| 1 | # Benchmark CLI |
| 2 | |
| 3 | This section guides you through running benchmark tests with the extensive datasets supported on vLLM. |
| 4 | |
| 5 | It's a living document, updated as new features and datasets become available. |
| 6 | |
| 7 | !!! tip |
| 8 | The benchmarks described on this page are mainly for evaluating specific vLLM features as well as regression testing. |
| 9 | |
| 10 | For benchmarking production vLLM servers, we recommend [GuideLLM](https://github.com/vllm-project/guidellm), an established performance benchmarking framework with live progress updates and automatic report generation. It is also more flexible than `vllm bench serve` in terms of dataset loading, request formatting, and workload patterns. |
| 11 | |
| 12 | ## Dataset Overview |
| 13 | |
| 14 | <style> |
| 15 | th { |
| 16 | min-width: 0 !important; |
| 17 | } |
| 18 | </style> |
| 19 | |
| 20 | | Dataset | Online | Offline | Data Path | |
| 21 | | ------- | ------ | ------- | --------- | |
| 22 | | ShareGPT | ✅ | ✅ | `wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json` | |
| 23 | | ShareGPT4V (Image) | ✅ | ✅ | `wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json`<br>Note that the images need to be downloaded separately. For example, to download COCO's 2017 Train images:<br>`wget http://images.cocodataset.org/zips/train2017.zip` | |
| 24 | | ShareGPT4Video (Video) | ✅ | ✅ | `git clone https://huggingface.co/datasets/ShareGPT4Video/ShareGPT4Video` | |
| 25 | | BurstGPT | ✅ | ✅ | `wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv` | |
| 26 | | Sonnet (deprecated) | ✅ | ✅ | Local file: `benchmarks/sonnet.txt` | |
| 27 | | Random | ✅ | ✅ | `synthetic` | |
| 28 | | RandomMultiModal (Image/Video) | ✅ | ✅ | `synthetic` | |
| 29 | | RandomForReranking | ✅ | ✅ | `synthetic` | |
| 30 | | Prefix Repetition | ✅ | ✅ | `synthetic` | |
| 31 | | HuggingFace-VisionArena | ✅ | ✅ | `lmarena-ai/VisionArena-Chat` | |
| 32 | | HuggingFace-MMVU | ✅ | ✅ | `yale-nlp/MMVU` | |
| 33 | | HuggingFace-InstructCoder | ✅ | ✅ | `likaixin/InstructCoder` | |
| 34 | | HuggingFace-AIMO | ✅ | ✅ | `AI-MO/aimo-validation-aime`, `AI-MO/NuminaMath-1.5`, `AI-MO/NuminaMath-CoT` | |
| 35 | | HuggingFace-Other | ✅ | ✅ | `lmms-lab/LLaVA-OneVision-Data`, `Aeala/ShareGPT_Vicuna_unfiltered` | |
| 36 | | HuggingFace-MTBench | ✅ | ✅ | `philschmid/mt-bench` | |
| 37 | | HuggingFace-HumanEval | ✅ | ✅ | `openai/openai_humaneval` | |
| 38 | | HuggingFace-GSM8K | ✅ | ✅ | `openai/gsm8k` | |
| 39 | | HuggingFace-Blazedit | ✅ | ✅ | `vdaita/edit_5k_char`, `vdaita/edit_10k_char` | |
| 40 | | HuggingFace-ASR | ✅ | ✅ | `openslr/librispeech_asr`, `facebook/voxpopuli`, `LIUM/tedlium`, `edinburghcstr/ami`, `speechcolab/gigaspeech`, `kensho/spgispeech`, `ArtificialAnalysis/Earnings22-Cleaned-AA`, `D4nt3/esb-datasets-earnings22-validation-tiny-filtered` | |
| 41 | | Spec Bench | ✅ | ✅ | `wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl` | |
| 42 | | SPEED-Bench | ✅ | ✅ | `curl -LsSf https://raw.githubusercontent.com/NVIDIA-NeMo/Skills/refs/heads/main/nemo_skills/dataset/speed-bench/prepare.py \| python3 -` | |
| 43 | | Custom | ✅ | ✅ | Local file: `data.jsonl` | |
| 44 | | Custom Audio | ✅ | ✅ | Local file: `audio_data.jsonl` | |
| 45 | | Custom Image | ✅ | ✅ | Local file: `image_data.jsonl` | |
| 46 | |
| 47 | Legend: |
| 48 | |
| 49 | - ✅ - supported |
| 50 | - 🟡 - Partial support |
| 51 | - 🚧 - to be supported |
| 52 | |
| 53 | !!! note |
| 54 | HuggingFace dataset's `dataset-name` should be set to `hf`. |
| 55 | For local `dataset-path`, please set `hf-name` to its Hugging Face ID like |
| 56 | |
| 57 | ```bash |
| 58 | --dataset-path /datasets/VisionArena-Chat/ --hf-name lmarena-ai/VisionArena-Chat |
| 59 | ``` |
| 60 | |
| 61 | ## Examples |
| 62 | |
| 63 | ### 🚀 Online Benchmark |
| 64 | |
| 65 | <details class="admonition abstract" markdown="1"> |
| 66 | <summary>Show more</summary> |
| 67 | |
| 68 | First start serving your model: |
| 69 | |
| 70 | ```bash |
| 71 | vllm serve NousResearch/Hermes-3-Llama-3.1-8B |
| 72 | ``` |
| 73 | |
| 74 | Then run the benchmarking script: |
| 75 | |
| 76 | ```bash |
| 77 | # download dataset |
| 78 | # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json |
| 79 | vllm bench serve \ |
| 80 | --backend vllm \ |
| 81 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 82 | --endpoint /v1/completions \ |
| 83 | --dataset-name sharegpt \ |
| 84 | --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \ |
| 85 | --num-prompts 10 |
| 86 | ``` |
| 87 | |
| 88 | If successful, you will see the following output: |
| 89 | |
| 90 | ```text |
| 91 | ============ Serving Benchmark Result ============ |
| 92 | Successful requests: 10 |
| 93 | Benchmark duration (s): 5.78 |
| 94 | Total input tokens: 1369 |
| 95 | Total generated tokens: 2212 |
| 96 | Request throughput (req/s): 1.73 |
| 97 | Output token throughput (tok/s): 382.89 |
| 98 | Total token throughput (tok/s): 619.85 |
| 99 | ---------------Time to First Token---------------- |
| 100 | Mean TTFT (ms): 71.54 |
| 101 | Median TTFT (ms): 73.88 |
| 102 | P99 TTFT (ms): 79.49 |
| 103 | -----Time per Output Token (excl. 1st token)------ |
| 104 | Mean TPOT (ms): 7.91 |
| 105 | Median TPOT (ms): 7.96 |
| 106 | P99 TPOT (ms): 8.03 |
| 107 | ---------------Inter-token Latency---------------- |
| 108 | Mean ITL (ms): 7.74 |
| 109 | Median ITL (ms): 7.70 |
| 110 | P99 ITL (ms): 8.39 |
| 111 | ================================================== |
| 112 | ``` |
| 113 | |
| 114 | !!! warning |
| 115 | Repeating `vllm bench serve` against the same server can reuse prompts left |
| 116 | in the prefix cache and inflate throughput. This can affect any reproducible |
| 117 | dataset; the synthetic random dataset is reproducible for a fixed `--seed`, |
| 118 | which defaults to `0`. Prefix cache hits can also come from shared prefixes |
| 119 | within the run, so interpret cache metrics in the context of the workload. |
| 120 | If cache reuse is not intended, vary `--seed`, reset or restart the server, |
| 121 | or use `vllm bench sweep serve`, which resets server caches between runs. |
| 122 | |
| 123 | #### Understanding the Latency Metrics |
| 124 | |
| 125 | `vllm bench serve` measures latency at the benchmark client: |
| 126 | |
| 127 | !!! note |
| 128 | Metric terminology is not standardized across benchmarking tools. When |
| 129 | comparing results, use the measurement points and formulas rather than the |
| 130 | metric names alone. This section explains how we refer to them in vLLM. |
| 131 | |
| 132 | - **Time to first token (TTFT)** is the time from sending a request to receiving |
| 133 | its first streamed output. |
| 134 | - **Inter-token latency (ITL)** records the time between consecutive streamed |
| 135 | outputs. The reported ITL statistics aggregate these individual gaps across |
| 136 | all successful requests. |
| 137 | - **Time per output token (TPOT)** is calculated once per request, excluding the |
| 138 | first token, and then aggregated across requests: |
| 139 | |
| 140 | $$ |
| 141 | \text{TPOT} = |
| 142 | \frac{\text{end-to-end latency} - \text{TTFT}} |
| 143 | {\text{number of output tokens} - 1} |
| 144 | $$ |
| 145 | |
| 146 | With standard decoding, each streamed output usually contains one token, so ITL |
| 147 | and TPOT are typically similar. |
| 148 | |
| 149 | With speculative decoding, one streamed output can contain multiple tokens, |
| 150 | such as several accepted draft tokens within a single engine step. ITL records |
| 151 | only the gaps between streamed outputs; it does not add zero-duration gaps for |
| 152 | tokens in the same output. TPOT instead amortizes the request's decoding time |
| 153 | over every output token. |
| 154 | |
| 155 |  |
| 156 |  |
| 157 | |
| 158 | In this example, the benchmark observes two 40 ms ITL samples. The three tokens |
| 159 | in the second streamed output do not create additional ITL samples, so mean ITL |
| 160 | is 40 ms. TPOT is `(180 ms - 100 ms) / (5 - 1) = 20 ms/token`. |
| 161 | |
| 162 | #### Results Visualization |
| 163 | |
| 164 | The `--plot-timeline` and `--plot-dataset-stats` can be used to generate respectively the requests completion timeline and dataset prompt and output tokens statistics, which can be useful for debugging purpose or for deeper analysis. |
| 165 | |
| 166 | ```bash |
| 167 | vllm bench serve \ |
| 168 | --backend vllm \ |
| 169 | --model meta-llama/Llama-3.1-8B-Instruct \ |
| 170 | --endpoint /v1/completions \ |
| 171 | --dataset-name sharegpt \ |
| 172 | --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \ |
| 173 | --num-prompts 100 \ |
| 174 | --plot-timeline \ |
| 175 | --timeline-itl-thresholds 2,5 \ |
| 176 | --plot-dataset-stats \ |
| 177 | --save-result |
| 178 | ``` |
| 179 | |
| 180 | ##### Interactive Timeline |
| 181 | |
| 182 | The generated timeline is an interactive visualization in the form of an HTML file that can be rendered in most browsers. To customize the ITL color thresholds, one can use `--timeline-itl-thresholds` flag (default: 25ms, 50ms) |
| 183 | |
| 184 | Example output: |
| 185 | |
| 186 | <iframe src="../assets/contributing/vllm_bench_serve_timeline.html" width="100%" height="600" frameborder="0"></iframe> |
| 187 | |
| 188 | ##### Dataset statistics |
| 189 | |
| 190 | The generated figure shows the input prompt and output tokens distribution. |
| 191 | |
| 192 | Example output:  |
| 193 | |
| 194 | #### Custom Dataset |
| 195 | |
| 196 | If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. At inference time, use the option `--dataset-name custom`. Your data needs to be in the `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl |
| 197 | |
| 198 | ```json |
| 199 | {"prompt": "What is the capital of India?"} |
| 200 | {"prompt": "What is the capital of Iran?"} |
| 201 | {"prompt": "What is the capital of China?"} |
| 202 | ``` |
| 203 | |
| 204 | ```bash |
| 205 | # start server |
| 206 | vllm serve meta-llama/Llama-3.1-8B-Instruct |
| 207 | ``` |
| 208 | |
| 209 | ```bash |
| 210 | # run benchmarking script |
| 211 | vllm bench serve --port 9001 --save-result --save-detailed \ |
| 212 | --backend vllm \ |
| 213 | --model meta-llama/Llama-3.1-8B-Instruct \ |
| 214 | --endpoint /v1/completions \ |
| 215 | --dataset-name custom \ |
| 216 | --dataset-path <path-to-your-data-jsonl> \ |
| 217 | --skip-chat-template \ |
| 218 | --num-prompts 80 \ |
| 219 | --max-concurrency 1 \ |
| 220 | --temperature=0.3 \ |
| 221 | --top-p=0.75 \ |
| 222 | --result-dir "./log/" |
| 223 | ``` |
| 224 | |
| 225 | You can skip applying chat template if your data already has it by using `--skip-chat-template`. |
| 226 | |
| 227 | #### Custom Audio Dataset |
| 228 | |
| 229 | If the audio dataset you want to benchmark is not supported yet in vLLM, then you can benchmark on it using `CustomAudioDataset`. At inference time, use the option `--dataset-name custom_audio`. Your data needs to be in the `.jsonl` format and needs to have "prompt" and "audio" fields per entry, e.g., `audio_data.jsonl`: |
| 230 | |
| 231 | ```json |
| 232 | {"prompt": "What does this audio say?", "audio": "/path/to/audio_1.wav"} |
| 233 | {"prompt": "Transcribe the audio.", "audio": "/path/to/audio_2.wav"} |
| 234 | ``` |
| 235 | |
| 236 | - **Supported models:** The `CustomAudioDataset` class supports two types of audio models: ASR models (e.g. Whisper) which do not require a "prompt" field; and multimodal audio-text chat models (e.g. Qwen2-Audio). Since these model types require different arguments at inference, we are giving two examples. |
| 237 | |
| 238 | - **Example 1: Whisper** |
| 239 | |
| 240 | Whisper is a dedicated ASR encoder-decoder model, so it uses `--backend openai-audio` and `--endpoint /v1/audio/transcriptions`. |
| 241 | |
| 242 | ```bash |
| 243 | # start server |
| 244 | vllm serve openai/whisper-tiny |
| 245 | ``` |
| 246 | |
| 247 | ```bash |
| 248 | vllm bench serve \ |
| 249 | --model openai/whisper-tiny \ |
| 250 | --backend openai-audio \ |
| 251 | --endpoint /v1/audio/transcriptions \ |
| 252 | --dataset-name custom_audio \ |
| 253 | --dataset-path audio_data.jsonl \ |
| 254 | --no-oversample \ |
| 255 | --custom-output-len 256 \ |
| 256 | --save-result \ |
| 257 | --save-detailed \ |
| 258 | --result-filename whisper_bench.json |
| 259 | ``` |
| 260 | |
| 261 | - **Example 2: Qwen2-Audio** |
| 262 | |
| 263 | Qwen2-Audio is a multimodal chat model that can do ASR and speech analysis, so it uses `--backend openai-chat`, and `--endpoint /v1/chat/completions`. It also requires `--enable-multimodal-chat` to enable multimodal chat transformation. |
| 264 | |
| 265 | ```bash |
| 266 | vllm bench serve \ |
| 267 | --model Qwen/Qwen2-Audio-7B-Instruct \ |
| 268 | --backend openai-chat \ |
| 269 | --endpoint /v1/chat/completions \ |
| 270 | --dataset-name custom_audio \ |
| 271 | --dataset-path audio_data.jsonl \ |
| 272 | --no-oversample \ |
| 273 | --custom-output-len 256 \ |
| 274 | --enable-multimodal-chat \ |
| 275 | --save-result \ |
| 276 | --save-detailed \ |
| 277 | --result-filename qwen_bench.json |
| 278 | ``` |
| 279 | |
| 280 | #### Custom Image Dataset |
| 281 | |
| 282 | If the image dataset you want to benchmark is not supported yet in vLLM, then you can benchmark on it using `CustomImageDataset`. At inference time, use the option `--dataset-name custom_image`. Your data needs to be in the `.jsonl` format and can use "prompt" and "image_files" fields per entry, e.g., `image_data.jsonl`: |
| 283 | |
| 284 | ```json |
| 285 | {"prompt": "How many animals are present in the given image?", "image_files": ["/path/to/image/folder/horsepony.jpg"]} |
| 286 | {"prompt": "What colour is the bird shown in the image?", "image_files": ["/path/to/image/folder/flycatcher.jpeg"]} |
| 287 | ``` |
| 288 | |
| 289 | Every image listed in "image_files" is added to the request in the listed order after the prompt text. To preserve an interleaved order of text and images, use a "content" field with OpenAI-compatible content parts: |
| 290 | |
| 291 | ```json |
| 292 | {"content": [{"type": "text", "text": "Compare "}, {"type": "image", "image": "/path/to/image/folder/chart_a.png"}, {"type": "text", "text": " with "}, {"type": "image_url", "image_url": {"url": "/path/to/image/folder/chart_b.png"}}]} |
| 293 | ``` |
| 294 | |
| 295 | The "image" shorthand accepts the same values as "image_files". The "image_url" field accepts either an OpenAI-style object with a "url" field or a URL string. |
| 296 | |
| 297 | By default, image references are sent to the serving endpoint as provided, with local image paths converted to `file://` URLs. |
| 298 | |
| 299 | If the benchmark client should load local and HTTP(S) images before sending requests, pass `--custom-ensure-client-side-data` to encode them as base64 data URLs on the client side. |
| 300 | |
| 301 | Existing `data:image/...` URLs are already self-contained and are kept unchanged. |
| 302 | |
| 303 | ```bash |
| 304 | # need a model with vision capability here |
| 305 | vllm serve Qwen/Qwen2-VL-7B-Instruct |
| 306 | ``` |
| 307 | |
| 308 | ```bash |
| 309 | # run benchmarking script |
| 310 | vllm bench serve --save-result --save-detailed \ |
| 311 | --backend openai-chat \ |
| 312 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 313 | --endpoint /v1/chat/completions \ |
| 314 | --dataset-name custom_image \ |
| 315 | --dataset-path <path-to-your-image-data-jsonl> \ |
| 316 | --custom-ensure-client-side-data |
| 317 | ``` |
| 318 | |
| 319 | Note that we need to use the `openai-chat` backend and `/v1/chat/completions` endpoint for multimodal inputs. |
| 320 | |
| 321 | #### VisionArena Benchmark for Vision Language Models |
| 322 | |
| 323 | ```bash |
| 324 | # need a model with vision capability here |
| 325 | vllm serve Qwen/Qwen2-VL-7B-Instruct |
| 326 | ``` |
| 327 | |
| 328 | ```bash |
| 329 | vllm bench serve \ |
| 330 | --backend openai-chat \ |
| 331 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 332 | --endpoint /v1/chat/completions \ |
| 333 | --dataset-name hf \ |
| 334 | --dataset-path lmarena-ai/VisionArena-Chat \ |
| 335 | --hf-split train \ |
| 336 | --num-prompts 1000 |
| 337 | ``` |
| 338 | |
| 339 | #### InstructCoder Benchmark with Speculative Decoding |
| 340 | |
| 341 | ``` bash |
| 342 | vllm serve meta-llama/Meta-Llama-3-8B-Instruct \ |
| 343 | --speculative-config $'{"method": "ngram", |
| 344 | "num_speculative_tokens": 5, "prompt_lookup_max": 5, |
| 345 | "prompt_lookup_min": 2}' |
| 346 | ``` |
| 347 | |
| 348 | ``` bash |
| 349 | vllm bench serve \ |
| 350 | --model meta-llama/Meta-Llama-3-8B-Instruct \ |
| 351 | --dataset-name hf \ |
| 352 | --dataset-path likaixin/InstructCoder \ |
| 353 | --num-prompts 2048 |
| 354 | ``` |
| 355 | |
| 356 | #### Spec Bench Benchmark with Speculative Decoding |
| 357 | |
| 358 | ``` bash |
| 359 | vllm serve meta-llama/Meta-Llama-3-8B-Instruct \ |
| 360 | --speculative-config $'{"method": "ngram", |
| 361 | "num_speculative_tokens": 5, "prompt_lookup_max": 5, |
| 362 | "prompt_lookup_min": 2}' |
| 363 | ``` |
| 364 | |
| 365 | [SpecBench dataset](https://github.com/hemingkx/Spec-Bench) |
| 366 | |
| 367 | Run all categories: |
| 368 | |
| 369 | ``` bash |
| 370 | # Download the dataset using: |
| 371 | # wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl |
| 372 | |
| 373 | vllm bench serve \ |
| 374 | --model meta-llama/Meta-Llama-3-8B-Instruct \ |
| 375 | --dataset-name spec_bench \ |
| 376 | --dataset-path "<YOUR_DOWNLOADED_PATH>/data/spec_bench/question.jsonl" \ |
| 377 | --num-prompts -1 |
| 378 | ``` |
| 379 | |
| 380 | Available categories include `[writing, roleplay, reasoning, math, coding, extraction, stem, humanities, translation, summarization, qa, math_reasoning, rag]`. |
| 381 | |
| 382 | Run only a specific category like "summarization": |
| 383 | |
| 384 | ``` bash |
| 385 | vllm bench serve \ |
| 386 | --model meta-llama/Meta-Llama-3-8B-Instruct \ |
| 387 | --dataset-name spec_bench \ |
| 388 | --dataset-path "<YOUR_DOWNLOADED_PATH>/data/spec_bench/question.jsonl" \ |
| 389 | --num-prompts -1 \ |
| 390 | --spec-bench-category "summarization" |
| 391 | ``` |
| 392 | |
| 393 | #### SPEED-Bench Benchmark with Speculative Decoding |
| 394 | |
| 395 | [SPEED-Bench](https://huggingface.co/datasets/nvidia/SPEED-Bench) is a unified and diverse dataset for speculative decoding, supporting acceptance rate and length measurements using the Qualitative split and throughput measurements using the Throughput splits in 5 configuration of input sequence length (1k, 2k, 8k, 16k, 32k). |
| 396 | |
| 397 | !!! note |
| 398 | This dataset is governed by the [NVIDIA Evaluation Dataset License Agreement](https://huggingface.co/datasets/nvidia/SPEED-Bench/blob/main/License.pdf). For each dataset a user elects to use, the user is responsible for checking if the dataset license is fit for the intended purpose. The `prepare.py` script automatically fetches data from all the source datasets. |
| 399 | |
| 400 | First, download the dataset to a folder, using this one liner: |
| 401 | |
| 402 | ```bash |
| 403 | curl -LsSf https://raw.githubusercontent.com/NVIDIA-NeMo/Skills/refs/heads/main/nemo_skills/dataset/speed-bench/prepare.py | python3 - |
| 404 | ``` |
| 405 | |
| 406 | The command supports also the following arguments: |
| 407 | |
| 408 | - `--config`: download only a subset of the dataset: `qualitative`, `throughput_1k`, `throughput_2k`, `throughput_8k`, `throughput_16k` and `throughput_32k`. By default, it will download all subsets. |
| 409 | - `--output_dir`: download to a specified folder. By default, it will download to the current directory. |
| 410 | |
| 411 | Start a server with speculative decoding: |
| 412 | |
| 413 | ```bash |
| 414 | vllm serve meta-llama/Llama-3.3-70B-Instruct \ |
| 415 | --speculative-config $'{"method": "eagle3", |
| 416 | "num_speculative_tokens": 3, |
| 417 | "model": "nvidia/Llama-3.3-70B-Instruct-Eagle3"}' |
| 418 | ``` |
| 419 | |
| 420 | Run all categories in the Qualitative split: |
| 421 | |
| 422 | ```bash |
| 423 | vllm bench serve \ |
| 424 | --model meta-llama/Llama-3.3-70B-Instruct \ |
| 425 | --dataset-name speed_bench \ |
| 426 | --dataset-path "<YOUR_DOWNLOADED_PATH>/data/speed_bench" \ |
| 427 | --num-prompts -1 |
| 428 | ``` |
| 429 | |
| 430 | Available categories include `[writing, roleplay, reasoning, math, coding, stem, humanities, multilingual, summarization, qa, rag]`. |
| 431 | |
| 432 | Run only a specific category like "multilingual": |
| 433 | |
| 434 | ```bash |
| 435 | vllm bench serve \ |
| 436 | --model meta-llama/Llama-3.3-70B-Instruct \ |
| 437 | --dataset-name speed_bench \ |
| 438 | --dataset-path "<YOUR_DOWNLOADED_PATH>/data/speed_bench" \ |
| 439 | --num-prompts -1 \ |
| 440 | --speed-bench-category "multilingual" |
| 441 | ``` |
| 442 | |
| 443 | Run all categories in the Throughput split (2k ISL): |
| 444 | |
| 445 | ```bash |
| 446 | vllm bench serve \ |
| 447 | --model meta-llama/Llama-3.3-70B-Instruct \ |
| 448 | --dataset-name speed_bench \ |
| 449 | --speed-bench-dataset-subset throughput_2k \ |
| 450 | --dataset-path "<YOUR_DOWNLOADED_PATH>/data/speed_bench/" \ |
| 451 | --num-prompts -1 |
| 452 | ``` |
| 453 | |
| 454 | Available categories include `[high_entropy, mixed, low_entropy]`, where high entropy data contains unstructured data such as creative writing while low entropy data contains more structured data such as coding, more details are in the dataset card. |
| 455 | |
| 456 | #### BFCL (Tool-Calling) Benchmark |
| 457 | |
| 458 | The Berkeley Function Calling Leaderboard (BFCL) dataset measures serving |
| 459 | latency and throughput on realistic tool-calling traffic. Each request |
| 460 | carries a per-sample `tools` schema and chat history, so the server must |
| 461 | expose `/v1/chat/completions` with an auto-tool-choice parser enabled. |
| 462 | The benchmark client always uses the `openai-chat` backend. |
| 463 | |
| 464 | Start a tool-parser-enabled server, then run the bench. For example, with |
| 465 | `gpt-oss-20b`: |
| 466 | |
| 467 | ```bash |
| 468 | # Server |
| 469 | vllm serve openai/gpt-oss-20b \ |
| 470 | --enable-auto-tool-choice \ |
| 471 | --tool-call-parser openai \ |
| 472 | --reasoning-parser openai_gptoss |
| 473 | |
| 474 | # Client |
| 475 | vllm bench serve \ |
| 476 | --backend openai-chat \ |
| 477 | --endpoint /v1/chat/completions \ |
| 478 | --model openai/gpt-oss-20b \ |
| 479 | --dataset-name hf \ |
| 480 | --dataset-path gorilla-llm/Berkeley-Function-Calling-Leaderboard \ |
| 481 | --bfcl-categories simple,live_simple,multiple \ |
| 482 | --num-prompts 200 |
| 483 | ``` |
| 484 | |
| 485 | `--bfcl-categories` is a comma-separated list of BFCL v3 category names |
| 486 | (without the `BFCL_v3_` prefix or `.json` suffix). Defaults to |
| 487 | `simple,live_simple,multiple`. Other supported non-multi-turn categories |
| 488 | include `parallel`, `live_parallel`, `parallel_multiple`, |
| 489 | `live_parallel_multiple`, `irrelevance`, `live_irrelevance`, |
| 490 | `live_relevance`, `java`, `javascript`, and `rest`. Multi-turn categories |
| 491 | are not yet supported. |
| 492 | |
| 493 | The dataset class normalizes BFCL's loose schema dialect (`dict` → |
| 494 | `object`, `float` → `number`, `tuple` → `array`, `any` → `string`) so |
| 495 | modern grammar backends accept the translated tool definitions. |
| 496 | |
| 497 | #### Other HuggingFaceDataset Examples |
| 498 | |
| 499 | ```bash |
| 500 | vllm serve Qwen/Qwen2-VL-7B-Instruct |
| 501 | ``` |
| 502 | |
| 503 | `lmms-lab/LLaVA-OneVision-Data`: |
| 504 | |
| 505 | ```bash |
| 506 | vllm bench serve \ |
| 507 | --backend openai-chat \ |
| 508 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 509 | --endpoint /v1/chat/completions \ |
| 510 | --dataset-name hf \ |
| 511 | --dataset-path lmms-lab/LLaVA-OneVision-Data \ |
| 512 | --hf-split train \ |
| 513 | --hf-subset "chart2text(cauldron)" \ |
| 514 | --num-prompts 10 |
| 515 | ``` |
| 516 | |
| 517 | `Aeala/ShareGPT_Vicuna_unfiltered`: |
| 518 | |
| 519 | ```bash |
| 520 | vllm bench serve \ |
| 521 | --backend openai-chat \ |
| 522 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 523 | --endpoint /v1/chat/completions \ |
| 524 | --dataset-name hf \ |
| 525 | --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ |
| 526 | --hf-split train \ |
| 527 | --num-prompts 10 |
| 528 | ``` |
| 529 | |
| 530 | `AI-MO/aimo-validation-aime`: |
| 531 | |
| 532 | ``` bash |
| 533 | vllm bench serve \ |
| 534 | --model Qwen/QwQ-32B \ |
| 535 | --dataset-name hf \ |
| 536 | --dataset-path AI-MO/aimo-validation-aime \ |
| 537 | --num-prompts 10 \ |
| 538 | --seed 42 |
| 539 | ``` |
| 540 | |
| 541 | `philschmid/mt-bench`: |
| 542 | |
| 543 | ``` bash |
| 544 | vllm bench serve \ |
| 545 | --model Qwen/QwQ-32B \ |
| 546 | --dataset-name hf \ |
| 547 | --dataset-path philschmid/mt-bench \ |
| 548 | --num-prompts 80 |
| 549 | ``` |
| 550 | |
| 551 | `openai/openai_humaneval`: |
| 552 | |
| 553 | ``` bash |
| 554 | vllm bench serve \ |
| 555 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 556 | --dataset-name hf \ |
| 557 | --dataset-path openai/openai_humaneval \ |
| 558 | --num-prompts 80 |
| 559 | ``` |
| 560 | |
| 561 | `openai/gsm8k`: |
| 562 | |
| 563 | ``` bash |
| 564 | vllm bench serve \ |
| 565 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 566 | --dataset-name hf \ |
| 567 | --dataset-path openai/gsm8k \ |
| 568 | --num-prompts 80 |
| 569 | ``` |
| 570 | |
| 571 | `vdaita/edit_5k_char` or `vdaita/edit_10k_char`: |
| 572 | |
| 573 | ``` bash |
| 574 | vllm bench serve \ |
| 575 | --model Qwen/QwQ-32B \ |
| 576 | --dataset-name hf \ |
| 577 | --dataset-path vdaita/edit_5k_char \ |
| 578 | --num-prompts 90 \ |
| 579 | --blazedit-min-distance 0.01 \ |
| 580 | --blazedit-max-distance 0.99 |
| 581 | ``` |
| 582 | |
| 583 | `openslr/librispeech_asr`, `facebook/voxpopuli`, `LIUM/tedlium`, `edinburghcstr/ami`, `speechcolab/gigaspeech`, `kensho/spgispeech`, `ArtificialAnalysis/Earnings22-Cleaned-AA`, `D4nt3/esb-datasets-earnings22-validation-tiny-filtered` |
| 584 | |
| 585 | ```bash |
| 586 | vllm bench serve \ |
| 587 | --model openai/whisper-large-v3-turbo \ |
| 588 | --backend openai-audio \ |
| 589 | --dataset-name hf \ |
| 590 | --dataset-path facebook/voxpopuli --hf-subset en --hf-split test --no-stream --trust-remote-code \ |
| 591 | --num-prompts 99999999 \ |
| 592 | --no-oversample \ |
| 593 | --endpoint /v1/audio/transcriptions \ |
| 594 | --ready-check-timeout-sec 600 \ |
| 595 | --save-result \ |
| 596 | --max-concurrency 512 |
| 597 | ``` |
| 598 | |
| 599 | #### Responses API Benchmark |
| 600 | |
| 601 | The `openai-responses` backend benchmarks vLLM's |
| 602 | [Responses API](../serving/online_serving/openai_compatible_server.md#responses-api) |
| 603 | directly, instead of routing the same workload through `/v1/chat/completions`. |
| 604 | |
| 605 | ```bash |
| 606 | # Server |
| 607 | vllm serve openai/gpt-oss-20b |
| 608 | |
| 609 | # Client |
| 610 | vllm bench serve \ |
| 611 | --backend openai-responses \ |
| 612 | --endpoint /v1/responses \ |
| 613 | --model openai/gpt-oss-20b \ |
| 614 | --dataset-name random \ |
| 615 | --random-input-len 1024 \ |
| 616 | --random-output-len 1024 \ |
| 617 | --num-prompts 200 |
| 618 | ``` |
| 619 | |
| 620 | Reasoning deltas (`response.reasoning_text.delta`) count towards TTFT and ITL, |
| 621 | because the server is already decoding tokens when it emits them. Only output |
| 622 | text (`response.output_text.delta`) is collected as the generated text, which |
| 623 | matches how the `openai-chat` backend treats `DeltaMessage.reasoning`. End-to-end |
| 624 | latency stops at the last token event, so `latency - ttft` equals `sum(itl)` and |
| 625 | TPOT is comparable with the other endpoints. The input and output token counts |
| 626 | come from the usage block on `response.completed`, which is still required: a |
| 627 | stream that ends without a terminal event is reported as a failed request. |
| 628 | |
| 629 | The backend measures one streamed text-generation request per prompt. Built-in |
| 630 | tools, MCP, and multi-turn state via `previous_response_id` are out of scope. |
| 631 | All sampling parameter flags are supported except `--min-p`, which the |
| 632 | Responses API does not accept. |
| 633 | |
| 634 | !!! warning |
| 635 | Do not pass `--extra-body '{"include_reasoning": false}'` when benchmarking |
| 636 | a reasoning model. The server still generates reasoning tokens but emits no |
| 637 | events for them, so the whole reasoning phase is absorbed into TTFT. |
| 638 | |
| 639 | #### Running With Sampling Parameters |
| 640 | |
| 641 | When using OpenAI-compatible backends such as `vllm`, optional sampling |
| 642 | parameters can be specified. Example client command: |
| 643 | |
| 644 | ```bash |
| 645 | vllm bench serve \ |
| 646 | --backend vllm \ |
| 647 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 648 | --endpoint /v1/completions \ |
| 649 | --dataset-name sharegpt \ |
| 650 | --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \ |
| 651 | --top-k 10 \ |
| 652 | --top-p 0.9 \ |
| 653 | --temperature 0.5 \ |
| 654 | --num-prompts 10 |
| 655 | ``` |
| 656 | |
| 657 | #### Running With Ramp-Up Request Rate |
| 658 | |
| 659 | The benchmark tool also supports ramping up the request rate over the |
| 660 | duration of the benchmark run. This can be useful for stress testing the |
| 661 | server or finding the maximum throughput that it can handle, given some latency budget. |
| 662 | |
| 663 | Two ramp-up strategies are supported: |
| 664 | |
| 665 | - `linear`: Increases the request rate linearly from a start value to an end value. |
| 666 | - `exponential`: Increases the request rate exponentially. |
| 667 | |
| 668 | The following arguments can be used to control the ramp-up: |
| 669 | |
| 670 | - `--ramp-up-strategy`: The ramp-up strategy to use (`linear` or `exponential`). |
| 671 | - `--ramp-up-start-rps`: The request rate at the beginning of the benchmark. |
| 672 | - `--ramp-up-end-rps`: The request rate at the end of the benchmark. |
| 673 | |
| 674 | #### Probe Requests |
| 675 | |
| 676 | The benchmark tool also supports sending probe requests alongside the main |
| 677 | workload. This can be useful for measuring how the main workload affects |
| 678 | unrelated traffic sharing the server, e.g. a few requests with large images |
| 679 | stalling a concurrent lightweight request while their multimodal preprocessing |
| 680 | occupies the frontend. |
| 681 | |
| 682 | Setting `--probe-request-rate` to a positive value sends single-token text-only |
| 683 | probe requests at that rate (requests per second) alongside the main workload. |
| 684 | Probes bypass `--max-concurrency` and their latency is reported separately, so |
| 685 | the probe percentiles directly measure the interference that the main workload |
| 686 | inflicts on unrelated requests. |
| 687 | |
| 688 | ```bash |
| 689 | vllm bench serve \ |
| 690 | --model Qwen/Qwen2.5-VL-3B-Instruct \ |
| 691 | --backend openai-chat \ |
| 692 | --endpoint /v1/chat/completions \ |
| 693 | --dataset-name random-mm \ |
| 694 | --random-mm-bucket-config '{(2048, 2048, 1): 1.0}' \ |
| 695 | --request-rate 4 \ |
| 696 | --probe-request-rate 20 |
| 697 | ``` |
| 698 | |
| 699 | #### Load Pattern Configuration |
| 700 | |
| 701 | vLLM's benchmark serving script provides sophisticated load pattern simulation capabilities through three key parameters that control request generation and concurrency behavior: |
| 702 | |
| 703 | ##### Load Pattern Control Parameters |
| 704 | |
| 705 | - `--request-rate`: Controls the target request generation rate (requests per second). Set to `inf` for maximum throughput testing or finite values for controlled load simulation. |
| 706 | - `--burstiness`: Controls traffic variability using a Gamma distribution (range: > 0). Lower values create bursty traffic, higher values create uniform traffic. |
| 707 | - `--max-concurrency`: Limits concurrent outstanding requests. If this argument is not provided, concurrency is unlimited. Set a value to simulate backpressure. When set, include `client_queue_time` in `--percentile-metrics` to report time spent waiting for the benchmark client's concurrency limit. With a finite `--request-rate`, `e2el_including_client_queue` reports schedule-relative end-to-end latency; it is omitted for `--request-rate=inf`, where all requests arrive at benchmark start. |
| 708 | |
| 709 | These parameters work together to create realistic load patterns with carefully chosen defaults. The `--request-rate` parameter defaults to `inf` (infinite), which sends all requests immediately for maximum throughput testing. When set to finite values, it uses either a Poisson process (default `--burstiness=1.0`) or Gamma distribution for realistic request timing. The `--burstiness` parameter only takes effect when `--request-rate` is not infinite - a value of 1.0 creates natural Poisson traffic, while lower values (0.1-0.5) create bursty patterns and higher values (2.0-5.0) create uniform spacing. The `--max-concurrency` parameter defaults to `None` (unlimited) but can be set to simulate real-world constraints where a load balancer or API gateway limits concurrent connections. When combined, these parameters allow you to simulate everything from unrestricted stress testing (`--request-rate=inf`) to production-like scenarios with realistic arrival patterns and resource constraints. |
| 710 | |
| 711 | The `--burstiness` parameter mathematically controls request arrival patterns using a Gamma distribution where: |
| 712 | |
| 713 | - Shape parameter: `burstiness` value |
| 714 | - Coefficient of Variation (CV): $\frac{1}{\sqrt{burstiness}}$ |
| 715 | - Traffic characteristics: |
| 716 | - `burstiness = 0.1`: Highly bursty traffic (CV ≈ 3.16) - stress testing |
| 717 | - `burstiness = 1.0`: Natural Poisson traffic (CV = 1.0) - realistic simulation |
| 718 | - `burstiness = 5.0`: Uniform traffic (CV ≈ 0.45) - controlled load testing |
| 719 | |
| 720 |  |
| 721 | |
| 722 | *Figure: Load pattern examples for each use case. Top row: Request arrival timelines showing cumulative requests over time. Bottom row: Inter-arrival time distributions showing traffic variability patterns. Each column represents a different use case with its specific parameter settings and resulting traffic characteristics.* |
| 723 | |
| 724 | Load Pattern Recommendations by Use Case: |
| 725 | |
| 726 | | Use Case | Burstiness | Request Rate | Max Concurrency | Description | |
| 727 | | --- | --- | --- | --- | --- | |
| 728 | | Maximum Throughput | N/A | Infinite | Limited | **Most common**: Simulates load balancer/gateway limits with unlimited user demand | |
| 729 | | Realistic Testing | 1.0 | Moderate (5-20) | Infinite | Natural Poisson traffic patterns for baseline performance | |
| 730 | | Stress Testing | 0.1-0.5 | High (20-100) | Infinite | Challenging burst patterns to test resilience | |
| 731 | | Latency Profiling | 2.0-5.0 | Low (1-10) | Infinite | Uniform load for consistent timing analysis | |
| 732 | | Capacity Planning | 1.0 | Variable | Limited | Test resource limits with realistic constraints | |
| 733 | | SLA Validation | 1.0 | Target rate | SLA limit | Production-like constraints for compliance testing | |
| 734 | |
| 735 | These load patterns help evaluate different aspects of your vLLM deployment, from basic performance characteristics to resilience under challenging traffic conditions. |
| 736 | |
| 737 | The **Maximum Throughput** pattern (`--request-rate=inf --max-concurrency=<limit>`) is the most commonly used configuration for production benchmarking. This simulates real-world deployment architectures where: |
| 738 | |
| 739 | - Users send requests as fast as they can (infinite rate) |
| 740 | - A load balancer or API gateway controls the maximum concurrent connections |
| 741 | - The system operates at its concurrency limit, revealing true throughput capacity |
| 742 | - `--burstiness` has no effect since request timing is not controlled when rate is infinite |
| 743 | |
| 744 | This pattern helps determine optimal concurrency settings for your production load balancer configuration. |
| 745 | |
| 746 | To effectively configure load patterns, especially for **Capacity Planning** and **SLA Validation** use cases, you need to understand your system's resource limits. During startup, vLLM reports KV cache configuration that directly impacts your load testing parameters: |
| 747 | |
| 748 | ```text |
| 749 | GPU KV cache size: 15,728,640 tokens |
| 750 | Maximum concurrency for 8,192 tokens per request: 1920 |
| 751 | ``` |
| 752 | |
| 753 | Where: |
| 754 | |
| 755 | - GPU KV cache size: Total tokens that can be cached across all concurrent requests |
| 756 | - Maximum concurrency: Theoretical maximum concurrent requests for the given `max_model_len` |
| 757 | - Calculation: `max_concurrency = kv_cache_size / max_model_len` |
| 758 | |
| 759 | Using KV cache metrics for load pattern configuration: |
| 760 | |
| 761 | - For Capacity Planning: Set `--max-concurrency` to 80-90% of the reported maximum to test realistic resource constraints |
| 762 | - For SLA Validation: Use the reported maximum as your SLA limit to ensure compliance testing matches production capacity |
| 763 | - For Realistic Testing: Monitor memory usage when approaching theoretical limits to understand sustainable request rates |
| 764 | - Request rate guidance: Use the KV cache size to estimate sustainable request rates for your specific workload and sequence lengths |
| 765 | |
| 766 | </details> |
| 767 | |
| 768 | ### 📈 Offline Throughput Benchmark |
| 769 | |
| 770 | <details class="admonition abstract" markdown="1"> |
| 771 | <summary>Show more</summary> |
| 772 | |
| 773 | ```bash |
| 774 | vllm bench throughput \ |
| 775 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 776 | --dataset-name sonnet \ |
| 777 | --dataset-path vllm/benchmarks/sonnet.txt \ |
| 778 | --num-prompts 10 |
| 779 | ``` |
| 780 | |
| 781 | If successful, you will see the following output |
| 782 | |
| 783 | ```text |
| 784 | Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s |
| 785 | Total num prompt tokens: 5014 |
| 786 | Total num output tokens: 1500 |
| 787 | ``` |
| 788 | |
| 789 | #### VisionArena Benchmark for Vision Language Models |
| 790 | |
| 791 | ```bash |
| 792 | vllm bench throughput \ |
| 793 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 794 | --backend vllm-chat \ |
| 795 | --dataset-name hf \ |
| 796 | --dataset-path lmarena-ai/VisionArena-Chat \ |
| 797 | --num-prompts 1000 \ |
| 798 | --hf-split train |
| 799 | ``` |
| 800 | |
| 801 | The `num prompt tokens` now includes image token counts |
| 802 | |
| 803 | ```text |
| 804 | Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s |
| 805 | Total num prompt tokens: 14527 |
| 806 | Total num output tokens: 1280 |
| 807 | ``` |
| 808 | |
| 809 | #### InstructCoder Benchmark with Speculative Decoding |
| 810 | |
| 811 | ``` bash |
| 812 | VLLM_WORKER_MULTIPROC_METHOD=spawn \ |
| 813 | vllm bench throughput \ |
| 814 | --dataset-name=hf \ |
| 815 | --dataset-path=likaixin/InstructCoder \ |
| 816 | --model=meta-llama/Meta-Llama-3-8B-Instruct \ |
| 817 | --input-len=1000 \ |
| 818 | --output-len=100 \ |
| 819 | --num-prompts=2048 \ |
| 820 | --async-engine \ |
| 821 | --speculative-config $'{"method": "ngram", |
| 822 | "num_speculative_tokens": 5, "prompt_lookup_max": 5, |
| 823 | "prompt_lookup_min": 2}' |
| 824 | ``` |
| 825 | |
| 826 | ```text |
| 827 | Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s |
| 828 | Total num prompt tokens: 261136 |
| 829 | Total num output tokens: 204800 |
| 830 | ``` |
| 831 | |
| 832 | #### Other HuggingFaceDataset Examples |
| 833 | |
| 834 | `lmms-lab/LLaVA-OneVision-Data`: |
| 835 | |
| 836 | ```bash |
| 837 | vllm bench throughput \ |
| 838 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 839 | --backend vllm-chat \ |
| 840 | --dataset-name hf \ |
| 841 | --dataset-path lmms-lab/LLaVA-OneVision-Data \ |
| 842 | --hf-split train \ |
| 843 | --hf-subset "chart2text(cauldron)" \ |
| 844 | --num-prompts 10 |
| 845 | ``` |
| 846 | |
| 847 | `Aeala/ShareGPT_Vicuna_unfiltered`: |
| 848 | |
| 849 | ```bash |
| 850 | vllm bench throughput \ |
| 851 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 852 | --backend vllm-chat \ |
| 853 | --dataset-name hf \ |
| 854 | --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ |
| 855 | --hf-split train \ |
| 856 | --num-prompts 10 |
| 857 | ``` |
| 858 | |
| 859 | `AI-MO/aimo-validation-aime`: |
| 860 | |
| 861 | ```bash |
| 862 | vllm bench throughput \ |
| 863 | --model Qwen/QwQ-32B \ |
| 864 | --backend vllm \ |
| 865 | --dataset-name hf \ |
| 866 | --dataset-path AI-MO/aimo-validation-aime \ |
| 867 | --hf-split train \ |
| 868 | --num-prompts 10 |
| 869 | ``` |
| 870 | |
| 871 | Benchmark with LoRA adapters: |
| 872 | |
| 873 | ``` bash |
| 874 | # download dataset |
| 875 | # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json |
| 876 | vllm bench throughput \ |
| 877 | --model meta-llama/Llama-2-7b-hf \ |
| 878 | --backend vllm \ |
| 879 | --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \ |
| 880 | --dataset_name sharegpt \ |
| 881 | --num-prompts 10 \ |
| 882 | --max-loras 2 \ |
| 883 | --max-lora-rank 8 \ |
| 884 | --enable-lora \ |
| 885 | --lora-path yard1/llama-2-7b-sql-lora-test |
| 886 | ``` |
| 887 | |
| 888 | #### Synthetic Random Multimodal (random-mm) |
| 889 | |
| 890 | Generate synthetic multimodal inputs for offline throughput testing without external datasets. |
| 891 | Use `--backend vllm-chat` so that image tokens are counted correctly. |
| 892 | |
| 893 | ```bash |
| 894 | vllm bench throughput \ |
| 895 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 896 | --backend vllm-chat \ |
| 897 | --dataset-name random-mm \ |
| 898 | --num-prompts 100 \ |
| 899 | --random-input-len 300 \ |
| 900 | --random-output-len 40 \ |
| 901 | --random-mm-base-items-per-request 2 \ |
| 902 | --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \ |
| 903 | --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}' |
| 904 | ``` |
| 905 | |
| 906 | </details> |
| 907 | |
| 908 | ### 🛠️ Structured Output Benchmark |
| 909 | |
| 910 | <details class="admonition abstract" markdown="1"> |
| 911 | <summary>Show more</summary> |
| 912 | |
| 913 | Benchmark the performance of structured output generation (JSON, grammar, regex). |
| 914 | |
| 915 | #### Server Setup |
| 916 | |
| 917 | ```bash |
| 918 | vllm serve NousResearch/Hermes-3-Llama-3.1-8B |
| 919 | ``` |
| 920 | |
| 921 | #### JSON Schema Benchmark |
| 922 | |
| 923 | ```bash |
| 924 | python3 benchmarks/benchmark_serving_structured_output.py \ |
| 925 | --backend vllm \ |
| 926 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 927 | --dataset json \ |
| 928 | --structured-output-ratio 1.0 \ |
| 929 | --request-rate 10 \ |
| 930 | --num-prompts 1000 |
| 931 | ``` |
| 932 | |
| 933 | #### Grammar-based Generation Benchmark |
| 934 | |
| 935 | ```bash |
| 936 | python3 benchmarks/benchmark_serving_structured_output.py \ |
| 937 | --backend vllm \ |
| 938 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 939 | --dataset grammar \ |
| 940 | --structure-type grammar \ |
| 941 | --request-rate 10 \ |
| 942 | --num-prompts 1000 |
| 943 | ``` |
| 944 | |
| 945 | #### Regex-based Generation Benchmark |
| 946 | |
| 947 | ```bash |
| 948 | python3 benchmarks/benchmark_serving_structured_output.py \ |
| 949 | --backend vllm \ |
| 950 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 951 | --dataset regex \ |
| 952 | --request-rate 10 \ |
| 953 | --num-prompts 1000 |
| 954 | ``` |
| 955 | |
| 956 | #### Choice-based Generation Benchmark |
| 957 | |
| 958 | ```bash |
| 959 | python3 benchmarks/benchmark_serving_structured_output.py \ |
| 960 | --backend vllm \ |
| 961 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 962 | --dataset choice \ |
| 963 | --request-rate 10 \ |
| 964 | --num-prompts 1000 |
| 965 | ``` |
| 966 | |
| 967 | #### XGrammar Benchmark Dataset |
| 968 | |
| 969 | ```bash |
| 970 | python3 benchmarks/benchmark_serving_structured_output.py \ |
| 971 | --backend vllm \ |
| 972 | --model NousResearch/Hermes-3-Llama-3.1-8B \ |
| 973 | --dataset xgrammar_bench \ |
| 974 | --request-rate 10 \ |
| 975 | --num-prompts 1000 |
| 976 | ``` |
| 977 | |
| 978 | </details> |
| 979 | |
| 980 | ### 📚 Long Document QA Benchmark |
| 981 | |
| 982 | <details class="admonition abstract" markdown="1"> |
| 983 | <summary>Show more</summary> |
| 984 | |
| 985 | Benchmark the performance of long document question-answering with prefix caching. |
| 986 | |
| 987 | #### Basic Long Document QA Test |
| 988 | |
| 989 | ```bash |
| 990 | python3 benchmarks/benchmark_long_document_qa_throughput.py \ |
| 991 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 992 | --enable-prefix-caching \ |
| 993 | --num-documents 16 \ |
| 994 | --document-length 2000 \ |
| 995 | --output-len 50 \ |
| 996 | --repeat-count 5 |
| 997 | ``` |
| 998 | |
| 999 | #### Different Repeat Modes |
| 1000 | |
| 1001 | ```bash |
| 1002 | # Random mode (default) - shuffle prompts randomly |
| 1003 | python3 benchmarks/benchmark_long_document_qa_throughput.py \ |
| 1004 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1005 | --enable-prefix-caching \ |
| 1006 | --num-documents 8 \ |
| 1007 | --document-length 3000 \ |
| 1008 | --repeat-count 3 \ |
| 1009 | --repeat-mode random |
| 1010 | |
| 1011 | # Tile mode - repeat entire prompt list in sequence |
| 1012 | python3 benchmarks/benchmark_long_document_qa_throughput.py \ |
| 1013 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1014 | --enable-prefix-caching \ |
| 1015 | --num-documents 8 \ |
| 1016 | --document-length 3000 \ |
| 1017 | --repeat-count 3 \ |
| 1018 | --repeat-mode tile |
| 1019 | |
| 1020 | # Interleave mode - repeat each prompt consecutively |
| 1021 | python3 benchmarks/benchmark_long_document_qa_throughput.py \ |
| 1022 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1023 | --enable-prefix-caching \ |
| 1024 | --num-documents 8 \ |
| 1025 | --document-length 3000 \ |
| 1026 | --repeat-count 3 \ |
| 1027 | --repeat-mode interleave |
| 1028 | ``` |
| 1029 | |
| 1030 | </details> |
| 1031 | |
| 1032 | ### 🗂️ Prefix Caching Benchmark |
| 1033 | |
| 1034 | <details class="admonition abstract" markdown="1"> |
| 1035 | <summary>Show more</summary> |
| 1036 | |
| 1037 | Benchmark the efficiency of automatic prefix caching. |
| 1038 | |
| 1039 | #### Fixed Prompt with Prefix Caching |
| 1040 | |
| 1041 | ```bash |
| 1042 | python3 benchmarks/benchmark_prefix_caching.py \ |
| 1043 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1044 | --enable-prefix-caching \ |
| 1045 | --num-prompts 1 \ |
| 1046 | --repeat-count 100 \ |
| 1047 | --input-length-range 128:256 |
| 1048 | ``` |
| 1049 | |
| 1050 | #### ShareGPT Dataset with Prefix Caching |
| 1051 | |
| 1052 | ```bash |
| 1053 | # download dataset |
| 1054 | # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json |
| 1055 | |
| 1056 | python3 benchmarks/benchmark_prefix_caching.py \ |
| 1057 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1058 | --dataset-path /path/ShareGPT_V3_unfiltered_cleaned_split.json \ |
| 1059 | --enable-prefix-caching \ |
| 1060 | --num-prompts 20 \ |
| 1061 | --repeat-count 5 \ |
| 1062 | --input-length-range 128:256 |
| 1063 | ``` |
| 1064 | |
| 1065 | ##### Prefix Repetition Dataset |
| 1066 | |
| 1067 | ```bash |
| 1068 | vllm bench serve \ |
| 1069 | --backend openai \ |
| 1070 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1071 | --dataset-name prefix_repetition \ |
| 1072 | --num-prompts 100 \ |
| 1073 | --prefix-repetition-prefix-len 512 \ |
| 1074 | --prefix-repetition-suffix-len 128 \ |
| 1075 | --prefix-repetition-num-prefixes 5 \ |
| 1076 | --prefix-repetition-output-len 128 |
| 1077 | ``` |
| 1078 | |
| 1079 | </details> |
| 1080 | |
| 1081 | ### Replay Timed Traces |
| 1082 | |
| 1083 | <details class="admonition abstract" markdown="1"> |
| 1084 | <summary>Show more</summary> |
| 1085 | |
| 1086 | Example of how to run traces which have timing information |
| 1087 | with them. |
| 1088 | |
| 1089 | #### Running MoonshotAI traces |
| 1090 | |
| 1091 | Start the server: |
| 1092 | |
| 1093 | ```bash |
| 1094 | vllm serve Qwen/Qwen3.5-2B \ |
| 1095 | --host 127.0.0.1 --port 8000 |
| 1096 | ``` |
| 1097 | |
| 1098 | Run the benchmark: |
| 1099 | |
| 1100 | ```bash |
| 1101 | # Download an example trace |
| 1102 | # curl -L -o conversation_trace.jsonl \ |
| 1103 | #https://raw.githubusercontent.com/kvcache-ai/Mooncake/main/FAST25-release/traces/conversation_trace.jsonl |
| 1104 | |
| 1105 | vllm bench serve --model Qwen/Qwen3.5-2B \ |
| 1106 | --dataset-name=timed_trace --num-prompts 100 --host 127.0.0.1 \ |
| 1107 | --port 8000 --dataset-path ./conversation_trace.jsonl \ |
| 1108 | --ignore-eos --self-timed --timed-trace-chunk-hash-size 512 \ |
| 1109 | --timed-trace-sec-multiplier 0.001 |
| 1110 | ``` |
| 1111 | |
| 1112 | This will replay the first 100 lines from the trace file `conversation.jsonl`. |
| 1113 | |
| 1114 | </details> |
| 1115 | |
| 1116 | ### 🧪 Hashing Benchmarks |
| 1117 | |
| 1118 | <details class="admonition abstract" markdown="1"> |
| 1119 | <summary>Show more</summary> |
| 1120 | |
| 1121 | Two helper scripts live in `benchmarks/` to compare hashing options used by prefix caching and related utilities. They are standalone (no server required) and help choose a hash algorithm before enabling prefix caching in production. |
| 1122 | |
| 1123 | - `benchmarks/benchmark_hash.py`: Micro-benchmark that measures per-call latency of three implementations on a representative `(bytes, tuple[int])` payload. |
| 1124 | |
| 1125 | ```bash |
| 1126 | python benchmarks/benchmark_hash.py --iterations 20000 --seed 42 |
| 1127 | ``` |
| 1128 | |
| 1129 | - `benchmarks/benchmark_prefix_block_hash.py`: End-to-end block hashing benchmark that runs the full prefix-cache hash pipeline (`hash_block_tokens`) across many fake blocks and reports throughput. |
| 1130 | |
| 1131 | ```bash |
| 1132 | python benchmarks/benchmark_prefix_block_hash.py --num-blocks 20000 --block-size 32 --trials 5 |
| 1133 | ``` |
| 1134 | |
| 1135 | Supported algorithms: `sha256`, `sha256_cbor`, `xxhash`, `xxhash_cbor`. Install optional deps to exercise all variants: |
| 1136 | |
| 1137 | ```bash |
| 1138 | uv pip install xxhash cbor2 |
| 1139 | ``` |
| 1140 | |
| 1141 | If an algorithm’s dependency is missing, the script will skip it and continue. |
| 1142 | |
| 1143 | </details> |
| 1144 | |
| 1145 | ### ⚡ Request Prioritization Benchmark |
| 1146 | |
| 1147 | <details class="admonition abstract" markdown="1"> |
| 1148 | <summary>Show more</summary> |
| 1149 | |
| 1150 | Benchmark the performance of request prioritization in vLLM. |
| 1151 | |
| 1152 | #### Basic Prioritization Test |
| 1153 | |
| 1154 | ```bash |
| 1155 | python3 benchmarks/benchmark_prioritization.py \ |
| 1156 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1157 | --input-len 128 \ |
| 1158 | --output-len 64 \ |
| 1159 | --num-prompts 100 \ |
| 1160 | --scheduling-policy priority |
| 1161 | ``` |
| 1162 | |
| 1163 | #### Multiple Sequences per Prompt |
| 1164 | |
| 1165 | ```bash |
| 1166 | python3 benchmarks/benchmark_prioritization.py \ |
| 1167 | --model meta-llama/Llama-2-7b-chat-hf \ |
| 1168 | --input-len 128 \ |
| 1169 | --output-len 64 \ |
| 1170 | --num-prompts 100 \ |
| 1171 | --scheduling-policy priority \ |
| 1172 | --n 2 |
| 1173 | ``` |
| 1174 | |
| 1175 | </details> |
| 1176 | |
| 1177 | ### 👁️ Multi-Modal Benchmark |
| 1178 | |
| 1179 | <details class="admonition abstract" markdown="1"> |
| 1180 | <summary>Show more</summary> |
| 1181 | |
| 1182 | Benchmark the performance of multi-modal requests in vLLM. |
| 1183 | |
| 1184 | #### Images (ShareGPT4V) |
| 1185 | |
| 1186 | Start vLLM: |
| 1187 | |
| 1188 | ```bash |
| 1189 | vllm serve Qwen/Qwen2.5-VL-7B-Instruct \ |
| 1190 | --dtype bfloat16 \ |
| 1191 | --limit-mm-per-prompt '{"image": 1}' \ |
| 1192 | --allowed-local-media-path /path/to/sharegpt4v/images |
| 1193 | ``` |
| 1194 | |
| 1195 | Send requests with images: |
| 1196 | |
| 1197 | ```bash |
| 1198 | vllm bench serve \ |
| 1199 | --backend openai-chat \ |
| 1200 | --model Qwen/Qwen2.5-VL-7B-Instruct \ |
| 1201 | --dataset-name sharegpt \ |
| 1202 | --dataset-path /path/to/ShareGPT4V/sharegpt4v_instruct_gpt4-vision_cap100k.json \ |
| 1203 | --num-prompts 100 \ |
| 1204 | --save-result \ |
| 1205 | --result-dir ~/vllm_benchmark_results \ |
| 1206 | --save-detailed \ |
| 1207 | --endpoint /v1/chat/completions |
| 1208 | ``` |
| 1209 | |
| 1210 | #### Videos (ShareGPT4Video) |
| 1211 | |
| 1212 | Start vLLM: |
| 1213 | |
| 1214 | ```bash |
| 1215 | vllm serve Qwen/Qwen2.5-VL-7B-Instruct \ |
| 1216 | --dtype bfloat16 \ |
| 1217 | --limit-mm-per-prompt '{"video": 1}' \ |
| 1218 | --allowed-local-media-path /path/to/sharegpt4video/videos |
| 1219 | ``` |
| 1220 | |
| 1221 | Send requests with videos: |
| 1222 | |
| 1223 | ```bash |
| 1224 | vllm bench serve \ |
| 1225 | --backend openai-chat \ |
| 1226 | --model Qwen/Qwen2.5-VL-7B-Instruct \ |
| 1227 | --dataset-name sharegpt \ |
| 1228 | --dataset-path /path/to/ShareGPT4Video/llava_v1_5_mix665k_with_video_chatgpt72k_share4video28k.json \ |
| 1229 | --num-prompts 100 \ |
| 1230 | --save-result \ |
| 1231 | --result-dir ~/vllm_benchmark_results \ |
| 1232 | --save-detailed \ |
| 1233 | --endpoint /v1/chat/completions |
| 1234 | ``` |
| 1235 | |
| 1236 | #### Synthetic Random Images (random-mm) |
| 1237 | |
| 1238 | Generate synthetic image inputs alongside random text prompts to stress-test vision models without external datasets. |
| 1239 | |
| 1240 | Notes: |
| 1241 | |
| 1242 | - For online benchmarks, use `--backend openai-chat` with endpoint `/v1/chat/completions`. |
| 1243 | - For offline benchmarks, use `--backend vllm-chat` (see [Offline Throughput Benchmark](#-offline-throughput-benchmark) for an example). |
| 1244 | |
| 1245 | Start the server (example): |
| 1246 | |
| 1247 | ```bash |
| 1248 | vllm serve Qwen/Qwen2.5-VL-3B-Instruct \ |
| 1249 | --dtype bfloat16 \ |
| 1250 | --max-model-len 16384 \ |
| 1251 | --limit-mm-per-prompt '{"image": 3, "video": 0}' \ |
| 1252 | --mm-processor-kwargs max_pixels=1003520 |
| 1253 | ``` |
| 1254 | |
| 1255 | Benchmark. It is recommended to use the flag `--ignore-eos` to simulate real responses. You can set the size of the output via the arg `random-output-len`. |
| 1256 | |
| 1257 | Ex.1: Fixed number of items and a single image resolution, enforcing generation of approx 40 tokens: |
| 1258 | |
| 1259 | ```bash |
| 1260 | vllm bench serve \ |
| 1261 | --backend openai-chat \ |
| 1262 | --model Qwen/Qwen2.5-VL-3B-Instruct \ |
| 1263 | --endpoint /v1/chat/completions \ |
| 1264 | --dataset-name random-mm \ |
| 1265 | --num-prompts 100 \ |
| 1266 | --max-concurrency 10 \ |
| 1267 | --random-prefix-len 25 \ |
| 1268 | --random-input-len 300 \ |
| 1269 | --random-output-len 40 \ |
| 1270 | --random-range-ratio 0.2 \ |
| 1271 | --random-mm-base-items-per-request 2 \ |
| 1272 | --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \ |
| 1273 | --random-mm-bucket-config '{(224, 224, 1): 1.0}' \ |
| 1274 | --request-rate inf \ |
| 1275 | --ignore-eos \ |
| 1276 | --seed 42 |
| 1277 | ``` |
| 1278 | |
| 1279 | The number of items per request can be controlled by passing multiple image buckets: |
| 1280 | |
| 1281 | ```bash |
| 1282 | --random-mm-base-items-per-request 2 \ |
| 1283 | --random-mm-num-mm-items-range-ratio 0.5 \ |
| 1284 | --random-mm-limit-mm-per-prompt '{"image": 4, "video": 0}' \ |
| 1285 | --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}' \ |
| 1286 | ``` |
| 1287 | |
| 1288 | Flags specific to `random-mm`: |
| 1289 | |
| 1290 | - `--random-mm-base-items-per-request`: base number of multimodal items per request. |
| 1291 | - `--random-mm-num-mm-items-range-ratio`: vary item count uniformly in the closed integer range [floor(n·(1−r)), ceil(n·(1+r))]. Set r=0 to keep it fixed; r=1 allows 0 items. |
| 1292 | - `--random-mm-limit-mm-per-prompt`: per-modality hard caps, e.g. '{"image": 3, "video": 0}'. |
| 1293 | - `--random-mm-bucket-config`: dict mapping (H, W, T) → probability. Entries with probability 0 are removed; remaining probabilities are renormalized to sum to 1. Use T=1 for images. Set any T>1 for videos (video sampling not yet supported). |
| 1294 | |
| 1295 | Behavioral notes: |
| 1296 | |
| 1297 | - If the requested base item count cannot be satisfied under the provided per-prompt limits, the tool raises an error rather than silently clamping. |
| 1298 | |
| 1299 | How sampling works: |
| 1300 | |
| 1301 | - Determine per-request item count k by sampling uniformly from the integer range defined by `--random-mm-base-items-per-request` and `--random-mm-num-mm-items-range-ratio`, then clamp k to at most the sum of per-modality limits. |
| 1302 | - For each of the k items, sample a bucket (H, W, T) according to the normalized probabilities in `--random-mm-bucket-config`, while tracking how many items of each modality have been added. |
| 1303 | - If a modality (e.g., image) reaches its limit from `--random-mm-limit-mm-per-prompt`, all buckets of that modality are excluded and the remaining bucket probabilities are renormalized before continuing. |
| 1304 | This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`. |
| 1305 | - The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`. |
| 1306 | |
| 1307 | </details> |
| 1308 | |
| 1309 | ### 🔬 Multimodal Processor Benchmark |
| 1310 | |
| 1311 | Benchmark per-stage latency of the multimodal (MM) input processor pipeline, including the encoder forward pass. This is useful for profiling preprocessing bottlenecks in vision-language models. |
| 1312 | |
| 1313 | <details class="admonition abstract" markdown="1"> |
| 1314 | <summary>Show more</summary> |
| 1315 | |
| 1316 | The benchmark measures the following stages for each request: |
| 1317 | |
| 1318 | | Stage | Description | |
| 1319 | | ----- | ----------- | |
| 1320 | | `get_mm_hashes_secs` | Time spent hashing multimodal inputs | |
| 1321 | | `get_cache_missing_items_secs` | Time spent looking up the processor cache | |
| 1322 | | `apply_hf_processor_secs` | Time spent in the HuggingFace processor | |
| 1323 | | `merge_mm_kwargs_secs` | Time spent merging multimodal kwargs | |
| 1324 | | `apply_prompt_updates_secs` | Time spent updating prompt tokens | |
| 1325 | | `preprocessor_total_secs` | Total preprocessing time | |
| 1326 | | `encoder_forward_secs` | Time spent in the encoder model forward pass | |
| 1327 | | `num_encoder_calls` | Number of encoder invocations per request | |
| 1328 | |
| 1329 | The benchmark also reports end-to-end latency (TTFT + decode time) per |
| 1330 | request. Use `--metric-percentiles` to select which percentiles to report |
| 1331 | (default: p99) and `--output-json` to save results. |
| 1332 | |
| 1333 | #### Basic Example with Synthetic Data (random-mm) |
| 1334 | |
| 1335 | ```bash |
| 1336 | vllm bench mm-processor \ |
| 1337 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 1338 | --dataset-name random-mm \ |
| 1339 | --num-prompts 50 \ |
| 1340 | --random-input-len 300 \ |
| 1341 | --random-output-len 40 \ |
| 1342 | --random-mm-base-items-per-request 2 \ |
| 1343 | --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \ |
| 1344 | --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}' |
| 1345 | ``` |
| 1346 | |
| 1347 | #### Using a HuggingFace Dataset |
| 1348 | |
| 1349 | ```bash |
| 1350 | vllm bench mm-processor \ |
| 1351 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 1352 | --dataset-name hf \ |
| 1353 | --dataset-path lmarena-ai/VisionArena-Chat \ |
| 1354 | --hf-split train \ |
| 1355 | --num-prompts 100 |
| 1356 | ``` |
| 1357 | |
| 1358 | #### Warmup, Custom Percentiles, and JSON Output |
| 1359 | |
| 1360 | ```bash |
| 1361 | vllm bench mm-processor \ |
| 1362 | --model Qwen/Qwen2-VL-7B-Instruct \ |
| 1363 | --dataset-name random-mm \ |
| 1364 | --num-prompts 200 \ |
| 1365 | --num-warmups 5 \ |
| 1366 | --random-input-len 300 \ |
| 1367 | --random-output-len 40 \ |
| 1368 | --random-mm-base-items-per-request 1 \ |
| 1369 | --metric-percentiles 50,90,95,99 \ |
| 1370 | --output-json results.json |
| 1371 | ``` |
| 1372 | |
| 1373 | See [`vllm bench mm-processor`](../cli/bench/mm_processor.md) for the full argument reference. |
| 1374 | |
| 1375 | </details> |
| 1376 | |
| 1377 | ### Embedding Benchmark |
| 1378 | |
| 1379 | Benchmark the performance of embedding requests in vLLM. |
| 1380 | |
| 1381 | <details class="admonition abstract" markdown="1"> |
| 1382 | <summary>Show more</summary> |
| 1383 | |
| 1384 | #### Text Embeddings |
| 1385 | |
| 1386 | Unlike generative models which use Completions API or Chat Completions API, |
| 1387 | you should set `--backend openai-embeddings` and `--endpoint /v1/embeddings` to use the Embeddings API. |
| 1388 | |
| 1389 | You can use any text dataset to benchmark the model, such as ShareGPT. |
| 1390 | |
| 1391 | Start the server: |
| 1392 | |
| 1393 | ```bash |
| 1394 | vllm serve jinaai/jina-embeddings-v3 --trust-remote-code |
| 1395 | ``` |
| 1396 | |
| 1397 | Run the benchmark: |
| 1398 | |
| 1399 | ```bash |
| 1400 | # download dataset |
| 1401 | # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json |
| 1402 | vllm bench serve \ |
| 1403 | --model jinaai/jina-embeddings-v3 \ |
| 1404 | --backend openai-embeddings \ |
| 1405 | --endpoint /v1/embeddings \ |
| 1406 | --dataset-name sharegpt \ |
| 1407 | --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json |
| 1408 | ``` |
| 1409 | |
| 1410 | #### Multi-modal Embeddings |
| 1411 | |
| 1412 | Unlike generative models which use Completions API or Chat Completions API, |
| 1413 | you should set `--endpoint /v1/embeddings` to use the Embeddings API. The backend to use depends on the model: |
| 1414 | |
| 1415 | - CLIP: `--backend openai-embeddings-clip` |
| 1416 | - VLM2Vec: `--backend openai-embeddings-vlm2vec` |
| 1417 | |
| 1418 | For other models, please add your own implementation inside [vllm/benchmarks/lib/endpoint_request_func.py](../../vllm/benchmarks/lib/endpoint_request_func.py) to match the expected instruction format. |
| 1419 | |
| 1420 | You can use any text or multi-modal dataset to benchmark the model, as long as the model supports it. |
| 1421 | For example, you can use ShareGPT and VisionArena to benchmark vision-language embeddings. |
| 1422 | |
| 1423 | Serve and benchmark CLIP: |
| 1424 | |
| 1425 | ```bash |
| 1426 | # Run this in another process |
| 1427 | vllm serve openai/clip-vit-base-patch32 |
| 1428 | |
| 1429 | # Run these one by one after the server is up |
| 1430 | # download dataset |
| 1431 | # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json |
| 1432 | vllm bench serve \ |
| 1433 | --model openai/clip-vit-base-patch32 \ |
| 1434 | --backend openai-embeddings-clip \ |
| 1435 | --endpoint /v1/embeddings \ |
| 1436 | --dataset-name sharegpt \ |
| 1437 | --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json |
| 1438 | |
| 1439 | vllm bench serve \ |
| 1440 | --model openai/clip-vit-base-patch32 \ |
| 1441 | --backend openai-embeddings-clip \ |
| 1442 | --endpoint /v1/embeddings \ |
| 1443 | --dataset-name hf \ |
| 1444 | --dataset-path lmarena-ai/VisionArena-Chat |
| 1445 | ``` |
| 1446 | |
| 1447 | Serve and benchmark VLM2Vec: |
| 1448 | |
| 1449 | ```bash |
| 1450 | # Run this in another process |
| 1451 | vllm serve TIGER-Lab/VLM2Vec-Full --runner pooling \ |
| 1452 | --trust-remote-code \ |
| 1453 | --chat-template examples/pooling/embed/template/vlm2vec_phi3v.jinja |
| 1454 | |
| 1455 | # Run these one by one after the server is up |
| 1456 | # download dataset |
| 1457 | # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json |
| 1458 | vllm bench serve \ |
| 1459 | --model TIGER-Lab/VLM2Vec-Full \ |
| 1460 | --backend openai-embeddings-vlm2vec \ |
| 1461 | --endpoint /v1/embeddings \ |
| 1462 | --dataset-name sharegpt \ |
| 1463 | --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json |
| 1464 | |
| 1465 | vllm bench serve \ |
| 1466 | --model TIGER-Lab/VLM2Vec-Full \ |
| 1467 | --backend openai-embeddings-vlm2vec \ |
| 1468 | --endpoint /v1/embeddings \ |
| 1469 | --dataset-name hf \ |
| 1470 | --dataset-path lmarena-ai/VisionArena-Chat |
| 1471 | ``` |
| 1472 | |
| 1473 | </details> |
| 1474 | |
| 1475 | ### Reranker Benchmark |
| 1476 | |
| 1477 | Benchmark the performance of rerank requests in vLLM. |
| 1478 | |
| 1479 | <details class="admonition abstract" markdown="1"> |
| 1480 | <summary>Show more</summary> |
| 1481 | |
| 1482 | Unlike generative models which use Completions API or Chat Completions API, |
| 1483 | you should set `--backend vllm-rerank` and `--endpoint /v1/rerank` to use the Reranker API. |
| 1484 | |
| 1485 | For reranking, the only supported dataset is `--dataset-name random-rerank` |
| 1486 | |
| 1487 | Start the server: |
| 1488 | |
| 1489 | ```bash |
| 1490 | vllm serve BAAI/bge-reranker-v2-m3 |
| 1491 | ``` |
| 1492 | |
| 1493 | Run the benchmark: |
| 1494 | |
| 1495 | ```bash |
| 1496 | vllm bench serve \ |
| 1497 | --model BAAI/bge-reranker-v2-m3 \ |
| 1498 | --backend vllm-rerank \ |
| 1499 | --endpoint /v1/rerank \ |
| 1500 | --dataset-name random-rerank \ |
| 1501 | --tokenizer BAAI/bge-reranker-v2-m3 \ |
| 1502 | --random-input-len 512 \ |
| 1503 | --num-prompts 10 \ |
| 1504 | --random-batch-size 5 |
| 1505 | ``` |
| 1506 | |
| 1507 | For reranker models, this will create `num_prompts / random_batch_size` requests with |
| 1508 | `random_batch_size` "documents" where each one has close to `random_input_len` tokens. |
| 1509 | In the example above, this results in 2 rerank requests with 5 "documents" each where |
| 1510 | each document has close to 512 tokens. |
| 1511 | |
| 1512 | Please note that the `/v1/rerank` is also supported by embedding models. So if you're running |
| 1513 | with an embedding model, also set `--no_reranker`. Because in this case the query is |
| 1514 | treated as an individual prompt by the server, here we send `random_batch_size - 1` documents |
| 1515 | to account for the extra prompt which is the query. The token accounting to report the |
| 1516 | throughput numbers correctly is also adjusted. |
| 1517 | |
| 1518 | </details> |