting CUDA kernels or other compute kernels. - `[Core]` for changes in the core vLLM logic (e.g., `LLMEngine`, `AsyncLLMEngine`, `Scheduler`, etc.) - `[Hardware][Vendor]` for hardware-specific changes. Vendor name should appear in the prefix (e.g., `[Hardware][AMD]`). - `[Misc]` for PRs that do not fit the above categories. Please use this sparingly. ```{note} If the PR spans more than one category, please include all relevant prefixes. ``` ### Code Quality The PR needs to meet the following code quality standards: - We adhere to [Google Python style guide](https://google.github.io/styleguide/pyguide.html) and [Google C++ style guide](https://google.github.io/styleguide/cppguide.html). - Pass all linter checks. Please use `pre-commit` to format your code. See if `pre-commit` is new to you. - The code needs to be well-documented to ensure future contributors can easily understand the code. - Include sufficient tests to ensure the project stays correct and robust. This includes both unit tests and integration tests. - Please add documentation to `docs/source/` if the PR modifies the user-facing behaviors of vLLM. It helps vLLM users understand and utilize the new features or changes. ### Adding or Changing Kernels Each custom kernel needs a schema and one or more implementations to be registered with PyTorch. - Make sure custom ops are registered following PyTorch guidelines: [Custom C++ and CUDA Operators](https://pytorch.org/tutorials/advanced/cpp_custom_ops.html#cpp-custom-ops-tutorial) and [The Custom Operators Manual](https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU). - Custom operations that return `Tensors` require meta-functions. Meta-functions should be implemented and registered in Python so that dynamic dims can be handled automatically. See above documents for a description of meta-functions. - Use [torch.library.opcheck()](https://pytorch.org/docs/stable/library.html#torch.library.opcheck) to test the function registration and meta-function for any registered ops. See `tests/kernels` for examples. - When changing the C++ signature of an existing op, the schema must be updated to reflect the changes. - If a new custom type is needed, see the following document: [Custom Class Support in PT2](https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA). ### Notes for Large Changes Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with `rfc-required` and might not go through the PR. ### What to Expect for the Reviews The goal of the vLLM team is to be a *transparent reviewing machine*. We would like to make the review process transparent and efficient and make sure no contributor feels confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process: - After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability. - After the PR is assigned, the reviewer will provide status updates every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team. - After the review, the reviewer will put an `action-required` label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR. - Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion. ## Thank You Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. All of your contributions help make vLLM a great tool and community for everyone! --- ### Source Design Arch Overview (`docs/source/design/arch_overview.md`) (arch-overview)= # Architecture Overview This document provides an overview of the vLLM architecture. ```{contents} Table of Contents :depth: 2 :local: true ``` ## Entrypoints vLLM provides a number of entrypoints for interacting with the system. The following diagram shows the relationship between them. ```{image} /assets/design/arch_overview/entrypoints.excalidraw.png :alt: Entrypoints Diagram ``` ### LLM Class The LLM class provides the primary Python interface for doing offline inference, which is interacting with a model without using a separate model inference server. Here is a sample of `LLM` class usage: ```python from vllm import LLM, SamplingParams # Define a list of input prompts prompts = [ "Hello, my name is", "The capital of France is", "The largest ocean is", ] # Define sampling parameters sampling_params = SamplingParams(temperature=0.8, top_p=0.95) # Initialize the LLM engine with the OPT-125M model llm = LLM(model="facebook/opt-125m") # Generate outputs for the input prompts outputs = llm.generate(prompts, sampling_params) # Print the generated outputs for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` More API details can be found in the {doc}`Offline Inference ` section of the API docs. The code for the `LLM` class can be found in . ### OpenAI-Compatible API Server The second primary interface to vLLM is via its OpenAI-compatible API server. This server can be started using the `vllm serve` command. ```bash vllm serve ``` The code for the `vllm` CLI can be found in . Sometimes you may see the API server entrypoint used directly instead of via the `vllm` CLI command. For example: ```bash python -m vllm.entrypoints.openai.api_server --model ``` That code can be found in . More details on the API server can be found in the [OpenAI-Compatible Server](#openai-compatible-server) document. ## LLM Engine The `LLMEngine` and `AsyncLLMEngine` classes are central to the functioning of the vLLM system, handling model inference and asynchronous request processing. ```{image} /assets/design/arch_overview/llm_engine.excalidraw.png :alt: LLMEngine Diagram ``` ### LLMEngine The `LLMEngine` class is the core component of the vLLM engine. It is responsible for receiving requests from clients and generating outputs from the model. The `LLMEngine` includes input processing, model execution (possibly distributed across multiple hosts and/or GPUs), scheduling, and output processing. - **Input Processing**: Handles tokenization of input text using the specified tokenizer. - **Scheduling**: Chooses which requests are processed in each step. - **Model Execution**: Manages the execution of the language model, including distributed execution across multiple GPUs. - **Output Processing**: Processes the outputs generated by the model, decoding the token IDs from a language model into human-readable text. The code for `LLMEngine` can be found in . ### AsyncLLMEngine The `AsyncLLMEngine` class is an asynchronous wrapper for the `LLMEngine` class. It uses `asyncio` to create a background loop that continuously processes incoming requests. The `AsyncLLMEngine` is designed for online serving, where it can handle multiple concurrent requests and stream outputs to clients. The OpenAI-compatible API server uses the `AsyncLLMEngine`. There is also a demo API server that serves as a simpler example in . The code for `AsyncLLMEngine` can be found in . ## Worker A worker is a process that runs the model inference. vLLM follows the common practice of using one process to control one accelerator device, such as GPUs. For example, if we use tensor parallelism of size 2 and pipeline parallelism of size 2, we will have 4 workers in total. Workers are identified by their `rank` and `local_rank`. `rank` is used for global orchestration, while `local_rank` is mainly used for assigning the accelerator device and accessing local resources such as the file system and shared memory. ## Model Runner Every worker has one model runner object, responsible for loading and running the model. Much of the model execution logic resides here, such as preparing input tensors and capturing cudagraphs. ## Model Every model runner object has one model object, which is the actual `torch.nn.Module` instance. See [huggingface_integration](#huggingface-integration) for how various configurations affect the class we ultimately get. ## Class Hierarchy The following figure shows the class hierarchy of vLLM: > ```{figure} /assets/design/hierarchy.png > :align: center > :alt: query > :width: 100% > ``` There are several important design choices behind this class hierarchy: 1\. **Extensibility**: All classes in the hierarchy accept a configuration object containing all the necessary information. The [VllmConfig](https://github.com/vllm-project/vllm/blob/d1c6799b8870e513bf4f2305cbf6cda9fc3d773b/vllm/config.py#L2036) class is the main configuration object that is passed around. The class hierarchy is quite deep, and every class needs to read the configuration it is interested in. By encapsulating all configurations in one object, we can easily pass the configuration object around and access the configuration we need. Suppose we want to add a new feature (this is often the case given how fast the field of LLM inference is evolving) that only touches the model runner. We will have to add a new configuration option in the `VllmConfig` class. Since we pass the whole config object around, we only need to add the configuration option to the `VllmConfig` class, and the model runner can access it directly. We don't need to change the constructor of the engine, worker, or model class to pass the new configuration option. 2\. **Uniformity**: The model runner needs a unified interface to create and initialize the model. vLLM supports more than 50 types of popular open-source models. Each model has its own initialization logic. If the constructor signature varies with models, the model runner does not know how to call the constructor accordingly, without complicated and error-prone inspection logic. By making the constructor of the model class uniform, the model runner can easily create and initialize the model without knowing the specific model type. This is also useful for composing models. Vision-language models often consist of a vision model and a language model. By making the constructor uniform, we can easily create a vision model and a language model and compose them into a vision-language model. ````{note} To support this change, all vLLM models' signatures have been updated to: ```python def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): ``` To avoid accidentally passing incorrect arguments, the constructor is now keyword-only. This ensures that the constructor will raise an error if old configurations are passed. vLLM developers have already made this change for all models within vLLM. For out-of-tree registered models, developers need to update their models, for example by adding shim code to adapt the old constructor signature to the new one: ```python class MyOldModel(nn.Module): def __init__( self, config, cache_config: Optional[CacheConfig] = None, quant_config: Optional[QuantizationConfig] = None, lora_config: Optional[LoRAConfig] = None, prefix: str = "", ) -> None: ... from vllm.config import VllmConfig class MyNewModel(MyOldModel): def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): config = vllm_config.model_config.hf_config cache_config = vllm_config.cache_config quant_config = vllm_config.quant_config lora_config = vllm_config.lora_config super().__init__(config, cache_config, quant_config, lora_config, prefix) if __version__ >= "0.6.4": MyModel = MyNewModel else: MyModel = MyOldModel ``` This way, the model can work with both old and new versions of vLLM. ```` 3\. **Sharding and Quantization at Initialization**: Certain features require changing the model weights. For example, tensor parallelism needs to shard the model weights, and quantization needs to quantize the model weights. There are two possible ways to implement this feature. One way is to change the model weights after the model is initialized. The other way is to change the model weights during the model initialization. vLLM chooses the latter. The first approach is not scalable to large models. Suppose we want to run a 405B model (with roughly 810GB weights) with 16 H100 80GB GPUs. Ideally, every GPU should only load 50GB weights. If we change the model weights after the model is initialized, we need to load the full 810GB weights to every GPU and then shard the weights, leading to a huge memory overhead. Instead, if we shard the weights during the model initialization, every layer will only create a shard of the weights it needs, leading to a much smaller memory overhead. The same idea applies to quantization. Note that we also add an additional argument `prefix` to the model's constructor so that the model can initialize itself differently based on the prefix. This is useful for non-uniform quantization, where different parts of the model are quantized differently. The `prefix` is usually an empty string for the top-level model and a string like `"vision"` or `"language"` for the sub-models. In general, it matches the name of the module's state dict in the checkpoint file. One disadvantage of this design is that it is hard to write unit tests for individual components in vLLM because every component needs to be initialized by a complete config object. We solve this problem by providing a default initialization function that creates a default config object with all fields set to `None`. If the component we want to test only cares about a few fields in the config object, we can create a default config object and set the fields we care about. This way, we can test the component in isolation. Note that many tests in vLLM are end-to-end tests that test the whole system, so this is not a big problem. In summary, the complete config object `VllmConfig` can be treated as an engine-level global state that is shared among all vLLM classes. --- ### Source Getting Started Quickstart (`docs/source/getting_started/quickstart.md`) (quickstart)= # Quickstart This guide will help you quickly get started with vLLM to perform: - [Offline batched inference](#quickstart-offline) - [Online serving using OpenAI-compatible server](#quickstart-online) ## Prerequisites - OS: Linux - Python: 3.9 -- 3.12 ## Installation If you are using NVIDIA GPUs, you can install vLLM using [pip](https://pypi.org/project/vllm/) directly. It's recommended to use [uv](https://docs.astral.sh/uv/), a very fast Python environment manager, to create and manage Python environments. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment and install vLLM using the following commands: ```console uv venv myenv --python 3.12 --seed source myenv/bin/activate uv pip install vllm ``` You can also use [conda](https://docs.conda.io/projects/conda/en/latest/user-guide/getting-started.html) to create and manage Python environments. ```console conda create -n myenv python=3.12 -y conda activate myenv pip install vllm ``` ```{note} For non-CUDA platforms, please refer [here](#installation-index) for specific instructions on how to install vLLM. ``` (quickstart-offline)= ## Offline Batched Inference With vLLM installed, you can start generating texts for list of input prompts (i.e. offline batch inferencing). See the example script: The first line of this example imports the classes {class}`~vllm.LLM` and {class}`~vllm.SamplingParams`: - {class}`~vllm.LLM` is the main class for running offline inference with vLLM engine. - {class}`~vllm.SamplingParams` specifies the parameters for the sampling process. ```python from vllm import LLM, SamplingParams ``` The next section defines a list of input prompts and sampling parameters for text generation. The [sampling temperature](https://arxiv.org/html/2402.05201v1) is set to `0.8` and the [nucleus sampling probability](https://en.wikipedia.org/wiki/Top-p_sampling) is set to `0.95`. You can find more information about the sampling parameters [here](#sampling-params). ```python prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) ``` The {class}`~vllm.LLM` class initializes vLLM's engine and the [OPT-125M model](https://arxiv.org/abs/2205.01068) for offline inference. The list of supported models can be found [here](#supported-models). ```python llm = LLM(model="facebook/opt-125m") ``` ```{note} By default, vLLM downloads models from [HuggingFace](https://huggingface.co/). If you would like to use models from [ModelScope](https://www.modelscope.cn), set the environment variable `VLLM_USE_MODELSCOPE` before initializing the engine. ``` Now, the fun part! The outputs are generated using `llm.generate`. It adds the input prompts to the vLLM engine's waiting queue and executes the vLLM engine to generate the outputs with high throughput. The outputs are returned as a list of `RequestOutput` objects, which include all of the output tokens. ```python outputs = llm.generate(prompts, sampling_params) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` (quickstart-online)= ## OpenAI-Compatible Server vLLM can be deployed as a server that implements the OpenAI API protocol. This allows vLLM to be used as a drop-in replacement for applications using OpenAI API. By default, it starts the server at `http://localhost:8000`. You can specify the address with `--host` and `--port` arguments. The server currently hosts one model at a time and implements endpoints such as [list models](https://platform.openai.com/docs/api-reference/models/list), [create chat completion](https://platform.openai.com/docs/api-reference/chat/completions/create), and [create completion](https://platform.openai.com/docs/api-reference/completions/create) endpoints. Run the following command to start the vLLM server with the [Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) model: ```console vllm serve Qwen/Qwen2.5-1.5B-Instruct ``` ```{note} By default, the server uses a predefined chat template stored in the tokenizer. You can learn about overriding it [here](#chat-template). ``` This server can be queried in the same format as OpenAI API. For example, to list the models: ```console curl http://localhost:8000/v1/models ``` You can pass in the argument `--api-key` or environment variable `VLLM_API_KEY` to enable the server to check for API key in the header. ### OpenAI Completions API with vLLM Once your server is started, you can query the model with input prompts: ```console curl http://localhost:8000/v1/completions \ -H "Content-Type: application/json" \ -d '{ "model": "Qwen/Qwen2.5-1.5B-Instruct", "prompt": "San Francisco is a", "max_tokens": 7, "temperature": 0 }' ``` Since this server is compatible with OpenAI API, you can use it as a drop-in replacement for any applications using OpenAI API. For example, another way to query the server is via the `openai` Python package: ```python from openai import OpenAI # Modify OpenAI's API key and API base to use vLLM's API server. openai_api_key = "EMPTY" openai_api_base = "http://localhost:8000/v1" client = OpenAI( api_key=openai_api_key, base_url=openai_api_base, ) completion = client.completions.create(model="Qwen/Qwen2.5-1.5B-Instruct", prompt="San Francisco is a") print("Completion result:", completion) ``` A more detailed client example can be found here: ### OpenAI Chat Completions API with vLLM vLLM is designed to also support the OpenAI Chat Completions API. The chat interface is a more dynamic, interactive way to communicate with the model, allowing back-and-forth exchanges that can be stored in the chat history. This is useful for tasks that require context or more detailed explanations. You can use the [create chat completion](https://platform.openai.com/docs/api-reference/chat/completions/create) endpoint to interact with the model: ```console curl http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "Qwen/Qwen2.5-1.5B-Instruct", "messages": [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Who won the world series in 2020?"} ] }' ``` Alternatively, you can use the `openai` Python package: ```python from openai import OpenAI # Set OpenAI's API key and API base to use vLLM's API server. openai_api_key = "EMPTY" openai_api_base = "http://localhost:8000/v1" client = OpenAI( api_key=openai_api_key, base_url=openai_api_base, ) chat_response = client.chat.completions.create( model="Qwen/Qwen2.5-1.5B-Instruct", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Tell me a joke."}, ] ) print("Chat response:", chat_response) ``` --- ### Source Api Engine Index (`docs/source/api/engine/index.md`) # vLLM Engine ```{eval-rst} .. automodule:: vllm.engine ``` ```{eval-rst} .. currentmodule:: vllm.engine ``` ```{toctree} :caption: Engines :maxdepth: 2 llm_engine async_llm_engine ``` --- ### Source Api Model Index (`docs/source/api/model/index.md`) # Model Development ## Submodules ```{toctree} :maxdepth: 1 interfaces_base interfaces adapters ``` --- ### Source Api Multimodal Index (`docs/source/api/multimodal/index.md`) (multi-modality)= # Multi-Modality vLLM provides experimental support for multi-modal models through the {mod}`vllm.multimodal` package. Multi-modal inputs can be passed alongside text and token prompts to [supported models](#supported-mm-models) via the `multi_modal_data` field in {class}`vllm.inputs.PromptType`. Looking to add your own multi-modal model? Please follow the instructions listed [here](#supports-multimodal). ## Module Contents ```{eval-rst} .. autodata:: vllm.multimodal.MULTIMODAL_REGISTRY ``` ## Submodules ```{toctree} :maxdepth: 1 inputs parse processing profiling registry ``` --- ### Source Api Offline Inference Index (`docs/source/api/offline_inference/index.md`) # Offline Inference ```{toctree} :caption: Contents :maxdepth: 1 llm llm_inputs ``` --- ### Source Contributing Model Index (`docs/source/contributing/model/index.md`) (new-model)= # Adding a New Model This section provides more information on how to integrate a [PyTorch](https://pytorch.org/) model into vLLM. ```{toctree} :caption: Contents :maxdepth: 1 basic registration tests multimodal ``` ```{note} The complexity of adding a new model depends heavily on the model's architecture. The process is considerably straightforward if the model shares a similar architecture with an existing model in vLLM. However, for models that include new operators (e.g., a new attention mechanism), the process can be a bit more complex. ``` ```{tip} If you are encountering issues while integrating your model into vLLM, feel free to open a [GitHub issue](https://github.com/vllm-project/vllm/issues) or ask on our [developer slack](https://slack.vllm.ai). We will be happy to help you out! ``` --- ### Source Contributing Profiling Profiling Index (`docs/source/contributing/profiling/profiling_index.md`) # Profiling vLLM We support tracing vLLM workers using the `torch.profiler` module. You can enable tracing by setting the `VLLM_TORCH_PROFILER_DIR` environment variable to the directory where you want to save the traces: `VLLM_TORCH_PROFILER_DIR=/mnt/traces/` The OpenAI server also needs to be started with the `VLLM_TORCH_PROFILER_DIR` environment variable set. When using `benchmarks/benchmark_serving.py`, you can enable profiling by passing the `--profile` flag. ```{warning} Only enable profiling in a development environment. ``` Traces can be visualized using . ```{tip} Only send a few requests through vLLM when profiling, as the traces can get quite large. Also, no need to untar the traces, they can be viewed directly. ``` ```{tip} To stop the profiler - it flushes out all the profile trace files to the directory. This takes time, for example for about 100 requests worth of data for a llama 70b, it takes about 10 minutes to flush out on a H100. Set the env variable VLLM_RPC_TIMEOUT to a big number before you start the server. Say something like 30 minutes. `export VLLM_RPC_TIMEOUT=1800000` ``` ## Example commands and usage ### Offline Inference Refer to for an example. ### OpenAI Server ```bash VLLM_TORCH_PROFILER_DIR=./vllm_profile python -m vllm.entrypoints.openai.api_server --model meta-llama/Meta-Llama-3-70B ``` benchmark_serving.py: ```bash python benchmarks/benchmark_serving.py --backend vllm --model meta-llama/Meta-Llama-3-70B --dataset-name sharegpt --dataset-path sharegpt.json --profile --num-prompts 2 ``` --- ### Source Deployment Frameworks Index (`docs/source/deployment/frameworks/index.md`) # Using other frameworks ```{toctree} :maxdepth: 1 bentoml cerebrium dstack helm lws modal skypilot triton ``` --- ### Source Deployment Integrations Index (`docs/source/deployment/integrations/index.md`) # External Integrations ```{toctree} :maxdepth: 1 kserve kubeai llamastack ``` --- ### Source Features Quantization Index (`docs/source/features/quantization/index.md`) (quantization-index)= # Quantization Quantization trades off model precision for smaller memory footprint, allowing large models to be run on a wider range of devices. ```{toctree} :caption: Contents :maxdepth: 1 supported_hardware auto_awq bnb gguf int8 fp8 fp8_e5m2_kvcache fp8_e4m3_kvcache ``` --- ### Source Getting Started Installation Index (`docs/source/getting_started/installation/index.md`) (installation-index)= # Installation vLLM supports the following hardware platforms: ```{toctree} :maxdepth: 1 gpu/index cpu/index ai_accelerator/index ``` --- ### Source Models Extensions Index (`docs/source/models/extensions/index.md`) # Built-in Extensions ```{toctree} :maxdepth: 1 runai_model_streamer tensorizer ``` --- ### Source Serving Integrations Index (`docs/source/serving/integrations/index.md`) # External Integrations ```{toctree} :maxdepth: 1 langchain llamaindex ``` --- ### Source Serving Integrations Llamaindex (`docs/source/serving/integrations/llamaindex.md`) (serving-llamaindex)= # LlamaIndex vLLM is also available via [LlamaIndex](https://github.com/run-llama/llama_index) . To install LlamaIndex, run ```console pip install llama-index-llms-vllm -q ``` To run inference on a single or multiple GPUs, use `Vllm` class from `llamaindex`. ```python from llama_index.llms.vllm import Vllm llm = Vllm( model="microsoft/Orca-2-7b", tensor_parallel_size=4, max_new_tokens=100, vllm_kwargs={"swap_space": 1, "gpu_memory_utilization": 0.5}, ) ``` Please refer to this [Tutorial](https://docs.llamaindex.ai/en/latest/examples/llm/vllm/) for more details. --- ### Source Getting Started Installation Ai Accelerator Index (`docs/source/getting_started/installation/ai_accelerator/index.md`) # Other AI accelerators vLLM is a Python library that supports the following AI accelerators. Select your AI accelerator type to see vendor specific instructions: ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} openvino.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::: ## Requirements ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "## Requirements" :end-before: "## Configure a new environment" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "## Requirements" :end-before: "## Configure a new environment" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "## Requirements" :end-before: "## Configure a new environment" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} openvino.inc.md :start-after: "## Requirements" :end-before: "## Set up using Python" ``` ::: :::: ## Configure a new environment ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "## Configure a new environment" :end-before: "## Set up using Python" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "## Configure a new environment" :end-before: "## Set up using Python" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "## Configure a new environment" :end-before: "## Set up using Python" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} ../python_env_setup.inc.md ``` ::: :::: ## Set up using Python ### Pre-built wheels ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "### Pre-built wheels" :end-before: "### Build wheel from source" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "### Pre-built wheels" :end-before: "### Build wheel from source" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "### Pre-built wheels" :end-before: "### Build wheel from source" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} openvino.inc.md :start-after: "### Pre-built wheels" :end-before: "### Build wheel from source" ``` ::: :::: ### Build wheel from source ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} openvino.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::: ## Set up using Docker ### Pre-built images ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "### Pre-built images" :end-before: "### Build image from source" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "### Pre-built images" :end-before: "### Build image from source" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "### Pre-built images" :end-before: "### Build image from source" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} openvino.inc.md :start-after: "### Pre-built images" :end-before: "### Build image from source" ``` ::: :::: ### Build image from source ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "### Build image from source" :end-before: "## Extra information" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "### Build image from source" :end-before: "## Extra information" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "### Build image from source" :end-before: "## Extra information" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} openvino.inc.md :start-after: "### Build image from source" :end-before: "## Extra information" ``` ::: :::: ## Extra information ::::{tab-set} :sync-group: device :::{tab-item} TPU :sync: tpu ```{include} tpu.inc.md :start-after: "## Extra information" ``` ::: :::{tab-item} Intel Gaudi :sync: hpu-gaudi ```{include} hpu-gaudi.inc.md :start-after: "## Extra information" ``` ::: :::{tab-item} Neuron :sync: neuron ```{include} neuron.inc.md :start-after: "## Extra information" ``` ::: :::{tab-item} OpenVINO :sync: openvino ```{include} openvino.inc.md :start-after: "## Extra information" ``` ::: :::: --- ### Source Getting Started Installation Cpu Index (`docs/source/getting_started/installation/cpu/index.md`) # CPU vLLM is a Python library that supports the following CPU variants. Select your CPU type to see vendor specific instructions: ::::{tab-set} :sync-group: device :::{tab-item} x86 :sync: x86 ```{include} x86.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::{tab-item} ARM :sync: arm ```{include} arm.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::{tab-item} Apple silicon :sync: apple ```{include} apple.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::: ## Requirements - Python: 3.9 -- 3.12 ::::{tab-set} :sync-group: device :::{tab-item} x86 :sync: x86 ```{include} x86.inc.md :start-after: "## Requirements" :end-before: "## Set up using Python" ``` ::: :::{tab-item} ARM :sync: arm ```{include} arm.inc.md :start-after: "## Requirements" :end-before: "## Set up using Python" ``` ::: :::{tab-item} Apple silicon :sync: apple ```{include} apple.inc.md :start-after: "## Requirements" :end-before: "## Set up using Python" ``` ::: :::: ## Set up using Python ### Create a new Python environment ```{include} ../python_env_setup.inc.md ``` ### Pre-built wheels Currently, there are no pre-built CPU wheels. ### Build wheel from source ::::{tab-set} :sync-group: device :::{tab-item} x86 :sync: x86 ```{include} x86.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::{tab-item} ARM :sync: arm ```{include} arm.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::{tab-item} Apple silicon :sync: apple ```{include} apple.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::: ## Set up using Docker ### Pre-built images Currently, there are no pre-build CPU images. ### Build image from source ```console $ docker build -f Dockerfile.cpu -t vllm-cpu-env --shm-size=4g . $ docker run -it \ --rm \ --network=host \ --cpuset-cpus= \ --cpuset-mems= \ vllm-cpu-env ``` :::{tip} For ARM or Apple silicon, use `Dockerfile.arm` ::: ## Supported features vLLM CPU backend supports the following vLLM features: - Tensor Parallel - Model Quantization (`INT8 W8A8, AWQ, GPTQ`) - Chunked-prefill - Prefix-caching - FP8-E5M2 KV-Caching (TODO) ## Related runtime environment variables - `VLLM_CPU_KVCACHE_SPACE`: specify the KV Cache size (e.g, `VLLM_CPU_KVCACHE_SPACE=40` means 40 GB space for KV cache), larger setting will allow vLLM running more requests in parallel. This parameter should be set based on the hardware configuration and memory management pattern of users. - `VLLM_CPU_OMP_THREADS_BIND`: specify the CPU cores dedicated to the OpenMP threads. For example, `VLLM_CPU_OMP_THREADS_BIND=0-31` means there will be 32 OpenMP threads bound on 0-31 CPU cores. `VLLM_CPU_OMP_THREADS_BIND=0-31|32-63` means there will be 2 tensor parallel processes, 32 OpenMP threads of rank0 are bound on 0-31 CPU cores, and the OpenMP threads of rank1 are bound on 32-63 CPU cores. ## Performance tips - We highly recommend to use TCMalloc for high performance memory allocation and better cache locality. For example, on Ubuntu 22.4, you can run: ```console sudo apt-get install libtcmalloc-minimal4 # install TCMalloc library find / -name *libtcmalloc* # find the dynamic link library path export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD # prepend the library to LD_PRELOAD python examples/offline_inference/basic.py # run vLLM ``` - When using the online serving, it is recommended to reserve 1-2 CPU cores for the serving framework to avoid CPU oversubscription. For example, on a platform with 32 physical CPU cores, reserving CPU 30 and 31 for the framework and using CPU 0-29 for OpenMP: ```console export VLLM_CPU_KVCACHE_SPACE=40 export VLLM_CPU_OMP_THREADS_BIND=0-29 vllm serve facebook/opt-125m ``` - If using vLLM CPU backend on a machine with hyper-threading, it is recommended to bind only one OpenMP thread on each physical CPU core using `VLLM_CPU_OMP_THREADS_BIND`. On a hyper-threading enabled platform with 16 logical CPU cores / 8 physical CPU cores: ``` /* Detailed source-code truncated for AI context efficiency. */ ``` - If using vLLM CPU backend on a multi-socket machine with NUMA, be aware to set CPU cores using `VLLM_CPU_OMP_THREADS_BIND` to avoid cross NUMA node memory access. ## Other considerations - The CPU backend significantly differs from the GPU backend since the vLLM architecture was originally optimized for GPU use. A number of optimizations are needed to enhance its performance. - Decouple the HTTP serving components from the inference components. In a GPU backend configuration, the HTTP serving and tokenization tasks operate on the CPU, while inference runs on the GPU, which typically does not pose a problem. However, in a CPU-based setup, the HTTP serving and tokenization can cause significant context switching and reduced cache efficiency. Therefore, it is strongly recommended to segregate these two components for improved performance. - On CPU based setup with NUMA enabled, the memory access performance may be largely impacted by the [topology](https://github.com/intel/intel-extension-for-pytorch/blob/main/docs/tutorials/performance_tuning/tuning_guide.inc.md#non-uniform-memory-access-numa). For NUMA architecture, two optimizations are to recommended: Tensor Parallel or Data Parallel. - Using Tensor Parallel for a latency constraints deployment: following GPU backend design, a Megatron-LM's parallel algorithm will be used to shard the model, based on the number of NUMA nodes (e.g. TP = 2 for a two NUMA node system). With [TP feature on CPU](gh-pr:6125) merged, Tensor Parallel is supported for serving and offline inferencing. In general each NUMA node is treated as one GPU card. Below is the example script to enable Tensor Parallel = 2 for serving: ```console VLLM_CPU_KVCACHE_SPACE=40 VLLM_CPU_OMP_THREADS_BIND="0-31|32-63" vllm serve meta-llama/Llama-2-7b-chat-hf -tp=2 --distributed-executor-backend mp ``` - Using Data Parallel for maximum throughput: to launch an LLM serving endpoint on each NUMA node along with one additional load balancer to dispatch the requests to those endpoints. Common solutions like [Nginx](#nginxloadbalancer) or HAProxy are recommended. Anyscale Ray project provides the feature on LLM [serving](https://docs.ray.io/en/latest/serve/index.html). Here is the example to setup a scalable LLM serving with [Ray Serve](https://github.com/intel/llm-on-ray/blob/main/docs/setup.inc.md). --- ### Source Getting Started Installation Gpu Index (`docs/source/getting_started/installation/gpu/index.md`) # GPU vLLM is a Python library that supports the following GPU variants. Select your GPU type to see vendor specific instructions: ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::{tab-item} ROCm :sync: rocm ```{include} rocm.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::{tab-item} XPU :sync: xpu ```{include} xpu.inc.md :start-after: "# Installation" :end-before: "## Requirements" ``` ::: :::: ## Requirements - OS: Linux - Python: 3.9 -- 3.12 ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "## Requirements" :end-before: "## Set up using Python" ``` ::: :::{tab-item} ROCm :sync: rocm ```{include} rocm.inc.md :start-after: "## Requirements" :end-before: "## Set up using Python" ``` ::: :::{tab-item} XPU :sync: xpu ```{include} xpu.inc.md :start-after: "## Requirements" :end-before: "## Set up using Python" ``` ::: :::: ## Set up using Python ### Create a new Python environment ```{include} ../python_env_setup.inc.md ``` ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "## Create a new Python environment" :end-before: "### Pre-built wheels" ``` ::: :::{tab-item} ROCm :sync: rocm There is no extra information on creating a new Python environment for this device. ::: :::{tab-item} XPU :sync: xpu There is no extra information on creating a new Python environment for this device. ::: :::: ### Pre-built wheels ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "### Pre-built wheels" :end-before: "### Build wheel from source" ``` ::: :::{tab-item} ROCm :sync: rocm ```{include} rocm.inc.md :start-after: "### Pre-built wheels" :end-before: "### Build wheel from source" ``` ::: :::{tab-item} XPU :sync: xpu ```{include} xpu.inc.md :start-after: "### Pre-built wheels" :end-before: "### Build wheel from source" ``` ::: :::: (build-from-source)= ### Build wheel from source ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::{tab-item} ROCm :sync: rocm ```{include} rocm.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::{tab-item} XPU :sync: xpu ```{include} xpu.inc.md :start-after: "### Build wheel from source" :end-before: "## Set up using Docker" ``` ::: :::: ## Set up using Docker ### Pre-built images ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "### Pre-built images" :end-before: "### Build image from source" ``` ::: :::{tab-item} ROCm :sync: rocm ```{include} rocm.inc.md :start-after: "### Pre-built images" :end-before: "### Build image from source" ``` ::: :::{tab-item} XPU :sync: xpu ```{include} xpu.inc.md :start-after: "### Pre-built images" :end-before: "### Build image from source" ``` ::: :::: ### Build image from source ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "### Build image from source" :end-before: "## Supported features" ``` ::: :::{tab-item} ROCm :sync: rocm ```{include} rocm.inc.md :start-after: "### Build image from source" :end-before: "## Supported features" ``` ::: :::{tab-item} XPU :sync: xpu ```{include} xpu.inc.md :start-after: "### Build image from source" :end-before: "## Supported features" ``` ::: :::: ## Supported features ::::{tab-set} :sync-group: device :::{tab-item} CUDA :sync: cuda ```{include} cuda.inc.md :start-after: "## Supported features" ``` ::: :::{tab-item} ROCm :sync: rocm ```{include} rocm.inc.md :start-after: "## Supported features" ``` ::: :::{tab-item} XPU :sync: xpu ```{include} xpu.inc.md :start-after: "## Supported features" ``` ::: :::: --- ### Requirements Docs (`docs/requirements-docs.txt`) sphinx==6.2.1 sphinx-book-theme==1.0.1 sphinx-copybutton==0.5.2 myst-parser==3.0.1 sphinx-argparse==0.4.0 sphinx-design==0.6.1 sphinx-togglebutton==0.3.2 msgspec cloudpickle # packages to install to build the documentation pydantic >= 2.8 -f https://download.pytorch.org/whl/cpu torch py-cpuinfo transformers mistral_common >= 1.5.0 aiohttp starlette openai # Required by docs/source/serving/openai_compatible_server.md's vllm.entrypoints.openai.cli_args fastapi # Required by docs/source/serving/openai_compatible_server.md's vllm.entrypoints.openai.cli_args partial-json-parser # Required by docs/source/serving/openai_compatible_server.md's vllm.entrypoints.openai.cli_args requests zmq --- ### Source Api Inference Params (`docs/source/api/inference_params.md`) # Inference Parameters Inference parameters for vLLM APIs. (sampling-params)= ## Sampling Parameters ```{eval-rst} .. autoclass:: vllm.SamplingParams :members: ``` (pooling-params)= ## Pooling Parameters ```{eval-rst} .. autoclass:: vllm.PoolingParams :members: ``` --- ### Source Community Meetups (`docs/source/community/meetups.md`) (meetups)= # vLLM Meetups We host regular meetups in San Francisco Bay Area every 2 months. We will share the project updates from the vLLM team and have guest speakers from the industry to share their experience and insights. Please find the materials of our previous meetups below: - [The seventh vLLM meetup](https://lu.ma/h0qvrajz), with Snowflake, November 14th 2024. [[Slides]](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing) - [The sixth vLLM meetup](https://lu.ma/87q3nvnh), with NVIDIA, September 9th 2024. [[Slides]](https://docs.google.com/presentation/d/1wrLGwytQfaOTd5wCGSPNhoaW3nq0E-9wqyP7ny93xRs/edit?usp=sharing) - [The fifth vLLM meetup](https://lu.ma/lp0gyjqr), with AWS, July 24th 2024. [[Slides]](https://docs.google.com/presentation/d/1RgUD8aCfcHocghoP3zmXzck9vX3RCI9yfUAB2Bbcl4Y/edit?usp=sharing) - [The fourth vLLM meetup](https://lu.ma/agivllm), with Cloudflare and BentoML, June 11th 2024. [[Slides]](https://docs.google.com/presentation/d/1iJ8o7V2bQEi0BFEljLTwc5G1S10_Rhv3beed5oB0NJ4/edit?usp=sharing) - [The third vLLM meetup](https://robloxandvllmmeetup2024.splashthat.com/), with Roblox, April 2nd 2024. [[Slides]](https://docs.google.com/presentation/d/1A--47JAK4BJ39t954HyTkvtfwn0fkqtsL8NGFuslReM/edit?usp=sharing) - [The second vLLM meetup](https://lu.ma/ygxbpzhl), with IBM Research, January 31st 2024. [[Slides]](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing) [[Video (vLLM Update)]](https://youtu.be/Y0C-DUvEnZQ) [[Video (IBM Research & torch.compile)]](https://youtu.be/m0dMtFLI-dg) - [The first vLLM meetup](https://lu.ma/first-vllm-meetup), with a16z, October 5th 2023. [[Slides]](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing) We are always looking for speakers and sponsors at San Francisco Bay Area and potentially other locations. If you are interested in speaking or sponsoring, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@lists.berkeley.edu). --- ### Source Community Sponsors (`docs/source/community/sponsors.md`) # Sponsors vLLM is a community project. Our compute resources for development and testing are supported by the following organizations. Thank you for your support! Cash Donations: - a16z - Dropbox - Sequoia Capital - Skywork AI - ZhenFund Compute Resources: - AMD - Anyscale - AWS - Crusoe Cloud - Databricks - DeepInfra - Google Cloud - Lambda Lab - Nebius - Novita AI - NVIDIA - Replicate - Roblox - RunPod - Trainy - UC Berkeley - UC San Diego Slack Sponsor: Anyscale We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). We plan to use the fund to support the development, maintenance, and adoption of vLLM. --- ### Source Contributing Vulnerability Management (`docs/source/contributing/vulnerability_management.md`) # Vulnerability Management ## Reporting Vulnerabilities As mentioned in the [security policy](https://github.com/vllm-project/vllm/tree/main/SECURITY.md), security vulnerabilities may be reported privately to the project via [GitHub](https://github.com/vllm-project/vllm/security/advisories/new). ## Vulnerability Management Team Once a vulnerability has been reported to the project, the Vulnerability Management Team (VMT) is responsible for managing the vulnerability. The VMT is responsible for: - Triaging the vulnerability. - Coordinating with reporters and project maintainers on vulnerability analysis and resolution. - Drafting of security advisories for confirmed vulnerabilities, as appropriate. - Coordination with project maintainers on a coordinated release of the fix and security advisory. ### Security Advisories Advisories are published via GitHub through the same system used to report vulnerabilities. More information on the process can be found in the [GitHub documentation](https://docs.github.com/en/code-security/security-advisories/working-with-repository-security-advisories/about-repository-security-advisories). ### Team Members We prefer to keep all vulnerability-related communication on the security report on GitHub. However, if you need to contact the VMT directly for an urgent issue, you may contact the following individuals: - Simon Mo - simon.mo@hey.com - Russell Bryant - rbryant@redhat.com ## Slack Discussion You may use the `#security` channel in the [VLLM Slack](https://slack.vllm.ai) to discuss security-related topics. However, please do not disclose any vulnerabilities in this channel. If you need to report a vulnerability, please use the GitHub security advisory system or contact a VMT member privately. --- ### Source Deployment Docker (`docs/source/deployment/docker.md`) (deployment-docker)= # Using Docker (deployment-docker-pre-built-image)= ## Use vLLM's Official Docker Image vLLM offers an official Docker image for deployment. The image can be used to run OpenAI compatible server and is available on Docker Hub as [vllm/vllm-openai](https://hub.docker.com/r/vllm/vllm-openai/tags). ```console $ docker run --runtime nvidia --gpus all \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HUGGING_FACE_HUB_TOKEN=" \ -p 8000:8000 \ --ipc=host \ vllm/vllm-openai:latest \ --model mistralai/Mistral-7B-v0.1 ``` You can add any other you need after the image tag (`vllm/vllm-openai:latest`). ```{note} You can either use the `ipc=host` flag or `--shm-size` flag to allow the container to access the host's shared memory. vLLM uses PyTorch, which uses shared memory to share data between processes under the hood, particularly for tensor parallel inference. ``` (deployment-docker-build-image-from-source)= ## Building vLLM's Docker Image from Source You can build and run vLLM from source via the provided . To build vLLM: ```console # optionally specifies: --build-arg max_jobs=8 --build-arg nvcc_threads=2 DOCKER_BUILDKIT=1 docker build . --target vllm-openai --tag vllm/vllm-openai ``` ```{note} By default vLLM will build for all GPU types for widest distribution. If you are just building for the current GPU type the machine is running on, you can add the argument `--build-arg torch_cuda_arch_list=""` for vLLM to find the current GPU type and build for that. If you are using Podman instead of Docker, you might need to disable SELinux labeling by adding `--security-opt label=disable` when running `podman build` command to avoid certain [existing issues](https://github.com/containers/buildah/discussions/4184). ``` ## Building for Arm64/aarch64 A docker container can be built for aarch64 systems such as the Nvidia Grace-Hopper. At time of this writing, this requires the use of PyTorch Nightly and should be considered **experimental**. Using the flag `--platform "linux/arm64"` will attempt to build for arm64. ```{note} Multiple modules must be compiled, so this process can take a while. Recommend using `--build-arg max_jobs=` & `--build-arg nvcc_threads=` flags to speed up build process. However, ensure your `max_jobs` is substantially larger than `nvcc_threads` to get the most benefits. Keep an eye on memory usage with parallel jobs as it can be substantial (see example below). ``` ```console # Example of building on Nvidia GH200 server. (Memory usage: ~15GB, Build time: ~1475s / ~25 min, Image size: 6.93GB) $ python3 use_existing_torch.py $ DOCKER_BUILDKIT=1 docker build . \ --target vllm-openai \ --platform "linux/arm64" \ -t vllm/vllm-gh200-openai:latest \ --build-arg max_jobs=66 \ --build-arg nvcc_threads=2 \ --build-arg torch_cuda_arch_list="9.0+PTX" \ --build-arg vllm_fa_cmake_gpu_arches="90-real" ``` ## Use the custom-built vLLM Docker image To run vLLM with the custom-built Docker image: ```console $ docker run --runtime nvidia --gpus all \ -v ~/.cache/huggingface:/root/.cache/huggingface \ -p 8000:8000 \ --env "HUGGING_FACE_HUB_TOKEN=" \ vllm/vllm-openai ``` The argument `vllm/vllm-openai` specifies the image to run, and should be replaced with the name of the custom-built image (the `-t` tag from the build command). ```{note} **For version 0.4.1 and 0.4.2 only** - the vLLM docker images under these versions are supposed to be run under the root user since a library under the root user's home directory, i.e. `/root/.config/vllm/nccl/cu12/libnccl.so.2.18.1` is required to be loaded during runtime. If you are running the container under a different user, you may need to first change the permissions of the library (and all the parent directories) to allow the user to access it, then run vLLM with environment variable `VLLM_NCCL_SO_PATH=/root/.config/vllm/nccl/cu12/libnccl.so.2.18.1` . ``` --- ### Source Deployment K8s (`docs/source/deployment/k8s.md`) (deployment-k8s)= # Using Kubernetes Using Kubernetes to deploy vLLM is a scalable and efficient way to serve machine learning models. This guide will walk you through the process of deploying vLLM with Kubernetes, including the necessary prerequisites, steps for deployment, and testing. ## Prerequisites Before you begin, ensure that you have the following: - A running Kubernetes cluster - NVIDIA Kubernetes Device Plugin (`k8s-device-plugin`): This can be found at `https://github.com/NVIDIA/k8s-device-plugin/` - Available GPU resources in your cluster ## Deployment Steps 1. Create a PVC, Secret and Deployment for vLLM PVC is used to store the model cache and it is optional, you can use hostPath or other storage options ```yaml apiVersion: v1 kind: PersistentVolumeClaim metadata: name: mistral-7b namespace: default spec: accessModes: - ReadWriteOnce resources: requests: storage: 50Gi storageClassName: default volumeMode: Filesystem ``` Secret is optional and only required for accessing gated models, you can skip this step if you are not using gated models ```yaml apiVersion: v1 kind: Secret metadata: name: hf-token-secret namespace: default type: Opaque stringData: token: "REPLACE_WITH_TOKEN" ``` Next to create the deployment file for vLLM to run the model server. The following example deploys the `Mistral-7B-Instruct-v0.3` model. Here are two examples for using NVIDIA GPU and AMD GPU. NVIDIA GPU: ``` /* Detailed source-code truncated for AI context efficiency. */ ``` AMD GPU: You can refer to the `deployment.yaml` below if using AMD ROCm GPU like MI300X. ``` /* Detailed source-code truncated for AI context efficiency. */ ``` You can get the full example with steps and sample yaml files from . 2. Create a Kubernetes Service for vLLM Next, create a Kubernetes Service file to expose the `mistral-7b` deployment: ```yaml apiVersion: v1 kind: Service metadata: name: mistral-7b namespace: default spec: ports: - name: http-mistral-7b port: 80 protocol: TCP targetPort: 8000 # The label selector should match the deployment labels & it is useful for prefix caching feature selector: app: mistral-7b sessionAffinity: None type: ClusterIP ``` 3. Deploy and Test Apply the deployment and service configurations using `kubectl apply -f `: ```console kubectl apply -f deployment.yaml kubectl apply -f service.yaml ``` To test the deployment, run the following `curl` command: ```console curl http://mistral-7b.default.svc.cluster.local/v1/completions \ -H "Content-Type: application/json" \ -d '{ "model": "mistralai/Mistral-7B-Instruct-v0.3", "prompt": "San Francisco is a", "max_tokens": 7, "temperature": 0 }' ``` If the service is correctly deployed, you should receive a response from the vLLM model. ## Conclusion Deploying vLLM with Kubernetes allows for efficient scaling and management of ML models leveraging GPU resources. By following the steps outlined above, you should be able to set up and test a vLLM deployment within your Kubernetes cluster. If you encounter any issues or have suggestions, please feel free to contribute to the documentation. --- ### Source Deployment Nginx (`docs/source/deployment/nginx.md`) (nginxloadbalancer)= # Using Nginx This document shows how to launch multiple vLLM serving containers and use Nginx to act as a load balancer between the servers. Table of contents: 1. [Build Nginx Container](#nginxloadbalancer-nginx-build) 2. [Create Simple Nginx Config file](#nginxloadbalancer-nginx-conf) 3. [Build vLLM Container](#nginxloadbalancer-nginx-vllm-container) 4. [Create Docker Network](#nginxloadbalancer-nginx-docker-network) 5. [Launch vLLM Containers](#nginxloadbalancer-nginx-launch-container) 6. [Launch Nginx](#nginxloadbalancer-nginx-launch-nginx) 7. [Verify That vLLM Servers Are Ready](#nginxloadbalancer-nginx-verify-nginx) (nginxloadbalancer-nginx-build)= ## Build Nginx Container This guide assumes that you have just cloned the vLLM project and you're currently in the vllm root directory. ```console export vllm_root=`pwd` ``` Create a file named `Dockerfile.nginx`: ```console FROM nginx:latest RUN rm /etc/nginx/conf.d/default.conf EXPOSE 80 CMD ["nginx", "-g", "daemon off;"] ``` Build the container: ```console docker build . -f Dockerfile.nginx --tag nginx-lb ``` (nginxloadbalancer-nginx-conf)= ## Create Simple Nginx Config file Create a file named `nginx_conf/nginx.conf`. Note that you can add as many servers as you'd like. In the below example we'll start with two. To add more, add another `server vllmN:8000 max_fails=3 fail_timeout=10000s;` entry to `upstream backend`. ```console upstream backend { least_conn; server vllm0:8000 max_fails=3 fail_timeout=10000s; server vllm1:8000 max_fails=3 fail_timeout=10000s; } server { listen 80; location / { proxy_pass http://backend; proxy_set_header Host $host; proxy_set_header X-Real-IP $remote_addr; proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for; proxy_set_header X-Forwarded-Proto $scheme; } } ``` (nginxloadbalancer-nginx-vllm-container)= ## Build vLLM Container ```console cd $vllm_root docker build -f Dockerfile . --tag vllm ``` If you are behind proxy, you can pass the proxy settings to the docker build command as shown below: ```console cd $vllm_root docker build -f Dockerfile . --tag vllm --build-arg http_proxy=$http_proxy --build-arg https_proxy=$https_proxy ``` (nginxloadbalancer-nginx-docker-network)= ## Create Docker Network ```console docker network create vllm_nginx ``` (nginxloadbalancer-nginx-launch-container)= ## Launch vLLM Containers Notes: - If you have your HuggingFace models cached somewhere else, update `hf_cache_dir` below. - If you don't have an existing HuggingFace cache you will want to start `vllm0` and wait for the model to complete downloading and the server to be ready. This will ensure that `vllm1` can leverage the model you just downloaded and it won't have to be downloaded again. - The below example assumes GPU backend used. If you are using CPU backend, remove `--gpus all`, add `VLLM_CPU_KVCACHE_SPACE` and `VLLM_CPU_OMP_THREADS_BIND` environment variables to the docker run command. - Adjust the model name that you want to use in your vLLM servers if you don't want to use `Llama-2-7b-chat-hf`. ```console mkdir -p ~/.cache/huggingface/hub/ hf_cache_dir=~/.cache/huggingface/ docker run -itd --ipc host --privileged --network vllm_nginx --gpus all --shm-size=10.24gb -v $hf_cache_dir:/root/.cache/huggingface/ -p 8081:8000 --name vllm0 vllm --model meta-llama/Llama-2-7b-chat-hf docker run -itd --ipc host --privileged --network vllm_nginx --gpus all --shm-size=10.24gb -v $hf_cache_dir:/root/.cache/huggingface/ -p 8082:8000 --name vllm1 vllm --model meta-llama/Llama-2-7b-chat-hf ``` ```{note} If you are behind proxy, you can pass the proxy settings to the docker run command via `-e http_proxy=$http_proxy -e https_proxy=$https_proxy`. ``` (nginxloadbalancer-nginx-launch-nginx)= ## Launch Nginx ```console docker run -itd -p 8000:80 --network vllm_nginx -v ./nginx_conf/:/etc/nginx/conf.d/ --name nginx-lb nginx-lb:latest ``` (nginxloadbalancer-nginx-verify-nginx)= ## Verify That vLLM Servers Are Ready ```console docker logs vllm0 | grep Uvicorn docker logs vllm1 | grep Uvicorn ``` Both outputs should look like this: ```console INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit) ``` --- ### Source Design Automatic Prefix Caching (`docs/source/design/automatic_prefix_caching.md`) (design-automatic-prefix-caching)= # Automatic Prefix Caching The core idea of [PagedAttention](https://blog.vllm.ai/2023/06/20/vllm.html) is to partition the KV cache of each request into KV Blocks. Each block contains the attention keys and values for a fixed number of tokens. The PagedAttention algorithm allows these blocks to be stored in non-contiguous physical memory so that we can eliminate memory fragmentation by allocating the memory on demand. To automatically cache the KV cache, we utilize the following key observation: Each KV block can be uniquely identified by the tokens within the block and the tokens in the prefix before the block. ```text Block 1 Block 2 Block 3 [A gentle breeze stirred] [the leaves as children] [laughed in the distance] Block 1: |<--- block tokens ---->| Block 2: |<------- prefix ------>| |<--- block tokens --->| Block 3: |<------------------ prefix -------------------->| |<--- block tokens ---->| ``` In the example above, the KV cache in the first block can be uniquely identified with the tokens “A gentle breeze stirred”. The third block can be uniquely identified with the tokens in the block “laughed in the distance”, along with the prefix tokens “A gentle breeze stirred the leaves as children”. Therefore, we can build the following one-to-one mapping: ```text hash(prefix tokens + block tokens) <--> KV Block ``` With this mapping, we can add another indirection in vLLM’s KV cache management. Previously, each sequence in vLLM maintained a mapping from their logical KV blocks to physical blocks. To achieve automatic caching of KV blocks, we map the logical KV blocks to their hash value and maintain a global hash table of all the physical blocks. In this way, all the KV blocks sharing the same hash value (e.g., shared prefix blocks across two requests) can be mapped to the same physical block and share the memory space. This design achieves automatic prefix caching without the need of maintaining a tree structure among the KV blocks. More specifically, all of the blocks are independent of each other and can be allocated and freed by itself, which enables us to manages the KV cache as ordinary caches in operating system. ## Generalized Caching Policy Keeping all the KV blocks in a hash table enables vLLM to cache KV blocks from earlier requests to save memory and accelerate the computation of future requests. For example, if a new request shares the system prompt with the previous request, the KV cache of the shared prompt can directly be used for the new request without recomputation. However, the total KV cache space is limited and we have to decide which KV blocks to keep or evict when the cache is full. Managing KV cache with a hash table allows us to implement flexible caching policies. As an example, in current vLLM, we implement the following eviction policy: * When there are no free blocks left, we will evict a KV block with reference count (i.e., number of current requests using the block) equals 0. * If there are multiple blocks with reference count equals to 0, we prioritize to evict the least recently used block (LRU). * If there are multiple blocks whose last access time are the same, we prioritize the eviction of the block that is at the end of the longest prefix (i.e., has the maximum number of blocks before it). Note that this eviction policy effectively implements the exact policy as in [RadixAttention](https://lmsys.org/blog/2024-01-17-sglang/) when applied to models with full attention, which prioritizes to evict reference count zero and least recent used leaf nodes in the prefix tree. However, the hash-based KV cache management gives us the flexibility to handle more complicated serving scenarios and implement more complicated eviction policies beyond the policy above: * Multi-LoRA serving. When serving requests for multiple LoRA adapters, we can simply let the hash of each KV block to also include the LoRA ID the request is querying for to enable caching for all adapters. In this way, we can jointly manage the KV blocks for different adapters, which simplifies the system implementation and improves the global cache hit rate and efficiency. * Multi-modal models. When the user input includes more than just discrete tokens, we can use different hashing methods to handle the caching of inputs of different modalities. For example, perceptual hashing for images to cache similar input images. --- ### Source Design Huggingface Integration (`docs/source/design/huggingface_integration.md`) (huggingface-integration)= # Integration with HuggingFace This document describes how vLLM integrates with HuggingFace libraries. We will explain step by step what happens under the hood when we run `vllm serve`. Let's say we want to serve the popular QWen model by running `vllm serve Qwen/Qwen2-7B`. 1. The `model` argument is `Qwen/Qwen2-7B`. vLLM determines whether this model exists by checking for the corresponding config file `config.json`. See this [code snippet](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L162-L182) for the implementation. Within this process: - If the `model` argument corresponds to an existing local path, vLLM will load the config file directly from this path. - If the `model` argument is a HuggingFace model ID consisting of a username and model name, vLLM will first try to use the config file from the HuggingFace local cache, using the `model` argument as the model name and the `--revision` argument as the revision. See [their website](https://huggingface.co/docs/huggingface_hub/en/package_reference/environment_variables#hfhome) for more information on how the HuggingFace cache works. - If the `model` argument is a HuggingFace model ID but it is not found in the cache, vLLM will download the config file from the HuggingFace model hub. Refer to [this function](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L91) for the implementation. The input arguments include the `model` argument as the model name, the `--revision` argument as the revision, and the environment variable `HF_TOKEN` as the token to access the model hub. In our case, vLLM will download the [config.json](https://huggingface.co/Qwen/Qwen2-7B/blob/main/config.json) file. 2. After confirming the existence of the model, vLLM loads its config file and converts it into a dictionary. See this [code snippet](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L185-L186) for the implementation. 3. Next, vLLM [inspects](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L189) the `model_type` field in the config dictionary to [generate](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#190-L216) the config object to use. There are some `model_type` values that vLLM directly supports; see [here](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/transformers_utils/config.py#L48) for the list. If the `model_type` is not in the list, vLLM will use [AutoConfig.from_pretrained](https://huggingface.co/docs/transformers/en/model_doc/auto#transformers.AutoConfig.from_pretrained) to load the config class, with `model`, `--revision`, and `--trust_remote_code` as the arguments. Please note that: - HuggingFace also has its own logic to determine the config class to use. It will again use the `model_type` field to search for the class name in the transformers library; see [here](https://github.com/huggingface/transformers/tree/main/src/transformers/models) for the list of supported models. If the `model_type` is not found, HuggingFace will use the `auto_map` field from the config JSON file to determine the class name. Specifically, it is the `AutoConfig` field under `auto_map`. See [DeepSeek](https://huggingface.co/deepseek-ai/DeepSeek-V2.5/blob/main/config.json) for an example. - The `AutoConfig` field under `auto_map` points to a module path in the model's repository. To create the config class, HuggingFace will import the module and use the `from_pretrained` method to load the config class. This can generally cause arbitrary code execution, so it is only executed when `--trust_remote_code` is enabled. 4. Subsequently, vLLM applies some historical patches to the config object. These are mostly related to RoPE configuration; see [here](https://github.com/vllm-project/vllm/blob/127c07480ecea15e4c2990820c457807ff78a057/vllm/transformers_utils/config.py#L244) for the implementation. 5. Finally, vLLM can reach the model class we want to initialize. vLLM uses the `architectures` field in the config object to determine the model class to initialize, as it maintains the mapping from architecture name to model class in [its registry](https://github.com/vllm-project/vllm/blob/127c07480ecea15e4c2990820c457807ff78a057/vllm/model_executor/models/registry.py#L80). If the architecture name is not found in the registry, it means this model architecture is not supported by vLLM. For `Qwen/Qwen2-7B`, the `architectures` field is `["Qwen2ForCausalLM"]`, which corresponds to the `Qwen2ForCausalLM` class in [vLLM's code](https://github.com/vllm-project/vllm/blob/127c07480ecea15e4c2990820c457807ff78a057/vllm/model_executor/models/qwen2.py#L364). This class will initialize itself depending on various configs. Beyond that, there are two more things vLLM depends on HuggingFace for. 1. **Tokenizer**: vLLM uses the tokenizer from HuggingFace to tokenize the input text. The tokenizer is loaded using [AutoTokenizer.from_pretrained](https://huggingface.co/docs/transformers/en/model_doc/auto#transformers.AutoTokenizer.from_pretrained) with the `model` argument as the model name and the `--revision` argument as the revision. It is also possible to use a tokenizer from another model by specifying the `--tokenizer` argument in the `vllm serve` command. Other relevant arguments are `--tokenizer-revision` and `--tokenizer-mode`. Please check HuggingFace's documentation for the meaning of these arguments. This part of the logic can be found in the [get_tokenizer](https://github.com/vllm-project/vllm/blob/127c07480ecea15e4c2990820c457807ff78a057/vllm/transformers_utils/tokenizer.py#L87) function. After obtaining the tokenizer, notably, vLLM will cache some expensive attributes of the tokenizer in [get_cached_tokenizer](https://github.com/vllm-project/vllm/blob/127c07480ecea15e4c2990820c457807ff78a057/vllm/transformers_utils/tokenizer.py#L24). 2. **Model weight**: vLLM downloads the model weight from the HuggingFace model hub using the `model` argument as the model name and the `--revision` argument as the revision. vLLM provides the argument `--load-format` to control what files to download from the model hub. By default, it will try to load the weights in the safetensors format and fall back to the PyTorch bin format if the safetensors format is not available. We can also pass `--load-format dummy` to skip downloading the weights. - It is recommended to use the safetensors format, as it is efficient for loading in distributed inference and also safe from arbitrary code execution. See the [documentation](https://huggingface.co/docs/safetensors/en/index) for more information on the safetensors format. This part of the logic can be found [here](https://github.com/vllm-project/vllm/blob/10b67d865d92e376956345becafc249d4c3c0ab7/vllm/model_executor/model_loader/loader.py#L385). Please note that: This completes the integration between vLLM and HuggingFace. In summary, vLLM reads the config file `config.json`, tokenizer, and model weight from the HuggingFace model hub or a local directory. It uses the config class from either vLLM, HuggingFace transformers, or loads the config class from the model's repository. --- ### Source Design Mm Processing (`docs/source/design/mm_processing.md`) (mm-processing)= # Multi-Modal Data Processing To enable various optimizations in vLLM such as [chunked prefill](#chunked-prefill) and [prefix caching](#automatic-prefix-caching), we use {class}`~vllm.multimodal.processing.BaseMultiModalProcessor` to provide the correspondence between placeholder feature tokens (e.g. ``) and multi-modal inputs (e.g. the raw input image) based on the outputs of HF processor. Here are the main features of {class}`~vllm.multimodal.processing.BaseMultiModalProcessor`: ## Prompt Replacement Detection One of the main responsibilies of HF processor is to replace input placeholder tokens (e.g. `` for a single image) with feature placeholder tokens (e.g. `...`, the number of which equals to the feature size). The information about which tokens have been replaced is key to finding the correspondence between placeholder feature tokens and multi-modal inputs. In vLLM, this information is specified using {class}`~vllm.multimodal.processing.PromptReplacement` in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`. Given this specification, we can automatically detect whether HF has replaced the input placeholder tokens by checking whether the feature placeholder tokens exist in the prompt. ## Tokenized Prompt Inputs To enable tokenization in a separate process, we support passing input token IDs alongside multi-modal data. ### The problem Consider that HF processors follow these main steps: 1. Tokenize the text 2. Process multi-modal inputs 3. Perform prompt replacement And we require that: - For text + multi-modal inputs, apply all steps 1--3. - For tokenized + multi-modal inputs, apply only steps 2--3. How can we achieve this without rewriting HF processors? We can try to call the HF processor several times on different inputs: - For text + multi-modal inputs, simply call the HF processor directly. - For tokenized + multi-modal inputs, call the processor only on the multi-modal inputs. While HF processors support text + multi-modal inputs natively, this is not so for tokenized + multi-modal inputs: an error is thrown if the number of input placeholder tokens do not correspond to the number of multi-modal inputs. Moreover, since the tokenized text has not passed through the HF processor, we have to apply Step 3 by ourselves to keep the output tokens and multi-modal data consistent with each other. (mm-dummy-text)= ### Dummy text We work around the first issue by requiring each model to define how to generate dummy text based on the number of multi-modal inputs, via {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`. This lets us generate dummy text corresponding to the multi-modal inputs and input them together to obtain the processed multi-modal data. (mm-automatic-prompt-replacement)= ### Automatic prompt replacement We address the second issue by implementing model-agnostic code in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_prompt_replacements` to automatically replace input placeholder tokens with feature placeholder tokens based on the specification outputted by {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`. ### Summary With the help of dummy text and automatic prompt replacement, our multi-modal processor can finally accept both text and token prompts with multi-modal data. The detailed logic is shown in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_main`. ## Processor Output Caching Some HF processors, such as the one for Qwen2-VL, are [very slow](gh-issue:9238). To alleviate this problem, we cache the multi-modal outputs of HF processor to avoid processing the same multi-modal input (e.g. image) again. When new data is passed in, we first check which items are in the cache, and which ones are missing. The missing items are passed into the HF processor in a single batch and cached, before being merged with the existing items in the cache. Since we only process the missing multi-modal data items, the number of input placeholder tokens no longer corresponds to the number of the multi-modal inputs, so they can't be passed alongside the text prompt to HF processor. Therefore, we process the text and multi-modal inputs separately, using [dummy text](#mm-dummy-text) to avoid HF errors. Since this skips HF's prompt replacement code, we apply [automatic prompt replacement](#mm-automatic-prompt-replacement) afterwards to keep the output tokens and multi-modal data consistent with each other. --- ### Source Design Multiprocessing (`docs/source/design/multiprocessing.md`) # Python Multiprocessing ## Debugging Please see the [Troubleshooting](#troubleshooting-python-multiprocessing) page for information on known issues and how to solve them. ## Introduction ```{important} The source code references are to the state of the code at the time of writing in December, 2024. ``` The use of Python multiprocessing in vLLM is complicated by: - The use of vLLM as a library and the inability to control the code using vLLM - Varying levels of incompatibilities between multiprocessing methods and vLLM dependencies This document describes how vLLM deals with these challenges. ## Multiprocessing Methods [Python multiprocessing methods](https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods) include: - `spawn` - spawn a new Python process. This will be the default as of Python 3.14. - `fork` - Use `os.fork()` to fork the Python interpreter. This is the default in Python versions prior to 3.14. - `forkserver` - Spawn a server process that will fork a new process on request. ### Tradeoffs `fork` is the fastest method, but is incompatible with dependencies that use threads. `spawn` is more compatible with dependencies, but can be problematic when vLLM is used as a library. If the consuming code does not use a `__main__` guard (`if __name__ == "__main__":`), the code will be inadvertently re-executed when vLLM spawns a new process. This can lead to infinite recursion, among other problems. `forkserver` will spawn a new server process that will fork new processes on demand. This unfortunately has the same problem as `spawn` when vLLM is used as a library. The server process is created as a spawned new process, which will re-execute code not protected by a `__main__` guard. For both `spawn` and `forkserver`, the process must not depend on inheriting any global state as would be the case with `fork`. ## Compatibility with Dependencies Multiple vLLM dependencies indicate either a preference or requirement for using `spawn`: - - - It is perhaps more accurate to say that there are known problems with using `fork` after initializing these dependencies. ## Current State (v0) The environment variable `VLLM_WORKER_MULTIPROC_METHOD` can be used to control which method is used by vLLM. The current default is `fork`. - When we know we own the process because the `vllm` command was used, we use `spawn` because it's the most widely compatible. - The `multiproc_xpu_executor` forces the use of `spawn`. - There are other miscellaneous places hard-coding the use of `spawn`: - - Related PRs: - ## Prior State in v1 There was an environment variable to control whether multiprocessing is used in the v1 engine core, `VLLM_ENABLE_V1_MULTIPROCESSING`. This defaulted to off. - When it was enabled, the v1 `LLMEngine` would create a new process to run the engine core. - - - It was off by default for all the reasons mentioned above - compatibility with dependencies and code using vLLM as a library. ### Changes Made in v1 There is not an easy solution with Python's `multiprocessing` that will work everywhere. As a first step, we can get v1 into a state where it does "best effort" choice of multiprocessing method to maximize compatibility. - Default to `fork`. - Use `spawn` when we know we control the main process (`vllm` was executed). - If we detect `cuda` was previously initialized, force `spawn` and emit a warning. We know `fork` will break, so this is the best we can do. The case that is known to still break in this scenario is code using vLLM as a library that initializes `cuda` before calling vLLM. The warning we emit should instruct users to either add a `__main__` guard or to disable multiprocessing. If that known-failure case occurs, the user will see two messages that explain what is happening. First, a log message from vLLM: ```console WARNING 12-11 14:50:37 multiproc_worker_utils.py:281] CUDA was previously initialized. We must use the `spawn` multiprocessing start method. Setting VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/getting_started/debugging.html#python-multiprocessing for more information. ``` Second, Python itself will raise an exception with a nice explanation: ```console RuntimeError: An attempt has been made to start a new process before the current process has finished its bootstrapping phase. This probably means that you are not using fork to start your child processes and you have forgotten to use the proper idiom in the main module: if __name__ == '__main__': freeze_support() ... The "freeze_support()" line can be omitted if the program is not going to be frozen to produce an executable. To fix this issue, refer to the "Safe importing of main module" section in https://docs.python.org/3/library/multiprocessing.html ``` ## Alternatives Considered ### Detect if a `__main__` guard is present It has been suggested that we could behave better if we could detect whether code using vLLM as a library has a `__main__` guard in place. This [post on stackoverflow](https://stackoverflow.com/questions/77220442/multiprocessing-pool-in-a-python-class-without-name-main-guard) was from a library author facing the same question. It is possible to detect whether we are in the original, `__main__` process, or a subsequent spawned process. However, it does not appear to be straight forward to detect whether a `__main__` guard is present in the code. This option has been discarded as impractical. ### Use `forkserver` At first it appears that `forkserver` is a nice solution to the problem. However, the way it works presents the same challenges that `spawn` does when vLLM is used as a library. ### Force `spawn` all the time One way to clean this up is to just force the use of `spawn` all the time and document that the use of a `__main__` guard is required when using vLLM as a library. This would unfortunately break existing code and make vLLM harder to use, violating the desire to make the `LLM` class as easy as possible to use. Instead of pushing this on our users, we will retain the complexity to do our best to make things work. ## Future Work We may want to consider a different worker management approach in the future that works around these challenges. 1. We could implement something `forkserver`-like, but have the process manager be something we initially launch by running our own subprocess and a custom entrypoint for worker management (launch a `vllm-manager` process). 2. We can explore other libraries that may better suit our needs. Examples to consider: - --- ### Source Design Plugin System (`docs/source/design/plugin_system.md`) (plugin-system)= # vLLM's Plugin System The community frequently requests the ability to extend vLLM with custom features. To facilitate this, vLLM includes a plugin system that allows users to add custom features without modifying the vLLM codebase. This document explains how plugins work in vLLM and how to create a plugin for vLLM. ## How Plugins Work in vLLM Plugins are user-registered code that vLLM executes. Given vLLM's architecture (see [](#arch-overview)), multiple processes may be involved, especially when using distributed inference with various parallelism techniques. To enable plugins successfully, every process created by vLLM needs to load the plugin. This is done by the [load_general_plugins](https://github.com/vllm-project/vllm/blob/c76ac49d266e27aa3fea84ef2df1f813d24c91c7/vllm/plugins/__init__.py#L16) function in the `vllm.plugins` module. This function is called for every process created by vLLM before it starts any work. ## How vLLM Discovers Plugins vLLM's plugin system uses the standard Python `entry_points` mechanism. This mechanism allows developers to register functions in their Python packages for use by other packages. An example of a plugin: ```python # inside `setup.py` file from setuptools import setup setup(name='vllm_add_dummy_model', version='0.1', packages=['vllm_add_dummy_model'], entry_points={ 'vllm.general_plugins': ["register_dummy_model = vllm_add_dummy_model:register"] }) # inside `vllm_add_dummy_model.py` file def register(): from vllm import ModelRegistry if "MyLlava" not in ModelRegistry.get_supported_archs(): ModelRegistry.register_model("MyLlava", "vllm_add_dummy_model.my_llava:MyLlava") ``` For more information on adding entry points to your package, please check the [official documentation](https://setuptools.pypa.io/en/latest/userguide/entry_point.html). Every plugin has three parts: 1. **Plugin group**: The name of the entry point group. vLLM uses the entry point group `vllm.general_plugins` to register general plugins. This is the key of `entry_points` in the `setup.py` file. Always use `vllm.general_plugins` for vLLM's general plugins. 2. **Plugin name**: The name of the plugin. This is the value in the dictionary of the `entry_points` dictionary. In the example above, the plugin name is `register_dummy_model`. Plugins can be filtered by their names using the `VLLM_PLUGINS` environment variable. To load only a specific plugin, set `VLLM_PLUGINS` to the plugin name. 3. **Plugin value**: The fully qualified name of the function to register in the plugin system. In the example above, the plugin value is `vllm_add_dummy_model:register`, which refers to a function named `register` in the `vllm_add_dummy_model` module. ## Types of supported plugins - **General plugins** (with group name `vllm.general_plugins`): The primary use case for these plugins is to register custom, out-of-the-tree models into vLLM. This is done by calling `ModelRegistry.register_model` to register the model inside the plugin function. - **Platform plugins** (with group name `vllm.platform_plugins`): The primary use case for these plugins is to register custom, out-of-the-tree platforms into vLLM. The plugin function should return `None` when the platform is not supported in the current environment, or the platform class's fully qualified name when the platform is supported. ## Guidelines for Writing Plugins - **Being re-entrant**: The function specified in the entry point should be re-entrant, meaning it can be called multiple times without causing issues. This is necessary because the function might be called multiple times in some processes. ## Compatibility Guarantee vLLM guarantees the interface of documented plugins, such as `ModelRegistry.register_model`, will always be available for plugins to register models. However, it is the responsibility of plugin developers to ensure their plugins are compatible with the version of vLLM they are targeting. For example, `"vllm_add_dummy_model.my_llava:MyLlava"` should be compatible with the version of vLLM that the plugin targets. The interface for the model may change during vLLM's development. --- ### Source Features Automatic Prefix Caching (`docs/source/features/automatic_prefix_caching.md`) (automatic-prefix-caching)= # Automatic Prefix Caching ## Introduction Automatic Prefix Caching (APC in short) caches the KV cache of existing queries, so that a new query can directly reuse the KV cache if it shares the same prefix with one of the existing queries, allowing the new query to skip the computation of the shared part. ```{note} Technical details on how vLLM implements APC can be found [here](#design-automatic-prefix-caching). ``` ## Enabling APC in vLLM Set `enable_prefix_caching=True` in vLLM engine to enable APC. Here is an example: ``` /* Detailed source-code truncated for AI context efficiency. */ ``` ## Example workloads We describe two example workloads, where APC can provide huge performance benefit: - Long document query, where the user repeatedly queries the same long document (e.g. software manual or annual report) with different queries. In this case, instead of processing the long document again and again, APC allows vLLM to process this long document *only once*, and all future requests can avoid recomputing this long document by reusing its KV cache. This allows vLLM to serve future requests with much higher throughput and much lower latency. - Multi-round conversation, where the user may chat with the application multiple times in the same chatting session. In this case, instead of processing the whole chatting history again and again, APC allows vLLM to reuse the processing results of the chat history across all future rounds of conversation, allowing vLLM to serve future requests with much higher throughput and much lower latency. ## Limits APC in general does not reduce the performance of vLLM. With that being said, APC only reduces the time of processing the queries (the prefilling phase) and does not reduce the time of generating new tokens (the decoding phase). So APC does not bring performance gain when vLLM spends most of the time generating answers to the queries (e.g. when the length of the answer is long), or new queries do not share the same prefix with any of existing queries (so that the computation cannot be reused). --- ### Source Features Compatibility Matrix (`docs/source/features/compatibility_matrix.md`) (compatibility-matrix)= # Compatibility Matrix The tables below show mutually exclusive features and the support on some hardware. ```{note} Check the '✗' with links to see tracking issue for unsupported feature/hardware combination. ``` ## Feature x Feature ```{raw} html ``` ``` /* Detailed source-code truncated for AI context efficiency. */ ``` (feature-x-hardware)= ## Feature x Hardware ``` /* Detailed source-code truncated for AI context efficiency. */ ``` --- ### Source Features Disagg Prefill (`docs/source/features/disagg_prefill.md`) (disagg-prefill)= # Disaggregated Prefilling (experimental) This page introduces you the disaggregated prefilling feature in vLLM. ```{note} This feature is experimental and subject to change. ``` ## Why disaggregated prefilling? Two main reasons: - **Tuning time-to-first-token (TTFT) and inter-token-latency (ITL) separately**. Disaggregated prefilling put prefill and decode phase of LLM inference inside different vLLM instances. This gives you the flexibility to assign different parallel strategies (e.g. `tp` and `pp`) to tune TTFT without affecting ITL, or to tune ITL without affecting TTFT. - **Controlling tail ITL**. Without disaggregated prefilling, vLLM may insert some prefill jobs during the decoding of one request. This results in higher tail latency. Disaggregated prefilling helps you solve this issue and control tail ITL. Chunked prefill with a proper chunk size also can achieve the same goal, but in practice it's hard to figure out the correct chunk size value. So disaggregated prefilling is a much more reliable way to control tail ITL. ```{note} Disaggregated prefill DOES NOT improve throughput. ``` ## Usage example Please refer to `examples/online_serving/disaggregated_prefill.sh` for the example usage of disaggregated prefilling. ## Benchmarks Please refer to `benchmarks/disagg_benchmarks/` for disaggregated prefilling benchmarks. ## Development We implement disaggregated prefilling by running 2 vLLM instances. One for prefill (we call it prefill instance) and one for decode (we call it decode instance), and then use a connector to transfer the prefill KV caches and results from prefill instance to decode instance. All disaggregated prefilling implementation is under `vllm/distributed/kv_transfer`. Key abstractions for disaggregated prefilling: - **Connector**: Connector allows **kv consumer** to retrieve the KV caches of a batch of request from **kv producer**. - **LookupBuffer**: LookupBuffer provides two API: `insert` KV cache and `drop_select` KV cache. The semantics of `insert` and `drop_select` are similar to SQL, where `insert` inserts a KV cache into the buffer, and `drop_select` returns the KV cache that matches the given condition and drop it from the buffer. - **Pipe**: A single-direction FIFO pipe for tensor transmission. It supports `send_tensor` and `recv_tensor`. ```{note} `insert` is non-blocking operation but `drop_select` is blocking operation. ``` Here is a figure illustrating how the above 3 abstractions are organized: ```{image} /assets/features/disagg_prefill/abstraction.jpg :alt: Disaggregated prefilling abstractions ``` The workflow of disaggregated prefilling is as follows: ```{image} /assets/features/disagg_prefill/overview.jpg :alt: Disaggregated prefilling workflow ``` The `buffer` corresponds to `insert` API in LookupBuffer, and the `drop_select` corresponds to `drop_select` API in LookupBuffer. ## Third-party contributions Disaggregated prefilling is highly related to infrastructure, so vLLM relies on third-party connectors for production-level disaggregated prefilling (and vLLM team will actively review and merge new PRs for third-party connectors). We recommend three ways of implementations: - **Fully-customized connector**: Implement your own `Connector`, and call third-party libraries to send and receive KV caches, and many many more (like editing vLLM's model input to perform customized prefilling, etc). This approach gives you the most control, but at the risk of being incompatible with future vLLM versions. - **Database-like connector**: Implement your own `LookupBuffer` and support the `insert` and `drop_select` APIs just like SQL. - **Distributed P2P connector**: Implement your own `Pipe` and support the `send_tensor` and `recv_tensor` APIs, just like `torch.distributed`. --- ### Source Features Lora (`docs/source/features/lora.md`) (lora-adapter)= # LoRA Adapters This document shows you how to use [LoRA adapters](https://arxiv.org/abs/2106.09685) with vLLM on top of a base model. LoRA adapters can be used with any vLLM model that implements {class}`~vllm.model_executor.models.interfaces.SupportsLoRA`. Adapters can be efficiently served on a per request basis with minimal overhead. First we download the adapter(s) and save them locally with ```python from huggingface_hub import snapshot_download sql_lora_path = snapshot_download(repo_id="yard1/llama-2-7b-sql-lora-test") ``` Then we instantiate the base model and pass in the `enable_lora=True` flag: ```python from vllm import LLM, SamplingParams from vllm.lora.request import LoRARequest llm = LLM(model="meta-llama/Llama-2-7b-hf", enable_lora=True) ``` We can now submit the prompts and call `llm.generate` with the `lora_request` parameter. The first parameter of `LoRARequest` is a human identifiable name, the second parameter is a globally unique ID for the adapter and the third parameter is the path to the LoRA adapter. ```python sampling_params = SamplingParams( temperature=0, max_tokens=256, stop=["[/assistant]"] ) prompts = [ "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_74 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]", "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_11 (nationality VARCHAR, elector VARCHAR)\n\n question: When Anchero Pantaleone was the elector what is under nationality? [/user] [assistant]", ] outputs = llm.generate( prompts, sampling_params, lora_request=LoRARequest("sql_adapter", 1, sql_lora_path) ) ``` Check out for an example of how to use LoRA adapters with the async engine and how to use more advanced configuration options. ## Serving LoRA Adapters LoRA adapted models can also be served with the Open-AI compatible vLLM server. To do so, we use `--lora-modules {name}={path} {name}={path}` to specify each LoRA module when we kickoff the server: ```bash vllm serve meta-llama/Llama-2-7b-hf \ --enable-lora \ --lora-modules sql-lora=$HOME/.cache/huggingface/hub/models--yard1--llama-2-7b-sql-lora-test/snapshots/0dfa347e8877a4d4ed19ee56c140fa518470028c/ ``` ```{note} The commit ID `0dfa347e8877a4d4ed19ee56c140fa518470028c` may change over time. Please check the latest commit ID in your environment to ensure you are using the correct one. ``` The server entrypoint accepts all other LoRA configuration parameters (`max_loras`, `max_lora_rank`, `max_cpu_loras`, etc.), which will apply to all forthcoming requests. Upon querying the `/models` endpoint, we should see our LoRA along with its base model: ```bash curl localhost:8000/v1/models | jq . { "object": "list", "data": [ { "id": "meta-llama/Llama-2-7b-hf", "object": "model", ... }, { "id": "sql-lora", "object": "model", ... } ] } ``` Requests can specify the LoRA adapter as if it were any other model via the `model` request parameter. The requests will be processed according to the server-wide LoRA configuration (i.e. in parallel with base model requests, and potentially other LoRA adapter requests if they were provided and `max_loras` is set high enough). The following is an example request ```bash curl http://localhost:8000/v1/completions \ -H "Content-Type: application/json" \ -d '{ "model": "sql-lora", "prompt": "San Francisco is a", "max_tokens": 7, "temperature": 0 }' | jq ``` ## Dynamically serving LoRA Adapters In addition to serving LoRA adapters at server startup, the vLLM server now supports dynamically loading and unloading LoRA adapters at runtime through dedicated API endpoints. This feature can be particularly useful when the flexibility to change models on-the-fly is needed. Note: Enabling this feature in production environments is risky as user may participate model adapter management. To enable dynamic LoRA loading and unloading, ensure that the environment variable `VLLM_ALLOW_RUNTIME_LORA_UPDATING` is set to `True`. When this option is enabled, the API server will log a warning to indicate that dynamic loading is active. ```bash export VLLM_ALLOW_RUNTIME_LORA_UPDATING=True ``` Loading a LoRA Adapter: To dynamically load a LoRA adapter, send a POST request to the `/v1/load_lora_adapter` endpoint with the necessary details of the adapter to be loaded. The request payload should include the name and path to the LoRA adapter. Example request to load a LoRA adapter: ```bash curl -X POST http://localhost:8000/v1/load_lora_adapter \ -H "Content-Type: application/json" \ -d '{ "lora_name": "sql_adapter", "lora_path": "/path/to/sql-lora-adapter" }' ``` Upon a successful request, the API will respond with a 200 OK status code. If an error occurs, such as if the adapter cannot be found or loaded, an appropriate error message will be returned. Unloading a LoRA Adapter: To unload a LoRA adapter that has been previously loaded, send a POST request to the `/v1/unload_lora_adapter` endpoint with the name or ID of the adapter to be unloaded. Example request to unload a LoRA adapter: ```bash curl -X POST http://localhost:8000/v1/unload_lora_adapter \ -H "Content-Type: application/json" \ -d '{ "lora_name": "sql_adapter" }' ``` ## New format for `--lora-modules` In the previous version, users would provide LoRA modules via the following format, either as a key-value pair or in JSON format. For example: ```bash --lora-modules sql-lora=$HOME/.cache/huggingface/hub/models--yard1--llama-2-7b-sql-lora-test/snapshots/0dfa347e8877a4d4ed19ee56c140fa518470028c/ ``` This would only include the `name` and `path` for each LoRA module, but did not provide a way to specify a `base_model_name`. Now, you can specify a base_model_name alongside the name and path using JSON format. For example: ```bash --lora-modules '{"name": "sql-lora", "path": "/path/to/lora", "base_model_name": "meta-llama/Llama-2-7b"}' ``` To provide the backward compatibility support, you can still use the old key-value format (name=path), but the `base_model_name` will remain unspecified in that case. ## Lora model lineage in model card The new format of `--lora-modules` is mainly to support the display of parent model information in the model card. Here's an explanation of how your current response supports this: - The `parent` field of LoRA model `sql-lora` now links to its base model `meta-llama/Llama-2-7b-hf`. This correctly reflects the hierarchical relationship between the base model and the LoRA adapter. - The `root` field points to the artifact location of the lora adapter. ```bash $ curl http://localhost:8000/v1/models { "object": "list", "data": [ { "id": "meta-llama/Llama-2-7b-hf", "object": "model", "created": 1715644056, "owned_by": "vllm", "root": "~/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf/snapshots/01c7f73d771dfac7d292323805ebc428287df4f9/", "parent": null, "permission": [ { ..... } ] }, { "id": "sql-lora", "object": "model", "created": 1715644056, "owned_by": "vllm", "root": "~/.cache/huggingface/hub/models--yard1--llama-2-7b-sql-lora-test/snapshots/0dfa347e8877a4d4ed19ee56c140fa518470028c/", "parent": meta-llama/Llama-2-7b-hf, "permission": [ { .... } ] } ] } ``` --- ### Source Features Spec Decode (`docs/source/features/spec_decode.md`) (spec-decode)= # Speculative Decoding ```{warning} Please note that speculative decoding in vLLM is not yet optimized and does not usually yield inter-token latency reductions for all prompt datasets or sampling parameters. The work to optimize it is ongoing and can be followed here: ``` ```{warning} Currently, speculative decoding in vLLM is not compatible with pipeline parallelism. ``` This document shows how to use [Speculative Decoding](https://x.com/karpathy/status/1697318534555336961) with vLLM. Speculative decoding is a technique which improves inter-token latency in memory-bound LLM inference. ## Speculating with a draft model The following code configures vLLM in an offline mode to use speculative decoding with a draft model, speculating 5 tokens at a time. ```python from vllm import LLM, SamplingParams prompts = [ "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM( model="facebook/opt-6.7b", tensor_parallel_size=1, speculative_model="facebook/opt-125m", num_speculative_tokens=5, ) outputs = llm.generate(prompts, sampling_params) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` To perform the same with an online mode launch the server: ```bash python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 8000 --model facebook/opt-6.7b \ --seed 42 -tp 1 --speculative_model facebook/opt-125m --use-v2-block-manager \ --num_speculative_tokens 5 --gpu_memory_utilization 0.8 ``` Then use a client: ```python from openai import OpenAI # Modify OpenAI's API key and API base to use vLLM's API server. openai_api_key = "EMPTY" openai_api_base = "http://localhost:8000/v1" client = OpenAI( # defaults to os.environ.get("OPENAI_API_KEY") api_key=openai_api_key, base_url=openai_api_base, ) models = client.models.list() model = models.data[0].id # Completion API stream = False completion = client.completions.create( model=model, prompt="The future of AI is", echo=False, n=1, stream=stream, ) print("Completion results:") if stream: for c in completion: print(c) else: print(completion) ``` ## Speculating by matching n-grams in the prompt The following code configures vLLM to use speculative decoding where proposals are generated by matching n-grams in the prompt. For more information read [this thread.](https://x.com/joao_gante/status/1747322413006643259) ```python from vllm import LLM, SamplingParams prompts = [ "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM( model="facebook/opt-6.7b", tensor_parallel_size=1, speculative_model="[ngram]", num_speculative_tokens=5, ngram_prompt_lookup_max=4, ) outputs = llm.generate(prompts, sampling_params) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` ## Speculating using MLP speculators The following code configures vLLM to use speculative decoding where proposals are generated by draft models that conditioning draft predictions on both context vectors and sampled tokens. For more information see [this blog](https://pytorch.org/blog/hitchhikers-guide-speculative-decoding/) or [this technical report](https://arxiv.org/abs/2404.19124). ```python from vllm import LLM, SamplingParams prompts = [ "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM( model="meta-llama/Meta-Llama-3.1-70B-Instruct", tensor_parallel_size=4, speculative_model="ibm-fms/llama3-70b-accelerator", speculative_draft_tensor_parallel_size=1, ) outputs = llm.generate(prompts, sampling_params) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` Note that these speculative models currently need to be run without tensor parallelism, although it is possible to run the main model using tensor parallelism (see example above). Since the speculative models are relatively small, we still see significant speedups. However, this limitation will be fixed in a future release. A variety of speculative models of this type are available on HF hub: - [llama-13b-accelerator](https://huggingface.co/ibm-fms/llama-13b-accelerator) - [llama3-8b-accelerator](https://huggingface.co/ibm-fms/llama3-8b-accelerator) - [codellama-34b-accelerator](https://huggingface.co/ibm-fms/codellama-34b-accelerator) - [llama2-70b-accelerator](https://huggingface.co/ibm-fms/llama2-70b-accelerator) - [llama3-70b-accelerator](https://huggingface.co/ibm-fms/llama3-70b-accelerator) - [granite-3b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-3b-code-instruct-accelerator) - [granite-8b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-8b-code-instruct-accelerator) - [granite-7b-instruct-accelerator](https://huggingface.co/ibm-granite/granite-7b-instruct-accelerator) - [granite-20b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-20b-code-instruct-accelerator) ## Speculating using EAGLE based draft models The following code configures vLLM to use speculative decoding where proposals are generated by an [EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency)](https://arxiv.org/pdf/2401.15077) based draft model. ```python from vllm import LLM, SamplingParams prompts = [ "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM( model="meta-llama/Meta-Llama-3-8B-Instruct", tensor_parallel_size=4, speculative_model="path/to/modified/eagle/model", speculative_draft_tensor_parallel_size=1, ) outputs = llm.generate(prompts, sampling_params) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` A few important things to consider when using the EAGLE based draft models: 1. The EAGLE draft models available in the [HF repository for EAGLE models](https://huggingface.co/yuhuili) cannot be used directly with vLLM due to differences in the expected layer names and model definition. To use these models with vLLM, use the [following script](https://gist.github.com/abhigoyal1997/1e7a4109ccb7704fbc67f625e86b2d6d) to convert them. Note that this script does not modify the model's weights. In the above example, use the script to first convert the [yuhuili/EAGLE-LLaMA3-Instruct-8B](https://huggingface.co/yuhuili/EAGLE-LLaMA3-Instruct-8B) model and then use the converted checkpoint as the draft model in vLLM. 2. The EAGLE based draft models need to be run without tensor parallelism (i.e. speculative_draft_tensor_parallel_size is set to 1), although it is possible to run the main model using tensor parallelism (see example above). 3. When using EAGLE-based speculators with vLLM, the observed speedup is lower than what is reported in the reference implementation [here](https://github.com/SafeAILab/EAGLE). This issue is under investigation and tracked here: [https://github.com/vllm-project/vllm/issues/9565](https://github.com/vllm-project/vllm/issues/9565). A variety of EAGLE draft models are available on the Hugging Face hub: | Base Model | EAGLE on Hugging Face | # EAGLE Parameters | |---------------------------------------------------------------------|-------------------------------------------|--------------------| | Vicuna-7B-v1.3 | yuhuili/EAGLE-Vicuna-7B-v1.3 | 0.24B | | Vicuna-13B-v1.3 | yuhuili/EAGLE-Vicuna-13B-v1.3 | 0.37B | | Vicuna-33B-v1.3 | yuhuili/EAGLE-Vicuna-33B-v1.3 | 0.56B | | LLaMA2-Chat 7B | yuhuili/EAGLE-llama2-chat-7B | 0.24B | | LLaMA2-Chat 13B | yuhuili/EAGLE-llama2-chat-13B | 0.37B | | LLaMA2-Chat 70B | yuhuili/EAGLE-llama2-chat-70B | 0.99B | | Mixtral-8x7B-Instruct-v0.1 | yuhuili/EAGLE-mixtral-instruct-8x7B | 0.28B | | LLaMA3-Instruct 8B | yuhuili/EAGLE-LLaMA3-Instruct-8B | 0.25B | | LLaMA3-Instruct 70B | yuhuili/EAGLE-LLaMA3-Instruct-70B | 0.99B | | Qwen2-7B-Instruct | yuhuili/EAGLE-Qwen2-7B-Instruct | 0.26B | | Qwen2-72B-Instruct | yuhuili/EAGLE-Qwen2-72B-Instruct | 1.05B | ## Lossless guarantees of Speculative Decoding In vLLM, speculative decoding aims to enhance inference efficiency while maintaining accuracy. This section addresses the lossless guarantees of speculative decoding, breaking down the guarantees into three key areas: 1. **Theoretical Losslessness** \- Speculative decoding sampling is theoretically lossless up to the precision limits of hardware numerics. Floating-point errors might cause slight variations in output distributions, as discussed in [Accelerating Large Language Model Decoding with Speculative Sampling](https://arxiv.org/pdf/2302.01318) 2. **Algorithmic Losslessness** \- vLLM’s implementation of speculative decoding is algorithmically validated to be lossless. Key validation tests include: > - **Rejection Sampler Convergence**: Ensures that samples from vLLM’s rejection sampler align with the target > distribution. [View Test Code](https://github.com/vllm-project/vllm/blob/47b65a550866c7ffbd076ecb74106714838ce7da/tests/samplers/test_rejection_sampler.py#L252) > - **Greedy Sampling Equality**: Confirms that greedy sampling with speculative decoding matches greedy sampling > without it. This verifies that vLLM's speculative decoding framework, when integrated with the vLLM forward pass and the vLLM rejection sampler, > provides a lossless guarantee. Almost all of the tests in . > verify this property using [this assertion implementation](https://github.com/vllm-project/vllm/blob/b67ae00cdbbe1a58ffc8ff170f0c8d79044a684a/tests/spec_decode/e2e/conftest.py#L291) 3. **vLLM Logprob Stability** \- vLLM does not currently guarantee stable token log probabilities (logprobs). This can result in different outputs for the same request across runs. For more details, see the FAQ section titled *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](#faq). While vLLM strives to ensure losslessness in speculative decoding, variations in generated outputs with and without speculative decoding can occur due to following factors: - **Floating-Point Precision**: Differences in hardware numerical precision may lead to slight discrepancies in the output distribution. - **Batch Size and Numerical Stability**: Changes in batch size may cause variations in logprobs and output probabilities, potentially due to non-deterministic behavior in batched operations or numerical instability. For mitigation strategies, please refer to the FAQ entry *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](#faq). ## Resources for vLLM contributors - [A Hacker's Guide to Speculative Decoding in vLLM](https://www.youtube.com/watch?v=9wNAgpX6z_4) - [What is Lookahead Scheduling in vLLM?](https://docs.google.com/document/d/1Z9TvqzzBPnh5WHcRwjvK2UEeFeq5zMZb5mFE8jR0HCs/edit#heading=h.1fjfb0donq5a) - [Information on batch expansion](https://docs.google.com/document/d/1T-JaS2T1NRfdP51qzqpyakoCXxSXTtORppiwaj5asxA/edit#heading=h.kk7dq05lc6q8) - [Dynamic speculative decoding](gh-issue:4565) --- ### Source Features Structured Outputs (`docs/source/features/structured_outputs.md`) (structured-outputs)= # Structured Outputs vLLM supports the generation of structured outputs using [outlines](https://github.com/dottxt-ai/outlines), [lm-format-enforcer](https://github.com/noamgat/lm-format-enforcer), or [xgrammar](https://github.com/mlc-ai/xgrammar) as backends for the guided decoding. This document shows you some examples of the different options that are available to generate structured outputs. ## Online Serving (OpenAI API) You can generate structured outputs using the OpenAI's [Completions](https://platform.openai.com/docs/api-reference/completions) and [Chat](https://platform.openai.com/docs/api-reference/chat) API. The following parameters are supported, which must be added as extra parameters: - `guided_choice`: the output will be exactly one of the choices. - `guided_regex`: the output will follow the regex pattern. - `guided_json`: the output will follow the JSON schema. - `guided_grammar`: the output will follow the context free grammar. - `guided_whitespace_pattern`: used to override the default whitespace pattern for guided json decoding. - `guided_decoding_backend`: used to select the guided decoding backend to use. You can see the complete list of supported parameters on the [OpenAI-Compatible Server](#openai-compatible-server)page. Now let´s see an example for each of the cases, starting with the `guided_choice`, as it´s the easiest one: ```python from openai import OpenAI client = OpenAI( base_url="http://localhost:8000/v1", api_key="-", ) completion = client.chat.completions.create( model="Qwen/Qwen2.5-3B-Instruct", messages=[ {"role": "user", "content": "Classify this sentiment: vLLM is wonderful!"} ], extra_body={"guided_choice": ["positive", "negative"]}, ) print(completion.choices[0].message.content) ``` The next example shows how to use the `guided_regex`. The idea is to generate an email address, given a simple regex template: ```python completion = client.chat.completions.create( model="Qwen/Qwen2.5-3B-Instruct", messages=[ { "role": "user", "content": "Generate an example email address for Alan Turing, who works in Enigma. End in .com and new line. Example result: alan.turing@enigma.com\n", } ], extra_body={"guided_regex": "\w+@\w+\.com\n", "stop": ["\n"]}, ) print(completion.choices[0].message.content) ``` One of the most relevant features in structured text generation is the option to generate a valid JSON with pre-defined fields and formats. For this we can use the `guided_json` parameter in two different ways: - Using directly a [JSON Schema](https://json-schema.org/) - Defining a [Pydantic model](https://docs.pydantic.dev/latest/) and then extracting the JSON Schema from it (which is normally an easier option). The next example shows how to use the `guided_json` parameter with a Pydantic model: ```python from pydantic import BaseModel from enum import Enum class CarType(str, Enum): sedan = "sedan" suv = "SUV" truck = "Truck" coupe = "Coupe" class CarDescription(BaseModel): brand: str model: str car_type: CarType json_schema = CarDescription.model_json_schema() completion = client.chat.completions.create( model="Qwen/Qwen2.5-3B-Instruct", messages=[ { "role": "user", "content": "Generate a JSON with the brand, model and car_type of the most iconic car from the 90's", } ], extra_body={"guided_json": json_schema}, ) print(completion.choices[0].message.content) ``` ```{tip} While not strictly necessary, normally it´s better to indicate in the prompt that a JSON needs to be generated and which fields and how should the LLM fill them. This can improve the results notably in most cases. ``` Finally we have the `guided_grammar`, which probably is the most difficult one to use but it´s really powerful, as it allows us to define complete languages like SQL queries. It works by using a context free EBNF grammar, which for example we can use to define a specific format of simplified SQL queries, like in the example below: ```python simplified_sql_grammar = """ ?start: select_statement ?select_statement: "SELECT " column_list " FROM " table_name ?column_list: column_name ("," column_name)* ?table_name: identifier ?column_name: identifier ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/ """ completion = client.chat.completions.create( model="Qwen/Qwen2.5-3B-Instruct", messages=[ { "role": "user", "content": "Generate an SQL query to show the 'username' and 'email' from the 'users' table.", } ], extra_body={"guided_grammar": simplified_sql_grammar}, ) print(completion.choices[0].message.content) ``` Full example: ## Experimental Automatic Parsing (OpenAI API) This section covers the OpenAI beta wrapper over the `client.chat.completions.create()` method that provides richer integrations with Python specific types. At the time of writing (`openai==1.54.4`), this is a "beta" feature in the OpenAI client library. Code reference can be found [here](https://github.com/openai/openai-python/blob/52357cff50bee57ef442e94d78a0de38b4173fc2/src/openai/resources/beta/chat/completions.py#L100-L104). For the following examples, vLLM was setup using `vllm serve meta-llama/Llama-3.1-8B-Instruct` Here is a simple example demonstrating how to get structured output using Pydantic models: ```python from pydantic import BaseModel from openai import OpenAI class Info(BaseModel): name: str age: int client = OpenAI(base_url="http://0.0.0.0:8000/v1", api_key="dummy") completion = client.beta.chat.completions.parse( model="meta-llama/Llama-3.1-8B-Instruct", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "My name is Cameron, I'm 28. What's my name and age?"}, ], response_format=Info, extra_body=dict(guided_decoding_backend="outlines"), ) message = completion.choices[0].message print(message) assert message.parsed print("Name:", message.parsed.name) print("Age:", message.parsed.age) ``` Output: ```console ParsedChatCompletionMessage[Testing](content='{"name": "Cameron", "age": 28}', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=[], parsed=Testing(name='Cameron', age=28)) Name: Cameron Age: 28 ``` Here is a more complex example using nested Pydantic models to handle a step-by-step math solution: ```python from typing import List from pydantic import BaseModel from openai import OpenAI class Step(BaseModel): explanation: str output: str class MathResponse(BaseModel): steps: List[Step] final_answer: str client = OpenAI(base_url="http://0.0.0.0:8000/v1", api_key="dummy") completion = client.beta.chat.completions.parse( model="meta-llama/Llama-3.1-8B-Instruct", messages=[ {"role": "system", "content": "You are a helpful expert math tutor."}, {"role": "user", "content": "Solve 8x + 31 = 2."}, ], response_format=MathResponse, extra_body=dict(guided_decoding_backend="outlines"), ) message = completion.choices[0].message print(message) assert message.parsed for i, step in enumerate(message.parsed.steps): print(f"Step #{i}:", step) print("Answer:", message.parsed.final_answer) ``` Output: ```console ParsedChatCompletionMessage[MathResponse](content='{ "steps": [{ "explanation": "First, let\'s isolate the term with the variable \'x\'. To do this, we\'ll subtract 31 from both sides of the equation.", "output": "8x + 31 - 31 = 2 - 31"}, { "explanation": "By subtracting 31 from both sides, we simplify the equation to 8x = -29.", "output": "8x = -29"}, { "explanation": "Next, let\'s isolate \'x\' by dividing both sides of the equation by 8.", "output": "8x / 8 = -29 / 8"}], "final_answer": "x = -29/8" }', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=[], parsed=MathResponse(steps=[Step(explanation="First, let's isolate the term with the variable 'x'. To do this, we'll subtract 31 from both sides of the equation.", output='8x + 31 - 31 = 2 - 31'), Step(explanation='By subtracting 31 from both sides, we simplify the equation to 8x = -29.', output='8x = -29'), Step(explanation="Next, let's isolate 'x' by dividing both sides of the equation by 8.", output='8x / 8 = -29 / 8')], final_answer='x = -29/8')) Step #0: explanation="First, let's isolate the term with the variable 'x'. To do this, we'll subtract 31 from both sides of the equation." output='8x + 31 - 31 = 2 - 31' Step #1: explanation='By subtracting 31 from both sides, we simplify the equation to 8x = -29.' output='8x = -29' Step #2: explanation="Next, let's isolate 'x' by dividing both sides of the equation by 8." output='8x / 8 = -29 / 8' Answer: x = -29/8 ``` ## Offline Inference Offline inference allows for the same types of guided decoding. To use it, we´ll need to configure the guided decoding using the class `GuidedDecodingParams` inside `SamplingParams`. The main available options inside `GuidedDecodingParams` are: - `json` - `regex` - `choice` - `grammar` - `backend` - `whitespace_pattern` These parameters can be used in the same way as the parameters from the Online Serving examples above. One example for the usage of the `choices` parameter is shown below: ```python from vllm import LLM, SamplingParams from vllm.sampling_params import GuidedDecodingParams llm = LLM(model="HuggingFaceTB/SmolLM2-1.7B-Instruct") guided_decoding_params = GuidedDecodingParams(choice=["Positive", "Negative"]) sampling_params = SamplingParams(guided_decoding=guided_decoding_params) outputs = llm.generate( prompts="Classify this sentiment: vLLM is wonderful!", sampling_params=sampling_params, ) print(outputs[0].outputs[0].text) ``` Full example: --- ### Source Features Tool Calling (`docs/source/features/tool_calling.md`) # Tool Calling vLLM currently supports named function calling, as well as the `auto` and `none` options for the `tool_choice` field in the chat completion API. The `tool_choice` option `required` is **not yet supported** but on the roadmap. ## Quickstart Start the server with tool calling enabled. This example uses Meta's Llama 3.1 8B model, so we need to use the llama3 tool calling chat template from the vLLM examples directory: ```bash vllm serve meta-llama/Llama-3.1-8B-Instruct \ --enable-auto-tool-choice \ --tool-call-parser llama3_json \ --chat-template examples/tool_chat_template_llama3.1_json.jinja ``` Next, make a request to the model that should result in it using the available tools: ```python from openai import OpenAI import json client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy") def get_weather(location: str, unit: str): return f"Getting the weather for {location} in {unit}..." tool_functions = {"get_weather": get_weather} tools = [{ "type": "function", "function": { "name": "get_weather", "description": "Get the current weather in a given location", "parameters": { "type": "object", "properties": { "location": {"type": "string", "description": "City and state, e.g., 'San Francisco, CA'"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]} }, "required": ["location", "unit"] } } }] response = client.chat.completions.create( model=client.models.list().data[0].id, messages=[{"role": "user", "content": "What's the weather like in San Francisco?"}], tools=tools, tool_choice="auto" ) tool_call = response.choices[0].message.tool_calls[0].function print(f"Function called: {tool_call.name}") print(f"Arguments: {tool_call.arguments}") print(f"Result: {get_weather(**json.loads(tool_call.arguments))}") ``` Example output: ```text Function called: get_weather Arguments: {"location": "San Francisco, CA", "unit": "fahrenheit"} Result: Getting the weather for San Francisco, CA in fahrenheit... ``` This example demonstrates: * Setting up the server with tool calling enabled * Defining an actual function to handle tool calls * Making a request with `tool_choice="auto"` * Handling the structured response and executing the corresponding function You can also specify a particular function using named function calling by setting `tool_choice={"type": "function", "function": {"name": "get_weather"}}`. Note that this will use the guided decoding backend - so the first time this is used, there will be several seconds of latency (or more) as the FSM is compiled for the first time before it is cached for subsequent requests. Remember that it's the callers responsibility to: 1. Define appropriate tools in the request 2. Include relevant context in the chat messages 3. Handle the tool calls in your application logic For more advanced usage, including parallel tool calls and different model-specific parsers, see the sections below. ## Named Function Calling vLLM supports named function calling in the chat completion API by default. It does so using Outlines through guided decoding, so this is enabled by default, and will work with any supported model. You are guaranteed a validly-parsable function call - not a high-quality one. vLLM will use guided decoding to ensure the response matches the tool parameter object defined by the JSON schema in the `tools` parameter. For best results, we recommend ensuring that the expected output format / schema is specified in the prompt to ensure that the model's intended generation is aligned with the schema that it's being forced to generate by the guided decoding backend. To use a named function, you need to define the functions in the `tools` parameter of the chat completion request, and specify the `name` of one of the tools in the `tool_choice` parameter of the chat completion request. ## Automatic Function Calling To enable this feature, you should set the following flags: * `--enable-auto-tool-choice` -- **mandatory** Auto tool choice. tells vLLM that you want to enable the model to generate its own tool calls when it deems appropriate. * `--tool-call-parser` -- select the tool parser to use (listed below). Additional tool parsers will continue to be added in the future, and also can register your own tool parsers in the `--tool-parser-plugin`. * `--tool-parser-plugin` -- **optional** tool parser plugin used to register user defined tool parsers into vllm, the registered tool parser name can be specified in `--tool-call-parser`. * `--chat-template` -- **optional** for auto tool choice. the path to the chat template which handles `tool`-role messages and `assistant`-role messages that contain previously generated tool calls. Hermes, Mistral and Llama models have tool-compatible chat templates in their `tokenizer_config.json` files, but you can specify a custom template. This argument can be set to `tool_use` if your model has a tool use-specific chat template configured in the `tokenizer_config.json`. In this case, it will be used per the `transformers` specification. More on this [here](https://huggingface.co/docs/transformers/en/chat_templating#why-do-some-models-have-multiple-templates) from HuggingFace; and you can find an example of this in a `tokenizer_config.json` [here](https://huggingface.co/NousResearch/Hermes-2-Pro-Llama-3-8B/blob/main/tokenizer_config.json) If your favorite tool-calling model is not supported, please feel free to contribute a parser & tool use chat template! ### Hermes Models (`hermes`) All Nous Research Hermes-series models newer than Hermes 2 Pro should be supported. * `NousResearch/Hermes-2-Pro-*` * `NousResearch/Hermes-2-Theta-*` * `NousResearch/Hermes-3-*` _Note that the Hermes 2 **Theta** models are known to have degraded tool call quality & capabilities due to the merge step in their creation_. Flags: `--tool-call-parser hermes` ### Mistral Models (`mistral`) Supported models: * `mistralai/Mistral-7B-Instruct-v0.3` (confirmed) * Additional mistral function-calling models are compatible as well. Known issues: 1. Mistral 7B struggles to generate parallel tool calls correctly. 2. Mistral's `tokenizer_config.json` chat template requires tool call IDs that are exactly 9 digits, which is much shorter than what vLLM generates. Since an exception is thrown when this condition is not met, the following additional chat templates are provided: * `examples/tool_chat_template_mistral.jinja` - this is the "official" Mistral chat template, but tweaked so that it works with vLLM's tool call IDs (provided `tool_call_id` fields are truncated to the last 9 digits) * `examples/tool_chat_template_mistral_parallel.jinja` - this is a "better" version that adds a tool-use system prompt when tools are provided, that results in much better reliability when working with parallel tool calling. Recommended flags: `--tool-call-parser mistral --chat-template examples/tool_chat_template_mistral_parallel.jinja` ### Llama Models (`llama3_json`) Supported models: * `meta-llama/Meta-Llama-3.1-8B-Instruct` * `meta-llama/Meta-Llama-3.1-70B-Instruct` * `meta-llama/Meta-Llama-3.1-405B-Instruct` * `meta-llama/Meta-Llama-3.1-405B-Instruct-FP8` The tool calling that is supported is the [JSON based tool calling](https://llama.meta.com/docs/model-cards-and-prompt-formats/llama3_1/#json-based-tool-calling). For [pythonic tool calling](https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/text_prompt_format.md#zero-shot-function-calling) in Llama-3.2 models, see the `pythonic` tool parser below. Other tool calling formats like the built in python tool calling or custom tool calling are not supported. Known issues: 1. Parallel tool calls are not supported. 2. The model can generate parameters with a wrong format, such as generating an array serialized as string instead of an array. The `tool_chat_template_llama3_json.jinja` file contains the "official" Llama chat template, but tweaked so that it works better with vLLM. Recommended flags: `--tool-call-parser llama3_json --chat-template examples/tool_chat_template_llama3_json.jinja` #### IBM Granite Supported models: * `ibm-granite/granite-3.0-8b-instruct` Recommended flags: `--tool-call-parser granite --chat-template examples/tool_chat_template_granite.jinja` `examples/tool_chat_template_granite.jinja`: this is a modified chat template from the original on Huggingface. Parallel function calls are supported. * `ibm-granite/granite-3.1-8b-instruct` Recommended flags: `--tool-call-parser granite` The chat template from Huggingface can be used directly. Parallel function calls are supported. * `ibm-granite/granite-20b-functioncalling` Recommended flags: `--tool-call-parser granite-20b-fc --chat-template examples/tool_chat_template_granite_20b_fc.jinja` `examples/tool_chat_template_granite_20b_fc.jinja`: this is a modified chat template from the original on Huggingface, which is not vLLM compatible. It blends function description elements from the Hermes template and follows the same system prompt as "Response Generation" mode from [the paper](https://arxiv.org/abs/2407.00121). Parallel function calls are supported. ### InternLM Models (`internlm`) Supported models: * `internlm/internlm2_5-7b-chat` (confirmed) * Additional internlm2.5 function-calling models are compatible as well Known issues: * Although this implementation also supports InternLM2, the tool call results are not stable when testing with the `internlm/internlm2-chat-7b` model. Recommended flags: `--tool-call-parser internlm --chat-template examples/tool_chat_template_internlm2_tool.jinja` ### Jamba Models (`jamba`) AI21's Jamba-1.5 models are supported. * `ai21labs/AI21-Jamba-1.5-Mini` * `ai21labs/AI21-Jamba-1.5-Large` Flags: `--tool-call-parser jamba` ### Models with Pythonic Tool Calls (`pythonic`) A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models. As a concrete example, these models may look up the weather in San Francisco and Seattle by generating: ```python [get_weather(city='San Francisco', metric='celsius'), get_weather(city='Seattle', metric='celsius')] ``` Limitations: * The model must not generate both text and tool calls in the same generation. This may not be hard to change for a specific model, but the community currently lacks consensus on which tokens to emit when starting and ending tool calls. (In particular, the Llama 3.2 models emit no such tokens.) * Llama's smaller models struggle to use tools effectively. Example supported models: * `meta-llama/Llama-3.2-1B-Instruct`\* (use with `examples/tool_chat_template_llama3.2_pythonic.jinja`) * `meta-llama/Llama-3.2-3B-Instruct`\* (use with `examples/tool_chat_template_llama3.2_pythonic.jinja`) * `Team-ACE/ToolACE-8B` (use with `examples/tool_chat_template_toolace.jinja`) * `fixie-ai/ultravox-v0_4-ToolACE-8B` (use with `examples/tool_chat_template_toolace.jinja`) Flags: `--tool-call-parser pythonic --chat-template {see_above}` --- **WARNING** Llama's smaller models frequently fail to emit tool calls in the correct format. Your mileage may vary. --- ## How to write a tool parser plugin A tool parser plugin is a Python file containing one or more ToolParser implementations. You can write a ToolParser similar to the `Hermes2ProToolParser` in vllm/entrypoints/openai/tool_parsers/hermes_tool_parser.py. Here is a summary of a plugin file: ```python # import the required packages # define a tool parser and register it to vllm # the name list in register_module can be used # in --tool-call-parser. you can define as many # tool parsers as you want here. @ToolParserManager.register_module(["example"]) class ExampleToolParser(ToolParser): def __init__(self, tokenizer: AnyTokenizer): super().__init__(tokenizer) # adjust request. e.g.: set skip special tokens # to False for tool call output. def adjust_request( self, request: ChatCompletionRequest) -> ChatCompletionRequest: return request # implement the tool call parse for stream call def extract_tool_calls_streaming( self, previous_text: str, current_text: str, delta_text: str, previous_token_ids: Sequence[int], current_token_ids: Sequence[int], delta_token_ids: Sequence[int], request: ChatCompletionRequest, ) -> Union[DeltaMessage, None]: return delta # implement the tool parse for non-stream call def extract_tool_calls( self, model_output: str, request: ChatCompletionRequest, ) -> ExtractedToolCallInformation: return ExtractedToolCallInformation(tools_called=False, tool_calls=[], content=text) ``` Then you can use this plugin in the command line like this. ```console --enable-auto-tool-choice \ --tool-parser-plugin --tool-call-parser example \ --chat-template \ ``` --- ### Source Getting Started Faq (`docs/source/getting_started/faq.md`) (faq)= # Frequently Asked Questions > Q: How can I serve multiple models on a single port using the OpenAI API? A: Assuming that you're referring to using OpenAI compatible server to serve multiple models at once, that is not currently supported, you can run multiple instances of the server (each serving a different model) at the same time, and have another layer to route the incoming request to the correct server accordingly. ______________________________________________________________________ > Q: Which model to use for offline inference embedding? A: You can try [e5-mistral-7b-instruct](https://huggingface.co/intfloat/e5-mistral-7b-instruct) and [BAAI/bge-base-en-v1.5](https://huggingface.co/BAAI/bge-base-en-v1.5); more are listed [here](#supported-models). By extracting hidden states, vLLM can automatically convert text generation models like [Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B), [Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) into embedding models, but they are expected be inferior to models that are specifically trained on embedding tasks. ______________________________________________________________________ > Q: Can the output of a prompt vary across runs in vLLM? A: Yes, it can. vLLM does not guarantee stable log probabilities (logprobs) for the output tokens. Variations in logprobs may occur due to numerical instability in Torch operations or non-deterministic behavior in batched Torch operations when batching changes. For more details, see the [Numerical Accuracy section](https://pytorch.org/docs/stable/notes/numerical_accuracy.html#batched-computations-or-slice-computations). In vLLM, the same requests might be batched differently due to factors such as other concurrent requests, changes in batch size, or batch expansion in speculative decoding. These batching variations, combined with numerical instability of Torch operations, can lead to slightly different logit/logprob values at each step. Such differences can accumulate, potentially resulting in different tokens being sampled. Once a different token is sampled, further divergence is likely. ## Mitigation Strategies - For improved stability and reduced variance, use `float32`. Note that this will require more memory. - If using `bfloat16`, switching to `float16` can also help. - Using request seeds can aid in achieving more stable generation for temperature > 0, but discrepancies due to precision differences may still occur. --- ### Source Getting Started Troubleshooting (`docs/source/getting_started/troubleshooting.md`) (troubleshooting)= # Troubleshooting This document outlines some troubleshooting strategies you can consider. If you think you've discovered a bug, please [search existing issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue) first to see if it has already been reported. If not, please [file a new issue](https://github.com/vllm-project/vllm/issues/new/choose), providing as much relevant information as possible. ```{note} Once you've debugged a problem, remember to turn off any debugging environment variables defined, or simply start a new shell to avoid being affected by lingering debugging settings. Otherwise, the system might be slow with debugging functionalities left activated. ``` ## Hangs downloading a model If the model isn't already downloaded to disk, vLLM will download it from the internet which can take time and depend on your internet connection. It's recommended to download the model first using the [huggingface-cli](https://huggingface.co/docs/huggingface_hub/en/guides/cli) and passing the local path to the model to vLLM. This way, you can isolate the issue. ## Hangs loading a model from disk If the model is large, it can take a long time to load it from disk. Pay attention to where you store the model. Some clusters have shared filesystems across nodes, e.g. a distributed filesystem or a network filesystem, which can be slow. It'd be better to store the model in a local disk. Additionally, have a look at the CPU memory usage, when the model is too large it might take a lot of CPU memory, slowing down the operating system because it needs to frequently swap between disk and memory. ```{note} To isolate the model downloading and loading issue, you can use the `--load-format dummy` argument to skip loading the model weights. This way, you can check if the model downloading and loading is the bottleneck. ``` ## Model is too large If the model is too large to fit in a single GPU, you might want to [consider tensor parallelism](#distributed-serving) to split the model across multiple GPUs. In that case, every process will read the whole model and split it into chunks, which makes the disk reading time even longer (proportional to the size of tensor parallelism). You can convert the model checkpoint to a sharded checkpoint using . The conversion process might take some time, but later you can load the sharded checkpoint much faster. The model loading time should remain constant regardless of the size of tensor parallelism. ## Enable more logging If other strategies don't solve the problem, it's likely that the vLLM instance is stuck somewhere. You can use the following environment variables to help debug the issue: - `export VLLM_LOGGING_LEVEL=DEBUG` to turn on more logging. - `export CUDA_LAUNCH_BLOCKING=1` to identify which CUDA kernel is causing the problem. - `export NCCL_DEBUG=TRACE` to turn on more logging for NCCL. - `export VLLM_TRACE_FUNCTION=1` to record all function calls for inspection in the log files to tell which function crashes or hangs. ## Incorrect network setup The vLLM instance cannot get the correct IP address if you have a complicated network config. You can find a log such as `DEBUG 06-10 21:32:17 parallel_state.py:88] world_size=8 rank=0 local_rank=0 distributed_init_method=tcp://xxx.xxx.xxx.xxx:54641 backend=nccl` and the IP address should be the correct one. If it's not, override the IP address using the environment variable `export VLLM_HOST_IP=`. You might also need to set `export NCCL_SOCKET_IFNAME=` and `export GLOO_SOCKET_IFNAME=` to specify the network interface for the IP address. ## Error near `self.graph.replay()` If vLLM crashes and the error trace captures it somewhere around `self.graph.replay()` in `vllm/worker/model_runner.py`, it is a CUDA error inside CUDAGraph. To identify the particular CUDA operation that causes the error, you can add `--enforce-eager` to the command line, or `enforce_eager=True` to the {class}`~vllm.LLM` class to disable the CUDAGraph optimization and isolate the exact CUDA operation that causes the error. (troubleshooting-incorrect-hardware-driver)= ## Incorrect hardware/driver If GPU/CPU communication cannot be established, you can use the following Python script and follow the instructions below to confirm whether the GPU/CPU communication is working correctly. ``` /* Detailed source-code truncated for AI context efficiency. */ ``` If you are testing with a single node, adjust `--nproc-per-node` to the number of GPUs you want to use: ```console NCCL_DEBUG=TRACE torchrun --nproc-per-node= test.py ``` If you are testing with multi-nodes, adjust `--nproc-per-node` and `--nnodes` according to your setup and set `MASTER_ADDR` to the correct IP address of the master node, reachable from all nodes. Then, run: ```console NCCL_DEBUG=TRACE torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=$MASTER_ADDR test.py ``` If the script runs successfully, you should see the message `sanity check is successful!`. If the test script hangs or crashes, usually it means the hardware/drivers are broken in some sense. You should try to contact your system administrator or hardware vendor for further assistance. As a common workaround, you can try to tune some NCCL environment variables, such as `export NCCL_P2P_DISABLE=1` to see if it helps. Please check [their documentation](https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html) for more information. Please only use these environment variables as a temporary workaround, as they might affect the performance of the system. The best solution is still to fix the hardware/drivers so that the test script can run successfully. ```{note} A multi-node environment is more complicated than a single-node one. If you see errors such as `torch.distributed.DistNetworkError`, it is likely that the network/DNS setup is incorrect. In that case, you can manually assign node rank and specify the IP via command line arguments: - In the first node, run `NCCL_DEBUG=TRACE torchrun --nnodes 2 --nproc-per-node=2 --node-rank 0 --master_addr $MASTER_ADDR test.py`. - In the second node, run `NCCL_DEBUG=TRACE torchrun --nnodes 2 --nproc-per-node=2 --node-rank 1 --master_addr $MASTER_ADDR test.py`. Adjust `--nproc-per-node`, `--nnodes`, and `--node-rank` according to your setup, being sure to execute different commands (with different `--node-rank`) on different nodes. ``` (troubleshooting-python-multiprocessing)= ## Python multiprocessing ### `RuntimeError` Exception If you have seen a warning in your logs like this: ```console WARNING 12-11 14:50:37 multiproc_worker_utils.py:281] CUDA was previously initialized. We must use the `spawn` multiprocessing start method. Setting VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/getting_started/troubleshooting.html#python-multiprocessing for more information. ``` or an error from Python that looks like this: ```console RuntimeError: An attempt has been made to start a new process before the current process has finished its bootstrapping phase. This probably means that you are not using fork to start your child processes and you have forgotten to use the proper idiom in the main module: if __name__ == '__main__': freeze_support() ... The "freeze_support()" line can be omitted if the program is not going to be frozen to produce an executable. To fix this issue, refer to the "Safe importing of main module" section in https://docs.python.org/3/library/multiprocessing.html ``` then you must update your Python code to guard usage of `vllm` behind a `if __name__ == '__main__':` block. For example, instead of this: ```python import vllm llm = vllm.LLM(...) ``` try this instead: ```python if __name__ == '__main__': import vllm llm = vllm.LLM(...) ``` ## Known Issues - In `v0.5.2`, `v0.5.3`, and `v0.5.3.post1`, there is a bug caused by [zmq](https://github.com/zeromq/pyzmq/issues/2000) , which can occasionally cause vLLM to hang depending on the machine configuration. The solution is to upgrade to the latest version of `vllm` to include the [fix](gh-pr:6759). - To circumvent a NCCL [bug](https://github.com/NVIDIA/nccl/issues/1234) , all vLLM processes will set an environment variable `NCCL_CUMEM_ENABLE=0` to disable NCCL's `cuMem` allocator. It does not affect performance but only gives memory benefits. When external processes want to set up a NCCL connection with vLLM's processes, they should also set this environment variable, otherwise, inconsistent environment setup will cause NCCL to hang or crash, as observed in the [RLHF integration](https://github.com/OpenRLHF/OpenRLHF/pull/604) and the [discussion](gh-issue:5723#issuecomment-2554389656) . --- ### Source Models Generative Models (`docs/source/models/generative_models.md`) (generative-models)= # Generative Models vLLM provides first-class support for generative models, which covers most of LLMs. In vLLM, generative models implement the {class}`~vllm.model_executor.models.VllmModelForTextGeneration` interface. Based on the final hidden states of the input, these models output log probabilities of the tokens to generate, which are then passed through {class}`~vllm.model_executor.layers.Sampler` to obtain the final text. For generative models, the only supported `--task` option is `"generate"`. Usually, this is automatically inferred so you don't have to specify it. ## Offline Inference The {class}`~vllm.LLM` class provides various methods for offline inference. See [Engine Arguments](#engine-args) for a list of options when initializing the model. ### `LLM.generate` The {class}`~vllm.LLM.generate` method is available to all generative models in vLLM. It is similar to [its counterpart in HF Transformers](https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationMixin.generate), except that tokenization and detokenization are also performed automatically. ```python llm = LLM(model="facebook/opt-125m") outputs = llm.generate("Hello, my name is") for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` You can optionally control the language generation by passing {class}`~vllm.SamplingParams`. For example, you can use greedy sampling by setting `temperature=0`: ```python llm = LLM(model="facebook/opt-125m") params = SamplingParams(temperature=0) outputs = llm.generate("Hello, my name is", params) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` A code example can be found here: ### `LLM.beam_search` The {class}`~vllm.LLM.beam_search` method implements [beam search](https://huggingface.co/docs/transformers/en/generation_strategies#beam-search-decoding) on top of {class}`~vllm.LLM.generate`. For example, to search using 5 beams and output at most 50 tokens: ```python llm = LLM(model="facebook/opt-125m") params = BeamSearchParams(beam_width=5, max_tokens=50) outputs = llm.generate("Hello, my name is", params) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` ### `LLM.chat` The {class}`~vllm.LLM.chat` method implements chat functionality on top of {class}`~vllm.LLM.generate`. In particular, it accepts input similar to [OpenAI Chat Completions API](https://platform.openai.com/docs/api-reference/chat) and automatically applies the model's [chat template](https://huggingface.co/docs/transformers/en/chat_templating) to format the prompt. ```{important} In general, only instruction-tuned models have a chat template. Base models may perform poorly as they are not trained to respond to the chat conversation. ``` ```python llm = LLM(model="meta-llama/Meta-Llama-3-8B-Instruct") conversation = [ { "role": "system", "content": "You are a helpful assistant" }, { "role": "user", "content": "Hello" }, { "role": "assistant", "content": "Hello! How can I assist you today?" }, { "role": "user", "content": "Write an essay about the importance of higher education.", }, ] outputs = llm.chat(conversation) for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` A code example can be found here: If the model doesn't have a chat template or you want to specify another one, you can explicitly pass a chat template: ```python from vllm.entrypoints.chat_utils import load_chat_template # You can find a list of existing chat templates under `examples/` custom_template = load_chat_template(chat_template="") print("Loaded chat template:", custom_template) outputs = llm.chat(conversation, chat_template=custom_template) ``` ## Online Serving Our [OpenAI-Compatible Server](#openai-compatible-server) provides endpoints that correspond to the offline APIs: - [Completions API](#completions-api) is similar to `LLM.generate` but only accepts text. - [Chat API](#chat-api) is similar to `LLM.chat`, accepting both text and [multi-modal inputs](#multimodal-inputs) for models with a chat template. --- ### Source Models Pooling Models (`docs/source/models/pooling_models.md`) (pooling-models)= # Pooling Models vLLM also supports pooling models, including embedding, reranking and reward models. In vLLM, pooling models implement the {class}`~vllm.model_executor.models.VllmModelForPooling` interface. These models use a {class}`~vllm.model_executor.layers.Pooler` to extract the final hidden states of the input before returning them. ```{note} We currently support pooling models primarily as a matter of convenience. As shown in the [Compatibility Matrix](#compatibility-matrix), most vLLM features are not applicable to pooling models as they only work on the generation or decode stage, so performance may not improve as much. ``` For pooling models, we support the following `--task` options. The selected option sets the default pooler used to extract the final hidden states: ```{list-table} :widths: 50 25 25 25 :header-rows: 1 * - Task - Pooling Type - Normalization - Softmax * - Embedding (`embed`) - `LAST` - ✅︎ - ✗ * - Classification (`classify`) - `LAST` - ✗ - ✅︎ * - Sentence Pair Scoring (`score`) - \* - \* - \* * - Reward Modeling (`reward`) - `ALL` - ✗ - ✗ ``` \*The default pooler is always defined by the model. ```{note} If the model's implementation in vLLM defines its own pooler, the default pooler is set to that instead of the one specified in this table. ``` When loading [Sentence Transformers](https://huggingface.co/sentence-transformers) models, we attempt to override the default pooler based on its Sentence Transformers configuration file (`modules.json`). ```{tip} You can customize the model's pooling method via the `--override-pooler-config` option, which takes priority over both the model's and Sentence Transformers's defaults. ``` ## Offline Inference The {class}`~vllm.LLM` class provides various methods for offline inference. See [Engine Arguments](#engine-args) for a list of options when initializing the model. ### `LLM.encode` The {class}`~vllm.LLM.encode` method is available to all pooling models in vLLM. It returns the extracted hidden states directly, which is useful for reward models. ```python llm = LLM(model="Qwen/Qwen2.5-Math-RM-72B", task="reward") (output,) = llm.encode("Hello, my name is") data = output.outputs.data print(f"Data: {data!r}") ``` ### `LLM.embed` The {class}`~vllm.LLM.embed` method outputs an embedding vector for each prompt. It is primarily designed for embedding models. ```python llm = LLM(model="intfloat/e5-mistral-7b-instruct", task="embed") (output,) = llm.embed("Hello, my name is") embeds = output.outputs.embedding print(f"Embeddings: {embeds!r} (size={len(embeds)})") ``` A code example can be found here: ### `LLM.classify` The {class}`~vllm.LLM.classify` method outputs a probability vector for each prompt. It is primarily designed for classification models. ```python llm = LLM(model="jason9693/Qwen2.5-1.5B-apeach", task="classify") (output,) = llm.classify("Hello, my name is") probs = output.outputs.probs print(f"Class Probabilities: {probs!r} (size={len(probs)})") ``` A code example can be found here: ### `LLM.score` The {class}`~vllm.LLM.score` method outputs similarity scores between sentence pairs. It is primarily designed for [cross-encoder models](https://www.sbert.net/examples/applications/cross-encoder/README.html). These types of models serve as rerankers between candidate query-document pairs in RAG systems. ```{note} vLLM can only perform the model inference component (e.g. embedding, reranking) of RAG. To handle RAG at a higher level, you should use integration frameworks such as [LangChain](https://github.com/langchain-ai/langchain). ``` ```python llm = LLM(model="BAAI/bge-reranker-v2-m3", task="score") (output,) = llm.score("What is the capital of France?", "The capital of Brazil is Brasilia.") score = output.outputs.score print(f"Score: {score}") ``` A code example can be found here: ## Online Serving Our [OpenAI-Compatible Server](#openai-compatible-server) provides endpoints that correspond to the offline APIs: - [Pooling API](#pooling-api) is similar to `LLM.encode`, being applicable to all types of pooling models. - [Embeddings API](#embeddings-api) is similar to `LLM.embed`, accepting both text and [multi-modal inputs](#multimodal-inputs) for embedding models. - [Score API](#score-api) is similar to `LLM.score` for cross-encoder models. --- ### Source Models Supported Models (`docs/source/models/supported_models.md`) (supported-models)= # List of Supported Models vLLM supports generative and pooling models across various tasks. If a model supports more than one task, you can set the task via the `--task` argument. For each task, we list the model architectures that have been implemented in vLLM. Alongside each architecture, we include some popular models that use it. ## Loading a Model ### HuggingFace Hub By default, vLLM loads models from [HuggingFace (HF) Hub](https://huggingface.co/models). To determine whether a given model is supported, you can check the `config.json` file inside the HF repository. If the `"architectures"` field contains a model architecture listed below, then it should be supported in theory. ````{tip} The easiest way to check if your model is really supported at runtime is to run the program below: ```python from vllm import LLM # For generative models (task=generate) only llm = LLM(model=..., task="generate") # Name or path of your model output = llm.generate("Hello, my name is") print(output) # For pooling models (task={embed,classify,reward,score}) only llm = LLM(model=..., task="embed") # Name or path of your model output = llm.encode("Hello, my name is") print(output) ``` If vLLM successfully returns text (for generative models) or hidden states (for pooling models), it indicates that your model is supported. ```` Otherwise, please refer to [Adding a New Model](#new-model) for instructions on how to implement your model in vLLM. Alternatively, you can [open an issue on GitHub](https://github.com/vllm-project/vllm/issues/new/choose) to request vLLM support. ### ModelScope To use models from [ModelScope](https://www.modelscope.cn) instead of HuggingFace Hub, set an environment variable: ```shell export VLLM_USE_MODELSCOPE=True ``` And use with `trust_remote_code=True`. ```python from vllm import LLM llm = LLM(model=..., revision=..., task=..., trust_remote_code=True) # For generative models (task=generate) only output = llm.generate("Hello, my name is") print(output) # For pooling models (task={embed,classify,reward,score}) only output = llm.encode("Hello, my name is") print(output) ``` ## List of Text-only Language Models ### Generative Models See [this page](#generative-models) for more information on how to use generative models. #### Text Generation (`--task generate`) ``` /* Detailed source-code truncated for AI context efficiency. */ ``` ```{note} Currently, the ROCm version of vLLM supports Mistral and Mixtral only for context lengths up to 4096. ``` ### Pooling Models See [this page](pooling-models) for more information on how to use pooling models. ```{important} Since some model architectures support both generative and pooling tasks, you should explicitly specify the task type to ensure that the model is used in pooling mode instead of generative mode. ``` #### Text Embedding (`--task embed`) ```{list-table} :widths: 25 25 50 5 5 :header-rows: 1 * - Architecture - Models - Example HF Models - [LoRA](#lora-adapter) - [PP](#distributed-serving) * - `BertModel` - BERT-based - `BAAI/bge-base-en-v1.5`, etc. - - * - `Gemma2Model` - Gemma2-based - `BAAI/bge-multilingual-gemma2`, etc. - - ✅︎ * - `GritLM` - GritLM - `parasail-ai/GritLM-7B-vllm`. - ✅︎ - ✅︎ * - `LlamaModel`, `LlamaForCausalLM`, `MistralModel`, etc. - Llama-based - `intfloat/e5-mistral-7b-instruct`, etc. - ✅︎ - ✅︎ * - `Qwen2Model`, `Qwen2ForCausalLM` - Qwen2-based - `ssmits/Qwen2-7B-Instruct-embed-base` (see note), `Alibaba-NLP/gte-Qwen2-7B-instruct` (see note), etc. - ✅︎ - ✅︎ * - `RobertaModel`, `RobertaForMaskedLM` - RoBERTa-based - `sentence-transformers/all-roberta-large-v1`, `sentence-transformers/all-roberta-large-v1`, etc. - - * - `XLMRobertaModel` - XLM-RoBERTa-based - `intfloat/multilingual-e5-large`, etc. - - ``` ```{note} `ssmits/Qwen2-7B-Instruct-embed-base` has an improperly defined Sentence Transformers config. You should manually set mean pooling by passing `--override-pooler-config '{"pooling_type": "MEAN"}'`. ``` ```{note} Unlike base Qwen2, `Alibaba-NLP/gte-Qwen2-7B-instruct` uses bi-directional attention. You can set `--hf-overrides '{"is_causal": false}'` to change the attention mask accordingly. On the other hand, its 1.5B variant (`Alibaba-NLP/gte-Qwen2-1.5B-instruct`) uses causal attention despite being described otherwise on its model card. Regardless of the variant, you need to enable `--trust-remote-code` for the correct tokenizer to be loaded. See [relevant issue on HF Transformers](https://github.com/huggingface/transformers/issues/34882). ``` If your model is not in the above list, we will try to automatically convert the model using {func}`~vllm.model_executor.models.adapters.as_embedding_model`. By default, the embeddings of the whole prompt are extracted from the normalized hidden state corresponding to the last token. #### Reward Modeling (`--task reward`) ```{list-table} :widths: 25 25 50 5 5 :header-rows: 1 * - Architecture - Models - Example HF Models - [LoRA](#lora-adapter) - [PP](#distributed-serving) * - `InternLM2ForRewardModel` - InternLM2-based - `internlm/internlm2-1_8b-reward`, `internlm/internlm2-7b-reward`, etc. - ✅︎ - ✅︎ * - `LlamaForCausalLM` - Llama-based - `peiyi9979/math-shepherd-mistral-7b-prm`, etc. - ✅︎ - ✅︎ * - `Qwen2ForRewardModel` - Qwen2-based - `Qwen/Qwen2.5-Math-RM-72B`, etc. - ✅︎ - ✅︎ * - `Qwen2ForProcessRewardModel` - Qwen2-based - `Qwen/Qwen2.5-Math-PRM-7B`, `Qwen/Qwen2.5-Math-PRM-72B`, etc. - ✅︎ - ✅︎ ``` If your model is not in the above list, we will try to automatically convert the model using {func}`~vllm.model_executor.models.adapters.as_reward_model`. By default, we return the hidden states of each token directly. ```{important} For process-supervised reward models such as `peiyi9979/math-shepherd-mistral-7b-prm`, the pooling config should be set explicitly, e.g.: `--override-pooler-config '{"pooling_type": "STEP", "step_tag_id": 123, "returned_token_ids": [456, 789]}'`. ``` #### Classification (`--task classify`) ```{list-table} :widths: 25 25 50 5 5 :header-rows: 1 * - Architecture - Models - Example HF Models - [LoRA](#lora-adapter) - [PP](#distributed-serving) * - `JambaForSequenceClassification` - Jamba - `ai21labs/Jamba-tiny-reward-dev`, etc. - ✅︎ - ✅︎ * - `Qwen2ForSequenceClassification` - Qwen2-based - `jason9693/Qwen2.5-1.5B-apeach`, etc. - ✅︎ - ✅︎ ``` If your model is not in the above list, we will try to automatically convert the model using {func}`~vllm.model_executor.models.adapters.as_classification_model`. By default, the class probabilities are extracted from the softmaxed hidden state corresponding to the last token. #### Sentence Pair Scoring (`--task score`) ```{list-table} :widths: 25 25 50 5 5 :header-rows: 1 * - Architecture - Models - Example HF Models - [LoRA](#lora-adapter) - [PP](#distributed-serving) * - `BertForSequenceClassification` - BERT-based - `cross-encoder/ms-marco-MiniLM-L-6-v2`, etc. - - * - `RobertaForSequenceClassification` - RoBERTa-based - `cross-encoder/quora-roberta-base`, etc. - - * - `XLMRobertaForSequenceClassification` - XLM-RoBERTa-based - `BAAI/bge-reranker-v2-m3`, etc. - - ``` (supported-mm-models)= ## List of Multimodal Language Models The following modalities are supported depending on the model: - **T**ext - **I**mage - **V**ideo - **A**udio Any combination of modalities joined by `+` are supported. - e.g.: `T + I` means that the model supports text-only, image-only, and text-with-image inputs. On the other hand, modalities separated by `/` are mutually exclusive. - e.g.: `T / I` means that the model supports text-only and image-only inputs, but not text-with-image inputs. See [this page](#multimodal-inputs) on how to pass multi-modal inputs to the model. ````{important} To enable multiple multi-modal items per text prompt, you have to set `limit_mm_per_prompt` (offline inference) or `--limit-mm-per-prompt` (online serving). For example, to enable passing up to 4 images per text prompt: Offline inference: ```python llm = LLM( model="Qwen/Qwen2-VL-7B-Instruct", limit_mm_per_prompt={"image": 4}, ) ``` Online serving: ```bash vllm serve Qwen/Qwen2-VL-7B-Instruct --limit-mm-per-prompt image=4 ``` ```` ```{note} vLLM currently only supports adding LoRA to the language backbone of multimodal models. ``` ### Generative Models See [this page](#generative-models) for more information on how to use generative models. #### Text Generation (`--task generate`) ``` /* Detailed source-code truncated for AI context efficiency. */ ``` E Pre-computed embeddings can be inputted for this modality. + Multiple items can be inputted per text prompt for this modality. ```{note} To use `DeepSeek-VL2` series models, you have to pass `--hf_overrides '{"architectures": ["DeepseekVLV2ForCausalLM"]}'` when running vLLM. ``` ```{note} To use `TIGER-Lab/Mantis-8B-siglip-llama3`, you have to pass `--hf_overrides '{"architectures": ["MantisForConditionalGeneration"]}'` when running vLLM. ``` ```{note} The official `openbmb/MiniCPM-V-2` doesn't work yet, so we need to use a fork (`HwwwH/MiniCPM-V-2`) for now. For more details, please see: ``` ```{note} The chat template for Pixtral-HF is incorrect (see [discussion](https://huggingface.co/mistral-community/pixtral-12b/discussions/22)). A corrected version is available at . ``` ### Pooling Models See [this page](pooling-models) for more information on how to use pooling models. ```{important} Since some model architectures support both generative and pooling tasks, you should explicitly specify the task type to ensure that the model is used in pooling mode instead of generative mode. ``` #### Text Embedding (`--task embed`) Any text generation model can be converted into an embedding model by passing `--task embed`. ```{note} To get the best results, you should use pooling models that are specifically trained as such. ``` The following table lists those that are tested in vLLM. ```{list-table} :widths: 25 25 15 25 5 5 :header-rows: 1 * - Architecture - Models - Inputs - Example HF Models - [LoRA](#lora-adapter) - [PP](#distributed-serving) * - `LlavaNextForConditionalGeneration` - LLaVA-NeXT-based - T / I - `royokong/e5-v` - - ✅︎ * - `Phi3VForCausalLM` - Phi-3-Vision-based - T + I - `TIGER-Lab/VLM2Vec-Full` - 🚧 - ✅︎ * - `Qwen2VLForConditionalGeneration` - Qwen2-VL-based - T + I - `MrLight/dse-qwen2-2b-mrl-v1` - - ✅︎ ``` _________________ ## Model Support Policy At vLLM, we are committed to facilitating the integration and support of third-party models within our ecosystem. Our approach is designed to balance the need for robustness and the practical limitations of supporting a wide range of models. Here’s how we manage third-party model support: 1. **Community-Driven Support**: We encourage community contributions for adding new models. When a user requests support for a new model, we welcome pull requests (PRs) from the community. These contributions are evaluated primarily on the sensibility of the output they generate, rather than strict consistency with existing implementations such as those in transformers. **Call for contribution:** PRs coming directly from model vendors are greatly appreciated! 2. **Best-Effort Consistency**: While we aim to maintain a level of consistency between the models implemented in vLLM and other frameworks like transformers, complete alignment is not always feasible. Factors like acceleration techniques and the use of low-precision computations can introduce discrepancies. Our commitment is to ensure that the implemented models are functional and produce sensible results. ```{tip} When comparing the output of `model.generate` from HuggingFace Transformers with the output of `llm.generate` from vLLM, note that the former reads the model's generation config file (i.e., [generation_config.json](https://github.com/huggingface/transformers/blob/19dabe96362803fb0a9ae7073d03533966598b17/src/transformers/generation/utils.py#L1945)) and applies the default parameters for generation, while the latter only uses the parameters passed to the function. Ensure all sampling parameters are identical when comparing outputs. ``` 3. **Issue Resolution and Model Updates**: Users are encouraged to report any bugs or issues they encounter with third-party models. Proposed fixes should be submitted via PRs, with a clear explanation of the problem and the rationale behind the proposed solution. If a fix for one model impacts another, we rely on the community to highlight and address these cross-model dependencies. Note: for bugfix PRs, it is good etiquette to inform the original author to seek their feedback. 4. **Monitoring and Updates**: Users interested in specific models should monitor the commit history for those models (e.g., by tracking changes in the main/vllm/model_executor/models directory). This proactive approach helps users stay informed about updates and changes that may affect the models they use. 5. **Selective Focus**: Our resources are primarily directed towards models with significant user interest and impact. Models that are less frequently used may receive less attention, and we rely on the community to play a more active role in their upkeep and improvement. Through this approach, vLLM fosters a collaborative environment where both the core development team and the broader community contribute to the robustness and diversity of the third-party models supported in our ecosystem. Note that, as an inference engine, vLLM does not introduce new models. Therefore, all models supported by vLLM are third-party models in this regard. We have the following levels of testing for models: 1. **Strict Consistency**: We compare the output of the model with the output of the model in the HuggingFace Transformers library under greedy decoding. This is the most stringent test. Please refer to [models tests](https://github.com/vllm-project/vllm/blob/main/tests/models) for the models that have passed this test. 2. **Output Sensibility**: We check if the output of the model is sensible and coherent, by measuring the perplexity of the output and checking for any obvious errors. This is a less stringent test. 3. **Runtime Functionality**: We check if the model can be loaded and run without errors. This is the least stringent test. Please refer to [functionality tests](gh-dir:tests) and [examples](gh-dir:main/examples) for the models that have passed this test. 4. **Community Feedback**: We rely on the community to provide feedback on the models. If a model is broken or not working as expected, we encourage users to raise issues to report it or open pull requests to fix it. The rest of the models fall under this category. --- ### Source Performance Benchmarks (`docs/source/performance/benchmarks.md`) (benchmarks)= # Benchmark Suites vLLM contains two sets of benchmarks: - [Performance benchmarks](#performance-benchmarks) - [Nightly benchmarks](#nightly-benchmarks) (performance-benchmarks)= ## Performance Benchmarks The performance benchmarks are used for development to confirm whether new changes improve performance under various workloads. They are triggered on every commit with both the `perf-benchmarks` and `ready` labels, and when a PR is merged into vLLM. The latest performance results are hosted on the public [vLLM Performance Dashboard](https://perf.vllm.ai). More information on the performance benchmarks and their parameters can be found [here](gh-file:.buildkite/nightly-benchmarks/performance-benchmarks-descriptions.md). (nightly-benchmarks)= ## Nightly Benchmarks These compare vLLM's performance against alternatives (`tgi`, `trt-llm`, and `lmdeploy`) when there are major updates of vLLM (e.g., bumping up to a new version). They are primarily intended for consumers to evaluate when to choose vLLM over other options and are triggered on every commit with both the `perf-benchmarks` and `nightly-benchmarks` labels. The latest nightly benchmark results are shared in major release blog posts such as [vLLM v0.6.0](https://blog.vllm.ai/2024/09/05/perf-update.html). More information on the nightly benchmarks and their parameters can be found [here](gh-file:.buildkite/nightly-benchmarks/nightly-descriptions.md). --- ### Source Performance Optimization (`docs/source/performance/optimization.md`) (optimization-and-tuning)= # Optimization and Tuning ## Preemption Due to the auto-regressive nature of transformer architecture, there are times when KV cache space is insufficient to handle all batched requests. The vLLM can preempt requests to free up KV cache space for other requests. Preempted requests are recomputed when sufficient KV cache space becomes available again. When this occurs, the following warning is printed: ```text WARNING 05-09 00:49:33 scheduler.py:1057 Sequence group 0 is preempted by PreemptionMode.SWAP mode because there is not enough KV cache space. This can affect the end-to-end performance. Increase gpu_memory_utilization or tensor_parallel_size to provide more KV cache memory. total_cumulative_preemption_cnt=1 ``` While this mechanism ensures system robustness, preemption and recomputation can adversely affect end-to-end latency. If you frequently encounter preemptions from the vLLM engine, consider the following actions: - Increase `gpu_memory_utilization`. The vLLM pre-allocates GPU cache by using gpu_memory_utilization% of memory. By increasing this utilization, you can provide more KV cache space. - Decrease `max_num_seqs` or `max_num_batched_tokens`. This can reduce the number of concurrent requests in a batch, thereby requiring less KV cache space. - Increase `tensor_parallel_size`. This approach shards model weights, so each GPU has more memory available for KV cache. You can also monitor the number of preemption requests through Prometheus metrics exposed by the vLLM. Additionally, you can log the cumulative number of preemption requests by setting disable_log_stats=False. (chunked-prefill)= ## Chunked Prefill vLLM supports an experimental feature chunked prefill. Chunked prefill allows to chunk large prefills into smaller chunks and batch them together with decode requests. You can enable the feature by specifying `--enable-chunked-prefill` in the command line or setting `enable_chunked_prefill=True` in the LLM constructor. ```python llm = LLM(model="meta-llama/Llama-2-7b-hf", enable_chunked_prefill=True) # Set max_num_batched_tokens to tune performance. # NOTE: 2048 is the default max_num_batched_tokens for chunked prefill. # llm = LLM(model="meta-llama/Llama-2-7b-hf", enable_chunked_prefill=True, max_num_batched_tokens=2048) ``` By default, vLLM scheduler prioritizes prefills and doesn't batch prefill and decode to the same batch. This policy optimizes the TTFT (time to the first token), but incurs slower ITL (inter token latency) and inefficient GPU utilization. Once chunked prefill is enabled, the policy is changed to prioritize decode requests. It batches all pending decode requests to the batch before scheduling any prefill. When there are available token_budget (`max_num_batched_tokens`), it schedules pending prefills. If a last pending prefill request cannot fit into `max_num_batched_tokens`, it chunks it. This policy has two benefits: - It improves ITL and generation decode because decode requests are prioritized. - It helps achieve better GPU utilization by locating compute-bound (prefill) and memory-bound (decode) requests to the same batch. You can tune the performance by changing `max_num_batched_tokens`. By default, it is set to 2048. Smaller `max_num_batched_tokens` achieves better ITL because there are fewer prefills interrupting decodes. Higher `max_num_batched_tokens` achieves better TTFT as you can put more prefill to the batch. - If `max_num_batched_tokens` is the same as `max_model_len`, that's almost the equivalent to the default scheduling policy (except that it still prioritizes decodes). - Note that the default value (2048) of `max_num_batched_tokens` is optimized for ITL, and it may have lower throughput than the default scheduler. We recommend you set `max_num_batched_tokens > 2048` for throughput. See related papers for more details ( or ). Please try out this feature and let us know your feedback via GitHub issues! --- ### Source Serving Distributed Serving (`docs/source/serving/distributed_serving.md`) (distributed-serving)= # Distributed Inference and Serving ## How to decide the distributed inference strategy? Before going into the details of distributed inference and serving, let's first make it clear when to use distributed inference and what are the strategies available. The common practice is: - **Single GPU (no distributed inference)**: If your model fits in a single GPU, you probably don't need to use distributed inference. Just use the single GPU to run the inference. - **Single-Node Multi-GPU (tensor parallel inference)**: If your model is too large to fit in a single GPU, but it can fit in a single node with multiple GPUs, you can use tensor parallelism. The tensor parallel size is the number of GPUs you want to use. For example, if you have 4 GPUs in a single node, you can set the tensor parallel size to 4. - **Multi-Node Multi-GPU (tensor parallel plus pipeline parallel inference)**: If your model is too large to fit in a single node, you can use tensor parallel together with pipeline parallelism. The tensor parallel size is the number of GPUs you want to use in each node, and the pipeline parallel size is the number of nodes you want to use. For example, if you have 16 GPUs in 2 nodes (8 GPUs per node), you can set the tensor parallel size to 8 and the pipeline parallel size to 2. In short, you should increase the number of GPUs and the number of nodes until you have enough GPU memory to hold the model. The tensor parallel size should be the number of GPUs in each node, and the pipeline parallel size should be the number of nodes. After adding enough GPUs and nodes to hold the model, you can run vLLM first, which will print some logs like `# GPU blocks: 790`. Multiply the number by `16` (the block size), and you can get roughly the maximum number of tokens that can be served on the current configuration. If this number is not satisfying, e.g. you want higher throughput, you can further increase the number of GPUs or nodes, until the number of blocks is enough. ```{note} There is one edge case: if the model fits in a single node with multiple GPUs, but the number of GPUs cannot divide the model size evenly, you can use pipeline parallelism, which splits the model along layers and supports uneven splits. In this case, the tensor parallel size should be 1 and the pipeline parallel size should be the number of GPUs. ``` ## Running vLLM on a single node vLLM supports distributed tensor-parallel and pipeline-parallel inference and serving. Currently, we support [Megatron-LM's tensor parallel algorithm](https://arxiv.org/pdf/1909.08053.pdf). We manage the distributed runtime with either [Ray](https://github.com/ray-project/ray) or python native multiprocessing. Multiprocessing can be used when deploying on a single node, multi-node inferencing currently requires Ray. Multiprocessing will be used by default when not running in a Ray placement group and if there are sufficient GPUs available on the same node for the configured `tensor_parallel_size`, otherwise Ray will be used. This default can be overridden via the `LLM` class `distributed_executor_backend` argument or `--distributed-executor-backend` API server argument. Set it to `mp` for multiprocessing or `ray` for Ray. It's not required for Ray to be installed for the multiprocessing case. To run multi-GPU inference with the `LLM` class, set the `tensor_parallel_size` argument to the number of GPUs you want to use. For example, to run inference on 4 GPUs: ```python from vllm import LLM llm = LLM("facebook/opt-13b", tensor_parallel_size=4) output = llm.generate("San Franciso is a") ``` To run multi-GPU serving, pass in the `--tensor-parallel-size` argument when starting the server. For example, to run API server on 4 GPUs: ```console vllm serve facebook/opt-13b \ --tensor-parallel-size 4 ``` You can also additionally specify `--pipeline-parallel-size` to enable pipeline parallelism. For example, to run API server on 8 GPUs with pipeline parallelism and tensor parallelism: ```console vllm serve gpt2 \ --tensor-parallel-size 4 \ --pipeline-parallel-size 2 ``` ## Running vLLM on multiple nodes If a single node does not have enough GPUs to hold the model, you can run the model using multiple nodes. It is important to make sure the execution environment is the same on all nodes, including the model path, the Python environment. The recommended way is to use docker images to ensure the same environment, and hide the heterogeneity of the host machines via mapping them into the same docker configuration. The first step, is to start containers and organize them into a cluster. We have provided the helper script to start the cluster. Please note, this script launches docker without administrative privileges that would be required to access GPU performance counters when running profiling and tracing tools. For that purpose, the script can have `CAP_SYS_ADMIN` to the docker container by using the `--cap-add` option in the docker run command. Pick a node as the head node, and run the following command: ```console bash run_cluster.sh \ vllm/vllm-openai \ ip_of_head_node \ --head \ /path/to/the/huggingface/home/in/this/node ``` On the rest of the worker nodes, run the following command: ```console bash run_cluster.sh \ vllm/vllm-openai \ ip_of_head_node \ --worker \ /path/to/the/huggingface/home/in/this/node ``` Then you get a ray cluster of containers. Note that you need to keep the shells running these commands alive to hold the cluster. Any shell disconnect will terminate the cluster. In addition, please note that the argument `ip_of_head_node` should be the IP address of the head node, which is accessible by all the worker nodes. A common misunderstanding is to use the IP address of the worker node, which is not correct. Then, on any node, use `docker exec -it node /bin/bash` to enter the container, execute `ray status` to check the status of the Ray cluster. You should see the right number of nodes and GPUs. After that, on any node, you can use vLLM as usual, just as you have all the GPUs on one node. The common practice is to set the tensor parallel size to the number of GPUs in each node, and the pipeline parallel size to the number of nodes. For example, if you have 16 GPUs in 2 nodes (8 GPUs per node), you can set the tensor parallel size to 8 and the pipeline parallel size to 2: ```console vllm serve /path/to/the/model/in/the/container \ --tensor-parallel-size 8 \ --pipeline-parallel-size 2 ``` You can also use tensor parallel without pipeline parallel, just set the tensor parallel size to the number of GPUs in the cluster. For example, if you have 16 GPUs in 2 nodes (8 GPUs per node), you can set the tensor parallel size to 16: ```console vllm serve /path/to/the/model/in/the/container \ --tensor-parallel-size 16 ``` To make tensor parallel performant, you should make sure the communication between nodes is efficient, e.g. using high-speed network cards like Infiniband. To correctly set up the cluster to use Infiniband, append additional arguments like `--privileged -e NCCL_IB_HCA=mlx5` to the `run_cluster.sh` script. Please contact your system administrator for more information on how to set up the flags. One way to confirm if the Infiniband is working is to run vLLM with `NCCL_DEBUG=TRACE` environment variable set, e.g. `NCCL_DEBUG=TRACE vllm serve ...` and check the logs for the NCCL version and the network used. If you find `[send] via NET/Socket` in the logs, it means NCCL uses raw TCP Socket, which is not efficient for cross-node tensor parallel. If you find `[send] via NET/IB/GDRDMA` in the logs, it means NCCL uses Infiniband with GPU-Direct RDMA, which is efficient. ```{warning} After you start the Ray cluster, you'd better also check the GPU-GPU communication between nodes. It can be non-trivial to set up. Please refer to the [sanity check script](#troubleshooting-incorrect-hardware-driver) for more information. If you need to set some environment variables for the communication configuration, you can append them to the `run_cluster.sh` script, e.g. `-e NCCL_SOCKET_IFNAME=eth0`. Note that setting environment variables in the shell (e.g. `NCCL_SOCKET_IFNAME=eth0 vllm serve ...`) only works for the processes in the same node, not for the processes in the other nodes. Setting environment variables when you create the cluster is the recommended way. See for more information. ``` ```{warning} Please make sure you downloaded the model to all the nodes (with the same path), or the model is downloaded to some distributed file system that is accessible by all nodes. When you use huggingface repo id to refer to the model, you should append your huggingface token to the `run_cluster.sh` script, e.g. `-e HF_TOKEN=`. The recommended way is to download the model first, and then use the path to refer to the model. ``` --- ### Source Serving Engine Args (`docs/source/serving/engine_args.md`) (engine-args)= # Engine Arguments Below, you can find an explanation of every engine argument for vLLM: ```{eval-rst} .. argparse:: :module: vllm.engine.arg_utils :func: _engine_args_parser :prog: vllm serve :nodefaultconst: ``` ## Async Engine Arguments Below are the additional arguments related to the asynchronous engine: ```{eval-rst} .. argparse:: :module: vllm.engine.arg_utils :func: _async_engine_args_parser :prog: vllm serve :nodefaultconst: ``` --- ### Source Serving Env Vars (`docs/source/serving/env_vars.md`) # Environment Variables vLLM uses the following environment variables to configure the system: ```{warning} Please note that `VLLM_PORT` and `VLLM_HOST_IP` set the port and ip for vLLM's **internal usage**. It is not the port and ip for the API server. If you use `--host $VLLM_HOST_IP` and `--port $VLLM_PORT` to start the API server, it will not work. All environment variables used by vLLM are prefixed with `VLLM_`. **Special care should be taken for Kubernetes users**: please do not name the service as `vllm`, otherwise environment variables set by Kubernetes might conflict with vLLM's environment variables, because [Kubernetes sets environment variables for each service with the capitalized service name as the prefix](https://kubernetes.io/docs/concepts/services-networking/service/#environment-variables). ``` ```{literalinclude} ../../../vllm/envs.py :end-before: end-env-vars-definition :language: python :start-after: begin-env-vars-definition ``` --- ### Source Serving Metrics (`docs/source/serving/metrics.md`) # Production Metrics vLLM exposes a number of metrics that can be used to monitor the health of the system. These metrics are exposed via the `/metrics` endpoint on the vLLM OpenAI compatible API server. You can start the server using Python, or using [Docker](#deployment-docker): ```console vllm serve unsloth/Llama-3.2-1B-Instruct ``` Then query the endpoint to get the latest metrics from the server: ```console $ curl http://0.0.0.0:8000/metrics # HELP vllm:iteration_tokens_total Histogram of number of tokens per engine_step. # TYPE vllm:iteration_tokens_total histogram vllm:iteration_tokens_total_sum{model_name="unsloth/Llama-3.2-1B-Instruct"} 0.0 vllm:iteration_tokens_total_bucket{le="1.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 vllm:iteration_tokens_total_bucket{le="8.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 vllm:iteration_tokens_total_bucket{le="16.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 vllm:iteration_tokens_total_bucket{le="32.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 vllm:iteration_tokens_total_bucket{le="64.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 vllm:iteration_tokens_total_bucket{le="128.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 vllm:iteration_tokens_total_bucket{le="256.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 vllm:iteration_tokens_total_bucket{le="512.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0 ... ``` The following metrics are exposed: ```{literalinclude} ../../../vllm/engine/metrics.py :end-before: end-metrics-definitions :language: python :start-after: begin-metrics-definitions ``` --- ### Source Serving Multimodal Inputs (`docs/source/serving/multimodal_inputs.md`) (multimodal-inputs)= # Multimodal Inputs This page teaches you how to pass multi-modal inputs to [multi-modal models](#supported-mm-models) in vLLM. ```{note} We are actively iterating on multi-modal support. See [this RFC](gh-issue:4194) for upcoming changes, and [open an issue on GitHub](https://github.com/vllm-project/vllm/issues/new/choose) if you have any feedback or feature requests. ``` ## Offline Inference To input multi-modal data, follow this schema in {class}`vllm.inputs.PromptType`: - `prompt`: The prompt should follow the format that is documented on HuggingFace. - `multi_modal_data`: This is a dictionary that follows the schema defined in {class}`vllm.multimodal.inputs.MultiModalDataDict`. ### Image You can pass a single image to the `'image'` field of the multi-modal dictionary, as shown in the following examples: ```python llm = LLM(model="llava-hf/llava-1.5-7b-hf") # Refer to the HuggingFace repo for the correct format to use prompt = "USER: \nWhat is the content of this image?\nASSISTANT:" # Load the image using PIL.Image image = PIL.Image.open(...) # Single prompt inference outputs = llm.generate({ "prompt": prompt, "multi_modal_data": {"image": image}, }) for o in outputs: generated_text = o.outputs[0].text print(generated_text) # Batch inference image_1 = PIL.Image.open(...) image_2 = PIL.Image.open(...) outputs = llm.generate( [ { "prompt": "USER: \nWhat is the content of this image?\nASSISTANT:", "multi_modal_data": {"image": image_1}, }, { "prompt": "USER: \nWhat's the color of this image?\nASSISTANT:", "multi_modal_data": {"image": image_2}, } ] ) for o in outputs: generated_text = o.outputs[0].text print(generated_text) ``` Full example: To substitute multiple images inside the same text prompt, you can pass in a list of images instead: ```python llm = LLM( model="microsoft/Phi-3.5-vision-instruct", trust_remote_code=True, # Required to load Phi-3.5-vision max_model_len=4096, # Otherwise, it may not fit in smaller GPUs limit_mm_per_prompt={"image": 2}, # The maximum number to accept ) # Refer to the HuggingFace repo for the correct format to use prompt = "<|user|>\n<|image_1|>\n<|image_2|>\nWhat is the content of each image?<|end|>\n<|assistant|>\n" # Load the images using PIL.Image image1 = PIL.Image.open(...) image2 = PIL.Image.open(...) outputs = llm.generate({ "prompt": prompt, "multi_modal_data": { "image": [image1, image2] }, }) for o in outputs: generated_text = o.outputs[0].text print(generated_text) ``` Full example: Multi-image input can be extended to perform video captioning. We show this with [Qwen2-VL](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct) as it supports videos: ```python # Specify the maximum number of frames per video to be 4. This can be changed. llm = LLM("Qwen/Qwen2-VL-2B-Instruct", limit_mm_per_prompt={"image": 4}) # Create the request payload. video_frames = ... # load your video making sure it only has the number of frames specified earlier. message = { "role": "user", "content": [ {"type": "text", "text": "Describe this set of frames. Consider the frames to be a part of the same video."}, ], } for i in range(len(video_frames)): base64_image = encode_image(video_frames[i]) # base64 encoding. new_image = {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}} message["content"].append(new_image) # Perform inference and log output. outputs = llm.chat([message]) for o in outputs: generated_text = o.outputs[0].text print(generated_text) ``` ### Video You can pass a list of NumPy arrays directly to the `'video'` field of the multi-modal dictionary instead of using multi-image input. Full example: ### Audio You can pass a tuple `(array, sampling_rate)` to the `'audio'` field of the multi-modal dictionary. Full example: ### Embedding To input pre-computed embeddings belonging to a data type (i.e. image, video, or audio) directly to the language model, pass a tensor of shape `(num_items, feature_size, hidden_size of LM)` to the corresponding field of the multi-modal dictionary. ```python # Inference with image embeddings as input llm = LLM(model="llava-hf/llava-1.5-7b-hf") # Refer to the HuggingFace repo for the correct format to use prompt = "USER: \nWhat is the content of this image?\nASSISTANT:" # Embeddings for single image # torch.Tensor of shape (1, image_feature_size, hidden_size of LM) image_embeds = torch.load(...) outputs = llm.generate({ "prompt": prompt, "multi_modal_data": {"image": image_embeds}, }) for o in outputs: generated_text = o.outputs[0].text print(generated_text) ``` For Qwen2-VL and MiniCPM-V, we accept additional parameters alongside the embeddings: ```python # Construct the prompt based on your model prompt = ... # Embeddings for multiple images # torch.Tensor of shape (num_images, image_feature_size, hidden_size of LM) image_embeds = torch.load(...) # Qwen2-VL llm = LLM("Qwen/Qwen2-VL-2B-Instruct", limit_mm_per_prompt={"image": 4}) mm_data = { "image": { "image_embeds": image_embeds, # image_grid_thw is needed to calculate positional encoding. "image_grid_thw": torch.load(...), # torch.Tensor of shape (1, 3), } } # MiniCPM-V llm = LLM("openbmb/MiniCPM-V-2_6", trust_remote_code=True, limit_mm_per_prompt={"image": 4}) mm_data = { "image": { "image_embeds": image_embeds, # image_size_list is needed to calculate details of the sliced image. "image_size_list": [image.size for image in images], # list of image sizes } } outputs = llm.generate({ "prompt": prompt, "multi_modal_data": mm_data, }) for o in outputs: generated_text = o.outputs[0].text print(generated_text) ``` ## Online Serving Our OpenAI-compatible server accepts multi-modal data via the [Chat Completions API](https://platform.openai.com/docs/api-reference/chat). ```{important} A chat template is **required** to use Chat Completions API. Although most models come with a chat template, for others you have to define one yourself. The chat template can be inferred based on the documentation on the model's HuggingFace repo. For example, LLaVA-1.5 (`llava-hf/llava-1.5-7b-hf`) requires a chat template that can be found here: ``` ### Image Image input is supported according to [OpenAI Vision API](https://platform.openai.com/docs/guides/vision). Here is a simple example using Phi-3.5-Vision. First, launch the OpenAI-compatible server: ```bash vllm serve microsoft/Phi-3.5-vision-instruct --task generate \ --trust-remote-code --max-model-len 4096 --limit-mm-per-prompt image=2 ``` Then, you can use the OpenAI client as follows: ``` /* Detailed source-code truncated for AI context efficiency. */ ``` Full example: ```{tip} Loading from local file paths is also supported on vLLM: You can specify the allowed local media path via `--allowed-local-media-path` when launching the API server/engine, and pass the file path as `url` in the API request. ``` ```{tip} There is no need to place image placeholders in the text content of the API request - they are already represented by the image content. In fact, you can place image placeholders in the middle of the text by interleaving text and image content. ``` ````{note} By default, the timeout for fetching images through HTTP URL is `5` seconds. You can override this by setting the environment variable: ```console $ export VLLM_IMAGE_FETCH_TIMEOUT= ``` ```` ### Video Instead of `image_url`, you can pass a video file via `video_url`. Here is a simple example using [LLaVA-OneVision](https://huggingface.co/llava-hf/llava-onevision-qwen2-0.5b-ov-hf). First, launch the OpenAI-compatible server: ```bash vllm serve llava-hf/llava-onevision-qwen2-0.5b-ov-hf --task generate --max-model-len 8192 ``` Then, you can use the OpenAI client as follows: ```python from openai import OpenAI openai_api_key = "EMPTY" openai_api_base = "http://localhost:8000/v1" client = OpenAI( api_key=openai_api_key, base_url=openai_api_base, ) video_url = "http://commondatastorage.googleapis.com/gtv-videos-bucket/sample/ForBiggerFun.mp4" ## Use video url in the payload chat_completion_from_url = client.chat.completions.create( messages=[{ "role": "user", "content": [ { "type": "text", "text": "What's in this video?" }, { "type": "video_url", "video_url": { "url": video_url }, }, ], }], model=model, max_completion_tokens=64, ) result = chat_completion_from_url.choices[0].message.content print("Chat completion output from image url:", result) ``` Full example: ````{note} By default, the timeout for fetching videos through HTTP URL is `30` seconds. You can override this by setting the environment variable: ```console $ export VLLM_VIDEO_FETCH_TIMEOUT= ``` ```` ### Audio Audio input is supported according to [OpenAI Audio API](https://platform.openai.com/docs/guides/audio?audio-generation-quickstart-example=audio-in). Here is a simple example using Ultravox-v0.3. First, launch the OpenAI-compatible server: ```bash vllm serve fixie-ai/ultravox-v0_3 ``` Then, you can use the OpenAI client as follows: ```python import base64 import requests from openai import OpenAI from vllm.assets.audio import AudioAsset def encode_base64_content_from_url(content_url: str) -> str: """Encode a content retrieved from a remote url to base64 format.""" with requests.get(content_url) as response: response.raise_for_status() result = base64.b64encode(response.content).decode('utf-8') return result openai_api_key = "EMPTY" openai_api_base = "http://localhost:8000/v1" client = OpenAI( api_key=openai_api_key, base_url=openai_api_base, ) # Any format supported by librosa is supported audio_url = AudioAsset("winning_call").url audio_base64 = encode_base64_content_from_url(audio_url) chat_completion_from_base64 = client.chat.completions.create( messages=[{ "role": "user", "content": [ { "type": "text", "text": "What's in this audio?" }, { "type": "input_audio", "input_audio": { "data": audio_base64, "format": "wav" }, }, ], }], model=model, max_completion_tokens=64, ) result = chat_completion_from_base64.choices[0].message.content print("Chat completion output from input audio:", result) ``` Alternatively, you can pass `audio_url`, which is the audio counterpart of `image_url` for image input: ```python chat_completion_from_url = client.chat.completions.create( messages=[{ "role": "user", "content": [ { "type": "text", "text": "What's in this audio?" }, { "type": "audio_url", "audio_url": { "url": audio_url }, }, ], }], model=model, max_completion_tokens=64, ) result = chat_completion_from_url.choices[0].message.content print("Chat completion output from audio url:", result) ``` Full example: ````{note} By default, the timeout for fetching audios through HTTP URL is `10` seconds. You can override this by setting the environment variable: ```console $ export VLLM_AUDIO_FETCH_TIMEOUT= ``` ```` ### Embedding vLLM's Embeddings API is a superset of OpenAI's [Embeddings API](https://platform.openai.com/docs/api-reference/embeddings), where a list of chat `messages` can be passed instead of batched `inputs`. This enables multi-modal inputs to be passed to embedding models. ```{tip} The schema of `messages` is exactly the same as in Chat Completions API. You can refer to the above tutorials for more details on how to pass each type of multi-modal data. ``` Usually, embedding models do not expect chat-based input, so we need to use a custom chat template to format the text and images. Refer to the examples below for illustration. Here is an end-to-end example using VLM2Vec. To serve the model: ```bash vllm serve TIGER-Lab/VLM2Vec-Full --task embed \ --trust-remote-code --max-model-len 4096 --chat-template examples/template_vlm2vec.jinja ``` ```{important} Since VLM2Vec has the same model architecture as Phi-3.5-Vision, we have to explicitly pass `--task embed` to run this model in embedding mode instead of text generation mode. The custom chat template is completely different from the original one for this model, and can be found here: ``` Since the request schema is not defined by OpenAI client, we post a request to the server using the lower-level `requests` library: ```python import requests image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" response = requests.post( "http://localhost:8000/v1/embeddings", json={ "model": "TIGER-Lab/VLM2Vec-Full", "messages": [{ "role": "user", "content": [ {"type": "image_url", "image_url": {"url": image_url}}, {"type": "text", "text": "Represent the given image."}, ], }], "encoding_format": "float", }, ) response.raise_for_status() response_json = response.json() print("Embedding output:", response_json["data"][0]["embedding"]) ``` Below is another example, this time using the `MrLight/dse-qwen2-2b-mrl-v1` model. ```bash vllm serve MrLight/dse-qwen2-2b-mrl-v1 --task embed \ --trust-remote-code --max-model-len 8192 --chat-template examples/template_dse_qwen2_vl.jinja ``` ```{important} Like with VLM2Vec, we have to explicitly pass `--task embed`. Additionally, `MrLight/dse-qwen2-2b-mrl-v1` requires an EOS token for embeddings, which is handled by a custom chat template: ``` ```{important} Also important, `MrLight/dse-qwen2-2b-mrl-v1` requires a placeholder image of the minimum image size for text query embeddings. See the full code example below for details. ``` Full example: --- ### Source Serving Offline Inference (`docs/source/serving/offline_inference.md`) (offline-inference)= # Offline Inference You can run vLLM in your own code on a list of prompts. The offline API is based on the {class}`~vllm.LLM` class. To initialize the vLLM engine, create a new instance of `LLM` and specify the model to run. For example, the following code downloads the [`facebook/opt-125m`](https://huggingface.co/facebook/opt-125m) model from HuggingFace and runs it in vLLM using the default configuration. ```python llm = LLM(model="facebook/opt-125m") ``` After initializing the `LLM` instance, you can perform model inference using various APIs. The available APIs depend on the type of model that is being run: - [Generative models](#generative-models) output logprobs which are sampled from to obtain the final output text. - [Pooling models](#pooling-models) output their hidden states directly. Please refer to the above pages for more details about each API. ```{seealso} [API Reference](/api/offline_inference/index) ``` ## Configuration Options This section lists the most common options for running the vLLM engine. For a full list, refer to the [Engine Arguments](#engine-args) page. ### Model resolution vLLM loads HuggingFace-compatible models by inspecting the `architectures` field in `config.json` of the model repository and finding the corresponding implementation that is registered to vLLM. Nevertheless, our model resolution may fail for the following reasons: - The `config.json` of the model repository lacks the `architectures` field. - Unofficial repositories refer to a model using alternative names which are not recorded in vLLM. - The same architecture name is used for multiple models, creating ambiguity as to which model should be loaded. In those cases, vLLM may throw an error like: ```text Traceback (most recent call last): ... File "vllm/model_executor/models/registry.py", line xxx, in inspect_model_cls for arch in architectures: TypeError: 'NoneType' object is not iterable ``` or: ```text File "vllm/model_executor/models/registry.py", line xxx, in _raise_for_unsupported raise ValueError( ValueError: Model architectures [''] are not supported for now. Supported architectures: [...] ``` :::{note} The above error is distinct from the following similar but different error: ```text File "vllm/model_executor/models/registry.py", line xxx, in _raise_for_unsupported raise ValueError( ValueError: Model architectures [''] failed to be inspected. Please check the logs for more details. ``` This error means that vLLM failed to import the model file. Usually, it is related to missing dependencies or outdated binaries in the vLLM build. Please read the logs carefully to determine the real cause of the error. ::: To fix this, explicitly specify the model architecture by passing `config.json` overrides to the `hf_overrides` option. For example: ```python model = LLM( model="cerebras/Cerebras-GPT-1.3B", hf_overrides={"architectures": ["GPT2LMHeadModel"]}, # GPT-2 ) ``` Our [list of supported models](#supported-models) shows the model architectures that are recognized by vLLM. ### Reducing memory usage Large models might cause your machine to run out of memory (OOM). Here are some options that help alleviate this problem. #### Tensor Parallelism (TP) Tensor parallelism (`tensor_parallel_size` option) can be used to split the model across multiple GPUs. The following code splits the model across 2 GPUs. ```python llm = LLM(model="ibm-granite/granite-3.1-8b-instruct", tensor_parallel_size=2) ``` ```{important} To ensure that vLLM initializes CUDA correctly, you should avoid calling related functions (e.g. {func}`torch.cuda.set_device`) before initializing vLLM. Otherwise, you may run into an error like `RuntimeError: Cannot re-initialize CUDA in forked subprocess`. To control which devices are used, please instead set the `CUDA_VISIBLE_DEVICES` environment variable. ``` #### Quantization Quantized models take less memory at the cost of lower precision. Statically quantized models can be downloaded from HF Hub (some popular ones are available at [Neural Magic](https://huggingface.co/neuralmagic)) and used directly without extra configuration. Dynamic quantization is also supported via the `quantization` option -- see [here](#quantization-index) for more details. #### Context length and batch size You can further reduce memory usage by limiting the context length of the model (`max_model_len` option) and the maximum batch size (`max_num_seqs` option). ```python llm = LLM(model="adept/fuyu-8b", max_model_len=2048, max_num_seqs=2) ``` ### Performance optimization and tuning You can potentially improve the performance of vLLM by finetuning various options. Please refer to [this guide](#optimization-and-tuning) for more details. --- ### Source Serving Openai Compatible Server (`docs/source/serving/openai_compatible_server.md`) (openai-compatible-server)= # OpenAI-Compatible Server vLLM provides an HTTP server that implements OpenAI's [Completions API](https://platform.openai.com/docs/api-reference/completions), [Chat API](https://platform.openai.com/docs/api-reference/chat), and more! You can start the server via the [`vllm serve`](#vllm-serve) command, or through [Docker](#deployment-docker): ```bash vllm serve NousResearch/Meta-Llama-3-8B-Instruct --dtype auto --api-key token-abc123 ``` To call the server, you can use the [official OpenAI Python client](https://github.com/openai/openai-python), or any other HTTP client. ```python from openai import OpenAI client = OpenAI( base_url="http://localhost:8000/v1", api_key="token-abc123", ) completion = client.chat.completions.create( model="NousResearch/Meta-Llama-3-8B-Instruct", messages=[ {"role": "user", "content": "Hello!"} ] ) print(completion.choices[0].message) ``` ## Supported APIs We currently support the following OpenAI APIs: - [Completions API](#completions-api) (`/v1/completions`) - Only applicable to [text generation models](../models/generative_models.md) (`--task generate`). - *Note: `suffix` parameter is not supported.* - [Chat Completions API](#chat-api) (`/v1/chat/completions`) - Only applicable to [text generation models](../models/generative_models.md) (`--task generate`) with a [chat template](#chat-template). - *Note: `parallel_tool_calls` and `user` parameters are ignored.* - [Embeddings API](#embeddings-api) (`/v1/embeddings`) - Only applicable to [embedding models](../models/pooling_models.md) (`--task embed`). In addition, we have the following custom APIs: - [Tokenizer API](#tokenizer-api) (`/tokenize`, `/detokenize`) - Applicable to any model with a tokenizer. - [Pooling API](#pooling-api) (`/pooling`) - Applicable to all [pooling models](../models/pooling_models.md). - [Score API](#score-api) (`/score`) - Only applicable to [cross-encoder models](../models/pooling_models.md) (`--task score`). (chat-template)= ## Chat Template In order for the language model to support chat protocol, vLLM requires the model to include a chat template in its tokenizer configuration. The chat template is a Jinja2 template that specifies how are roles, messages, and other chat-specific tokens are encoded in the input. An example chat template for `NousResearch/Meta-Llama-3-8B-Instruct` can be found [here](https://github.com/meta-llama/llama3?tab=readme-ov-file#instruction-tuned-models) Some models do not provide a chat template even though they are instruction/chat fine-tuned. For those model, you can manually specify their chat template in the `--chat-template` parameter with the file path to the chat template, or the template in string form. Without a chat template, the server will not be able to process chat and all chat requests will error. ```bash vllm serve --chat-template ./path-to-chat-template.jinja ``` vLLM community provides a set of chat templates for popular models. You can find them under the directory. With the inclusion of multi-modal chat APIs, the OpenAI spec now accepts chat messages in a new format which specifies both a `type` and a `text` field. An example is provided below: ```python completion = client.chat.completions.create( model="NousResearch/Meta-Llama-3-8B-Instruct", messages=[ {"role": "user", "content": [{"type": "text", "text": "Classify this sentiment: vLLM is wonderful!"}]} ] ) ``` Most chat templates for LLMs expect the `content` field to be a string, but there are some newer models like `meta-llama/Llama-Guard-3-1B` that expect the content to be formatted according to the OpenAI schema in the request. vLLM provides best-effort support to detect this automatically, which is logged as a string like *"Detected the chat template content format to be..."*, and internally converts incoming requests to match the detected format, which can be one of: - `"string"`: A string. - Example: `"Hello world"` - `"openai"`: A list of dictionaries, similar to OpenAI schema. - Example: `[{"type": "text", "text": "Hello world!"}]` If the result is not what you expect, you can set the `--chat-template-content-format` CLI argument to override which format to use. ## Extra Parameters vLLM supports a set of parameters that are not part of the OpenAI API. In order to use them, you can pass them as extra parameters in the OpenAI client. Or directly merge them into the JSON payload if you are using HTTP call directly. ```python completion = client.chat.completions.create( model="NousResearch/Meta-Llama-3-8B-Instruct", messages=[ {"role": "user", "content": "Classify this sentiment: vLLM is wonderful!"} ], extra_body={ "guided_choice": ["positive", "negative"] } ) ``` ## Extra HTTP Headers Only `X-Request-Id` HTTP request header is supported for now. It can be enabled with `--enable-request-id-headers`. > Note that enablement of the headers can impact performance significantly at high QPS > rates. We recommend implementing HTTP headers at the router level (e.g. via Istio), > rather than within the vLLM layer for this reason. > See [this PR](https://github.com/vllm-project/vllm/pull/11529) for more details. ```python completion = client.chat.completions.create( model="NousResearch/Meta-Llama-3-8B-Instruct", messages=[ {"role": "user", "content": "Classify this sentiment: vLLM is wonderful!"} ], extra_headers={ "x-request-id": "sentiment-classification-00001", } ) print(completion._request_id) completion = client.completions.create( model="NousResearch/Meta-Llama-3-8B-Instruct", prompt="A robot may not injure a human being", extra_headers={ "x-request-id": "completion-test", } ) print(completion._request_id) ``` ## CLI Reference (vllm-serve)= ### `vllm serve` The `vllm serve` command is used to launch the OpenAI-compatible server. ```{argparse} :module: vllm.entrypoints.openai.cli_args :func: create_parser_for_docs :prog: vllm serve ``` #### Configuration file You can load CLI arguments via a [YAML](https://yaml.org/) config file. The argument names must be the long form of those outlined [above](#vllm-serve). For example: ```yaml # config.yaml host: "127.0.0.1" port: 6379 uvicorn-log-level: "info" ``` To use the above config file: ```bash vllm serve SOME_MODEL --config config.yaml ``` ```{note} In case an argument is supplied simultaneously using command line and the config file, the value from the command line will take precedence. The order of priorities is `command line > config file values > defaults`. ``` ## API Reference (completions-api)= ### Completions API Our Completions API is compatible with [OpenAI's Completions API](https://platform.openai.com/docs/api-reference/completions); you can use the [official OpenAI Python client](https://github.com/openai/openai-python) to interact with it. Code example: #### Extra parameters The following [sampling parameters](#sampling-params) are supported. ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-completion-sampling-params :end-before: end-completion-sampling-params ``` The following extra parameters are supported: ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-completion-extra-params :end-before: end-completion-extra-params ``` (chat-api)= ### Chat API Our Chat API is compatible with [OpenAI's Chat Completions API](https://platform.openai.com/docs/api-reference/chat); you can use the [official OpenAI Python client](https://github.com/openai/openai-python) to interact with it. We support both [Vision](https://platform.openai.com/docs/guides/vision)- and [Audio](https://platform.openai.com/docs/guides/audio?audio-generation-quickstart-example=audio-in)-related parameters; see our [Multimodal Inputs](#multimodal-inputs) guide for more information. - *Note: `image_url.detail` parameter is not supported.* Code example: #### Extra parameters The following [sampling parameters](#sampling-params) are supported. ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-chat-completion-sampling-params :end-before: end-chat-completion-sampling-params ``` The following extra parameters are supported: ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-chat-completion-extra-params :end-before: end-chat-completion-extra-params ``` (embeddings-api)= ### Embeddings API Our Embeddings API is compatible with [OpenAI's Embeddings API](https://platform.openai.com/docs/api-reference/embeddings); you can use the [official OpenAI Python client](https://github.com/openai/openai-python) to interact with it. If the model has a [chat template](#chat-template), you can replace `inputs` with a list of `messages` (same schema as [Chat API](#chat-api)) which will be treated as a single prompt to the model. ```{tip} This enables multi-modal inputs to be passed to embedding models, see [this page](#multimodal-inputs) for details. ``` Code example: #### Extra parameters The following [pooling parameters](#pooling-params) are supported. ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-embedding-pooling-params :end-before: end-embedding-pooling-params ``` The following extra parameters are supported by default: ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-embedding-extra-params :end-before: end-embedding-extra-params ``` For chat-like input (i.e. if `messages` is passed), these extra parameters are supported instead: ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-chat-embedding-extra-params :end-before: end-chat-embedding-extra-params ``` (tokenizer-api)= ### Tokenizer API Our Tokenizer API is a simple wrapper over [HuggingFace-style tokenizers](https://huggingface.co/docs/transformers/en/main_classes/tokenizer). It consists of two endpoints: - `/tokenize` corresponds to calling `tokenizer.encode()`. - `/detokenize` corresponds to calling `tokenizer.decode()`. (pooling-api)= ### Pooling API Our Pooling API encodes input prompts using a [pooling model](../models/pooling_models.md) and returns the corresponding hidden states. The input format is the same as [Embeddings API](#embeddings-api), but the output data can contain an arbitrary nested list, not just a 1-D list of floats. Code example: (score-api)= ### Score API Our Score API applies a cross-encoder model to predict scores for sentence pairs. Usually, the score for a sentence pair refers to the similarity between two sentences, on a scale of 0 to 1. You can find the documentation for these kind of models at [sbert.net](https://www.sbert.net/docs/package_reference/cross_encoder/cross_encoder.html). Code example: #### Single inference You can pass a string to both `text_1` and `text_2`, forming a single sentence pair. Request: ```bash curl -X 'POST' \ 'http://127.0.0.1:8000/score' \ -H 'accept: application/json' \ -H 'Content-Type: application/json' \ -d '{ "model": "BAAI/bge-reranker-v2-m3", "encoding_format": "float", "text_1": "What is the capital of France?", "text_2": "The capital of France is Paris." }' ``` Response: ```bash { "id": "score-request-id", "object": "list", "created": 693447, "model": "BAAI/bge-reranker-v2-m3", "data": [ { "index": 0, "object": "score", "score": 1 } ], "usage": {} } ``` #### Batch inference You can pass a string to `text_1` and a list to `text_2`, forming multiple sentence pairs where each pair is built from `text_1` and a string in `text_2`. The total number of pairs is `len(text_2)`. Request: ```bash curl -X 'POST' \ 'http://127.0.0.1:8000/score' \ -H 'accept: application/json' \ -H 'Content-Type: application/json' \ -d '{ "model": "BAAI/bge-reranker-v2-m3", "text_1": "What is the capital of France?", "text_2": [ "The capital of Brazil is Brasilia.", "The capital of France is Paris." ] }' ``` Response: ```bash { "id": "score-request-id", "object": "list", "created": 693570, "model": "BAAI/bge-reranker-v2-m3", "data": [ { "index": 0, "object": "score", "score": 0.001094818115234375 }, { "index": 1, "object": "score", "score": 1 } ], "usage": {} } ``` You can pass a list to both `text_1` and `text_2`, forming multiple sentence pairs where each pair is built from a string in `text_1` and the corresponding string in `text_2` (similar to `zip()`). The total number of pairs is `len(text_2)`. Request: ```bash curl -X 'POST' \ 'http://127.0.0.1:8000/score' \ -H 'accept: application/json' \ -H 'Content-Type: application/json' \ -d '{ "model": "BAAI/bge-reranker-v2-m3", "encoding_format": "float", "text_1": [ "What is the capital of Brazil?", "What is the capital of France?" ], "text_2": [ "The capital of Brazil is Brasilia.", "The capital of France is Paris." ] }' ``` Response: ```bash { "id": "score-request-id", "object": "list", "created": 693447, "model": "BAAI/bge-reranker-v2-m3", "data": [ { "index": 0, "object": "score", "score": 1 }, { "index": 1, "object": "score", "score": 1 } ], "usage": {} } ``` #### Extra parameters The following [pooling parameters](#pooling-params) are supported. ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-score-pooling-params :end-before: end-score-pooling-params ``` The following extra parameters are supported: ```{literalinclude} ../../../vllm/entrypoints/openai/protocol.py :language: python :start-after: begin-score-extra-params :end-before: end-score-extra-params ``` --- ### Source Serving Usage Stats (`docs/source/serving/usage_stats.md`) # Usage Stats Collection vLLM collects anonymous usage data by default to help the engineering team better understand which hardware and model configurations are widely used. This data allows them to prioritize their efforts on the most common workloads. The collected data is transparent, does not contain any sensitive information, and will be publicly released for the community's benefit. ## What data is collected? The list of data collected by the latest version of vLLM can be found here: Here is an example as of v0.4.0: ```json { "uuid": "fbe880e9-084d-4cab-a395-8984c50f1109", "provider": "GCP", "num_cpu": 24, "cpu_type": "Intel(R) Xeon(R) CPU @ 2.20GHz", "cpu_family_model_stepping": "6,85,7", "total_memory": 101261135872, "architecture": "x86_64", "platform": "Linux-5.10.0-28-cloud-amd64-x86_64-with-glibc2.31", "gpu_count": 2, "gpu_type": "NVIDIA L4", "gpu_memory_per_device": 23580639232, "model_architecture": "OPTForCausalLM", "vllm_version": "0.3.2+cu123", "context": "LLM_CLASS", "log_time": 1711663373492490000, "source": "production", "dtype": "torch.float16", "tensor_parallel_size": 1, "block_size": 16, "gpu_memory_utilization": 0.9, "quantization": null, "kv_cache_dtype": "auto", "enable_lora": false, "enable_prefix_caching": false, "enforce_eager": false, "disable_custom_all_reduce": true } ``` You can preview the collected data by running the following command: ```bash tail ~/.config/vllm/usage_stats.json ``` ## Opting out You can opt-out of usage stats collection by setting the `VLLM_NO_USAGE_STATS` or `DO_NOT_TRACK` environment variable, or by creating a `~/.config/vllm/do_not_track` file: ```bash # Any of the following methods can disable usage stats collection export VLLM_NO_USAGE_STATS=1 export DO_NOT_TRACK=1 mkdir -p ~/.config/vllm && touch ~/.config/vllm/do_not_track ``` --- ### Source Api Engine Async Llm Engine (`docs/source/api/engine/async_llm_engine.md`) # AsyncLLMEngine ```{eval-rst} .. autoclass:: vllm.AsyncLLMEngine :members: :show-inheritance: ``` --- ### Source Api Engine Llm Engine (`docs/source/api/engine/llm_engine.md`) # LLMEngine ```{eval-rst} .. autoclass:: vllm.LLMEngine :members: :show-inheritance: ``` --- ### Source Api Model Adapters (`docs/source/api/model/adapters.md`) # Model Adapters ## Module Contents ```{eval-rst} .. automodule:: vllm.model_executor.models.adapters :members: :member-order: bysource ``` --- ### Source Api Model Interfaces (`docs/source/api/model/interfaces.md`) # Optional Interfaces ## Module Contents ```{eval-rst} .. automodule:: vllm.model_executor.models.interfaces :members: :member-order: bysource ``` --- ### Source Api Model Interfaces Base (`docs/source/api/model/interfaces_base.md`) # Base Model Interfaces ## Module Contents ```{eval-rst} .. automodule:: vllm.model_executor.models.interfaces_base :members: :member-order: bysource ``` --- ### Source Api Multimodal Inputs (`docs/source/api/multimodal/inputs.md`) # Input Definitions ## User-facing inputs ```{eval-rst} .. autodata:: vllm.multimodal.inputs.MultiModalDataDict ``` ## Internal data structures ```{eval-rst} .. autoclass:: vllm.multimodal.inputs.PlaceholderRange :members: :show-inheritance: ``` ```{eval-rst} .. autodata:: vllm.multimodal.inputs.NestedTensors ``` ```{eval-rst} .. autoclass:: vllm.multimodal.inputs.MultiModalFieldElem :members: :show-inheritance: ``` ```{eval-rst} .. autoclass:: vllm.multimodal.inputs.MultiModalFieldConfig :members: :show-inheritance: ``` ```{eval-rst} .. autoclass:: vllm.multimodal.inputs.MultiModalKwargsItem :members: :show-inheritance: ``` ```{eval-rst} .. autoclass:: vllm.multimodal.inputs.MultiModalKwargs :members: :show-inheritance: ``` ```{eval-rst} .. autoclass:: vllm.multimodal.inputs.MultiModalInputs :members: :show-inheritance: ``` --- ### Source Api Multimodal Parse (`docs/source/api/multimodal/parse.md`) # Data Parsing ## Module Contents ```{eval-rst} .. automodule:: vllm.multimodal.parse :members: :member-order: bysource ``` --- ### Source Api Multimodal Processing (`docs/source/api/multimodal/processing.md`) # Data Processing ## Module Contents ```{eval-rst} .. automodule:: vllm.multimodal.processing :members: :member-order: bysource ``` --- ### Source Api Multimodal Profiling (`docs/source/api/multimodal/profiling.md`) # Memory Profiling ## Module Contents ```{eval-rst} .. automodule:: vllm.multimodal.profiling :members: :member-order: bysource ``` --- ### Source Api Multimodal Registry (`docs/source/api/multimodal/registry.md`) # Registry ## Module Contents ```{eval-rst} .. automodule:: vllm.multimodal.registry :members: :member-order: bysource ``` --- ### Source Api Offline Inference Llm (`docs/source/api/offline_inference/llm.md`) # LLM Class ```{eval-rst} .. autoclass:: vllm.LLM :members: :show-inheritance: ``` --- ### Source Api Offline Inference Llm Inputs (`docs/source/api/offline_inference/llm_inputs.md`) # LLM Inputs ```{eval-rst} .. autodata:: vllm.inputs.PromptType ``` ```{eval-rst} .. autoclass:: vllm.inputs.TextPrompt :show-inheritance: :members: :member-order: bysource ``` ```{eval-rst} .. autoclass:: vllm.inputs.TokensPrompt :show-inheritance: :members: :member-order: bysource ``` --- ### Source Contributing Dockerfile Dockerfile (`docs/source/contributing/dockerfile/dockerfile.md`) # Dockerfile We provide a to construct the image for running an OpenAI compatible server with vLLM. More information about deploying with Docker can be found [here](#deployment-docker). Below is a visual representation of the multi-stage Dockerfile. The build graph contains the following nodes: - All build stages - The default build target (highlighted in grey) - External images (with dashed borders) The edges of the build graph represent: - `FROM ...` dependencies (with a solid line and a full arrow head) - `COPY --from=...` dependencies (with a dashed line and an empty arrow head) - `RUN --mount=(.\*)from=...` dependencies (with a dotted line and an empty diamond arrow head) > ```{figure} /assets/contributing/dockerfile-stages-dependency.png > :align: center > :alt: query > :width: 100% > ``` > > Made using: > > Commands to regenerate the build graph (make sure to run it **from the \`root\` directory of the vLLM repository** where the dockerfile is present): > > ```bash > dockerfilegraph -o png --legend --dpi 200 --max-label-length 50 --filename Dockerfile > ``` > > or in case you want to run it directly with the docker image: > > ```bash > docker run \ > --rm \ > --user "$(id -u):$(id -g)" \ > --workdir /workspace \ > --volume "$(pwd)":/workspace \ > ghcr.io/patrickhoefler/dockerfilegraph:alpine \ > --output png \ > --dpi 200 \ > --max-label-length 50 \ > --filename Dockerfile \ > --legend > ``` > > (To run it for a different file, you can pass in a different argument to the flag `--filename`.) --- ### Source Contributing Model Basic (`docs/source/contributing/model/basic.md`) (new-model-basic)= # Implementing a Basic Model This guide walks you through the steps to implement a basic vLLM model. ## 1. Bring your model code First, clone the PyTorch model code from the source repository. For instance, vLLM's [OPT model](gh-file:vllm/model_executor/models/opt.py) was adapted from HuggingFace's [modeling_opt.py](https://github.com/huggingface/transformers/blob/main/src/transformers/models/opt/modeling_opt.py) file. ```{warning} Make sure to review and adhere to the original code's copyright and licensing terms! ``` ## 2. Make your code compatible with vLLM To ensure compatibility with vLLM, your model must meet the following requirements: ### Initialization Code All vLLM modules within the model must include a `prefix` argument in their constructor. This `prefix` is typically the full name of the module in the model's state dictionary and is crucial for: - Runtime support: vLLM's attention operators are registered in a model's state by their full names. Each attention operator must have a unique prefix as its layer name to avoid conflicts. - Non-uniform quantization support: A quantized checkpoint can selectively quantize certain layers while keeping others in full precision. By providing the `prefix` during initialization, vLLM can match the current layer's `prefix` with the quantization configuration to determine if the layer should be initialized in quantized mode. The initialization code should look like this: ```python from torch import nn from vllm.config import VllmConfig from vllm.attention import Attention class MyAttention(nn.Module): def __init__(self, vllm_config: VllmConfig, prefix: str): super().__init__() self.attn = Attention(prefix=f"{prefix}.attn") class MyDecoderLayer(nn.Module): def __init__(self, vllm_config: VllmConfig, prefix: str): super().__init__() self.self_attn = MyAttention(prefix=f"{prefix}.self_attn") class MyModel(nn.Module): def __init__(self, vllm_config: VllmConfig, prefix: str): super().__init__() self.layers = nn.ModuleList( [MyDecoderLayer(vllm_config, prefix=f"{prefix}.layers.{i}") for i in range(vllm_config.model_config.hf_config.num_hidden_layers)] ) class MyModelForCausalLM(nn.Module): def __init__(self, vllm_config: VllmConfig, prefix: str = ""): super().__init__() self.model = MyModel(vllm_config, prefix=f"{prefix}.model") ``` ### Computation Code - Add a `get_input_embeddings` method inside `MyModel` module that returns the text embeddings given `input_ids`. This is equivalent to directly calling the text embedding layer, but provides a unified interface in case `MyModel` is used within a composite multimodal model. ```python class MyModel(nn.Module): ... def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: ... ``` - Rewrite the {meth}`~torch.nn.Module.forward` method of your model to remove any unnecessary code, such as training-specific code. Modify the input parameters to treat `input_ids` and `positions` as flattened tensors with a single batch size dimension, without a max-sequence length dimension. ```python def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, kv_caches: List[torch.Tensor], attn_metadata: AttentionMetadata, ) -> torch.Tensor: ... ``` ```{note} Currently, vLLM supports the basic multi-head attention mechanism and its variant with rotary positional embeddings. If your model employs a different attention mechanism, you will need to implement a new attention layer in vLLM. ``` For reference, check out our [Llama implementation](gh-file:vllm/model_executor/models/llama.py). vLLM already supports a large number of models. It is recommended to find a model similar to yours and adapt it to your model's architecture. Check out for more examples. ## 3. (Optional) Implement tensor parallelism and quantization support If your model is too large to fit into a single GPU, you can use tensor parallelism to manage it. To do this, substitute your model's linear and embedding layers with their tensor-parallel versions. For the embedding layer, you can simply replace {class}`torch.nn.Embedding` with `VocabParallelEmbedding`. For the output LM head, you can use `ParallelLMHead`. When it comes to the linear layers, we provide the following options to parallelize them: - `ReplicatedLinear`: Replicates the inputs and weights across multiple GPUs. No memory saving. - `RowParallelLinear`: The input tensor is partitioned along the hidden dimension. The weight matrix is partitioned along the rows (input dimension). An *all-reduce* operation is performed after the matrix multiplication to reduce the results. Typically used for the second FFN layer and the output linear transformation of the attention layer. - `ColumnParallelLinear`: The input tensor is replicated. The weight matrix is partitioned along the columns (output dimension). The result is partitioned along the column dimension. Typically used for the first FFN layer and the separated QKV transformation of the attention layer in the original Transformer. - `MergedColumnParallelLinear`: Column-parallel linear that merges multiple `ColumnParallelLinear` operators. Typically used for the first FFN layer with weighted activation functions (e.g., SiLU). This class handles the sharded weight loading logic of multiple weight matrices. - `QKVParallelLinear`: Parallel linear layer for the query, key, and value projections of the multi-head and grouped-query attention mechanisms. When number of key/value heads are less than the world size, this class replicates the key/value heads properly. This class handles the weight loading and replication of the weight matrices. Note that all the linear layers above take `linear_method` as an input. vLLM will set this parameter according to different quantization schemes to support weight quantization. ## 4. Implement the weight loading logic You now need to implement the `load_weights` method in your `*ForCausalLM` class. This method should load the weights from the HuggingFace's checkpoint file and assign them to the corresponding layers in your model. Specifically, for `MergedColumnParallelLinear` and `QKVParallelLinear` layers, if the original model has separated weight matrices, you need to load the different parts separately. ## 5. Register your model See [this page](#new-model-registration) for instructions on how to register your new model to be used by vLLM. ## Frequently Asked Questions ### How to support models with interleaving sliding windows? For models with interleaving sliding windows (e.g. `google/gemma-2-2b-it` and `mistralai/Ministral-8B-Instruct-2410`), the scheduler will treat the model as a full-attention model, i.e., kv-cache of all tokens will not be dropped. This is to make sure prefix caching works with these models. Sliding window only appears as a parameter to the attention kernel computation. To support a model with interleaving sliding windows, we need to take care of the following details: - Make sure [this line](https://github.com/vllm-project/vllm/blob/996357e4808ca5eab97d4c97c7d25b3073f46aab/vllm/config.py#L308) evaluates `has_interleaved_attention` to `True` for this model, and set `self.hf_text_config.interleaved_sliding_window` to the format of interleaving sliding windows the model can understand. Then, `self.hf_text_config.sliding_window` will be deleted, and the model will be treated as a full-attention model. - In the modeling code, parse the correct sliding window value for every layer, and pass it to the attention layer's `per_layer_sliding_window` argument. For reference, check [this line](https://github.com/vllm-project/vllm/blob/996357e4808ca5eab97d4c97c7d25b3073f46aab/vllm/model_executor/models/llama.py#L171). With these two steps, interleave sliding windows should work with the model. --- ### Source Contributing Model Multimodal (`docs/source/contributing/model/multimodal.md`) (supports-multimodal)= # Multi-Modal Support This document walks you through the steps to extend a basic model so that it accepts [multi-modal inputs](#multimodal-inputs). ## 1. Update the base vLLM model It is assumed that you have already implemented the model in vLLM according to [these steps](#new-model-basic). Further update the model as follows: - Reserve a keyword parameter in {meth}`~torch.nn.Module.forward` for each input tensor that corresponds to a multi-modal input, as shown in the following example: ```diff def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, kv_caches: List[torch.Tensor], attn_metadata: AttentionMetadata, + pixel_values: torch.Tensor, ) -> SamplerOutput: ``` More conveniently, you can simply pass `**kwargs` to the {meth}`~torch.nn.Module.forward` method and retrieve the keyword parameters for multimodal inputs from it. - Implement {meth}`~vllm.model_executor.models.interfaces.SupportsMultiModal.get_multimodal_embeddings` that returns the embeddings from running the multimodal inputs through the multimodal tokenizer of the model. Below we provide a boilerplate of a typical implementation pattern, but feel free to adjust it to your own needs. ```python class YourModelForImage2Seq(nn.Module): ... def _process_image_input(self, image_input: YourModelImageInputs) -> torch.Tensor: assert self.vision_encoder is not None image_features = self.vision_encoder(image_input) return self.multi_modal_projector(image_features) def get_multimodal_embeddings(self, **kwargs: object) -> Optional[NestedTensors]: # Validate the multimodal input keyword arguments image_input = self._parse_and_validate_image_input(**kwargs) if image_input is None: return None # Run multimodal inputs through encoder and projector vision_embeddings = self._process_image_input(image_input) return vision_embeddings ``` ```{important} The returned `multimodal_embeddings` must be either a **3D {class}`torch.Tensor`** of shape `(num_items, feature_size, hidden_size)`, or a **list / tuple of 2D {class}`torch.Tensor`'s** of shape `(feature_size, hidden_size)`, so that `multimodal_embeddings[i]` retrieves the embeddings generated from the `i`-th multimodal data item (e.g, image) of the request. ``` - Implement {meth}`~vllm.model_executor.models.interfaces.SupportsMultiModal.get_input_embeddings` to merge `multimodal_embeddings` with text embeddings from the `input_ids`. If input processing for the model is implemented correctly (see sections below), then you can leverage the utility function we provide to easily merge the embeddings. ```python from .utils import merge_multimodal_embeddings class YourModelForImage2Seq(nn.Module): ... def get_input_embeddings( self, input_ids: torch.Tensor, multimodal_embeddings: Optional[NestedTensors] = None, ) -> torch.Tensor: # `get_input_embeddings` should already be implemented for the language # model as one of the requirements of basic vLLM model implementation. inputs_embeds = self.language_model.get_input_embeddings(input_ids) if multimodal_embeddings is not None: inputs_embeds = merge_multimodal_embeddings( input_ids=input_ids, inputs_embeds=inputs_embeds, multimodal_embeddings=multimodal_embeddings, placeholder_token_id=self.config.image_token_index) return inputs_embeds ``` - Once the above steps are done, update the model class with the {class}`~vllm.model_executor.models.interfaces.SupportsMultiModal` interface. ```diff + from vllm.model_executor.models.interfaces import SupportsMultiModal - class YourModelForImage2Seq(nn.Module): + class YourModelForImage2Seq(nn.Module, SupportsMultiModal): ``` ```{note} The model class does not have to be named {code}`*ForCausalLM`. Check out [the HuggingFace Transformers documentation](https://huggingface.co/docs/transformers/model_doc/auto#multimodal) for some examples. ``` ## 2. Specify processing information Next, create a subclass of {class}`~vllm.multimodal.processing.BaseProcessingInfo` to provide basic information related to HF processing. ### Maximum number of input items You need to override the abstract method {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_supported_mm_limits` to return the maximum number of input items for each modality supported by the model. For example, if the model supports any number of images but only one video per prompt: ```python def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]: return {"image": None, "video": 1} ``` ### Maximum number of placeholder feature tokens Also, override the abstract method {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_mm_max_tokens_per_item` to return the maximum number of placeholder feature tokens per input item for each modality. When calling the model, the output embeddings from the visual encoder are assigned to the input positions containing placeholder feature tokens. Therefore, the number of placeholder feature tokens should be equal to the size of the output embeddings. ::::{tab-set} :::{tab-item} Basic example: LLaVA :sync: llava Looking at the code of HF's `LlavaForConditionalGeneration`: ```python # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L530-L544 n_image_tokens = (input_ids == self.config.image_token_index).sum().item() n_image_features = image_features.shape[0] * image_features.shape[1] if n_image_tokens != n_image_features: raise ValueError( f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}" ) special_image_mask = ( (input_ids == self.config.image_token_index) .unsqueeze(-1) .expand_as(inputs_embeds) .to(inputs_embeds.device) ) image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype) inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features) ``` The number of placeholder feature tokens per image is `image_features.shape[1]`. `image_features` is calculated inside the `get_image_features` method: ```python # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L290-L300 image_outputs = self.vision_tower(pixel_values, output_hidden_states=True) selected_image_feature = image_outputs.hidden_states[vision_feature_layer] if vision_feature_select_strategy == "default": selected_image_feature = selected_image_feature[:, 1:] elif vision_feature_select_strategy == "full": selected_image_feature = selected_image_feature else: raise ValueError(f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}") image_features = self.multi_modal_projector(selected_image_feature) return image_features ``` We can infer that `image_features.shape[1]` is based on `image_outputs.hidden_states.shape[1]` from the vision tower (`CLIPVisionModel` for the [`llava-hf/llava-1.5-7b-hf`](https://huggingface.co/llava-hf/llava-1.5-7b-hf) model). Moreover, we only need the sequence length (the second dimension of the tensor) to get `image_features.shape[1]`. The sequence length is determined by the initial hidden states in `CLIPVisionTransformer` since the attention mechanism doesn't change the sequence length of the output hidden states. ```python # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/modeling_clip.py#L1094-L1102 hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding) hidden_states = self.pre_layrnorm(hidden_states) encoder_outputs = self.encoder( inputs_embeds=hidden_states, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) ``` To find the sequence length, we turn to the code of `CLIPVisionEmbeddings`: ```python # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/modeling_clip.py#L247-L257 target_dtype = self.patch_embedding.weight.dtype patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid] patch_embeds = patch_embeds.flatten(2).transpose(1, 2) class_embeds = self.class_embedding.expand(batch_size, 1, -1) embeddings = torch.cat([class_embeds, patch_embeds], dim=1) if interpolate_pos_encoding: embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width) else: embeddings = embeddings + self.position_embedding(self.position_ids) return embeddings ``` We can infer that `embeddings.shape[1] == self.num_positions`, where ```python # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/modeling_clip.py#L195-L196 self.num_patches = (self.image_size // self.patch_size) ** 2 self.num_positions = self.num_patches + 1 ``` Overall, the number of placeholder feature tokens for an image can be calculated as: ```python def get_num_image_tokens( self, *, image_width: int, image_height: int, ) -> int: hf_config = self.get_hf_config() hf_processor = self.get_hf_processor() image_size = hf_config.vision_config.image_size patch_size = hf_config.vision_config.patch_size num_image_tokens = (image_size // patch_size) ** 2 + 1 if hf_processor.vision_feature_select_strategy == "default": num_image_tokens -= 1 return num_image_tokens ``` Notice that the number of image tokens doesn't depend on the image width and height. So, we can calculate the maximum number of image tokens using any image size: ```python def get_image_size_with_most_features(self) -> ImageSize: hf_config = self.get_hf_config() width = height = hf_config.image_size return ImageSize(width=width, height=height) def get_max_image_tokens(self) -> int: target_width, target_height = self.get_image_size_with_most_features() return self.get_num_image_tokens( image_width=target_width, image_height=target_height, ) ``` And thus, we can override the method as: ```python def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]: return {"image": self.get_max_image_tokens()} ``` ```{note} Our [actual code](gh-file:vllm/model_executor/models/llava.py) is more abstracted to support vision encoders other than CLIP. ``` ::: :::: ## 3. Specify dummy inputs Then, inherit {class}`~vllm.multimodal.profiling.BaseDummyInputsBuilder` to construct dummy inputs for HF processing as well as memory profiling. ### For memory profiling Override the abstract method {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs` to construct dummy inputs for memory profiling. This dummy input should result in the worst-case memory usage of the model so that vLLM can reserve the correct amount of memory for it. Assuming that the memory usage increases with the number of tokens, the dummy input can be constructed based on the code for {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_mm_max_tokens_per_item`. ::::{tab-set} :::{tab-item} Basic example: LLaVA :sync: llava Making use of the `get_image_size_with_most_features` method implemented in the previous section: ```python def get_dummy_processor_inputs( self, seq_len: int, mm_counts: Mapping[str, int], ) -> ProcessorInputs: num_images = mm_counts.get("image", 0) processor = self.info.get_hf_processor() image_token = processor.image_token hf_config = self.get_hf_config() target_width, target_height = self.info.get_image_size_with_most_features() mm_data = { "image": self._get_dummy_images(width=target_width, height=target_height, num_images=num_images) } return ProcessorInputs( prompt_text=image_token * num_images, mm_data=mm_data, ) ``` ::: :::: ## 4. Specify processing details Afterwards, create a subclass of {class}`~vllm.multimodal.processing.BaseMultiModalProcessor` to fill in the missing details about HF processing. ```{seealso} [Multi-Modal Data Processing](#mm-processing) ``` ### Multi-modal fields Override {class}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_mm_fields_config` to return a schema of the tensors outputted by the HF processor that are related to the input multi-modal items. ::::{tab-set} :::{tab-item} Basic example: LLaVA :sync: llava Looking at the model's `forward` method: ```python # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L387-L404 def forward( self, input_ids: torch.LongTensor = None, pixel_values: torch.FloatTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, vision_feature_layer: Optional[int] = None, vision_feature_select_strategy: Optional[str] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, num_logits_to_keep: int = 0, ) -> Union[Tuple, LlavaCausalLMOutputWithPast]: ``` The only related keyword argument is `pixel_values` which directly corresponds to input images. The shape of `pixel_values` is `(N, C, H, W)` where `N` is the number of images. So, we override the method as follows: ```python def _get_mm_fields_config( self, hf_inputs: BatchFeature, hf_processor_mm_kwargs: Mapping[str, object], ) -> Mapping[str, MultiModalFieldConfig]: return dict( pixel_values=MultiModalFieldConfig.batched("image"), ) ``` ```{note} Our [actual code](gh-file:vllm/model_executor/models/llava.py) additionally supports pre-computed image embeddings, which can be passed to be model via the `image_embeds` argument. ``` ::: :::: ### Prompt replacements Override {class}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements` to return a list of {class}`~vllm.multimodal.processing.PromptReplacement` instances. Each {class}`~vllm.multimodal.processing.PromptReplacement` instance specifies a find-and-replace operation performed by the HF processor. ::::{tab-set} :::{tab-item} Basic example: LLaVA :sync: llava Looking at HF's `LlavaProcessor`: ```python # https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/processing_llava.py#L167-L170 prompt_strings = [] for sample in text: sample = sample.replace(self.image_token, self.image_token * num_image_tokens) prompt_strings.append(sample) ``` It simply repeats each input `image_token` a number of times equal to the number of placeholder feature tokens (`num_image_tokens`). Based on this, we override the method as follows: ```python def _get_prompt_replacements( self, mm_items: MultiModalDataItems, hf_processor_mm_kwargs: Mapping[str, object], out_mm_kwargs: MultiModalKwargs, ) -> list[PromptReplacement]: hf_config = self.info.get_hf_config() image_token_id = hf_config.image_token_index def get_replacement(item_idx: int): images = mm_items.get_items("image", ImageProcessorItems) image_size = images.get_image_size(item_idx) num_image_tokens = self.info.get_num_image_tokens( image_width=image_size.width, image_height=image_size.height, ) return [image_token_id] * num_image_tokens return [ PromptReplacement( modality="image", target=[image_token_id], replacement=get_replacement, ), ] ``` ::: :::: ## 5. Register processor-related classes After you have defined {class}`~vllm.multimodal.processing.BaseProcessingInfo` (Step 2), {class}`~vllm.multimodal.profiling.BaseDummyInputsBuilder` (Step 3), and {class}`~vllm.multimodal.processing.BaseMultiModalProcessor` (Step 4), decorate the model class with {meth}`MULTIMODAL_REGISTRY.register_processor ` to register them to the multi-modal registry: ```diff from vllm.model_executor.models.interfaces import SupportsMultiModal + from vllm.multimodal import MULTIMODAL_REGISTRY + @MULTIMODAL_REGISTRY.register_processor(YourMultiModalProcessor, + info=YourProcessingInfo, + dummy_inputs=YourDummyInputsBuilder) class YourModelForImage2Seq(nn.Module, SupportsMultiModal): ``` --- ### Source Contributing Model Registration (`docs/source/contributing/model/registration.md`) (new-model-registration)= # Registering a Model to vLLM vLLM relies on a model registry to determine how to run each model. A list of pre-registered architectures can be found [here](#supported-models). If your model is not on this list, you must register it to vLLM. This page provides detailed instructions on how to do so. ## Built-in models To add a model directly to the vLLM library, start by forking our [GitHub repository](https://github.com/vllm-project/vllm) and then [build it from source](#build-from-source). This gives you the ability to modify the codebase and test your model. After you have implemented your model (see [tutorial](#new-model-basic)), put it into the directory. Then, add your model class to `_VLLM_MODELS` in so that it is automatically registered upon importing vLLM. Finally, update our [list of supported models](#supported-models) to promote your model! ```{important} The list of models in each section should be maintained in alphabetical order. ``` ## Out-of-tree models You can load an external model using a plugin without modifying the vLLM codebase. ```{seealso} [vLLM's Plugin System](#plugin-system) ``` To register the model, use the following code: ```python from vllm import ModelRegistry from your_code import YourModelForCausalLM ModelRegistry.register_model("YourModelForCausalLM", YourModelForCausalLM) ``` If your model imports modules that initialize CUDA, consider lazy-importing it to avoid errors like `RuntimeError: Cannot re-initialize CUDA in forked subprocess`: ```python from vllm import ModelRegistry ModelRegistry.register_model("YourModelForCausalLM", "your_code:YourModelForCausalLM") ``` ```{important} If your model is a multimodal model, ensure the model class implements the {class}`~vllm.model_executor.models.interfaces.SupportsMultiModal` interface. Read more about that [here](#supports-multimodal). ``` ```{note} Although you can directly put these code snippets in your script using `vllm.LLM`, the recommended way is to place these snippets in a vLLM plugin. This ensures compatibility with various vLLM features like distributed inference and the API server. ``` --- ### Source Contributing Model Tests (`docs/source/contributing/model/tests.md`) (new-model-tests)= # Writing Unit Tests This page explains how to write unit tests to verify the implementation of your model. ## Required Tests These tests are necessary to get your PR merged into vLLM library. Without them, the CI for your PR will fail. ### Model loading Include an example HuggingFace repository for your model in . This enables a unit test that loads dummy weights to ensure that the model can be initialized in vLLM. ```{important} The list of models in each section should be maintained in alphabetical order. ``` ```{tip} If your model requires a development version of HF Transformers, you can set `min_transformers_version` to skip the test in CI until the model is released. ``` ## Optional Tests These tests are optional to get your PR merged into vLLM library. Passing these tests provides more confidence that your implementation is correct, and helps avoid future regressions. ### Model correctness These tests compare the model outputs of vLLM against [HF Transformers](https://github.com/huggingface/transformers). You can add new tests under the subdirectories of . #### Generative models For [generative models](#generative-models), there are two levels of correctness tests, as defined in : - Exact correctness (`check_outputs_equal`): The text outputted by vLLM should exactly match the text outputted by HF. - Logprobs similarity (`check_logprobs_close`): The logprobs outputted by vLLM should be in the top-k logprobs outputted by HF, and vice versa. #### Pooling models For [pooling models](#pooling-models), we simply check the cosine similarity, as defined in . (mm-processing-tests)= ### Multi-modal processing #### Common tests Adding your model to verifies that the following input combinations result in the same outputs: - Text + multi-modal data - Tokens + multi-modal data - Text + cached multi-modal data - Tokens + cached multi-modal data #### Model-specific tests You can add a new file under to run tests that only apply to your model. For example, if the HF processor for your model accepts user-specified keyword arguments, you can verify that the keyword arguments are being applied correctly, such as in . --- ### Source Deployment Frameworks Bentoml (`docs/source/deployment/frameworks/bentoml.md`) (deployment-bentoml)= # BentoML [BentoML](https://github.com/bentoml/BentoML) allows you to deploy a large language model (LLM) server with vLLM as the backend, which exposes OpenAI-compatible endpoints. You can serve the model locally or containerize it as an OCI-compliant image and deploy it on Kubernetes. For details, see the tutorial [vLLM inference in the BentoML documentation](https://docs.bentoml.com/en/latest/use-cases/large-language-models/vllm.html). --- ### Source Deployment Frameworks Cerebrium (`docs/source/deployment/frameworks/cerebrium.md`) (deployment-cerebrium)= # Cerebrium ```{raw} html

vLLM_plus_cerebrium

``` vLLM can be run on a cloud based GPU machine with [Cerebrium](https://www.cerebrium.ai/), a serverless AI infrastructure platform that makes it easier for companies to build and deploy AI based applications. To install the Cerebrium client, run: ```console pip install cerebrium cerebrium login ``` Next, create your Cerebrium project, run: ```console cerebrium init vllm-project ``` Next, to install the required packages, add the following to your cerebrium.toml: ```toml [cerebrium.deployment] docker_base_image_url = "nvidia/cuda:12.1.1-runtime-ubuntu22.04" [cerebrium.dependencies.pip] vllm = "latest" ``` Next, let us add our code to handle inference for the LLM of your choice (`mistralai/Mistral-7B-Instruct-v0.1` for this example), add the following code to your `main.py`: ```python from vllm import LLM, SamplingParams llm = LLM(model="mistralai/Mistral-7B-Instruct-v0.1") def run(prompts: list[str], temperature: float = 0.8, top_p: float = 0.95): sampling_params = SamplingParams(temperature=temperature, top_p=top_p) outputs = llm.generate(prompts, sampling_params) # Print the outputs. results = [] for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text results.append({"prompt": prompt, "generated_text": generated_text}) return {"results": results} ``` Then, run the following code to deploy it to the cloud: ```console cerebrium deploy ``` If successful, you should be returned a CURL command that you can call inference against. Just remember to end the url with the function name you are calling (in our case`/run`) ```python curl -X POST https://api.cortex.cerebrium.ai/v4/p-xxxxxx/vllm/run \ -H 'Content-Type: application/json' \ -H 'Authorization: ' \ --data '{ "prompts": [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is" ] }' ``` You should get a response like: ```python { "run_id": "52911756-3066-9ae8-bcc9-d9129d1bd262", "result": { "result": [ { "prompt": "Hello, my name is", "generated_text": " Sarah, and I'm a teacher. I teach elementary school students. One of" }, { "prompt": "The president of the United States is", "generated_text": " elected every four years. This is a democratic system.\n\n5. What" }, { "prompt": "The capital of France is", "generated_text": " Paris.\n" }, { "prompt": "The future of AI is", "generated_text": " bright, but it's important to approach it with a balanced and nuanced perspective." } ] }, "run_time_ms": 152.53663063049316 } ``` You now have an autoscaling endpoint where you only pay for the compute you use! --- ### Source Deployment Frameworks Dstack (`docs/source/deployment/frameworks/dstack.md`) (deployment-dstack)= # dstack ```{raw} html

vLLM_plus_dstack

``` vLLM can be run on a cloud based GPU machine with [dstack](https://dstack.ai/), an open-source framework for running LLMs on any cloud. This tutorial assumes that you have already configured credentials, gateway, and GPU quotas on your cloud environment. To install dstack client, run: ```console pip install "dstack[all] dstack server ``` Next, to configure your dstack project, run: ```console mkdir -p vllm-dstack cd vllm-dstack dstack init ``` Next, to provision a VM instance with LLM of your choice (`NousResearch/Llama-2-7b-chat-hf` for this example), create the following `serve.dstack.yml` file for the dstack `Service`: ```yaml type: service python: "3.11" env: - MODEL=NousResearch/Llama-2-7b-chat-hf port: 8000 resources: gpu: 24GB commands: - pip install vllm - vllm serve $MODEL --port 8000 model: format: openai type: chat name: NousResearch/Llama-2-7b-chat-hf ``` Then, run the following CLI for provisioning: ```console $ dstack run . -f serve.dstack.yml ⠸ Getting run plan... Configuration serve.dstack.yml Project deep-diver-main User deep-diver Min resources 2..xCPU, 8GB.., 1xGPU (24GB) Max price - Max duration - Spot policy auto Retry policy no # BACKEND REGION INSTANCE RESOURCES SPOT PRICE 1 gcp us-central1 g2-standard-4 4xCPU, 16GB, 1xL4 (24GB), 100GB (disk) yes $0.223804 2 gcp us-east1 g2-standard-4 4xCPU, 16GB, 1xL4 (24GB), 100GB (disk) yes $0.223804 3 gcp us-west1 g2-standard-4 4xCPU, 16GB, 1xL4 (24GB), 100GB (disk) yes $0.223804 ... Shown 3 of 193 offers, $5.876 max Continue? [y/n]: y ⠙ Submitting run... ⠏ Launching spicy-treefrog-1 (pulling) spicy-treefrog-1 provisioning completed (running) Service is published at ... ``` After the provisioning, you can interact with the model by using the OpenAI SDK: ```python from openai import OpenAI client = OpenAI( base_url="https://gateway.", api_key="" ) completion = client.chat.completions.create( model="NousResearch/Llama-2-7b-chat-hf", messages=[ { "role": "user", "content": "Compose a poem that explains the concept of recursion in programming.", } ] ) print(completion.choices[0].message.content) ``` ```{note} dstack automatically handles authentication on the gateway using dstack's tokens. Meanwhile, if you don't want to configure a gateway, you can provision dstack `Task` instead of `Service`. The `Task` is for development purpose only. If you want to know more about hands-on materials how to serve vLLM using dstack, check out [this repository](https://github.com/dstackai/dstack-examples/tree/main/deployment/vllm) ``` --- ### Source Deployment Frameworks Helm (`docs/source/deployment/frameworks/helm.md`) (deployment-helm)= # Helm A Helm chart to deploy vLLM for Kubernetes Helm is a package manager for Kubernetes. It will help you to deploy vLLM on k8s and automate the deployment of vLLMm Kubernetes applications. With Helm, you can deploy the same framework architecture with different configurations to multiple namespaces by overriding variables values. This guide will walk you through the process of deploying vLLM with Helm, including the necessary prerequisites, steps for helm install and documentation on architecture and values file. ## Prerequisites Before you begin, ensure that you have the following: - A running Kubernetes cluster - NVIDIA Kubernetes Device Plugin (`k8s-device-plugin`): This can be found at [https://github.com/NVIDIA/k8s-device-plugin](https://github.com/NVIDIA/k8s-device-plugin) - Available GPU resources in your cluster - S3 with the model which will be deployed ## Installing the chart To install the chart with the release name `test-vllm`: ```console helm upgrade --install --create-namespace --namespace=ns-vllm test-vllm . -f values.yaml --set secrets.s3endpoint=$ACCESS_POINT --set secrets.s3bucketname=$BUCKET --set secrets.s3accesskeyid=$ACCESS_KEY --set secrets.s3accesskey=$SECRET_KEY ``` ## Uninstalling the Chart To uninstall the `test-vllm` deployment: ```console helm uninstall test-vllm --namespace=ns-vllm ``` The command removes all the Kubernetes components associated with the chart **including persistent volumes** and deletes the release. ## Architecture ```{image} /assets/deployment/architecture_helm_deployment.png ``` ## Values ``` /* Detailed source-code truncated for AI context efficiency. */ ``` --- ### Source Deployment Frameworks Lws (`docs/source/deployment/frameworks/lws.md`) (deployment-lws)= # LWS LeaderWorkerSet (LWS) is a Kubernetes API that aims to address common deployment patterns of AI/ML inference workloads. A major use case is for multi-host/multi-node distributed inference. vLLM can be deployed with [LWS](https://github.com/kubernetes-sigs/lws) on Kubernetes for distributed model serving. Please see [this guide](https://github.com/kubernetes-sigs/lws/tree/main/docs/examples/vllm) for more details on deploying vLLM on Kubernetes using LWS. --- ### Source Deployment Frameworks Modal (`docs/source/deployment/frameworks/modal.md`) (deployment-modal)= # Modal vLLM can be run on cloud GPUs with [Modal](https://modal.com), a serverless computing platform designed for fast auto-scaling. For details on how to deploy vLLM on Modal, see [this tutorial in the Modal documentation](https://modal.com/docs/examples/vllm_inference). --- ### Source Deployment Frameworks Skypilot (`docs/source/deployment/frameworks/skypilot.md`) (deployment-skypilot)= # SkyPilot ```{raw} html

vLLM

``` vLLM can be **run and scaled to multiple service replicas on clouds and Kubernetes** with [SkyPilot](https://github.com/skypilot-org/skypilot), an open-source framework for running LLMs on any cloud. More examples for various open models, such as Llama-3, Mixtral, etc, can be found in [SkyPilot AI gallery](https://skypilot.readthedocs.io/en/latest/gallery/index.html). ## Prerequisites - Go to the [HuggingFace model page](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) and request access to the model `meta-llama/Meta-Llama-3-8B-Instruct`. - Check that you have installed SkyPilot ([docs](https://skypilot.readthedocs.io/en/latest/getting-started/installation.html)). - Check that `sky check` shows clouds or Kubernetes are enabled. ```console pip install skypilot-nightly sky check ``` ## Run on a single instance See the vLLM SkyPilot YAML for serving, [serving.yaml](https://github.com/skypilot-org/skypilot/blob/master/llm/vllm/serve.yaml). ```yaml resources: accelerators: {L4, A10g, A10, L40, A40, A100, A100-80GB} # We can use cheaper accelerators for 8B model. use_spot: True disk_size: 512 # Ensure model checkpoints can fit. disk_tier: best ports: 8081 # Expose to internet traffic. envs: MODEL_NAME: meta-llama/Meta-Llama-3-8B-Instruct HF_TOKEN: # Change to your own huggingface token, or use --env to pass. setup: | conda create -n vllm python=3.10 -y conda activate vllm pip install vllm==0.4.0.post1 # Install Gradio for web UI. pip install gradio openai pip install flash-attn==2.5.7 run: | conda activate vllm echo 'Starting vllm api server...' python -u -m vllm.entrypoints.openai.api_server \ --port 8081 \ --model $MODEL_NAME \ --trust-remote-code \ --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \ 2>&1 | tee api_server.log & echo 'Waiting for vllm api server to start...' while ! `cat api_server.log | grep -q 'Uvicorn running on'`; do sleep 1; done echo 'Starting gradio server...' git clone https://github.com/vllm-project/vllm.git || true python vllm/examples/online_serving/gradio_openai_chatbot_webserver.py \ -m $MODEL_NAME \ --port 8811 \ --model-url http://localhost:8081/v1 \ --stop-token-ids 128009,128001 ``` Start the serving the Llama-3 8B model on any of the candidate GPUs listed (L4, A10g, ...): ```console HF_TOKEN="your-huggingface-token" sky launch serving.yaml --env HF_TOKEN ``` Check the output of the command. There will be a shareable gradio link (like the last line of the following). Open it in your browser to use the LLaMA model to do the text completion. ```console (task, pid=7431) Running on public URL: https://.gradio.live ``` **Optional**: Serve the 70B model instead of the default 8B and use more GPU: ```console HF_TOKEN="your-huggingface-token" sky launch serving.yaml --gpus A100:8 --env HF_TOKEN --env MODEL_NAME=meta-llama/Meta-Llama-3-70B-Instruct ``` ## Scale up to multiple replicas SkyPilot can scale up the service to multiple service replicas with built-in autoscaling, load-balancing and fault-tolerance. You can do it by adding a services section to the YAML file. ```yaml service: replicas: 2 # An actual request for readiness probe. readiness_probe: path: /v1/chat/completions post_data: model: $MODEL_NAME messages: - role: user content: Hello! What is your name? max_completion_tokens: 1 ``` ```{raw} html
Click to see the full recipe YAML ``` ```yaml service: replicas: 2 # An actual request for readiness probe. readiness_probe: path: /v1/chat/completions post_data: model: $MODEL_NAME messages: - role: user content: Hello! What is your name? max_completion_tokens: 1 resources: accelerators: {L4, A10g, A10, L40, A40, A100, A100-80GB} # We can use cheaper accelerators for 8B model. use_spot: True disk_size: 512 # Ensure model checkpoints can fit. disk_tier: best ports: 8081 # Expose to internet traffic. envs: MODEL_NAME: meta-llama/Meta-Llama-3-8B-Instruct HF_TOKEN: # Change to your own huggingface token, or use --env to pass. setup: | conda create -n vllm python=3.10 -y conda activate vllm pip install vllm==0.4.0.post1 # Install Gradio for web UI. pip install gradio openai pip install flash-attn==2.5.7 run: | conda activate vllm echo 'Starting vllm api server...' python -u -m vllm.entrypoints.openai.api_server \ --port 8081 \ --model $MODEL_NAME \ --trust-remote-code \ --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \ 2>&1 | tee api_server.log ``` ```{raw} html
``` Start the serving the Llama-3 8B model on multiple replicas: ```console HF_TOKEN="your-huggingface-token" sky serve up -n vllm serving.yaml --env HF_TOKEN ``` Wait until the service is ready: ```console watch -n10 sky serve status vllm ``` ```{raw} html
Example outputs: ``` ```console Services NAME VERSION UPTIME STATUS REPLICAS ENDPOINT vllm 1 35s READY 2/2 xx.yy.zz.100:30001 Service Replicas SERVICE_NAME ID VERSION IP LAUNCHED RESOURCES STATUS REGION vllm 1 1 xx.yy.zz.121 18 mins ago 1x GCP([Spot]{'L4': 1}) READY us-east4 vllm 2 1 xx.yy.zz.245 18 mins ago 1x GCP([Spot]{'L4': 1}) READY us-east4 ``` ```{raw} html
``` After the service is READY, you can find a single endpoint for the service and access the service with the endpoint: ```console ENDPOINT=$(sky serve status --endpoint 8081 vllm) curl -L http://$ENDPOINT/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "meta-llama/Meta-Llama-3-8B-Instruct", "messages": [ { "role": "system", "content": "You are a helpful assistant." }, { "role": "user", "content": "Who are you?" } ], "stop_token_ids": [128009, 128001] }' ``` To enable autoscaling, you could replace the `replicas` with the following configs in `service`: ```yaml service: replica_policy: min_replicas: 2 max_replicas: 4 target_qps_per_replica: 2 ``` This will scale the service up to when the QPS exceeds 2 for each replica. ```{raw} html
Click to see the full recipe YAML ``` ```yaml service: replica_policy: min_replicas: 2 max_replicas: 4 target_qps_per_replica: 2 # An actual request for readiness probe. readiness_probe: path: /v1/chat/completions post_data: model: $MODEL_NAME messages: - role: user content: Hello! What is your name? max_completion_tokens: 1 resources: accelerators: {L4, A10g, A10, L40, A40, A100, A100-80GB} # We can use cheaper accelerators for 8B model. use_spot: True disk_size: 512 # Ensure model checkpoints can fit. disk_tier: best ports: 8081 # Expose to internet traffic. envs: MODEL_NAME: meta-llama/Meta-Llama-3-8B-Instruct HF_TOKEN: # Change to your own huggingface token, or use --env to pass. setup: | conda create -n vllm python=3.10 -y conda activate vllm pip install vllm==0.4.0.post1 # Install Gradio for web UI. pip install gradio openai pip install flash-attn==2.5.7 run: | conda activate vllm echo 'Starting vllm api server...' python -u -m vllm.entrypoints.openai.api_server \ --port 8081 \ --model $MODEL_NAME \ --trust-remote-code \ --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \ 2>&1 | tee api_server.log ``` ```{raw} html
``` To update the service with the new config: ```console HF_TOKEN="your-huggingface-token" sky serve update vllm serving.yaml --env HF_TOKEN ``` To stop the service: ```console sky serve down vllm ``` ### **Optional**: Connect a GUI to the endpoint It is also possible to access the Llama-3 service with a separate GUI frontend, so the user requests send to the GUI will be load-balanced across replicas. ```{raw} html
Click to see the full GUI YAML ``` ```yaml envs: MODEL_NAME: meta-llama/Meta-Llama-3-8B-Instruct ENDPOINT: x.x.x.x:3031 # Address of the API server running vllm. resources: cpus: 2 setup: | conda create -n vllm python=3.10 -y conda activate vllm # Install Gradio for web UI. pip install gradio openai run: | conda activate vllm export PATH=$PATH:/sbin echo 'Starting gradio server...' git clone https://github.com/vllm-project/vllm.git || true python vllm/examples/online_serving/gradio_openai_chatbot_webserver.py \ -m $MODEL_NAME \ --port 8811 \ --model-url http://$ENDPOINT/v1 \ --stop-token-ids 128009,128001 | tee ~/gradio.log ``` ```{raw} html
``` 1. Start the chat web UI: ```console sky launch -c gui ./gui.yaml --env ENDPOINT=$(sky serve status --endpoint vllm) ``` 2. Then, we can access the GUI at the returned gradio link: ```console | INFO | stdout | Running on public URL: https://6141e84201ce0bb4ed.gradio.live ``` --- ### Source Deployment Frameworks Triton (`docs/source/deployment/frameworks/triton.md`) (deployment-triton)= # NVIDIA Triton The [Triton Inference Server](https://github.com/triton-inference-server) hosts a tutorial demonstrating how to quickly deploy a simple [facebook/opt-125m](https://huggingface.co/facebook/opt-125m) model using vLLM. Please see [Deploying a vLLM model in Triton](https://github.com/triton-inference-server/tutorials/blob/main/Quick_Deploy/vLLM/README.md#deploying-a-vllm-model-in-triton) for more details. --- ### Source Deployment Integrations Kserve (`docs/source/deployment/integrations/kserve.md`) (deployment-kserve)= # KServe vLLM can be deployed with [KServe](https://github.com/kserve/kserve) on Kubernetes for highly scalable distributed model serving. Please see [this guide](https://kserve.github.io/website/latest/modelserving/v1beta1/llm/huggingface/) for more details on using vLLM with KServe. --- ### Source Deployment Integrations Kubeai (`docs/source/deployment/integrations/kubeai.md`) (deployment-kubeai)= # KubeAI [KubeAI](https://github.com/substratusai/kubeai) is a Kubernetes operator that enables you to deploy and manage AI models on Kubernetes. It provides a simple and scalable way to deploy vLLM in production. Functionality such as scale-from-zero, load based autoscaling, model caching, and much more is provided out of the box with zero external dependencies. Please see the Installation Guides for environment specific instructions: - [Any Kubernetes Cluster](https://www.kubeai.org/installation/any/) - [EKS](https://www.kubeai.org/installation/eks/) - [GKE](https://www.kubeai.org/installation/gke/) Once you have KubeAI installed, you can [configure text generation models](https://www.kubeai.org/how-to/configure-text-generation-models/) using vLLM. --- ### Source Deployment Integrations Llamastack (`docs/source/deployment/integrations/llamastack.md`) (deployment-llamastack)= # Llama Stack vLLM is also available via [Llama Stack](https://github.com/meta-llama/llama-stack) . To install Llama Stack, run ```console pip install llama-stack -q ``` ## Inference using OpenAI Compatible API Then start Llama Stack server pointing to your vLLM server with the following configuration: ```yaml inference: - provider_id: vllm0 provider_type: remote::vllm config: url: http://127.0.0.1:8000 ``` Please refer to [this guide](https://llama-stack.readthedocs.io/en/latest/distributions/self_hosted_distro/remote-vllm.html) for more details on this remote vLLM provider. ## Inference via Embedded vLLM An [inline vLLM provider](https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/inline/inference/vllm) is also available. This is a sample of configuration using that method: ```yaml inference - provider_type: vllm config: model: Llama3.1-8B-Instruct tensor_parallel_size: 4 ``` --- ### Source Design Kernel Paged Attention (`docs/source/design/kernel/paged_attention.md`) (design-paged-attention)= # vLLM Paged Attention - Currently, vLLM utilizes its own implementation of a multi-head query attention kernel (`csrc/attention/attention_kernels.cu`). This kernel is designed to be compatible with vLLM's paged KV caches, where the key and value cache are stored in separate blocks (note that this block concept differs from the GPU thread block. So in a later document, I will refer to vLLM paged attention block as "block", while refer to GPU thread block as "thread block"). - To achieve high performance, this kernel relies on a specially designed memory layout and access method, specifically when threads read data from global memory to shared memory. The purpose of this document is to provide a high-level explanation of the kernel implementation step by step, aiding those who wish to learn about the vLLM multi-head query attention kernel. After going through this document, users will likely have a better understanding and feel easier to follow the actual implementation. - Please note that this document may not cover all details, such as how to calculate the correct index for the corresponding data or the dot multiplication implementation. However, after reading this document and becoming familiar with the high-level logic flow, it should be easier for you to read the actual code and understand the details. ## Inputs - The kernel function takes a list of arguments for the current thread to perform its assigned work. The three most important arguments are the input pointers `q`, `k_cache`, and `v_cache`, which point to query, key, and value data on global memory that need to be read and processed. The output pointer `out` points to global memory where the result should be written. These four pointers actually refer to multi-dimensional arrays, but each thread only accesses the portion of data assigned to it. I have omitted all other runtime parameters here for simplicity. ```cpp template< typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int NUM_THREADS, int PARTITION_SIZE = 0> __device__ void paged_attention_kernel( ... // Other side args. const scalar_t* __restrict__ out, // [num_seqs, num_heads, max_num_partitions, head_size] const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size] const scalar_t* __restrict__ k_cache, // [num_blocks, num_kv_heads, head_size/x, block_size, x] const scalar_t* __restrict__ v_cache, // [num_blocks, num_kv_heads, head_size, block_size] ... // Other side args. ) ``` - There are also a list of template arguments above the function signature that are determined during compilation time. `scalar_t` represents the data type of the query, key, and value data elements, such as FP16. `HEAD_SIZE` indicates the number of elements in each head. `BLOCK_SIZE` refers to the number of tokens in each block. `NUM_THREADS` denotes the number of threads in each thread block. `PARTITION_SIZE` represents the number of tensor parallel GPUs (For simplicity, we assume this is 0 and tensor parallel is disabled). - With these arguments, we need to perform a sequence of preparations. This includes calculating the current head index, block index, and other necessary variables. However, for now, we can ignore these preparations and proceed directly to the actual calculations. It will be easier to understand them once we grasp the entire flow. ## Concepts - Just before we dive into the calculation flow, I want to describe a few concepts that are needed for later sections. However, you may skip this section and return later if you encounter any confusing terminologies. - **Sequence**: A sequence represents a client request. For example, the data pointed to by `q` has a shape of `[num_seqs, num_heads, head_size]`. That represents there are total `num_seqs` of query sequence data are pointed by `q`. Since this kernel is a single query attention kernel, each sequence only has one query token. Hence, the `num_seqs` equals the total number of tokens that are processed in the batch. - **Context**: The context consists of the generated tokens from the sequence. For instance, `["What", "is", "your"]` are the context tokens, and the input query token is `"name"`. The model might generate the token `"?"`. - **Vec**: The vec is a list of elements that are fetched and calculated together. For query and key data, the vec size (`VEC_SIZE`) is determined so that each thread group can fetch and calculate 16 bytes of data at a time. For value data, the vec size (`V_VEC_SIZE`) is determined so that each thread can fetch and calculate 16 bytes of data at a time. For example, if the `scalar_t` is FP16 (2 bytes) and `THREAD_GROUP_SIZE` is 2, the `VEC_SIZE` will be 4, while the `V_VEC_SIZE` will be 8. - **Thread group**: The thread group is a small group of threads(`THREAD_GROUP_SIZE`) that fetches and calculates one query token and one key token at a time. Each thread handles only a portion of the token data. The total number of elements processed by one thread group is referred as `x`. For example, if the thread group contains 2 threads and the head size is 8, then thread 0 handles the query and key elements at index 0, 2, 4, 6, while thread 1 handles the elements at index 1, 3, 5, 7. - **Block**: The key and value cache data in vLLM are split into blocks. Each block stores data for a fixed number(`BLOCK_SIZE`) of tokens at one head. Each block may contain only a portion of the whole context tokens. For example, if the block size is 16 and the head size is 128, then for one head, one block can store 16 * 128 = 2048 elements. - **Warp**: A warp is a group of 32 threads(`WARP_SIZE`) that execute simultaneously on a stream multiprocessor (SM). In this kernel, each warp processes the calculation between one query token and key tokens of one entire block at a time (it may process multiple blocks in multiple iterations). For example, if there are 4 warps and 6 blocks for one context, the assignment would be like warp 0 handles the 0th, 4th blocks, warp 1 handles the 1st, 5th blocks, warp 2 handles the 2nd block and warp 3 handles the 3rd block. - **Thread block**: A thread block is a group of threads(`NUM_THREADS`) that can access the same shared memory. Each thread block contains multiple warps(`NUM_WARPS`), and in this kernel, each thread block processes the calculation between one query token and key tokens of a whole context. - **Grid**: A grid is a collection of thread blocks and defines the shape of the collection. In this kernel, the shape is `(num_heads, num_seqs, max_num_partitions)`. Therefore, each thread block only handles the calculation for one head, one sequence, and one partition. ## Query - This section will introduce how query data is stored in memory and fetched by each thread. As mentioned above, each thread group fetches one query token data, while each thread itself only handles a part of one query token data. Within each warp, every thread group will fetch the same query token data, but will multiply it with different key token data. ```cpp const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE; ``` ```{figure} ../../assets/kernel/query.png :align: center :alt: query :width: 70% Query data of one token at one head ``` - Each thread defines its own `q_ptr` which points to the assigned query token data on global memory. For example, if `VEC_SIZE` is 4 and `HEAD_SIZE` is 128, the `q_ptr` points to data that contains total of 128 elements divided into 128 / 4 = 32 vecs. ```{figure} ../../assets/kernel/q_vecs.png :align: center :alt: q_vecs :width: 70% `q_vecs` for one thread group ``` ```cpp __shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD]; ``` - Next, we need to read the global memory data pointed to by `q_ptr` into shared memory as `q_vecs`. It is important to note that each vecs is assigned to a different row. For example, if the `THREAD_GROUP_SIZE` is 2, thread 0 will handle the 0th row vecs, while thread 1 handles the 1st row vecs. By reading the query data in this way, neighboring threads like thread 0 and thread 1 can read neighbor memory, achieving the memory coalescing to improve performance. ## Key - Similar to the "Query" section, this section introduces memory layout and assignment for keys. While each thread group only handle one query token one kernel run, it may handle multiple key tokens across multiple iterations. Meanwhile, each warp will process multiple blocks of key tokens in multiple iterations, ensuring that all context tokens are processed by the entire thread group after the kernel run. In this context, "handle" refers to performing the dot multiplication between query data and key data. ```cpp const scalar_t* k_ptr = k_cache + physical_block_number * kv_block_stride + kv_head_idx * kv_head_stride + physical_block_offset * x; ``` - Unlike to `q_ptr`, `k_ptr` in each thread will point to different key token at different iterations. As shown above, that `k_ptr` points to key token data based on `k_cache` at assigned block, assigned head and assigned token. ```{figure} ../../assets/kernel/key.png :align: center :alt: key :width: 70% Key data of all context tokens at one head ``` - The diagram above illustrates the memory layout for key data. It assumes that the `BLOCK_SIZE` is 16, `HEAD_SIZE` is 128, `x` is 8, `THREAD_GROUP_SIZE` is 2, and there are a total of 4 warps. Each rectangle represents all the elements for one key token at one head, which will be processed by one thread group. The left half shows the total 16 blocks of key token data for warp 0, while the right half represents the remaining key token data for other warps or iterations. Inside each rectangle, there are a total 32 vecs (128 elements for one token) that will be processed by 2 threads (one thread group) separately. ```{figure} ../../assets/kernel/k_vecs.png :align: center :alt: k_vecs :width: 70% `k_vecs` for one thread ``` ```cpp K_vec k_vecs[NUM_VECS_PER_THREAD] ``` - Next, we need to read the key token data from `k_ptr` and store them on register memory as `k_vecs`. We use register memory for `k_vecs` because it will only be accessed by one thread once, whereas `q_vecs` will be accessed by multiple threads multiple times. Each `k_vecs` will contain multiple vectors for later calculation. Each vec will be set at each inner iteration. The assignment of vecs allows neighboring threads in a warp to read neighboring memory together, which again promotes the memory coalescing. For instance, thread 0 will read vec 0, while thread 1 will read vec 1. In the next inner loop, thread 0 will read vec 2, while thread 1 will read vec 3, and so on. - You may still be a little confused about the overall flow. Don't worry, please keep reading the next "QK" section. It will illustrate the query and key calculation flow in a clearer and higher-level manner. ## QK - As shown the pseudo code below, before the entire for loop block, we fetch the query data for one token and store it in `q_vecs`. Then, in the outer for loop, we iterate through different `k_ptrs` that point to different tokens and prepare the `k_vecs` in the inner for loop. Finally, we perform the dot multiplication between the `q_vecs` and each `k_vecs`. ```cpp q_vecs = ... for ... { k_ptr = ... for ... { k_vecs[i] = ... } ... float qk = scale * Qk_dot::dot(q_vecs[thread_group_offset], k_vecs); } ``` - As mentioned before, for each thread, it only fetches part of the query and key token data at a time. However, there will be a cross thread group reduction happen in the `Qk_dot<>::dot` . So `qk` returned here is not just between part of the query and key token dot multiplication, but actually a full result between entire query and key token data. - For example, if the value of `HEAD_SIZE` is 128 and `THREAD_GROUP_SIZE` is 2, each thread's `k_vecs` will contain total 64 elements. However, the returned `qk` is actually the result of dot multiplication between 128 query elements and 128 key elements. If you want to learn more about the details of the dot multiplication and reduction, you may refer to the implementation of `Qk_dot<>::dot`. However, for the sake of simplicity, I will not cover it in this document. ## Softmax - Next, we need to calculate the normalized softmax for all `qk`s, as shown above, where each $x$ represents a `qk`. To do this, we must obtain the reduced value of `qk_max`($m(x)$) and the `exp_sum`($\ell(x)$) of all `qk`s. The reduction should be performed across the entire thread block, encompassing results between the query token and all context key tokens. ```{math} :nowrap: true \begin{gather*} m(x):=\max _i \quad x_i \\ \quad f(x):=\left[\begin{array}{lll}e^{x_1-m(x)} & \ldots & e^{x_B-m(x)}\end{array}\right]\\ \quad \ell(x):=\sum_i f(x)_i \\ \quad \operatorname{softmax}(x):=\frac{f(x)}{\ell(x)} \end{gather*} ``` ### `qk_max` and `logits` - Just right after we get the `qk` result, we can set the temporary `logits` result with `qk` (In the end, the `logits` should store the normalized softmax result). Also we can compare and collect the `qk_max` for all `qk`s that are calculated by current thread group. ```cpp if (thread_group_offset == 0) { const bool mask = token_idx >= context_len; logits[token_idx - start_token_idx] = mask ? 0.f : qk; qk_max = mask ? qk_max : fmaxf(qk_max, qk); } ``` - Please note that the `logits` here is on shared memory, so each thread group will set the fields for its own assigned context tokens. Overall, the size of logits should be number of context tokens. ```cpp for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) { qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask)); } if (lane == 0) { red_smem[warp_idx] = qk_max; } ``` - Then we need to get the reduced `qk_max` across each warp. The main idea is to make threads in warp to communicate with each other and get the final max `qk` . ```cpp for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) { qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask)); } qk_max = VLLM_SHFL_SYNC(qk_max, 0); ``` - Finally, we can get the reduced `qk_max` from whole thread block by compare the `qk_max` from all warps in this thread block. Then we need to broadcast the final result to each thread. ### `exp_sum` - Similar to `qk_max`, we need to get the reduced sum value from the entire thread block too. ```cpp for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) { float val = __expf(logits[i] - qk_max); logits[i] = val; exp_sum += val; } ... exp_sum = block_sum(&red_smem[NUM_WARPS], exp_sum); ``` - Firstly, sum all exp values from each thread group, and meanwhile, convert each entry of `logits` from `qk` to `exp(qk - qk_max)`. Please note, the `qk_max` here is already the max `qk` across the whole thread block. And then we can do reduction for `exp_sum` across whole thread block just like the `qk_max`. ```cpp const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f); for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) { logits[i] *= inv_sum; } ``` - Finally, with the reduced `qk_max` and `exp_sum`, we can obtain the final normalized softmax result as `logits`. This `logits` variable will be used for dot multiplication with the value data in later steps. Now, it should store the normalized softmax result of `qk` for all assigned context tokens. ## Value ```{figure} ../../assets/kernel/value.png :align: center :alt: value :width: 70% Value data of all context tokens at one head ``` ```{figure} ../../assets/kernel/logits_vec.png :align: center :alt: logits_vec :width: 50% `logits_vec` for one thread ``` ```{figure} ../../assets/kernel/v_vec.png :align: center :alt: v_vec :width: 70% List of `v_vec` for one thread ``` - Now we need to retrieve the value data and perform dot multiplication with `logits`. Unlike query and key, there is no thread group concept for value data. As shown in diagram, different from key token memory layout, elements from the same column correspond to the same value token. For one block of value data, there are `HEAD_SIZE` of rows and `BLOCK_SIZE` of columns that are split into multiple `v_vecs`. - Each thread always fetches `V_VEC_SIZE` elements from the same `V_VEC_SIZE` of tokens at a time. As a result, a single thread retrieves multiple `v_vec`s from different rows and the same columns through multiple inner iterations. For each `v_vec`, it needs to be dot multiplied with the corresponding `logits_vec`, which is also `V_VEC_SIZE` elements from `logits`. Overall, with multiple inner iterations, each warp will process one block of value tokens. And with multiple outer iterations, the whole context value tokens are processd ```cpp float accs[NUM_ROWS_PER_THREAD]; for ... { // Iteration over different blocks. logits_vec = ... for ... { // Iteration over different rows. v_vec = ... ... accs[i] += dot(logits_vec, v_vec); } } ``` - As shown in the above pseudo code, in the outer loop, similar to `k_ptr`, `logits_vec` iterates over different blocks and reads `V_VEC_SIZE` elements from `logits`. In the inner loop, each thread reads `V_VEC_SIZE` elements from the same tokens as a `v_vec` and performs dot multiplication. It is important to note that in each inner iteration, the thread fetches different head position elements for the same tokens. The dot result is then accumulated in `accs`. Therefore, each entry of `accs` is mapped to a head position assigned to the current thread. - For example, if `BLOCK_SIZE` is 16 and `V_VEC_SIZE` is 8, each thread fetches 8 value elements for 8 tokens at a time. Each element is from different tokens at the same head position. If `HEAD_SIZE` is 128 and `WARP_SIZE` is 32, for each inner loop, a warp needs to fetch `WARP_SIZE * V_VEC_SIZE = 256` elements. This means there are a total of 128 * 16 / 256 = 8 inner iterations for a warp to handle a whole block of value tokens. And each `accs` in each thread contains 8 elements that accumulated at 8 different head positions. For the thread 0, the `accs` variable will have 8 elements, which are 0th, 32th … 224th elements of a value head that are accumulated from all assigned 8 tokens. ## LV - Now, we need to perform reduction for `accs` within each warp. This process allows each thread to accumulate the `accs` for the assigned head positions of all tokens in one block. ```cpp for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) { float acc = accs[i]; for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) { acc += VLLM_SHFL_XOR_SYNC(acc, mask); } accs[i] = acc; } ``` - Next, we perform reduction for `accs` across all warps, allowing each thread to have the accumulation of `accs` for the assigned head positions of all context tokens. Please note that each `accs` in every thread only stores the accumulation for a portion of elements of the entire head for all context tokens. However, overall, all results for output have been calculated but are just stored in different thread register memory. ```cpp float* out_smem = reinterpret_cast(shared_mem); for (int i = NUM_WARPS; i > 1; i /= 2) { // Upper warps write to shared memory. ... float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE]; for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) { ... dst[row_idx] = accs[i]; } // Lower warps update the output. const float* src = &out_smem[warp_idx * HEAD_SIZE]; for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) { ... accs[i] += src[row_idx]; } // Write out the accs. } ``` ## Output - Now we can write all of calculated result from local register memory to final output global memory. ```cpp scalar_t* out_ptr = out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE + head_idx * max_num_partitions * HEAD_SIZE + partition_idx * HEAD_SIZE; ``` - First, we need to define the `out_ptr` variable, which points to the start address of the assigned sequence and assigned head. ```cpp for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) { const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER; if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) { from_float(*(out_ptr + row_idx), accs[i]); } } ``` - Finally, we need to iterate over different assigned head positions and write out the corresponding accumulated result based on the `out_ptr`. --- ### Source Features Quantization Auto Awq (`docs/source/features/quantization/auto_awq.md`) (auto-awq)= # AutoAWQ ```{warning} Please note that AWQ support in vLLM is under-optimized at the moment. We would recommend using the unquantized version of the model for better accuracy and higher throughput. Currently, you can use AWQ as a way to reduce memory footprint. As of now, it is more suitable for low latency inference with small number of concurrent requests. vLLM's AWQ implementation have lower throughput than unquantized version. ``` To create a new 4-bit quantized model, you can leverage [AutoAWQ](https://github.com/casper-hansen/AutoAWQ). Quantizing reduces the model's precision from FP16 to INT4 which effectively reduces the file size by ~70%. The main benefits are lower latency and memory usage. You can quantize your own models by installing AutoAWQ or picking one of the [400+ models on Huggingface](https://huggingface.co/models?sort=trending&search=awq). ```console pip install autoawq ``` After installing AutoAWQ, you are ready to quantize a model. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`: ```python from awq import AutoAWQForCausalLM from transformers import AutoTokenizer model_path = 'mistralai/Mistral-7B-Instruct-v0.2' quant_path = 'mistral-instruct-v0.2-awq' quant_config = { "zero_point": True, "q_group_size": 128, "w_bit": 4, "version": "GEMM" } # Load model model = AutoAWQForCausalLM.from_pretrained( model_path, **{"low_cpu_mem_usage": True, "use_cache": False} ) tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) # Quantize model.quantize(tokenizer, quant_config=quant_config) # Save quantized model model.save_quantized(quant_path) tokenizer.save_pretrained(quant_path) print(f'Model is quantized and saved at "{quant_path}"') ``` To run an AWQ model with vLLM, you can use [TheBloke/Llama-2-7b-Chat-AWQ](https://huggingface.co/TheBloke/Llama-2-7b-Chat-AWQ) with the following command: ```console python examples/offline_inference/llm_engine_example.py --model TheBloke/Llama-2-7b-Chat-AWQ --quantization awq ``` AWQ models are also supported directly through the LLM entrypoint: ```python from vllm import LLM, SamplingParams # Sample prompts. prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] # Create a sampling params object. sampling_params = SamplingParams(temperature=0.8, top_p=0.95) # Create an LLM. llm = LLM(model="TheBloke/Llama-2-7b-Chat-AWQ", quantization="AWQ") # Generate texts from the prompts. The output is a list of RequestOutput objects # that contain the prompt, generated text, and other information. outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` --- ### Source Features Quantization Bnb (`docs/source/features/quantization/bnb.md`) (bits-and-bytes)= # BitsAndBytes vLLM now supports [BitsAndBytes](https://github.com/TimDettmers/bitsandbytes) for more efficient model inference. BitsAndBytes quantizes models to reduce memory usage and enhance performance without significantly sacrificing accuracy. Compared to other quantization methods, BitsAndBytes eliminates the need for calibrating the quantized model with input data. Below are the steps to utilize BitsAndBytes with vLLM. ```console pip install bitsandbytes>=0.45.0 ``` vLLM reads the model's config file and supports both in-flight quantization and pre-quantized checkpoint. You can find bitsandbytes quantized models on . And usually, these repositories have a config.json file that includes a quantization_config section. ## Read quantized checkpoint ```python from vllm import LLM import torch # unsloth/tinyllama-bnb-4bit is a pre-quantized checkpoint. model_id = "unsloth/tinyllama-bnb-4bit" llm = LLM(model=model_id, dtype=torch.bfloat16, trust_remote_code=True, \ quantization="bitsandbytes", load_format="bitsandbytes") ``` ## Inflight quantization: load as 4bit quantization ```python from vllm import LLM import torch model_id = "huggyllama/llama-7b" llm = LLM(model=model_id, dtype=torch.bfloat16, trust_remote_code=True, \ quantization="bitsandbytes", load_format="bitsandbytes") ``` ## OpenAI Compatible Server Append the following to your 4bit model arguments: ```console --quantization bitsandbytes --load-format bitsandbytes ``` --- ### Source Features Quantization Fp8 (`docs/source/features/quantization/fp8.md`) (fp8)= # FP8 W8A8 vLLM supports FP8 (8-bit floating point) weight and activation quantization using hardware acceleration on GPUs such as Nvidia H100 and AMD MI300x. Currently, only Hopper and Ada Lovelace GPUs are officially supported for W8A8. Ampere GPUs are supported for W8A16 (weight-only FP8) utilizing Marlin kernels. Quantization of models with FP8 allows for a 2x reduction in model memory requirements and up to a 1.6x improvement in throughput with minimal impact on accuracy. Please visit the HF collection of [quantized FP8 checkpoints of popular LLMs ready to use with vLLM](https://huggingface.co/collections/neuralmagic/fp8-llms-for-vllm-666742ed2b78b7ac8df13127). The FP8 types typically supported in hardware have two distinct representations, each useful in different scenarios: - **E4M3**: Consists of 1 sign bit, 4 exponent bits, and 3 bits of mantissa. It can store values up to +/-448 and `nan`. - **E5M2**: Consists of 1 sign bit, 5 exponent bits, and 2 bits of mantissa. It can store values up to +/-57344, +/- `inf`, and `nan`. The tradeoff for the increased dynamic range is lower precision of the stored values. ```{note} FP8 computation is supported on NVIDIA GPUs with compute capability > 8.9 (Ada Lovelace, Hopper). FP8 models will run on compute capability > 8.0 (Ampere) as weight-only W8A16, utilizing FP8 Marlin. ``` ## Quick Start with Online Dynamic Quantization Dynamic quantization of an original precision BF16/FP16 model to FP8 can be achieved with vLLM without any calibration data required. You can enable the feature by specifying `--quantization="fp8"` in the command line or setting `quantization="fp8"` in the LLM constructor. In this mode, all Linear modules (except for the final `lm_head`) have their weights quantized down to FP8_E4M3 precision with a per-tensor scale. Activations have their minimum and maximum values calculated during each forward pass to provide a dynamic per-tensor scale for high accuracy. As a result, latency improvements are limited in this mode. ```python from vllm import LLM model = LLM("facebook/opt-125m", quantization="fp8") # INFO 06-10 17:55:42 model_runner.py:157] Loading model weights took 0.1550 GB result = model.generate("Hello, my name is") ``` ```{warning} Currently, we load the model at original precision before quantizing down to 8-bits, so you need enough memory to load the whole model. ``` ## Installation To produce performant FP8 quantized models with vLLM, you'll need to install the [llm-compressor](https://github.com/vllm-project/llm-compressor/) library: ```console pip install llmcompressor ``` ## Quantization Process The quantization process involves three main steps: 1. Loading the model 2. Applying quantization 3. Evaluating accuracy in vLLM ### 1. Loading the Model Load your model and tokenizer using the standard `transformers` AutoModel classes: ```python from transformers import AutoTokenizer, AutoModelForCausalLM MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct" model = AutoModelForCausalLM.from_pretrained( MODEL_ID, device_map="auto", torch_dtype="auto", ) tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) ``` ### 2. Applying Quantization For FP8 quantization, we can recover accuracy with simple RTN quantization. We recommend targeting all `Linear` layers using the `FP8_DYNAMIC` scheme, which uses: - Static, per-channel quantization on the weights - Dynamic, per-token quantization on the activations Since simple RTN does not require data for weight quantization and the activations are quantized dynamically, we do not need any calibration data for this quantization flow. ```python from llmcompressor.transformers import oneshot from llmcompressor.modifiers.quantization import QuantizationModifier # Configure the simple PTQ quantization recipe = QuantizationModifier( targets="Linear", scheme="FP8_DYNAMIC", ignore=["lm_head"]) # Apply the quantization algorithm. oneshot(model=model, recipe=recipe) # Save the model. SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic" model.save_pretrained(SAVE_DIR) tokenizer.save_pretrained(SAVE_DIR) ``` ### 3. Evaluating Accuracy Install `vllm` and `lm-evaluation-harness`: ```console pip install vllm lm-eval==0.4.4 ``` Load and run the model in `vllm`: ```python from vllm import LLM model = LLM("./Meta-Llama-3-8B-Instruct-FP8-Dynamic") model.generate("Hello my name is") ``` Evaluate accuracy with `lm_eval` (for example on 250 samples of `gsm8k`): ```{note} Quantized models can be sensitive to the presence of the `bos` token. `lm_eval` does not add a `bos` token by default, so make sure to include the `add_bos_token=True` argument when running your evaluations. ``` ```console $ MODEL=$PWD/Meta-Llama-3-8B-Instruct-FP8-Dynamic $ lm_eval \ --model vllm \ --model_args pretrained=$MODEL,add_bos_token=True \ --tasks gsm8k --num_fewshot 5 --batch_size auto --limit 250 ``` Here's an example of the resulting scores: ```text |Tasks|Version| Filter |n-shot| Metric | |Value| |Stderr| |-----|------:|----------------|-----:|-----------|---|----:|---|-----:| |gsm8k| 3|flexible-extract| 5|exact_match|↑ |0.768|± |0.0268| | | |strict-match | 5|exact_match|↑ |0.768|± |0.0268| ``` ## Troubleshooting and Support If you encounter any issues or have feature requests, please open an issue on the `vllm-project/llm-compressor` GitHub repository. ## Deprecated Flow ```{note} The following information is preserved for reference and search purposes. The quantization method described below is deprecated in favor of the `llmcompressor` method described above. ``` For static per-tensor offline quantization to FP8, please install the [AutoFP8 library](https://github.com/neuralmagic/autofp8). ```bash git clone https://github.com/neuralmagic/AutoFP8.git pip install -e AutoFP8 ``` This package introduces the `AutoFP8ForCausalLM` and `BaseQuantizeConfig` objects for managing how your model will be compressed. ## Offline Quantization with Static Activation Scaling Factors You can use AutoFP8 with calibration data to produce per-tensor static scales for both the weights and activations by enabling the `activation_scheme="static"` argument. ```python from datasets import load_dataset from transformers import AutoTokenizer from auto_fp8 import AutoFP8ForCausalLM, BaseQuantizeConfig pretrained_model_dir = "meta-llama/Meta-Llama-3-8B-Instruct" quantized_model_dir = "Meta-Llama-3-8B-Instruct-FP8" tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True) tokenizer.pad_token = tokenizer.eos_token # Load and tokenize 512 dataset samples for calibration of activation scales ds = load_dataset("mgoin/ultrachat_2k", split="train_sft").select(range(512)) examples = [tokenizer.apply_chat_template(batch["messages"], tokenize=False) for batch in ds] examples = tokenizer(examples, padding=True, truncation=True, return_tensors="pt").to("cuda") # Define quantization config with static activation scales quantize_config = BaseQuantizeConfig(quant_method="fp8", activation_scheme="static") # Load the model, quantize, and save checkpoint model = AutoFP8ForCausalLM.from_pretrained(pretrained_model_dir, quantize_config) model.quantize(examples) model.save_quantized(quantized_model_dir) ``` Your model checkpoint with quantized weights and activations should be available at `Meta-Llama-3-8B-Instruct-FP8/`. Finally, you can load the quantized model checkpoint directly in vLLM. ```python from vllm import LLM model = LLM(model="Meta-Llama-3-8B-Instruct-FP8/") # INFO 06-10 21:15:41 model_runner.py:159] Loading model weights took 8.4596 GB result = model.generate("Hello, my name is") ``` --- ### Source Features Quantization Fp8 E4m3 Kvcache (`docs/source/features/quantization/fp8_e4m3_kvcache.md`) (fp8-e4m3-kvcache)= # FP8 E4M3 KV Cache Quantizing the KV cache to FP8 reduces its memory footprint. This increases the number of tokens that can be stored in the cache, improving throughput. OCP (Open Compute Project www.opencompute.org) specifies two common 8-bit floating point data formats: E5M2 (5 exponent bits and 2 mantissa bits) and E4M3FN (4 exponent bits and 3 mantissa bits), often shortened as E4M3. One benefit of the E4M3 format over E5M2 is that floating point numbers are represented in higher precision. However, the small dynamic range of FP8 E4M3 (±240.0 can be represented) typically necessitates the use of a higher-precision (typically FP32) scaling factor alongside each quantized tensor. For now, only per-tensor (scalar) scaling factors are supported. Development is ongoing to support scaling factors of a finer granularity (e.g. per-channel). These scaling factors can be specified by passing an optional quantization param JSON to the LLM engine at load time. If this JSON is not specified, scaling factors default to 1.0. These scaling factors are typically obtained when running an unquantized model through a quantizer tool (e.g. AMD quantizer or NVIDIA AMMO). To install AMMO (AlgorithMic Model Optimization): ```console pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com nvidia-ammo ``` Studies have shown that FP8 E4M3 quantization typically only minimally degrades inference accuracy. The most recent silicon offerings e.g. AMD MI300, NVIDIA Hopper or later support native hardware conversion to and from fp32, fp16, bf16, etc. Thus, LLM inference is greatly accelerated with minimal accuracy loss. Here is an example of how to enable this feature: ```python # two float8_e4m3fn kv cache scaling factor files are provided under tests/fp8_kv, please refer to # https://github.com/vllm-project/vllm/blob/main/examples/other/fp8/README.md to generate kv_cache_scales.json of your own. from vllm import LLM, SamplingParams sampling_params = SamplingParams(temperature=1.3, top_p=0.8) llm = LLM(model="meta-llama/Llama-2-7b-chat-hf", kv_cache_dtype="fp8", quantization_param_path="./tests/fp8_kv/llama2-7b-fp8-kv/kv_cache_scales.json") prompt = "London is the capital of" out = llm.generate(prompt, sampling_params)[0].outputs[0].text print(out) # output w/ scaling factors: England, the United Kingdom, and one of the world's leading financial, # output w/o scaling factors: England, located in the southeastern part of the country. It is known ``` --- ### Source Features Quantization Fp8 E5m2 Kvcache (`docs/source/features/quantization/fp8_e5m2_kvcache.md`) (fp8-kv-cache)= # FP8 E5M2 KV Cache The int8/int4 quantization scheme requires additional scale GPU memory storage, which reduces the expected GPU memory benefits. The FP8 data format retains 2~3 mantissa bits and can convert float/fp16/bfloat16 and fp8 to each other. Here is an example of how to enable this feature: ```python from vllm import LLM, SamplingParams # Sample prompts. prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] # Create a sampling params object. sampling_params = SamplingParams(temperature=0.8, top_p=0.95) # Create an LLM. llm = LLM(model="facebook/opt-125m", kv_cache_dtype="fp8") # Generate texts from the prompts. The output is a list of RequestOutput objects # that contain the prompt, generated text, and other information. outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` --- ### Source Features Quantization Gguf (`docs/source/features/quantization/gguf.md`) (gguf)= # GGUF ```{warning} Please note that GGUF support in vLLM is highly experimental and under-optimized at the moment, it might be incompatible with other features. Currently, you can use GGUF as a way to reduce memory footprint. If you encounter any issues, please report them to the vLLM team. ``` ```{warning} Currently, vllm only supports loading single-file GGUF models. If you have a multi-files GGUF model, you can use [gguf-split](https://github.com/ggerganov/llama.cpp/pull/6135) tool to merge them to a single-file model. ``` To run a GGUF model with vLLM, you can download and use the local GGUF model from [TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF](https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF) with the following command: ```console wget https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf # We recommend using the tokenizer from base model to avoid long-time and buggy tokenizer conversion. vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlama-1.1B-Chat-v1.0 ``` You can also add `--tensor-parallel-size 2` to enable tensor parallelism inference with 2 GPUs: ```console # We recommend using the tokenizer from base model to avoid long-time and buggy tokenizer conversion. vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlama-1.1B-Chat-v1.0 --tensor-parallel-size 2 ``` ```{warning} We recommend using the tokenizer from base model instead of GGUF model. Because the tokenizer conversion from GGUF is time-consuming and unstable, especially for some models with large vocab size. ``` You can also use the GGUF model directly through the LLM entrypoint: ```python from vllm import LLM, SamplingParams # In this script, we demonstrate how to pass input to the chat method: conversation = [ { "role": "system", "content": "You are a helpful assistant" }, { "role": "user", "content": "Hello" }, { "role": "assistant", "content": "Hello! How can I assist you today?" }, { "role": "user", "content": "Write an essay about the importance of higher education.", }, ] # Create a sampling params object. sampling_params = SamplingParams(temperature=0.8, top_p=0.95) # Create an LLM. llm = LLM(model="./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf", tokenizer="TinyLlama/TinyLlama-1.1B-Chat-v1.0") # Generate texts from the prompts. The output is a list of RequestOutput objects # that contain the prompt, generated text, and other information. outputs = llm.chat(conversation, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` --- ### Source Features Quantization Int8 (`docs/source/features/quantization/int8.md`) (int8)= # INT8 W8A8 vLLM supports quantizing weights and activations to INT8 for memory savings and inference acceleration. This quantization method is particularly useful for reducing model size while maintaining good performance. Please visit the HF collection of [quantized INT8 checkpoints of popular LLMs ready to use with vLLM](https://huggingface.co/collections/neuralmagic/int8-llms-for-vllm-668ec32c049dca0369816415). ```{note} INT8 computation is supported on NVIDIA GPUs with compute capability > 7.5 (Turing, Ampere, Ada Lovelace, Hopper). ``` ## Prerequisites To use INT8 quantization with vLLM, you'll need to install the [llm-compressor](https://github.com/vllm-project/llm-compressor/) library: ```console pip install llmcompressor ``` ## Quantization Process The quantization process involves four main steps: 1. Loading the model 2. Preparing calibration data 3. Applying quantization 4. Evaluating accuracy in vLLM ### 1. Loading the Model Load your model and tokenizer using the standard `transformers` AutoModel classes: ```python from transformers import AutoTokenizer, AutoModelForCausalLM MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct" model = AutoModelForCausalLM.from_pretrained( MODEL_ID, device_map="auto", torch_dtype="auto", ) tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) ``` ### 2. Preparing Calibration Data When quantizing activations to INT8, you need sample data to estimate the activation scales. It's best to use calibration data that closely matches your deployment data. For a general-purpose instruction-tuned model, you can use a dataset like `ultrachat`: ```python from datasets import load_dataset NUM_CALIBRATION_SAMPLES = 512 MAX_SEQUENCE_LENGTH = 2048 # Load and preprocess the dataset ds = load_dataset("HuggingFaceH4/ultrachat_200k", split="train_sft") ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES)) def preprocess(example): return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)} ds = ds.map(preprocess) def tokenize(sample): return tokenizer(sample["text"], padding=False, max_length=MAX_SEQUENCE_LENGTH, truncation=True, add_special_tokens=False) ds = ds.map(tokenize, remove_columns=ds.column_names) ``` ### 3. Applying Quantization Now, apply the quantization algorithms: ```python from llmcompressor.transformers import oneshot from llmcompressor.modifiers.quantization import GPTQModifier from llmcompressor.modifiers.smoothquant import SmoothQuantModifier # Configure the quantization algorithms recipe = [ SmoothQuantModifier(smoothing_strength=0.8), GPTQModifier(targets="Linear", scheme="W8A8", ignore=["lm_head"]), ] # Apply quantization oneshot( model=model, dataset=ds, recipe=recipe, max_seq_length=MAX_SEQUENCE_LENGTH, num_calibration_samples=NUM_CALIBRATION_SAMPLES, ) # Save the compressed model SAVE_DIR = MODEL_ID.split("/")[1] + "-W8A8-Dynamic-Per-Token" model.save_pretrained(SAVE_DIR, save_compressed=True) tokenizer.save_pretrained(SAVE_DIR) ``` This process creates a W8A8 model with weights and activations quantized to 8-bit integers. ### 4. Evaluating Accuracy After quantization, you can load and run the model in vLLM: ```python from vllm import LLM model = LLM("./Meta-Llama-3-8B-Instruct-W8A8-Dynamic-Per-Token") ``` To evaluate accuracy, you can use `lm_eval`: ```console $ lm_eval --model vllm \ --model_args pretrained="./Meta-Llama-3-8B-Instruct-W8A8-Dynamic-Per-Token",add_bos_token=true \ --tasks gsm8k \ --num_fewshot 5 \ --limit 250 \ --batch_size 'auto' ``` ```{note} Quantized models can be sensitive to the presence of the `bos` token. Make sure to include the `add_bos_token=True` argument when running evaluations. ``` ## Best Practices - Start with 512 samples for calibration data (increase if accuracy drops) - Use a sequence length of 2048 as a starting point - Employ the chat template or instruction template that the model was trained with - If you've fine-tuned a model, consider using a sample of your training data for calibration ## Troubleshooting and Support If you encounter any issues or have feature requests, please open an issue on the `vllm-project/llm-compressor` GitHub repository. --- ### Source Features Quantization Supported Hardware (`docs/source/features/quantization/supported_hardware.md`) (quantization-supported-hardware)= # Supported Hardware The table below shows the compatibility of various quantization implementations with different hardware platforms in vLLM: ```{list-table} :header-rows: 1 :widths: 20 8 8 8 8 8 8 8 8 8 8 * - Implementation - Volta - Turing - Ampere - Ada - Hopper - AMD GPU - Intel GPU - x86 CPU - AWS Inferentia - Google TPU * - AWQ - ✗ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✅︎ - ✅︎ - ✗ - ✗ * - GPTQ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✅︎ - ✅︎ - ✗ - ✗ * - Marlin (GPTQ/AWQ/FP8) - ✗ - ✗ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✗ - ✗ - ✗ - ✗ * - INT8 (W8A8) - ✗ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✗ - ✅︎ - ✗ - ✗ * - FP8 (W8A8) - ✗ - ✗ - ✗ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✗ - ✗ - ✗ * - AQLM - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✗ - ✗ - ✗ - ✗ * - bitsandbytes - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✗ - ✗ - ✗ - ✗ * - DeepSpeedFP - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✗ - ✗ - ✗ - ✗ * - GGUF - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✅︎ - ✗ - ✗ - ✗ - ✗ ``` - Volta refers to SM 7.0, Turing to SM 7.5, Ampere to SM 8.0/8.6, Ada to SM 8.9, and Hopper to SM 9.0. - "✅︎" indicates that the quantization method is supported on the specified hardware. - "✗" indicates that the quantization method is not supported on the specified hardware. ```{note} This compatibility chart is subject to change as vLLM continues to evolve and expand its support for different hardware platforms and quantization methods. For the most up-to-date information on hardware support and quantization methods, please refer to or consult with the vLLM development team. ``` --- ### Source Getting Started Installation Device.Template (`docs/source/getting_started/installation/device.template.md`) # Installation ## Requirements ## Set up using Python ### Pre-built wheels ### Build wheel from source ## Set up using Docker ### Pre-built images ### Build image from source ## Extra information --- ### Source Getting Started Installation Python Env Setup.Inc (`docs/source/getting_started/installation/python_env_setup.inc.md`) You can create a new Python environment using `conda`: ```console # (Recommended) Create a new conda environment. conda create -n myenv python=3.12 -y conda activate myenv ``` ```{note} [PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages. ``` Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command: ```console # (Recommended) Create a new uv environment. Use `--seed` to install `pip` and `setuptools` in the environment. uv venv myenv --python 3.12 --seed source myenv/bin/activate ``` --- ### Source Models Extensions Runai Model Streamer (`docs/source/models/extensions/runai_model_streamer.md`) (runai-model-streamer)= # Loading models with Run:ai Model Streamer Run:ai Model Streamer is a library to read tensors in concurrency, while streaming it to GPU memory. Further reading can be found in [Run:ai Model Streamer Documentation](https://github.com/run-ai/runai-model-streamer/blob/master/docs/README.md). vLLM supports loading weights in Safetensors format using the Run:ai Model Streamer. You first need to install vLLM RunAI optional dependency: ```console pip3 install vllm[runai] ``` To run it as an OpenAI-compatible server, add the `--load-format runai_streamer` flag: ```console vllm serve /home/meta-llama/Llama-3.2-3B-Instruct --load-format runai_streamer ``` To run model from AWS S3 object store run: ```console vllm serve s3://core-llm/Llama-3-8b --load-format runai_streamer ``` To run model from a S3 compatible object store run: ```console RUNAI_STREAMER_S3_USE_VIRTUAL_ADDRESSING=0 AWS_EC2_METADATA_DISABLED=true AWS_ENDPOINT_URL=https://storage.googleapis.com vllm serve s3://core-llm/Llama-3-8b --load-format runai_streamer ``` ## Tunable parameters You can tune parameters using `--model-loader-extra-config`: You can tune `concurrency` that controls the level of concurrency and number of OS threads reading tensors from the file to the CPU buffer. For reading from S3, it will be the number of client instances the host is opening to the S3 server. ```console vllm serve /home/meta-llama/Llama-3.2-3B-Instruct --load-format runai_streamer --model-loader-extra-config '{"concurrency":16}' ``` You can control the size of the CPU Memory buffer to which tensors are read from the file, and limit this size. You can read further about CPU buffer memory limiting [here](https://github.com/run-ai/runai-model-streamer/blob/master/docs/src/env-vars.md#runai_streamer_memory_limit). ```console vllm serve /home/meta-llama/Llama-3.2-3B-Instruct --load-format runai_streamer --model-loader-extra-config '{"memory_limit":5368709120}' ``` ```{note} For further instructions about tunable parameters and additional parameters configurable through environment variables, read the [Environment Variables Documentation](https://github.com/run-ai/runai-model-streamer/blob/master/docs/src/env-vars.md). ``` --- ### Source Models Extensions Tensorizer (`docs/source/models/extensions/tensorizer.md`) (tensorizer)= # Loading models with CoreWeave's Tensorizer vLLM supports loading models with [CoreWeave's Tensorizer](https://docs.coreweave.com/coreweave-machine-learning-and-ai/inference/tensorizer). vLLM model tensors that have been serialized to disk, an HTTP/HTTPS endpoint, or S3 endpoint can be deserialized at runtime extremely quickly directly to the GPU, resulting in significantly shorter Pod startup times and CPU memory usage. Tensor encryption is also supported. For more information on CoreWeave's Tensorizer, please refer to [CoreWeave's Tensorizer documentation](https://github.com/coreweave/tensorizer). For more information on serializing a vLLM model, as well a general usage guide to using Tensorizer with vLLM, see the [vLLM example script](https://docs.vllm.ai/en/stable/getting_started/examples/offline_inference/tensorize_vllm_model.html). ```{note} Note that to use this feature you will need to install `tensorizer` by running `pip install vllm[tensorizer]`. ``` --- ### Source Serving Integrations Langchain (`docs/source/serving/integrations/langchain.md`) (serving-langchain)= # LangChain vLLM is also available via [LangChain](https://github.com/langchain-ai/langchain) . To install LangChain, run ```console pip install langchain langchain_community -q ``` To run inference on a single or multiple GPUs, use `VLLM` class from `langchain`. ```python from langchain_community.llms import VLLM llm = VLLM(model="mosaicml/mpt-7b", trust_remote_code=True, # mandatory for hf models max_new_tokens=128, top_k=10, top_p=0.95, temperature=0.8, # tensor_parallel_size=... # for distributed inference ) print(llm("What is the capital of France ?")) ``` Please refer to this [Tutorial](https://python.langchain.com/docs/integrations/llms/vllm) for more details. --- ### Source Getting Started Installation Ai Accelerator Hpu Gaudi.Inc (`docs/source/getting_started/installation/ai_accelerator/hpu-gaudi.inc.md`) # Installation This tab provides instructions on running vLLM with Intel Gaudi devices. ## Requirements - OS: Ubuntu 22.04 LTS - Python: 3.10 - Intel Gaudi accelerator - Intel Gaudi software version 1.18.0 Please follow the instructions provided in the [Gaudi Installation Guide](https://docs.habana.ai/en/latest/Installation_Guide/index.html) to set up the execution environment. To achieve the best performance, please follow the methods outlined in the [Optimizing Training Platform Guide](https://docs.habana.ai/en/latest/PyTorch/Model_Optimization_PyTorch/Optimization_in_Training_Platform.html). ## Configure a new environment ### Environment verification To verify that the Intel Gaudi software was correctly installed, run: ```console hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core, habanalabs-thunk and habanalabs-container-runtime are installed pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed pip list | grep neural # verify that neural_compressor is installed ``` Refer to [Intel Gaudi Software Stack Verification](https://docs.habana.ai/en/latest/Installation_Guide/SW_Verification.html#platform-upgrade) for more details. ### Run Docker Image It is highly recommended to use the latest Docker image from Intel Gaudi vault. Refer to the [Intel Gaudi documentation](https://docs.habana.ai/en/latest/Installation_Guide/Bare_Metal_Fresh_OS.html#pull-prebuilt-containers) for more details. Use the following commands to run a Docker image: ```console docker pull vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest ``` ## Set up using Python ### Pre-built wheels Currently, there are no pre-built Intel Gaudi wheels. ### Build wheel from source To build and install vLLM from source, run: ```console git clone https://github.com/vllm-project/vllm.git cd vllm python setup.py develop ``` Currently, the latest features and performance optimizations are developed in Gaudi's [vLLM-fork](https://github.com/HabanaAI/vllm-fork) and we periodically upstream them to vLLM main repo. To install latest [HabanaAI/vLLM-fork](https://github.com/HabanaAI/vllm-fork), run the following: ```console git clone https://github.com/HabanaAI/vllm-fork.git cd vllm-fork git checkout habana_main python setup.py develop ``` ## Set up using Docker ### Pre-built images Currently, there are no pre-built Intel Gaudi images. ### Build image from source ```console docker build -f Dockerfile.hpu -t vllm-hpu-env . docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --rm vllm-hpu-env ``` ```{tip} If you're observing the following error: `docker: Error response from daemon: Unknown runtime specified habana.`, please refer to "Install Using Containers" section of [Intel Gaudi Software Stack and Driver Installation](https://docs.habana.ai/en/v1.18.0/Installation_Guide/Bare_Metal_Fresh_OS.html). Make sure you have `habana-container-runtime` package installed and that `habana` container runtime is registered. ``` ## Extra information ## Supported features - [Offline inference](#offline-inference) - Online serving via [OpenAI-Compatible Server](#openai-compatible-server) - HPU autodetection - no need to manually select device within vLLM - Paged KV cache with algorithms enabled for Intel Gaudi accelerators - Custom Intel Gaudi implementations of Paged Attention, KV cache ops, prefill attention, Root Mean Square Layer Normalization, Rotary Positional Encoding - Tensor parallelism support for multi-card inference - Inference with [HPU Graphs](https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html) for accelerating low-batch latency and throughput - Attention with Linear Biases (ALiBi) ## Unsupported features - Beam search - LoRA adapters - Quantization - Prefill chunking (mixed-batch inferencing) ## Supported configurations The following configurations have been validated to be function with Gaudi2 devices. Configurations that are not listed may or may not work. - [meta-llama/Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b) on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B) on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3.1-8B](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B) on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct) on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Llama-2-70b](https://huggingface.co/meta-llama/Llama-2-70b) with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Llama-2-70b-chat-hf](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf) with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B) with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3.1-70B](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B) with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling - [meta-llama/Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-Instruct) with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling ## Performance tuning ### Execution modes Currently in vLLM for HPU we support four execution modes, depending on selected HPU PyTorch Bridge backend (via `PT_HPU_LAZY_MODE` environment variable), and `--enforce-eager` flag. ```{list-table} vLLM execution modes :widths: 25 25 50 :header-rows: 1 * - `PT_HPU_LAZY_MODE` - `enforce_eager` - execution mode * - 0 - 0 - torch.compile * - 0 - 1 - PyTorch eager mode * - 1 - 0 - HPU Graphs * - 1 - 1 - PyTorch lazy mode ``` ```{warning} In 1.18.0, all modes utilizing `PT_HPU_LAZY_MODE=0` are highly experimental and should be only used for validating functional correctness. Their performance will be improved in the next releases. For obtaining the best performance in 1.18.0, please use HPU Graphs, or PyTorch lazy mode. ``` (gaudi-bucketing-mechanism)= ### Bucketing mechanism Intel Gaudi accelerators work best when operating on models with fixed tensor shapes. [Intel Gaudi Graph Compiler](https://docs.habana.ai/en/latest/Gaudi_Overview/Intel_Gaudi_Software_Suite.html#graph-compiler-and-runtime) is responsible for generating optimized binary code that implements the given model topology on Gaudi. In its default configuration, the produced binary code may be heavily dependent on input and output tensor shapes, and can require graph recompilation when encountering differently shaped tensors within the same topology. While the resulting binaries utilize Gaudi efficiently, the compilation itself may introduce a noticeable overhead in end-to-end execution. In a dynamic inference serving scenario, there is a need to minimize the number of graph compilations and reduce the risk of graph compilation occurring during server runtime. Currently it is achieved by "bucketing" model's forward pass across two dimensions - `batch_size` and `sequence_length`. ```{note} Bucketing allows us to reduce the number of required graphs significantly, but it does not handle any graph compilation and device code generation - this is done in warmup and HPUGraph capture phase. ``` Bucketing ranges are determined with 3 parameters - `min`, `step` and `max`. They can be set separately for prompt and decode phase, and for batch size and sequence length dimension. These parameters can be observed in logs during vLLM startup: ```text INFO 08-01 21:37:59 hpu_model_runner.py:493] Prompt bucket config (min, step, max_warmup) bs:[1, 32, 4], seq:[128, 128, 1024] INFO 08-01 21:37:59 hpu_model_runner.py:499] Generated 24 prompt buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024)] INFO 08-01 21:37:59 hpu_model_runner.py:504] Decode bucket config (min, step, max_warmup) bs:[1, 128, 4], seq:[128, 128, 2048] INFO 08-01 21:37:59 hpu_model_runner.py:509] Generated 48 decode buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (1, 1152), (1, 1280), (1, 1408), (1, 1536), (1, 1664), (1, 1792), (1, 1920), (1, 2048), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (2, 1152), (2, 1280), (2, 1408), (2, 1536), (2, 1664), (2, 1792), (2, 1920), (2, 2048), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024), (4, 1152), (4, 1280), (4, 1408), (4, 1536), (4, 1664), (4, 1792), (4, 1920), (4, 2048)] ``` `min` determines the lowest value of the bucket. `step` determines the interval between buckets, and `max` determines the upper bound of the bucket. Furthermore, interval between `min` and `step` has special handling -- `min` gets multiplied by consecutive powers of two, until `step` gets reached. We call this the ramp-up phase and it is used for handling lower batch sizes with minimum wastage, while allowing larger padding on larger batch sizes. Example (with ramp-up) ```text min = 2, step = 32, max = 64 => ramp_up = (2, 4, 8, 16) => stable = (32, 64) => buckets = ramp_up + stable => (2, 4, 8, 16, 32, 64) ``` Example (without ramp-up) ```text min = 128, step = 128, max = 512 => ramp_up = () => stable = (128, 256, 384, 512) => buckets = ramp_up + stable => (128, 256, 384, 512) ``` In the logged scenario, 24 buckets were generated for prompt (prefill) runs, and 48 buckets for decode runs. Each bucket corresponds to a separate optimized device binary for a given model with specified tensor shapes. Whenever a batch of requests is processed, it is padded across batch and sequence length dimension to the smallest possible bucket. ```{warning} If a request exceeds maximum bucket size in any dimension, it will be processed without padding, and its processing may require a graph compilation, potentially significantly increasing end-to-end latency. The boundaries of the buckets are user-configurable via environment variables, and upper bucket boundaries can be increased to avoid such scenario. ``` As an example, if a request of 3 sequences, with max sequence length of 412 comes in to an idle vLLM server, it will be padded executed as `(4, 512)` prefill bucket, as `batch_size` (number of sequences) will be padded to 4 (closest batch_size dimension higher than 3), and max sequence length will be padded to 512 (closest sequence length dimension higher than 412). After prefill stage, it will be executed as `(4, 512)` decode bucket and will continue as that bucket until either batch dimension changes (due to request being finished) - in which case it will become a `(2, 512)` bucket, or context length increases above 512 tokens, in which case it will become `(4, 640)` bucket. ```{note} Bucketing is transparent to a client -- padding in sequence length dimension is never returned to the client, and padding in batch dimension does not create new requests. ``` ### Warmup Warmup is an optional, but highly recommended step occurring before vLLM server starts listening. It executes a forward pass for each bucket with dummy data. The goal is to pre-compile all graphs and not incur any graph compilation overheads within bucket boundaries during server runtime. Each warmup step is logged during vLLM startup: ```text INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][1/24] batch_size:4 seq_len:1024 free_mem:79.16 GiB INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][2/24] batch_size:4 seq_len:896 free_mem:55.43 GiB INFO 08-01 22:26:48 hpu_model_runner.py:1066] [Warmup][Prompt][3/24] batch_size:4 seq_len:768 free_mem:55.43 GiB ... INFO 08-01 22:26:59 hpu_model_runner.py:1066] [Warmup][Prompt][24/24] batch_size:1 seq_len:128 free_mem:55.43 GiB INFO 08-01 22:27:00 hpu_model_runner.py:1066] [Warmup][Decode][1/48] batch_size:4 seq_len:2048 free_mem:55.43 GiB INFO 08-01 22:27:00 hpu_model_runner.py:1066] [Warmup][Decode][2/48] batch_size:4 seq_len:1920 free_mem:55.43 GiB INFO 08-01 22:27:01 hpu_model_runner.py:1066] [Warmup][Decode][3/48] batch_size:4 seq_len:1792 free_mem:55.43 GiB ... INFO 08-01 22:27:16 hpu_model_runner.py:1066] [Warmup][Decode][47/48] batch_size:2 seq_len:128 free_mem:55.43 GiB INFO 08-01 22:27:16 hpu_model_runner.py:1066] [Warmup][Decode][48/48] batch_size:1 seq_len:128 free_mem:55.43 GiB ``` This example uses the same buckets as in the [Bucketing Mechanism](#gaudi-bucketing-mechanism) section. Each output line corresponds to execution of a single bucket. When bucket is executed for the first time, its graph is compiled and can be reused later on, skipping further graph compilations. ```{tip} Compiling all the buckets might take some time and can be turned off with `VLLM_SKIP_WARMUP=true` environment variable. Keep in mind that if you do that, you may face graph compilations once executing a given bucket for the first time. It is fine to disable warmup for development, but it's highly recommended to enable it in deployment. ``` ### HPU Graph capture [HPU Graphs](https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html) are currently the most performant execution method of vLLM on Intel Gaudi. When HPU Graphs are enabled, execution graphs will be traced (recorded) ahead of time (after performing warmup), to be later replayed during inference, significantly reducing host overheads. Recording can take large amounts of memory, which needs to be taken into account when allocating KV cache. Enabling HPU Graphs will impact the number of available KV cache blocks, but vLLM provides user-configurable variables to control memory management. When HPU Graphs are being used, they share the common memory pool ("usable memory") as KV cache, determined by `gpu_memory_utilization` flag (`0.9` by default). Before KV cache gets allocated, model weights are loaded onto the device, and a forward pass of the model is executed on dummy data, to estimate memory usage. Only after that, `gpu_memory_utilization` flag is utilized - at its default value, will mark 90% of free device memory at that point as usable. Next, KV cache gets allocated, model is warmed up, and HPU Graphs are captured. Environment variable `VLLM_GRAPH_RESERVED_MEM` defines the ratio of memory reserved for HPU Graphs capture. With its default value (`VLLM_GRAPH_RESERVED_MEM=0.1`), 10% of usable memory will be reserved for graph capture (later referred to as "usable graph memory"), and the remaining 90% will be utilized for KV cache. Environment variable `VLLM_GRAPH_PROMPT_RATIO` determines the ratio of usable graph memory reserved for prefill and decode graphs. By default (`VLLM_GRAPH_PROMPT_RATIO=0.3`), both stages have equal memory constraints. Lower value corresponds to less usable graph memory reserved for prefill stage, e.g. `VLLM_GRAPH_PROMPT_RATIO=0.2` will reserve 20% of usable graph memory for prefill graphs, and 80% of usable graph memory for decode graphs. ```{note} `gpu_memory_utilization` does not correspond to the absolute memory usage across HPU. It specifies the memory margin after loading the model and performing a profile run. If device has 100 GiB of total memory, and 50 GiB of free memory after loading model weights and executing profiling run, `gpu_memory_utilization` at its default value will mark 90% of 50 GiB as usable, leaving 5 GiB of margin, regardless of total device memory. ``` User can also configure the strategy for capturing HPU Graphs for prompt and decode stages separately. Strategy affects the order of capturing graphs. There are two strategies implemented: \- `max_bs` - graph capture queue will sorted in descending order by their batch sizes. Buckets with equal batch sizes are sorted by sequence length in ascending order (e.g. `(64, 128)`, `(64, 256)`, `(32, 128)`, `(32, 256)`, `(1, 128)`, `(1,256)`), default strategy for decode \- `min_tokens` - graph capture queue will be sorted in ascending order by the number of tokens each graph processes (`batch_size*sequence_length`), default strategy for prompt When there's large amount of requests pending, vLLM scheduler will attempt to fill the maximum batch size for decode as soon as possible. When a request is finished, decode batch size decreases. When that happens, vLLM will attempt to schedule a prefill iteration for requests in the waiting queue, to fill the decode batch size to its previous state. This means that in a full load scenario, decode batch size is often at its maximum, which makes large batch size HPU Graphs crucial to capture, as reflected by `max_bs` strategy. On the other hand, prefills will be executed most frequently with very low batch sizes (1-4), which is reflected in `min_tokens` strategy. ```{note} `VLLM_GRAPH_PROMPT_RATIO` does not set a hard limit on memory taken by graphs for each stage (prefill and decode). vLLM will first attempt to use up entirety of usable prefill graph memory (usable graph memory * `VLLM_GRAPH_PROMPT_RATIO`) for capturing prefill HPU Graphs, next it will attempt do the same for decode graphs and usable decode graph memory pool. If one stage is fully captured, and there is unused memory left within usable graph memory pool, vLLM will attempt further graph capture for the other stage, until no more HPU Graphs can be captured without exceeding reserved memory pool. The behavior on that mechanism can be observed in the example below. ``` Each described step is logged by vLLM server, as follows (negative values correspond to memory being released): ``` /* Detailed source-code truncated for AI context efficiency. */ ``` ### Recommended vLLM Parameters - We recommend running inference on Gaudi 2 with `block_size` of 128 for BF16 data type. Using default values (16, 32) might lead to sub-optimal performance due to Matrix Multiplication Engine under-utilization (see [Gaudi Architecture](https://docs.habana.ai/en/latest/Gaudi_Overview/Gaudi_Architecture.html)). - For max throughput on Llama 7B, we recommend running with batch size of 128 or 256 and max context length of 2048 with HPU Graphs enabled. If you encounter out-of-memory issues, see troubleshooting section. ### Environment variables **Diagnostic and profiling knobs:** - `VLLM_PROFILER_ENABLED`: if `true`, high level profiler will be enabled. Resulting JSON traces can be viewed in [perfetto.habana.ai](https://perfetto.habana.ai/#!/viewer). Disabled by default. - `VLLM_HPU_LOG_STEP_GRAPH_COMPILATION`: if `true`, will log graph compilations per each vLLM engine step, only when there was any - highly recommended to use alongside `PT_HPU_METRICS_GC_DETAILS=1`. Disabled by default. - `VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL`: if `true`, will log graph compilations per each vLLM engine step, always, even if there were none. Disabled by default. - `VLLM_HPU_LOG_STEP_CPU_FALLBACKS`: if `true`, will log cpu fallbacks per each vLLM engine step, only when there was any. Disabled by default. - `VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL`: if `true`, will log cpu fallbacks per each vLLM engine step, always, even if there were none. Disabled by default. **Performance tuning knobs:** - `VLLM_SKIP_WARMUP`: if `true`, warmup will be skipped, `false` by default - `VLLM_GRAPH_RESERVED_MEM`: percentage of memory dedicated for HPUGraph capture, `0.1` by default - `VLLM_GRAPH_PROMPT_RATIO`: percentage of reserved graph memory dedicated for prompt graphs, `0.3` by default - `VLLM_GRAPH_PROMPT_STRATEGY`: strategy determining order of prompt graph capture, `min_tokens` or `max_bs`, `min_tokens` by default - `VLLM_GRAPH_DECODE_STRATEGY`: strategy determining order of decode graph capture, `min_tokens` or `max_bs`, `max_bs` by default - `VLLM_{phase}_{dim}_BUCKET_{param}` - collection of 12 environment variables configuring ranges of bucketing mechanism - `{phase}` is either `PROMPT` or `DECODE` - `{dim}` is either `BS`, `SEQ` or `BLOCK` - `{param}` is either `MIN`, `STEP` or `MAX` - Default values: - Prompt: - batch size min (`VLLM_PROMPT_BS_BUCKET_MIN`): `1` - batch size step (`VLLM_PROMPT_BS_BUCKET_STEP`): `min(max_num_seqs, 32)` - batch size max (`VLLM_PROMPT_BS_BUCKET_MAX`): `min(max_num_seqs, 64)` - sequence length min (`VLLM_PROMPT_SEQ_BUCKET_MIN`): `block_size` - sequence length step (`VLLM_PROMPT_SEQ_BUCKET_STEP`): `block_size` - sequence length max (`VLLM_PROMPT_SEQ_BUCKET_MAX`): `max_model_len` - Decode: - batch size min (`VLLM_DECODE_BS_BUCKET_MIN`): `1` - batch size step (`VLLM_DECODE_BS_BUCKET_STEP`): `min(max_num_seqs, 32)` - batch size max (`VLLM_DECODE_BS_BUCKET_MAX`): `max_num_seqs` - sequence length min (`VLLM_DECODE_BLOCK_BUCKET_MIN`): `block_size` - sequence length step (`VLLM_DECODE_BLOCK_BUCKET_STEP`): `block_size` - sequence length max (`VLLM_DECODE_BLOCK_BUCKET_MAX`): `max(128, (max_num_seqs*max_model_len)/block_size)` Additionally, there are HPU PyTorch Bridge environment variables impacting vLLM execution: - `PT_HPU_LAZY_MODE`: if `0`, PyTorch Eager backend for Gaudi will be used, if `1` PyTorch Lazy backend for Gaudi will be used, `1` is default - `PT_HPU_ENABLE_LAZY_COLLECTIVES`: required to be `true` for tensor parallel inference with HPU Graphs ## Troubleshooting: tweaking HPU graphs If you experience device out-of-memory issues or want to attempt inference at higher batch sizes, try tweaking HPU Graphs by following the below: - Tweak `gpu_memory_utilization` knob. It will decrease the allocation of KV cache, leaving some headroom for capturing graphs with larger batch size. By default `gpu_memory_utilization` is set to 0.9. It attempts to allocate ~90% of HBM left for KV cache after short profiling run. Note that decreasing reduces the number of KV cache blocks you have available, and therefore reduces the effective maximum number of tokens you can handle at a given time. - If this method is not efficient, you can disable `HPUGraph` completely. With HPU Graphs disabled, you are trading latency and throughput at lower batches for potentially higher throughput on higher batches. You can do that by adding `--enforce-eager` flag to server (for online serving), or by passing `enforce_eager=True` argument to LLM constructor (for offline inference). --- ### Source Getting Started Installation Ai Accelerator Neuron.Inc (`docs/source/getting_started/installation/ai_accelerator/neuron.inc.md`) # Installation vLLM 0.3.3 onwards supports model inferencing and serving on AWS Trainium/Inferentia with Neuron SDK with continuous batching. Paged Attention and Chunked Prefill are currently in development and will be available soon. Data types currently supported in Neuron SDK are FP16 and BF16. ## Requirements - OS: Linux - Python: 3.9 -- 3.11 - Accelerator: NeuronCore_v2 (in trn1/inf2 instances) - Pytorch 2.0.1/2.1.1 - AWS Neuron SDK 2.16/2.17 (Verified on python 3.8) ## Configure a new environment ### Launch Trn1/Inf2 instances Here are the steps to launch trn1/inf2 instances, in order to install [PyTorch Neuron ("torch-neuronx") Setup on Ubuntu 22.04 LTS](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/general/setup/neuron-setup/pytorch/neuronx/ubuntu/torch-neuronx-ubuntu22.html). - Please follow the instructions at [launch an Amazon EC2 Instance](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EC2_GetStarted.html#ec2-launch-instance) to launch an instance. When choosing the instance type at the EC2 console, please make sure to select the correct instance type. - To get more information about instances sizes and pricing see: [Trn1 web page](https://aws.amazon.com/ec2/instance-types/trn1/), [Inf2 web page](https://aws.amazon.com/ec2/instance-types/inf2/) - Select Ubuntu Server 22.04 TLS AMI - When launching a Trn1/Inf2, please adjust your primary EBS volume size to a minimum of 512GB. - After launching the instance, follow the instructions in [Connect to your instance](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/AccessingInstancesLinux.html) to connect to the instance ### Install drivers and tools The installation of drivers and tools wouldn't be necessary, if [Deep Learning AMI Neuron](https://docs.aws.amazon.com/dlami/latest/devguide/appendix-ami-release-notes.html) is installed. In case the drivers and tools are not installed on the operating system, follow the steps below: ```console # Configure Linux for Neuron repository updates . /etc/os-release sudo tee /etc/apt/sources.list.d/neuron.list > /dev/null <= 0.5.3`. You may see an error `cannot import name 'default_dump_dir...`. To work around this, run a `pip install --upgrade triton==3.0.0` after installing the vLLM wheel. ``` Following instructions are applicable to Neuron SDK 2.16 and beyond. #### Install transformers-neuronx and its dependencies [transformers-neuronx](https://github.com/aws-neuron/transformers-neuronx) will be the backend to support inference on trn1/inf2 instances. Follow the steps below to install transformer-neuronx package and its dependencies. ```console # Install Python venv sudo apt-get install -y python3.10-venv g++ # Create Python venv python3.10 -m venv aws_neuron_venv_pytorch # Activate Python venv source aws_neuron_venv_pytorch/bin/activate # Install Jupyter notebook kernel pip install ipykernel python3.10 -m ipykernel install --user --name aws_neuron_venv_pytorch --display-name "Python (torch-neuronx)" pip install jupyter notebook pip install environment_kernels # Set pip repository pointing to the Neuron repository python -m pip config set global.extra-index-url https://pip.repos.neuron.amazonaws.com # Install wget, awscli python -m pip install wget python -m pip install awscli # Update Neuron Compiler and Framework python -m pip install --upgrade neuronx-cc==2.* --pre torch-neuronx==2.1.* torchvision transformers-neuronx ``` #### Install vLLM from source Once neuronx-cc and transformers-neuronx packages are installed, we will be able to install vllm as follows: ```console git clone https://github.com/vllm-project/vllm.git cd vllm pip install -U -r requirements-neuron.txt VLLM_TARGET_DEVICE="neuron" pip install . ``` If neuron packages are detected correctly in the installation process, `vllm-0.3.0+neuron212` will be installed. ## Set up using Docker ### Pre-built images Currently, there are no pre-built Neuron images. ### Build image from source See for instructions on building the Docker image. Make sure to use in place of the default Dockerfile. ## Extra information There is no extra information for this device. --- ### Source Getting Started Installation Ai Accelerator Openvino.Inc (`docs/source/getting_started/installation/ai_accelerator/openvino.inc.md`) # Installation vLLM powered by OpenVINO supports all LLM models from [vLLM supported models list](#supported-models) and can perform optimal model serving on all x86-64 CPUs with, at least, AVX2 support, as well as on both integrated and discrete Intel® GPUs ([the list of supported GPUs](https://docs.openvino.ai/2024/about-openvino/release-notes-openvino/system-requirements.html#gpu)). ## Requirements - OS: Linux - Instruction set architecture (ISA) requirement: at least AVX2. ## Set up using Python ### Pre-built wheels Currently, there are no pre-built OpenVINO wheels. ### Build wheel from source First, install Python. For example, on Ubuntu 22.04, you can run: ```console sudo apt-get update -y sudo apt-get install python3 ``` Second, install prerequisites vLLM OpenVINO backend installation: ```console pip install --upgrade pip pip install -r requirements-build.txt --extra-index-url https://download.pytorch.org/whl/cpu ``` Finally, install vLLM with OpenVINO backend: ```console PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE=openvino python -m pip install -v . ``` :::{tip} To use vLLM OpenVINO backend with a GPU device, ensure your system is properly set up. Follow the instructions provided here: [https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html](https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html). ::: ## Set up using Docker ### Pre-built images Currently, there are no pre-built OpenVINO images. ### Build image from source ```console docker build -f Dockerfile.openvino -t vllm-openvino-env . docker run -it --rm vllm-openvino-env ``` ## Extra information ## Supported features OpenVINO vLLM backend supports the following advanced vLLM features: - Prefix caching (`--enable-prefix-caching`) - Chunked prefill (`--enable-chunked-prefill`) ## Performance tips ### vLLM OpenVINO backend environment variables - `VLLM_OPENVINO_DEVICE` to specify which device utilize for the inference. If there are multiple GPUs in the system, additional indexes can be used to choose the proper one (e.g, `VLLM_OPENVINO_DEVICE=GPU.1`). If the value is not specified, CPU device is used by default. - `VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON` to enable U8 weights compression during model loading stage. By default, compression is turned off. You can also export model with different compression techniques using `optimum-cli` and pass exported folder as `` ### CPU performance tips CPU uses the following environment variables to control behavior: - `VLLM_OPENVINO_KVCACHE_SPACE` to specify the KV Cache size (e.g, `VLLM_OPENVINO_KVCACHE_SPACE=40` means 40 GB space for KV cache), larger setting will allow vLLM running more requests in parallel. This parameter should be set based on the hardware configuration and memory management pattern of users. - `VLLM_OPENVINO_CPU_KV_CACHE_PRECISION=u8` to control KV cache precision. By default, FP16 / BF16 is used depending on platform. To enable better TPOT / TTFT latency, you can use vLLM's chunked prefill feature (`--enable-chunked-prefill`). Based on the experiments, the recommended batch size is `256` (`--max-num-batched-tokens`) OpenVINO best known configuration for CPU is: ```console $ VLLM_OPENVINO_KVCACHE_SPACE=100 VLLM_OPENVINO_CPU_KV_CACHE_PRECISION=u8 VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON \ python3 vllm/benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-chat-hf --dataset vllm/benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json --enable-chunked-prefill --max-num-batched-tokens 256 ``` ### GPU performance tips GPU device implements the logic for automatic detection of available GPU memory and, by default, tries to reserve as much memory as possible for the KV cache (taking into account `gpu_memory_utilization` option). However, this behavior can be overridden by explicitly specifying the desired amount of memory for the KV cache using `VLLM_OPENVINO_KVCACHE_SPACE` environment variable (e.g, `VLLM_OPENVINO_KVCACHE_SPACE=8` means 8 GB space for KV cache). Currently, the best performance using GPU can be achieved with the default vLLM execution parameters for models with quantized weights (8 and 4-bit integer data types are supported) and `preemption-mode=swap`. OpenVINO best known configuration for GPU is: ```console $ VLLM_OPENVINO_DEVICE=GPU VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON \ python3 vllm/benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-chat-hf --dataset vllm/benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json ``` ## Limitations - LoRA serving is not supported. - Only LLM models are currently supported. LLaVa and encoder-decoder models are not currently enabled in vLLM OpenVINO integration. - Tensor and pipeline parallelism are not currently enabled in vLLM integration. --- ### Source Getting Started Installation Ai Accelerator Tpu.Inc (`docs/source/getting_started/installation/ai_accelerator/tpu.inc.md`) # Installation Tensor Processing Units (TPUs) are Google's custom-developed application-specific integrated circuits (ASICs) used to accelerate machine learning workloads. TPUs are available in different versions each with different hardware specifications. For more information about TPUs, see [TPU System Architecture](https://cloud.google.com/tpu/docs/system-architecture-tpu-vm). For more information on the TPU versions supported with vLLM, see: - [TPU v6e](https://cloud.google.com/tpu/docs/v6e) - [TPU v5e](https://cloud.google.com/tpu/docs/v5e) - [TPU v5p](https://cloud.google.com/tpu/docs/v5p) - [TPU v4](https://cloud.google.com/tpu/docs/v4) These TPU versions allow you to configure the physical arrangements of the TPU chips. This can improve throughput and networking performance. For more information see: - [TPU v6e topologies](https://cloud.google.com/tpu/docs/v6e#configurations) - [TPU v5e topologies](https://cloud.google.com/tpu/docs/v5e#tpu-v5e-config) - [TPU v5p topologies](https://cloud.google.com/tpu/docs/v5p#tpu-v5p-config) - [TPU v4 topologies](https://cloud.google.com/tpu/docs/v4#tpu-v4-config) In order for you to use Cloud TPUs you need to have TPU quota granted to your Google Cloud Platform project. TPU quotas specify how many TPUs you can use in a GPC project and are specified in terms of TPU version, the number of TPU you want to use, and quota type. For more information, see [TPU quota](https://cloud.google.com/tpu/docs/quota#tpu_quota). For TPU pricing information, see [Cloud TPU pricing](https://cloud.google.com/tpu/pricing). You may need additional persistent storage for your TPU VMs. For more information, see [Storage options for Cloud TPU data](https://cloud.devsite.corp.google.com/tpu/docs/storage-options). ## Requirements - Google Cloud TPU VM - TPU versions: v6e, v5e, v5p, v4 - Python: 3.10 or newer ### Provision Cloud TPUs You can provision Cloud TPUs using the [Cloud TPU API](https://cloud.google.com/tpu/docs/reference/rest) or the [queued resources](https://cloud.google.com/tpu/docs/queued-resources) API. This section shows how to create TPUs using the queued resource API. For more information about using the Cloud TPU API, see [Create a Cloud TPU using the Create Node API](https://cloud.google.com/tpu/docs/managing-tpus-tpu-vm#create-node-api). Queued resources enable you to request Cloud TPU resources in a queued manner. When you request queued resources, the request is added to a queue maintained by the Cloud TPU service. When the requested resource becomes available, it's assigned to your Google Cloud project for your immediate exclusive use. ```{note} In all of the following commands, replace the ALL CAPS parameter names with appropriate values. See the parameter descriptions table for more information. ``` ### Provision Cloud TPUs with GKE For more information about using TPUs with GKE, see: - - - ## Configure a new environment ### Provision a Cloud TPU with the queued resource API Create a TPU v5e with 4 TPU chips: ```console gcloud alpha compute tpus queued-resources create QUEUED_RESOURCE_ID \ --node-id TPU_NAME \ --project PROJECT_ID \ --zone ZONE \ --accelerator-type ACCELERATOR_TYPE \ --runtime-version RUNTIME_VERSION \ --service-account SERVICE_ACCOUNT ``` ```{list-table} Parameter descriptions :header-rows: 1 * - Parameter name - Description * - QUEUED_RESOURCE_ID - The user-assigned ID of the queued resource request. * - TPU_NAME - The user-assigned name of the TPU which is created when the queued resource request is allocated. * - PROJECT_ID - Your Google Cloud project * - ZONE - The GCP zone where you want to create your Cloud TPU. The value you use depends on the version of TPUs you are using. For more information, see `TPU regions and zones `_ * - ACCELERATOR_TYPE - The TPU version you want to use. Specify the TPU version, for example `v5litepod-4` specifies a v5e TPU with 4 cores. For more information, see `TPU versions `_. * - RUNTIME_VERSION - The TPU VM runtime version to use. For more information see `TPU VM images `_. * - SERVICE_ACCOUNT - The email address for your service account. You can find it in the IAM Cloud Console under *Service Accounts*. For example: `tpu-service-account@.iam.gserviceaccount.com` ``` Connect to your TPU using SSH: ```bash gcloud compute tpus tpu-vm ssh TPU_NAME --zone ZONE ``` ## Set up using Python ### Pre-built wheels Currently, there are no pre-built TPU wheels. ### Build wheel from source Install Miniconda: ```bash wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh bash Miniconda3-latest-Linux-x86_64.sh source ~/.bashrc ``` Create and activate a Conda environment for vLLM: ```bash conda create -n vllm python=3.10 -y conda activate vllm ``` Clone the vLLM repository and go to the vLLM directory: ```bash git clone https://github.com/vllm-project/vllm.git && cd vllm ``` Uninstall the existing `torch` and `torch_xla` packages: ```bash pip uninstall torch torch-xla -y ``` Install build dependencies: ```bash pip install -r requirements-tpu.txt sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev ``` Run the setup script: ```bash VLLM_TARGET_DEVICE="tpu" python setup.py develop ``` ## Set up using Docker ### Pre-built images See for instructions on using the official Docker image, making sure to substitute the image name `vllm/vllm-openai` with `vllm/vllm-tpu`. ### Build image from source You can use to build a Docker image with TPU support. ```console docker build -f Dockerfile.tpu -t vllm-tpu . ``` Run the Docker image with the following command: ```console # Make sure to add `--privileged --net host --shm-size=16G`. docker run --privileged --net host --shm-size=16G -it vllm-tpu ``` ```{note} Since TPU relies on XLA which requires static shapes, vLLM bucketizes the possible input shapes and compiles an XLA graph for each shape. The compilation time may take 20~30 minutes in the first run. However, the compilation time reduces to ~5 minutes afterwards because the XLA graphs are cached in the disk (in {code}`VLLM_XLA_CACHE_PATH` or {code}`~/.cache/vllm/xla_cache` by default). ``` ````{tip} If you encounter the following error: ```console from torch._C import * # noqa: F403 ImportError: libopenblas.so.0: cannot open shared object file: No such file or directory ``` Install OpenBLAS with the following command: ```console $ sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev ``` ```` ## Extra information There is no extra information for this device. --- ### Source Getting Started Installation Cpu Apple.Inc (`docs/source/getting_started/installation/cpu/apple.inc.md`) # Installation vLLM has experimental support for macOS with Apple silicon. For now, users shall build from the source vLLM to natively run on macOS. Currently the CPU implementation for macOS supports FP32 and FP16 datatypes. ## Requirements - OS: `macOS Sonoma` or later - SDK: `XCode 15.4` or later with Command Line Tools - Compiler: `Apple Clang >= 15.0.0` ## Set up using Python ### Pre-built wheels ### Build wheel from source After installation of XCode and the Command Line Tools, which include Apple Clang, execute the following commands to build and install vLLM from the source. ```console git clone https://github.com/vllm-project/vllm.git cd vllm pip install -r requirements-cpu.txt pip install -e . ``` ```{note} On macOS the `VLLM_TARGET_DEVICE` is automatically set to `cpu`, which currently is the only supported device. ``` #### Troubleshooting If the build has error like the following snippet where standard C++ headers cannot be found, try to remove and reinstall your [Command Line Tools for Xcode](https://developer.apple.com/download/all/). ```text [...] fatal error: 'map' file not found 1 | #include | ^~~~~ 1 error generated. [2/8] Building CXX object CMakeFiles/_C.dir/csrc/cpu/pos_encoding.cpp.o [...] fatal error: 'cstddef' file not found 10 | #include | ^~~~~~~~~ 1 error generated. ``` ## Set up using Docker ### Pre-built images ### Build image from source ## Extra information --- ### Source Getting Started Installation Cpu Arm.Inc (`docs/source/getting_started/installation/cpu/arm.inc.md`) # Installation vLLM has been adapted to work on ARM64 CPUs with NEON support, leveraging the CPU backend initially developed for the x86 platform. ARM CPU backend currently supports Float32, FP16 and BFloat16 datatypes. ## Requirements - OS: Linux - Compiler: `gcc/g++ >= 12.3.0` (optional, recommended) - Instruction Set Architecture (ISA): NEON support is required ## Set up using Python ### Pre-built wheels ### Build wheel from source :::{include} build.inc.md ::: Testing has been conducted on AWS Graviton3 instances for compatibility. ## Set up using Docker ### Pre-built images ### Build image from source ## Extra information --- ### Source Getting Started Installation Cpu Build.Inc (`docs/source/getting_started/installation/cpu/build.inc.md`) First, install recommended compiler. We recommend to use `gcc/g++ >= 12.3.0` as the default compiler to avoid potential problems. For example, on Ubuntu 22.4, you can run: ```console sudo apt-get update -y sudo apt-get install -y gcc-12 g++-12 libnuma-dev sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12 ``` Second, install Python packages for vLLM CPU backend building: ```console pip install --upgrade pip pip install cmake>=3.26 wheel packaging ninja "setuptools-scm>=8" numpy pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu ``` Finally, build and install vLLM CPU backend: ```console VLLM_TARGET_DEVICE=cpu python setup.py install ``` --- ### Source Getting Started Installation Cpu X86.Inc (`docs/source/getting_started/installation/cpu/x86.inc.md`) # Installation vLLM initially supports basic model inferencing and serving on x86 CPU platform, with data types FP32, FP16 and BF16. ## Requirements - OS: Linux - Compiler: `gcc/g++ >= 12.3.0` (optional, recommended) - Instruction Set Architecture (ISA): AVX512 (optional, recommended) ## Set up using Python ### Pre-built wheels ### Build wheel from source :::{include} build.inc.md ::: ```{note} - AVX512_BF16 is an extension ISA provides native BF16 data type conversion and vector product instructions, will brings some performance improvement compared with pure AVX512. The CPU backend build script will check the host CPU flags to determine whether to enable AVX512_BF16. - If you want to force enable AVX512_BF16 for the cross-compilation, please set environment variable `VLLM_CPU_AVX512BF16=1` before the building. ``` ## Set up using Docker ### Pre-built images ### Build image from source ## Extra information ## Intel Extension for PyTorch - [Intel Extension for PyTorch (IPEX)](https://github.com/intel/intel-extension-for-pytorch) extends PyTorch with up-to-date features optimizations for an extra performance boost on Intel hardware. --- ### Source Getting Started Installation Gpu Cuda.Inc (`docs/source/getting_started/installation/gpu/cuda.inc.md`) # Installation vLLM contains pre-compiled C++ and CUDA (12.1) binaries. ## Requirements - GPU: compute capability 7.0 or higher (e.g., V100, T4, RTX20xx, A100, L4, H100, etc.) ## Set up using Python ### Create a new Python environment ```{note} PyTorch installed via `conda` will statically link `NCCL` library, which can cause issues when vLLM tries to use `NCCL`. See for more details. ``` In order to be performant, vLLM has to compile many cuda kernels. The compilation unfortunately introduces binary incompatibility with other CUDA versions and PyTorch versions, even for the same PyTorch version with different building configurations. Therefore, it is recommended to install vLLM with a **fresh new** environment. If either you have a different CUDA version or you want to use an existing PyTorch installation, you need to build vLLM from source. See [below](#build-from-source) for more details. ### Pre-built wheels You can install vLLM using either `pip` or `uv pip`: ```console # Install vLLM with CUDA 12.1. pip install vllm # If you are using pip. uv pip install vllm # If you are using uv. ``` As of now, vLLM's binaries are compiled with CUDA 12.1 and public PyTorch release versions by default. We also provide vLLM binaries compiled with CUDA 11.8 and public PyTorch release versions: ```console # Install vLLM with CUDA 11.8. export VLLM_VERSION=0.6.1.post1 export PYTHON_VERSION=310 pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu118-cp${PYTHON_VERSION}-cp${PYTHON_VERSION}-manylinux1_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu118 ``` (install-the-latest-code)= #### Install the latest code LLM inference is a fast-evolving field, and the latest code may contain bug fixes, performance improvements, and new features that are not released yet. To allow users to try the latest code without waiting for the next release, vLLM provides wheels for Linux running on a x86 platform with CUDA 12 for every commit since `v0.5.3`. ##### Install the latest code using `pip` ```console pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly ``` `--pre` is required for `pip` to consider pre-released versions. If you want to access the wheels for previous commits (e.g. to bisect the behavior change, performance regression), due to the limitation of `pip`, you have to specify the full URL of the wheel file by embedding the commit hash in the URL: ```console export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch pip install https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl ``` Note that the wheels are built with Python 3.8 ABI (see [PEP 425](https://peps.python.org/pep-0425/) for more details about ABI), so **they are compatible with Python 3.8 and later**. The version string in the wheel file name (`1.0.0.dev`) is just a placeholder to have a unified URL for the wheels, the actual versions of wheels are contained in the wheel metadata (the wheels listed in the extra index url have correct versions). Although we don't support Python 3.8 any more (because PyTorch 2.5 dropped support for Python 3.8), the wheels are still built with Python 3.8 ABI to keep the same wheel name as before. ##### Install the latest code using `uv` Another way to install the latest code is to use `uv`: ```console uv pip install vllm --extra-index-url https://wheels.vllm.ai/nightly ``` If you want to access the wheels for previous commits (e.g. to bisect the behavior change, performance regression), you can specify the commit hash in the URL: ```console export VLLM_COMMIT=72d9c316d3f6ede485146fe5aabd4e61dbc59069 # use full commit hash from the main branch uv pip install vllm --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT} ``` The `uv` approach works for vLLM `v0.6.6` and later and offers an easy-to-remember command. A unique feature of `uv` is that packages in `--extra-index-url` have [higher priority than the default index](https://docs.astral.sh/uv/pip/compatibility/#packages-that-exist-on-multiple-indexes). If the latest public release is `v0.6.6.post1`, `uv`'s behavior allows installing a commit before `v0.6.6.post1` by specifying the `--extra-index-url`. In contrast, `pip` combines packages from `--extra-index-url` and the default index, choosing only the latest version, which makes it difficult to install a development version prior to the released version. ### Build wheel from source #### Set up using Python-only build (without compilation) If you only need to change Python code, you can build and install vLLM without compilation. Using `pip`'s [`--editable` flag](https://pip.pypa.io/en/stable/topics/local-project-installs/#editable-installs), changes you make to the code will be reflected when you run vLLM: ```console git clone https://github.com/vllm-project/vllm.git cd vllm VLLM_USE_PRECOMPILED=1 pip install --editable . ``` This will download the [latest nightly wheel](https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl) and use the compiled libraries from there in the installation. The `VLLM_PRECOMPILED_WHEEL_LOCATION` environment variable can be used instead of `VLLM_USE_PRECOMPILED` to specify a custom path or URL to the wheel file. For example, to use the [0.6.1.post1 PyPi wheel](https://pypi.org/project/vllm/#files): ```console export VLLM_PRECOMPILED_WHEEL_LOCATION=https://files.pythonhosted.org/packages/4a/4c/ee65ba33467a4c0de350ce29fbae39b9d0e7fcd887cc756fa993654d1228/vllm-0.6.3.post1-cp38-abi3-manylinux1_x86_64.whl pip install --editable . ``` You can find more information about vLLM's wheels in . ```{note} There is a possibility that your source code may have a different commit ID compared to the latest vLLM wheel, which could potentially lead to unknown errors. It is recommended to use the same commit ID for the source code as the vLLM wheel you have installed. Please refer to for instructions on how to install a specified wheel. ``` #### Full build (with compilation) If you want to modify C++ or CUDA code, you'll need to build vLLM from source. This can take several minutes: ```console git clone https://github.com/vllm-project/vllm.git cd vllm pip install -e . ``` ```{tip} Building from source requires a lot of compilation. If you are building from source repeatedly, it's more efficient to cache the compilation results. For example, you can install [ccache](https://github.com/ccache/ccache) using `conda install ccache` or `apt install ccache` . As long as `which ccache` command can find the `ccache` binary, it will be used automatically by the build system. After the first build, subsequent builds will be much faster. [sccache](https://github.com/mozilla/sccache) works similarly to `ccache`, but has the capability to utilize caching in remote storage environments. The following environment variables can be set to configure the vLLM `sccache` remote: `SCCACHE_BUCKET=vllm-build-sccache SCCACHE_REGION=us-west-2 SCCACHE_S3_NO_CREDENTIALS=1`. We also recommend setting `SCCACHE_IDLE_TIMEOUT=0`. ``` ##### Use an existing PyTorch installation There are scenarios where the PyTorch dependency cannot be easily installed via pip, e.g.: - Building vLLM with PyTorch nightly or a custom PyTorch build. - Building vLLM with aarch64 and CUDA (GH200), where the PyTorch wheels are not available on PyPI. Currently, only the PyTorch nightly has wheels for aarch64 with CUDA. You can run `pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124` to [install PyTorch nightly](https://pytorch.org/get-started/locally/), and then build vLLM on top of it. To build vLLM using an existing PyTorch installation: ```console git clone https://github.com/vllm-project/vllm.git cd vllm python use_existing_torch.py pip install -r requirements-build.txt pip install -e . --no-build-isolation ``` ##### Use the local cutlass for compilation Currently, before starting the build process, vLLM fetches cutlass code from GitHub. However, there may be scenarios where you want to use a local version of cutlass instead. To achieve this, you can set the environment variable VLLM_CUTLASS_SRC_DIR to point to your local cutlass directory. ```console git clone https://github.com/vllm-project/vllm.git cd vllm VLLM_CUTLASS_SRC_DIR=/path/to/cutlass pip install -e . ``` ##### Troubleshooting To avoid your system being overloaded, you can limit the number of compilation jobs to be run simultaneously, via the environment variable `MAX_JOBS`. For example: ```console export MAX_JOBS=6 pip install -e . ``` This is especially useful when you are building on less powerful machines. For example, when you use WSL it only [assigns 50% of the total memory by default](https://learn.microsoft.com/en-us/windows/wsl/wsl-config#main-wsl-settings), so using `export MAX_JOBS=1` can avoid compiling multiple files simultaneously and running out of memory. A side effect is a much slower build process. Additionally, if you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image. ```console # Use `--ipc=host` to make sure the shared memory is large enough. docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:23.10-py3 ``` If you don't want to use docker, it is recommended to have a full installation of CUDA Toolkit. You can download and install it from [the official website](https://developer.nvidia.com/cuda-toolkit-archive). After installation, set the environment variable `CUDA_HOME` to the installation path of CUDA Toolkit, and make sure that the `nvcc` compiler is in your `PATH`, e.g.: ```console export CUDA_HOME=/usr/local/cuda export PATH="${CUDA_HOME}/bin:$PATH" ``` Here is a sanity check to verify that the CUDA Toolkit is correctly installed: ```console nvcc --version # verify that nvcc is in your PATH ${CUDA_HOME}/bin/nvcc --version # verify that nvcc is in your CUDA_HOME ``` #### Unsupported OS build vLLM can fully run only on Linux but for development purposes, you can still build it on other systems (for example, macOS), allowing for imports and a more convenient development environment. The binaries will not be compiled and won't work on non-Linux systems. Simply disable the `VLLM_TARGET_DEVICE` environment variable before installing: ```console export VLLM_TARGET_DEVICE=empty pip install -e . ``` ## Set up using Docker ### Pre-built images See for instructions on using the official Docker image. Another way to access the latest code is to use the docker images: ```console export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch docker pull public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:${VLLM_COMMIT} ``` These docker images are used for CI and testing only, and they are not intended for production use. They will be expired after several days. The latest code can contain bugs and may not be stable. Please use it with caution. ### Build image from source See for instructions on building the Docker image. ## Supported features See compatibility matrix for feature support information. --- ### Source Getting Started Installation Gpu Rocm.Inc (`docs/source/getting_started/installation/gpu/rocm.inc.md`) # Installation vLLM supports AMD GPUs with ROCm 6.2. ## Requirements - GPU: MI200s (gfx90a), MI300 (gfx942), Radeon RX 7900 series (gfx1100) - ROCm 6.2 ## Set up using Python ### Pre-built wheels Currently, there are no pre-built ROCm wheels. However, the [AMD Infinity hub for vLLM](https://hub.docker.com/r/rocm/vllm/tags) offers a prebuilt, optimized docker image designed for validating inference performance on the AMD Instinct™ MI300X accelerator. ```{tip} Please check [LLM inference performance validation on AMD Instinct MI300X](https://rocm.docs.amd.com/en/latest/how-to/performance-validation/mi300x/vllm-benchmark.html) for instructions on how to use this prebuilt docker image. ``` ### Build wheel from source 0. Install prerequisites (skip if you are already in an environment/docker with the following installed): - [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/index.html) - [PyTorch](https://pytorch.org/) For installing PyTorch, you can start from a fresh docker image, e.g, `rocm/pytorch:rocm6.2_ubuntu20.04_py3.9_pytorch_release_2.3.0`, `rocm/pytorch-nightly`. Alternatively, you can install PyTorch using PyTorch wheels. You can check PyTorch installation guide in PyTorch [Getting Started](https://pytorch.org/get-started/locally/) 1. Install [Triton flash attention for ROCm](https://github.com/ROCm/triton) Install ROCm's Triton flash attention (the default triton-mlir branch) following the instructions from [ROCm/triton](https://github.com/ROCm/triton/blob/triton-mlir/README.md) ```console python3 -m pip install ninja cmake wheel pybind11 pip uninstall -y triton git clone https://github.com/OpenAI/triton.git cd triton git checkout e192dba cd python pip3 install . cd ../.. ``` ```{note} - If you see HTTP issue related to downloading packages during building triton, please try again as the HTTP error is intermittent. ``` 2. Optionally, if you choose to use CK flash attention, you can install [flash attention for ROCm](https://github.com/ROCm/flash-attention/tree/ck_tile) Install ROCm's flash attention (v2.5.9.post1) following the instructions from [ROCm/flash-attention](https://github.com/ROCm/flash-attention/tree/ck_tile#amd-gpurocm-support) Alternatively, wheels intended for vLLM use can be accessed under the releases. For example, for ROCm 6.2, suppose your gfx arch is `gfx90a`. To get your gfx architecture, run `rocminfo |grep gfx`. ```console git clone https://github.com/ROCm/flash-attention.git cd flash-attention git checkout 3cea2fb git submodule update --init GPU_ARCHS="gfx90a" python3 setup.py install cd .. ``` ```{note} - You might need to downgrade the "ninja" version to 1.10 it is not used when compiling flash-attention-2 (e.g. `pip install ninja==1.10.2.4`) ``` 3. Build vLLM. For example, vLLM on ROCM 6.2 can be built with the following steps: ```bash $ pip install --upgrade pip # Install PyTorch $ pip uninstall torch -y $ pip install --no-cache-dir --pre torch --index-url https://download.pytorch.org/whl/rocm6.2 # Build & install AMD SMI $ pip install /opt/rocm/share/amd_smi # Install dependencies $ pip install --upgrade numba scipy huggingface-hub[cli] $ pip install "numpy<2" $ pip install -r requirements-rocm.txt # Build vLLM for MI210/MI250/MI300. $ export PYTORCH_ROCM_ARCH="gfx90a;gfx942" $ python3 setup.py develop ``` This may take 5-10 minutes. Currently, `pip install .` does not work for ROCm installation. ```{tip} - Triton flash attention is used by default. For benchmarking purposes, it is recommended to run a warm up step before collecting perf numbers. - Triton flash attention does not currently support sliding window attention. If using half precision, please use CK flash-attention for sliding window support. - To use CK flash-attention or PyTorch naive attention, please use this flag `export VLLM_USE_TRITON_FLASH_ATTN=0` to turn off triton flash attention. - The ROCm version of PyTorch, ideally, should match the ROCm driver version. ``` ```{tip} - For MI300x (gfx942) users, to achieve optimal performance, please refer to [MI300x tuning guide](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/index.html) for performance optimization and tuning tips on system and workflow level. For vLLM, please refer to [vLLM performance optimization](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/workload.html#vllm-performance-optimization). ``` ## Set up using Docker ### Pre-built images Currently, there are no pre-built ROCm images. ### Build image from source Building the Docker image from source is the recommended way to use vLLM with ROCm. First, build a docker image from and launch a docker container from the image. It is important that the user kicks off the docker build using buildkit. Either the user put DOCKER_BUILDKIT=1 as environment variable when calling docker build command, or the user needs to setup buildkit in the docker daemon configuration /etc/docker/daemon.json as follows and restart the daemon: ```console { "features": { "buildkit": true } } ``` uses ROCm 6.2 by default, but also supports ROCm 5.7, 6.0 and 6.1 in older vLLM branches. It provides flexibility to customize the build of docker image using the following arguments: - `BASE_IMAGE`: specifies the base image used when running `docker build`. The default value `rocm/vllm-dev:base` is an image published and maintained by AMD. It is being built using - `USE_CYTHON`: An option to run cython compilation on a subset of python files upon docker build - `BUILD_RPD`: Include RocmProfileData profiling tool in the image - `ARG_PYTORCH_ROCM_ARCH`: Allows to override the gfx architecture values from the base docker image Their values can be passed in when running `docker build` with `--build-arg` options. To build vllm on ROCm 6.2 for MI200 and MI300 series, you can use the default: ```console DOCKER_BUILDKIT=1 docker build -f Dockerfile.rocm -t vllm-rocm . ``` To build vllm on ROCm 6.2 for Radeon RX7900 series (gfx1100), you should pick the alternative base image: ```console DOCKER_BUILDKIT=1 docker build --build-arg BASE_IMAGE="rocm/vllm-dev:navi_base" -f Dockerfile.rocm -t vllm-rocm . ``` To run the above docker image `vllm-rocm`, use the below command: ```console docker run -it \ --network=host \ --group-add=video \ --ipc=host \ --cap-add=SYS_PTRACE \ --security-opt seccomp=unconfined \ --device /dev/kfd \ --device /dev/dri \ -v :/app/model \ vllm-rocm \ bash ``` Where the `` is the location where the model is stored, for example, the weights for llama2 or llama3 models. ## Supported features See compatibility matrix for feature support information. --- ### Source Getting Started Installation Gpu Xpu.Inc (`docs/source/getting_started/installation/gpu/xpu.inc.md`) # Installation vLLM initially supports basic model inferencing and serving on Intel GPU platform. ## Requirements - Supported Hardware: Intel Data Center GPU, Intel ARC GPU - OneAPI requirements: oneAPI 2024.2 ## Set up using Python ### Pre-built wheels Currently, there are no pre-built XPU wheels. ### Build wheel from source - First, install required driver and intel OneAPI 2024.2 or later. - Second, install Python packages for vLLM XPU backend building: ```console source /opt/intel/oneapi/setvars.sh pip install --upgrade pip pip install -v -r requirements-xpu.txt ``` - Finally, build and install vLLM XPU backend: ```console VLLM_TARGET_DEVICE=xpu python setup.py install ``` ```{note} - FP16 is the default data type in the current XPU backend. The BF16 data type will be supported in the future. ``` ## Set up using Docker ### Pre-built images Currently, there are no pre-built XPU images. ### Build image from source ```console $ docker build -f Dockerfile.xpu -t vllm-xpu-env --shm-size=4g . $ docker run -it \ --rm \ --network=host \ --device /dev/dri \ -v /dev/dri/by-path:/dev/dri/by-path \ vllm-xpu-env ``` ## Supported features XPU platform supports tensor-parallel inference/serving and also supports pipeline parallel as a beta feature for online serving. We requires Ray as the distributed runtime backend. For example, a reference execution likes following: ```console python -m vllm.entrypoints.openai.api_server \ --model=facebook/opt-13b \ --dtype=bfloat16 \ --device=xpu \ --max_model_len=1024 \ --distributed-executor-backend=ray \ --pipeline-parallel-size=2 \ -tp=8 ``` By default, a ray instance will be launched automatically if no existing one is detected in system, with `num-gpus` equals to `parallel_config.world_size`. We recommend properly starting a ray cluster before execution, referring to the helper script. --- --- METRICS --- - Files Extracted: 118 - Estimated Token Budget: ~112939 tokens - Canonical Reference: https://codewiki.google/github.com/OpenBMB/vllm