|
| 1 | +## Backend: SGLang Runtime (SRT) |
| 2 | +The SGLang Runtime (SRT) is an efficient serving engine. |
| 3 | + |
| 4 | +### Quick Start |
| 5 | +Launch a server |
| 6 | +``` |
| 7 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 |
| 8 | +``` |
| 9 | + |
| 10 | +Send a request |
| 11 | +``` |
| 12 | +curl http://localhost:30000/generate \ |
| 13 | + -H "Content-Type: application/json" \ |
| 14 | + -d '{ |
| 15 | + "text": "Once upon a time,", |
| 16 | + "sampling_params": { |
| 17 | + "max_new_tokens": 16, |
| 18 | + "temperature": 0 |
| 19 | + } |
| 20 | + }' |
| 21 | +``` |
| 22 | +Learn more about the argument format [here](docs/en/sampling_params.md). |
| 23 | + |
| 24 | +### OpenAI Compatible API |
| 25 | +In addition, the server supports OpenAI-compatible APIs. |
| 26 | + |
| 27 | +```python |
| 28 | +import openai |
| 29 | +client = openai.Client( |
| 30 | + base_url="http://127.0.0.1:30000/v1", api_key="EMPTY") |
| 31 | + |
| 32 | +# Text completion |
| 33 | +response = client.completions.create( |
| 34 | + model="default", |
| 35 | + prompt="The capital of France is", |
| 36 | + temperature=0, |
| 37 | + max_tokens=32, |
| 38 | +) |
| 39 | +print(response) |
| 40 | + |
| 41 | +# Chat completion |
| 42 | +response = client.chat.completions.create( |
| 43 | + model="default", |
| 44 | + messages=[ |
| 45 | + {"role": "system", "content": "You are a helpful AI assistant"}, |
| 46 | + {"role": "user", "content": "List 3 countries and their capitals."}, |
| 47 | + ], |
| 48 | + temperature=0, |
| 49 | + max_tokens=64, |
| 50 | +) |
| 51 | +print(response) |
| 52 | + |
| 53 | +# Text embedding |
| 54 | +response = client.embeddings.create( |
| 55 | + model="default", |
| 56 | + input="How are you today", |
| 57 | +) |
| 58 | +print(response) |
| 59 | +``` |
| 60 | + |
| 61 | +It supports streaming, vision, and most features of the Chat/Completions/Models/Batch endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/). |
| 62 | + |
| 63 | +### Additional Server Arguments |
| 64 | +- Add `--tp 2` to enable multi-GPU tensor parallelism. If it reports the error "peer access is not supported between these two devices", add `--enable-p2p-check` to the server launch command. |
| 65 | +``` |
| 66 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --tp 2 |
| 67 | +``` |
| 68 | +- Add `--dp 2` to enable multi-GPU data parallelism. Data parallelism is better for throughput if there is enough memory. It can also be used together with tensor parallelism. The following command uses 4 GPUs in total. |
| 69 | +``` |
| 70 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --dp 2 --tp 2 |
| 71 | +``` |
| 72 | +- If you see out-of-memory errors during serving, try to reduce the memory usage of the KV cache pool by setting a smaller value of `--mem-fraction-static`. The default value is `0.9`. |
| 73 | +``` |
| 74 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --mem-fraction-static 0.7 |
| 75 | +``` |
| 76 | +- See [hyperparameter_tuning.md](docs/en/hyperparameter_tuning.md) on tuning hyperparameters for better performance. |
| 77 | +- If you see out-of-memory errors during prefill for long prompts, try to set a smaller chunked prefill size. |
| 78 | +``` |
| 79 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --chunked-prefill-size 4096 |
| 80 | +``` |
| 81 | +- To enable torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes. |
| 82 | +- To enable fp8 weight quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments. |
| 83 | +- To enable fp8 kv cache quanzation, you can add `--kv-cache-dtype fp8_e5m2`. |
| 84 | +- If the model does not have a template in the Hugging Face tokenizer, you can specify a [custom chat template](docs/en/custom_chat_template.md). |
| 85 | +- Add `--nnodes 2` to run tensor parallelism on multiple nodes. If you have two nodes with two GPUs on each node and want to run TP=4, let `sgl-dev-0` be the hostname of the first node and `50000` be an available port. |
| 86 | +``` |
| 87 | +# Node 0 |
| 88 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --tp 4 --nccl-init sgl-dev-0:50000 --nnodes 2 --node-rank 0 |
| 89 | +
|
| 90 | +# Node 1 |
| 91 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --tp 4 --nccl-init sgl-dev-0:50000 --nnodes 2 --node-rank 1 |
| 92 | +``` |
| 93 | + |
| 94 | +### Supported Models |
| 95 | + |
| 96 | +**Generative Models** |
| 97 | +- Llama / Llama 2 / Llama 3 / Llama 3.1 |
| 98 | +- Mistral / Mixtral / Mistral NeMo |
| 99 | +- Gemma / Gemma 2 |
| 100 | +- Qwen / Qwen 2 / Qwen 2 MoE |
| 101 | +- DeepSeek / DeepSeek 2 |
| 102 | +- [LLaVA-OneVision](https://llava-vl.github.io/blog/2024-08-05-llava-onevision/) |
| 103 | + - `python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-7b-ov --port=30000 --chat-template=chatml-llava` |
| 104 | + - `python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-72b-ov --port=30000 --tp-size=8 --chat-template=chatml-llava` |
| 105 | + - Query the server with the [OpenAI Vision API](https://platform.openai.com/docs/guides/vision). See examples at [test/srt/test_vision_openai_server.py](test/srt/test_vision_openai_server.py) |
| 106 | +- LLaVA 1.5 / 1.6 / NeXT |
| 107 | + - `python -m sglang.launch_server --model-path lmms-lab/llama3-llava-next-8b --port=30000 --tp-size=1 --chat-template=llava_llama_3` |
| 108 | + - `python -m sglang.launch_server --model-path lmms-lab/llava-next-72b --port=30000 --tp-size=8 --chat-template=chatml-llava` |
| 109 | + - Query the server with the [OpenAI Vision API](https://platform.openai.com/docs/guides/vision). See examples at [test/srt/test_vision_openai_server.py](test/srt/test_vision_openai_server.py) |
| 110 | +- Yi-VL |
| 111 | +- StableLM |
| 112 | +- Command-R |
| 113 | +- DBRX |
| 114 | +- Grok |
| 115 | +- ChatGLM |
| 116 | +- InternLM 2 |
| 117 | +- Exaone 3 |
| 118 | + |
| 119 | +**Embedding Models** |
| 120 | + |
| 121 | +- e5-mistral |
| 122 | +- gte-Qwen2 |
| 123 | + - `python -m sglang.launch_server --model-path Alibaba-NLP/gte-Qwen2-7B-instruct --is-embedding` |
| 124 | + |
| 125 | +Instructions for supporting a new model are [here](https://github.com/sgl-project/sglang/blob/main/docs/en/model_support.md). |
| 126 | + |
| 127 | +#### Use Models From ModelScope |
| 128 | +<details> |
| 129 | +<summary>More</summary> |
| 130 | + |
| 131 | +To use a model from [ModelScope](https://www.modelscope.cn), set the environment variable SGLANG_USE_MODELSCOPE. |
| 132 | +``` |
| 133 | +export SGLANG_USE_MODELSCOPE=true |
| 134 | +``` |
| 135 | +Launch [Qwen2-7B-Instruct](https://www.modelscope.cn/models/qwen/qwen2-7b-instruct) Server |
| 136 | +``` |
| 137 | +SGLANG_USE_MODELSCOPE=true python -m sglang.launch_server --model-path qwen/Qwen2-7B-Instruct --port 30000 |
| 138 | +``` |
| 139 | + |
| 140 | +</details> |
| 141 | + |
| 142 | +#### Run Llama 3.1 405B |
| 143 | +<details> |
| 144 | +<summary>More</summary> |
| 145 | + |
| 146 | +```bash |
| 147 | +# Run 405B (fp8) on a single node |
| 148 | +python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8 |
| 149 | + |
| 150 | +# Run 405B (fp16) on two nodes |
| 151 | +## on the first node, replace the `172.16.4.52:20000` with your own first node ip address and port |
| 152 | +GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 0 --disable-cuda-graph |
| 153 | + |
| 154 | +## on the first node, replace the `172.16.4.52:20000` with your own first node ip address and port |
| 155 | +GLOO_SOCKET_IFNAME=eth0 python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct --tp 16 --nccl-init-addr 172.16.4.52:20000 --nnodes 2 --node-rank 1 --disable-cuda-graph |
| 156 | +``` |
| 157 | + |
| 158 | +</details> |
| 159 | + |
| 160 | +### Benchmark Performance |
| 161 | + |
| 162 | +- Benchmark a single static batch by running the following command without launching a server. The arguments are the same as for `launch_server.py`. |
| 163 | + Note that this is not a dynamic batching server, so it may run out of memory for a batch size that a real server can handle. |
| 164 | + A real server truncates the prefill into several batches, while this unit test does not. For accurate large batch testing, please use `sglang.bench_serving` instead. |
| 165 | + ``` |
| 166 | + python -m sglang.bench_latency --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch 32 --input-len 256 --output-len 32 |
| 167 | + ``` |
| 168 | +- Benchmark online serving. Launch a server first and run the following command. |
| 169 | + ``` |
| 170 | + python3 -m sglang.bench_serving --backend sglang --num-prompt 10 |
| 171 | + ``` |
0 commit comments