快速上手：5 分钟跑通第一个请求

# 1) 装 uv（如果还没有）
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2) 建一个干净 venv
mkdir sglang-play && cd sglang-play
uv venv --python 3.11
source .venv/bin/activate

# 3) 装 sglang（含 srt 后端）
uv pip install "sglang[all]" --upgrade

python -m sglang.check_env

python -m sglang.launch_server \
    --model-path Qwen/Qwen2.5-0.5B-Instruct \
    --host 0.0.0.0 \
    --port 30000 \
    --mem-fraction-static 0.5

INFO ... server_args=ServerArgs(...)            ← 全部启动参数
INFO ... Attention backend not specified.        ← 自动选了 attention 后端
       Use ... backend by default.
INFO ... Init torch distributed begin.           ← 单卡也会建 process group
INFO ... Memory pool end. avail mem=XX.X GB      ← KV pool 留多少显存
INFO ... Use cuda graph for decoding mode.       ← CUDA Graph 开启了
INFO ... Capture cuda graph begin. ...           ← 在录 CUDA Graph
INFO ... Capture cuda graph end. ...             ← 录完，启动完成
INFO ... Application startup complete.
INFO ... The server is fired up and ready to roll!

curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen2.5-0.5B-Instruct",
    "messages": [{"role": "user", "content": "1+1 等于几？"}],
    "max_tokens": 32
  }'

{
  "id": "...",
  "model": "Qwen/Qwen2.5-0.5B-Instruct",
  "choices": [{
    "index": 0,
    "message": {"role": "assistant", "content": "1+1=2。"},
    "finish_reason": "stop"
  }],
  "usage": {"prompt_tokens": ..., "completion_tokens": ...}
}

curl -N http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen2.5-0.5B-Instruct",
    "messages": [{"role": "user", "content": "讲一个短笑话"}],
    "stream": true
  }'

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")

# 非流式
resp = client.chat.completions.create(
    model="Qwen/Qwen2.5-0.5B-Instruct",
    messages=[{"role": "user", "content": "1+1?"}],
    max_tokens=32,
)
print(resp.choices[0].message.content)

# 流式
stream = client.chat.completions.create(
    model="Qwen/Qwen2.5-0.5B-Instruct",
    messages=[{"role": "user", "content": "说三个颜色"}],
    stream=True,
    max_tokens=64,
)
for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)
print()

import sglang as sgl

sgl.set_default_backend(sgl.RuntimeEndpoint("http://localhost:30000"))

@sgl.function
def hello(s, name):
    s += "User: 你好，我叫 " + name + "，请用一句话欢迎我。\nAssistant: "
    s += sgl.gen("reply", max_tokens=64, stop="\n")

state = hello.run(name="Zoe")
print(state["reply"])

@sgl.function
def multi_choice(s, q):
    s += f"问题：{q}\n答案是："
    s += sgl.select("ans", choices=["A", "B", "C", "D"])

print(multi_choice.run(q="天空是什么颜色？A.红 B.蓝 C.绿 D.黄")["ans"])

curl http://localhost:30000/metrics | head -50

sglang:num_requests_total{...} 5
sglang:num_input_tokens_total{...} 42
sglang:num_output_tokens_total{...} 137
sglang:e2e_request_latency_seconds_bucket{le="0.05",...} 2
sglang:e2e_request_latency_seconds_bucket{le="0.1",...} 5
sglang:gen_throughput{...} 23.4
sglang:cache_hit_rate{...} 0.62          ← 重要：前缀命中率
...

python -m sglang.launch_server --help | head -50   # 完整列表

# Ctrl-C 优雅退出（会等当前请求结束）
# 或者
pkill -f sglang.launch_server