Backend: SGLang Runtime (SRT)

Backend: SGLang Runtime (SRT)#

The SGLang Runtime (SRT) is an efficient serving engine.

Quick Start#

Launch a server

python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000

Send a request

curl http://localhost:30000/generate \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Once upon a time,",
    "sampling_params": {
      "max_new_tokens": 16,
      "temperature": 0
    }
  }'

Learn more about the argument specification, streaming, and multi-modal support here.

OpenAI Compatible API#

In addition, the server supports OpenAI-compatible APIs.

import openai
client = openai.Client(
    base_url="http://127.0.0.1:30000/v1", api_key="EMPTY")

# Text completion
response = client.completions.create(
	model="default",
	prompt="The capital of France is",
	temperature=0,
	max_tokens=32,
)
print(response)

# Chat completion
response = client.chat.completions.create(
    model="default",
    messages=[
        {"role": "system", "content": "You are a helpful AI assistant"},
        {"role": "user", "content": "List 3 countries and their capitals."},
    ],
    temperature=0,
    max_tokens=64,
)
print(response)

# Text embedding
response = client.embeddings.create(
    model="default",
    input="How are you today",
)
print(response)

It supports streaming, vision, and almost all features of the Chat/Completions/Models/Batch endpoints specified by the OpenAI API Reference.

Additional Server Arguments#

  • To enable multi-GPU tensor parallelism, add --tp 2. If it reports the error “peer access is not supported between these two devices”, add --enable-p2p-check to the server launch command.

python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --tp 2

  • To enable multi-GPU data parallelism, add --dp 2. 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.

python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --dp 2 --tp 2

  • 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.

python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --mem-fraction-static 0.7

  • See hyperparameter tuning on tuning hyperparameters for better performance.

  • If you see out-of-memory errors during prefill for long prompts, try to set a smaller chunked prefill size.

python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --chunked-prefill-size 4096

  • To enable the experimental overlapped scheduler, add --enable-overlap-scheduler. It overlaps CPU scheduler with GPU computation and can accelerate almost all workloads. This does not work for constrained decoding currenly.

  • To enable torch.compile acceleration, add --enable-torch-compile. It accelerates small models on small batch sizes. This does not work for FP8 currenly.

  • To enable torchao quantization, add --torchao-config int4wo-128. It supports various quantization strategies.

  • To enable fp8 weight quantization, add --quantization fp8 on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.

  • To enable fp8 kv cache quantization, add --kv-cache-dtype fp8_e5m2.

  • If the model does not have a chat template in the Hugging Face tokenizer, you can specify a custom chat template.

  • To run tensor parallelism on multiple nodes, add --nnodes 2. 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, you can use the following commands. If you meet deadlock, please try to add --disable-cuda-graph

# Node 0
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

# Node 1
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

Engine Without HTTP Server#

We also provide an inference engine without a HTTP server. For example,

import sglang as sgl

def main():
    prompts = [
        "Hello, my name is",
        "The president of the United States is",
        "The capital of France is",
        "The future of AI is",
    ]
    sampling_params = {"temperature": 0.8, "top_p": 0.95}
    llm = sgl.Engine(model_path="meta-llama/Meta-Llama-3.1-8B-Instruct")

    outputs = llm.generate(prompts, sampling_params)
    for prompt, output in zip(prompts, outputs):
        print("===============================")
        print(f"Prompt: {prompt}\nGenerated text: {output['text']}")

if __name__ == "__main__":
    main()

This can be used for offline batch inference and building custom servers. You can view the full example here.

Supported Models#

Generative Models

  • Llama / Llama 2 / Llama 3 / Llama 3.1

  • Mistral / Mixtral / Mistral NeMo

  • Gemma / Gemma 2

  • Qwen / Qwen 2 / Qwen 2 MoE / Qwen 2 VL

  • DeepSeek / DeepSeek 2

  • OLMoE

  • LLaVA-OneVision

    • python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-7b-ov --port=30000 --chat-template=chatml-llava

    • python3 -m sglang.launch_server --model-path lmms-lab/llava-onevision-qwen2-72b-ov --port=30000 --tp-size=8 --chat-template=chatml-llava

    • Query the server with the OpenAI Vision API. See examples at test/srt/test_vision_openai_server.py

  • LLaVA 1.5 / 1.6 / NeXT

    • python -m sglang.launch_server --model-path lmms-lab/llama3-llava-next-8b --port=30000 --tp-size=1 --chat-template=llava_llama_3

    • python -m sglang.launch_server --model-path lmms-lab/llava-next-72b --port=30000 --tp-size=8 --chat-template=chatml-llava

    • Query the server with the OpenAI Vision API. See examples at test/srt/test_vision_openai_server.py

  • Yi-VL

  • StableLM

  • Command-R

  • DBRX

  • Grok

  • ChatGLM

  • InternLM 2

  • Exaone 3

  • BaiChuan2

  • MiniCPM / MiniCPM 3

  • XVERSE / XVERSE MoE

  • SmolLM

  • GLM-4

Embedding Models

  • e5-mistral

  • gte-Qwen2

    • python -m sglang.launch_server --model-path Alibaba-NLP/gte-Qwen2-7B-instruct --is-embedding

Instructions for supporting a new model are here.

Use Models From ModelScope#

More

To use a model from ModelScope, set the environment variable SGLANG_USE_MODELSCOPE.

export SGLANG_USE_MODELSCOPE=true

Launch Qwen2-7B-Instruct Server

SGLANG_USE_MODELSCOPE=true python -m sglang.launch_server --model-path qwen/Qwen2-7B-Instruct --port 30000

Or start it by docker.

docker run --gpus all \
    -p 30000:30000 \
    -v ~/.cache/modelscope:/root/.cache/modelscope \
    --env "SGLANG_USE_MODELSCOPE=true" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server --model-path Qwen/Qwen2.5-7B-Instruct --host 0.0.0.0 --port 30000

Run Llama 3.1 405B#

More 
# Run 405B (fp8) on a single node
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 --tp 8

# Run 405B (fp16) on two nodes
## on the first node, replace the `172.16.4.52:20000` with your own first node ip address and port
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

## on the first node, replace the `172.16.4.52:20000` with your own first node ip address and port
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

Benchmark Performance#

  • Benchmark a single static batch by running the following command without launching a server. The arguments are the same as for launch_server.py. 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. 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.

    python -m sglang.bench_latency --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch 32 --input-len 256 --output-len 32

  • Benchmark online serving. Launch a server first and run the following command.

    python3 -m sglang.bench_serving --backend sglang --num-prompt 10