README.md 20.8 KB
Newer Older
Lianmin Zheng's avatar
Lianmin Zheng committed
1
<div align="center">
2
<img src="https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png" alt="logo" width="400"></img>
Lianmin Zheng's avatar
Lianmin Zheng committed
3

Yineng Zhang's avatar
Yineng Zhang committed
4
5
6
7
8
9
[![PyPI](https://img.shields.io/pypi/v/sglang)](https://pypi.org/project/sglang)
![PyPI - Downloads](https://img.shields.io/pypi/dm/sglang)
[![license](https://img.shields.io/github/license/sgl-project/sglang.svg)](https://github.com/sgl-project/sglang/tree/main/LICENSE)
[![issue resolution](https://img.shields.io/github/issues-closed-raw/sgl-project/sglang)](https://github.com/sgl-project/sglang/issues)
[![open issues](https://img.shields.io/github/issues-raw/sgl-project/sglang)](https://github.com/sgl-project/sglang/issues)

Yineng Zhang's avatar
Yineng Zhang committed
10
11
</div>

Lianmin Zheng's avatar
Lianmin Zheng committed
12
13
--------------------------------------------------------------------------------

Ying Sheng's avatar
Ying Sheng committed
14
| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) | [**Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2ngly9muu-t37XiH87qvD~6rVBTkTEHw) |
Lianmin Zheng's avatar
Lianmin Zheng committed
15

Ying Sheng's avatar
Ying Sheng committed
16
17
SGLang is a fast serving framework for large language models and vision language models.
It makes your interaction with models faster and more controllable by co-designing the backend runtime and frontend language.
Lianmin Zheng's avatar
Lianmin Zheng committed
18

19
The core features include:
Ying Sheng's avatar
Ying Sheng committed
20
- **Fast Backend Runtime**: Efficient serving with RadixAttention for prefix caching, jump-forward constrained decoding, continuous batching, token attention (paged attention), tensor parallelism, flashinfer kernels, and quantization (AWQ/FP8/GPTQ/Marlin).
Lianmin Zheng's avatar
Lianmin Zheng committed
21
- **Flexible Frontend Language**: Enables easy programming of LLM applications with chained generation calls, advanced prompting, control flow, multiple modalities, parallelism, and external interactions.
Lianmin Zheng's avatar
Lianmin Zheng committed
22

Ying Sheng's avatar
Ying Sheng committed
23
## News
Ying Sheng's avatar
Ying Sheng committed
24
25
26
- [2024/07] 🔥 Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
- [2024/04] SGLang is used by the official **LLaVA-NeXT (video)** release ([blog](https://llava-vl.github.io/blog/2024-04-30-llava-next-video/)).
- [2024/02] SGLang enables **3x faster JSON decoding** with compressed finite state machine ([blog](https://lmsys.org/blog/2024-02-05-compressed-fsm/)).
Ying Sheng's avatar
Ying Sheng committed
27

Ying Sheng's avatar
Ying Sheng committed
28
29
30
<details>
<summary>More</summary>

Ying Sheng's avatar
Ying Sheng committed
31
- [2024/01] SGLang provides up to **5x faster inference** with RadixAttention ([blog](https://lmsys.org/blog/2024-01-17-sglang/)).
Ying Sheng's avatar
Ying Sheng committed
32
33
34
35
- [2024/01] SGLang powers the serving of the official **LLaVA v1.6** release demo ([usage](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#demo)).

</details>

Lianmin Zheng's avatar
Lianmin Zheng committed
36
37
38
## Contents
- [Install](#install)
- [Backend: SGLang Runtime (SRT)](#backend-sglang-runtime-srt)
Ying Sheng's avatar
Ying Sheng committed
39
- [Frontend: Structured Generation Language (SGLang)](#frontend-structured-generation-language-sglang)
Lianmin Zheng's avatar
Lianmin Zheng committed
40
41
42
43
44
45
- [Benchmark And Performance](#benchmark-and-performance)
- [Roadmap](#roadmap)
- [Citation And Acknowledgment](#citation-and-acknowledgment)

## Install

Lianmin Zheng's avatar
Lianmin Zheng committed
46
47
### Method 1: With pip
```
48
pip install --upgrade pip
Lianmin Zheng's avatar
Lianmin Zheng committed
49
pip install "sglang[all]"
Lianmin Zheng's avatar
Lianmin Zheng committed
50

Lianmin Zheng's avatar
Lianmin Zheng committed
51
# Install FlashInfer CUDA kernels
52
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
Lianmin Zheng's avatar
Lianmin Zheng committed
53
```
54

Lianmin Zheng's avatar
Lianmin Zheng committed
55
### Method 2: From source
Lianmin Zheng's avatar
Lianmin Zheng committed
56
```
Ying Sheng's avatar
Ying Sheng committed
57
# Use the last release branch
Yineng Zhang's avatar
Yineng Zhang committed
58
git clone -b v0.2.13 https://github.com/sgl-project/sglang.git
Lianmin Zheng's avatar
Lianmin Zheng committed
59
60
cd sglang

Lianmin Zheng's avatar
Lianmin Zheng committed
61
pip install --upgrade pip
Lianmin Zheng's avatar
Lianmin Zheng committed
62
63
pip install -e "python[all]"

Lianmin Zheng's avatar
Lianmin Zheng committed
64
# Install FlashInfer CUDA kernels
65
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/
Lianmin Zheng's avatar
Lianmin Zheng committed
66
```
67

Lianmin Zheng's avatar
Lianmin Zheng committed
68
### Method 3: Using docker
Ying Sheng's avatar
Ying Sheng committed
69
The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags), built from [Dockerfile](docker).
70
Replace `<secret>` below with your huggingface hub [token](https://huggingface.co/docs/hub/en/security-tokens).
Ying Sheng's avatar
Ying Sheng committed
71

Liangsheng Yin's avatar
Liangsheng Yin committed
72
73
74
75
```bash
docker run --gpus all \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
76
    --env "HF_TOKEN=<secret>" \
Liangsheng Yin's avatar
Liangsheng Yin committed
77
78
    --ipc=host \
    lmsysorg/sglang:latest \
79
    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --host 0.0.0.0 --port 30000
Liangsheng Yin's avatar
Liangsheng Yin committed
80
81
```

82
83
### Method 4: Using docker compose

84
<details>
85
86
87
88
89
> This method is recommended if you plan to serve it as a service.
> A better approach is to use the [k8s-sglang-service.yaml](./docker/k8s-sglang-service.yaml).

1. Copy the [compose.yml](./docker/compose.yaml) to your local machine
2. Execute the command `docker compose up -d` in your terminal.
90
</details>
91

92
93
### Method 5: Run on Kubernetes or Clouds with SkyPilot

94
<details>
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
To deploy on Kubernetes or 12+ clouds, you can use [SkyPilot](https://github.com/skypilot-org/skypilot).

1. Install SkyPilot and set up Kubernetes cluster or cloud access: see [SkyPilot's documentation](https://skypilot.readthedocs.io/en/latest/getting-started/installation.html).
2. Deploy on your own infra with a single command and get the HTTP API endpoint:
<details>
<summary>SkyPilot YAML: <code>sglang.yaml</code></summary>

```yaml
# sglang.yaml
envs:
  HF_TOKEN: null

resources:
  image_id: docker:lmsysorg/sglang:latest
  accelerators: A100
  ports: 30000

run: |
  conda deactivate
  python3 -m sglang.launch_server \
    --model-path meta-llama/Meta-Llama-3.1-8B-Instruct \
    --host 0.0.0.0 \
    --port 30000
```

```bash
# Deploy on any cloud or Kubernetes cluster. Use --cloud <cloud> to select a specific cloud provider.
HF_TOKEN=<secret> sky launch -c sglang --env HF_TOKEN sglang.yaml

# Get the HTTP API endpoint
sky status --endpoint 30000 sglang
```
3. To further scale up your deployment with autoscaling and failure recovery, check out the [SkyServe + SGLang guide](https://github.com/skypilot-org/skypilot/tree/master/llm/sglang#serving-llama-2-with-sglang-for-more-traffic-using-skyserve).
128
</details>
129
130


Lianmin Zheng's avatar
Lianmin Zheng committed
131
### Common Notes
Yineng Zhang's avatar
Yineng Zhang committed
132
- [FlashInfer](https://github.com/flashinfer-ai/flashinfer) is currently one of the dependencies that must be installed for SGLang. If you are using NVIDIA GPU devices below sm80, such as T4, you can't use SGLang for the time being. We expect to resolve this issue soon, so please stay tuned. If you encounter any FlashInfer-related issues on sm80+ devices (e.g., A100, L40S, H100), consider using Triton's kernel by `--disable-flashinfer --disable-flashinfer-sampling` and raise a issue.
Lianmin Zheng's avatar
Lianmin Zheng committed
133
- If you only need to use the OpenAI backend, you can avoid installing other dependencies by using `pip install "sglang[openai]"`.
Ying Sheng's avatar
Ying Sheng committed
134

Ying Sheng's avatar
Ying Sheng committed
135
136
137
## Backend: SGLang Runtime (SRT)
The SGLang Runtime (SRT) is an efficient serving engine.

Ying Sheng's avatar
Ying Sheng committed
138
### Quick Start
Ying Sheng's avatar
Ying Sheng committed
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
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
    }
  }'
```
156
Learn more about the argument format [here](docs/en/sampling_params.md).
Ying Sheng's avatar
Ying Sheng committed
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187

### OpenAI Compatible API
In addition, the server supports OpenAI-compatible APIs.

```python
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)
```

Ying Sheng's avatar
Ying Sheng committed
188
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/).
Ying Sheng's avatar
Ying Sheng committed
189
190

### Additional Server Arguments
191
- 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.
Ying Sheng's avatar
Ying Sheng committed
192
193
194
```
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --tp 2
```
195
- Add `--dp 2` to enable multi-GPU data parallelism. It can also be used together with tensor parallelism. Data parallelism is better for throughput if there is enough memory.
Ying Sheng's avatar
Ying Sheng committed
196
197
198
```
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --dp 2 --tp 2
```
199
- 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`.
Ying Sheng's avatar
Ying Sheng committed
200
201
202
```
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --mem-fraction-static 0.7
```
203
204
- See [hyperparameter_tuning.md](docs/en/hyperparameter_tuning.md) 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.
Ying Sheng's avatar
Ying Sheng committed
205
```
206
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --chunked-prefill-size 4096
Ying Sheng's avatar
Ying Sheng committed
207
```
208
- 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.
Ying Sheng's avatar
Ying Sheng committed
209
210
```
# Node 0
211
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
Ying Sheng's avatar
Ying Sheng committed
212
213

# Node 1
214
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
Ying Sheng's avatar
Ying Sheng committed
215
```
216
- 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).
217
- To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
218
- To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
Lianmin Zheng's avatar
Lianmin Zheng committed
219
 
Ying Sheng's avatar
Ying Sheng committed
220
221
### Supported Models

222
- Llama / Llama 2 / Llama 3 / Llama 3.1
223
- Mistral / Mixtral / Mistral NeMo
Ying Sheng's avatar
Ying Sheng committed
224
225
- Gemma / Gemma 2
- Qwen / Qwen 2 / Qwen 2 MoE
226
- DeepSeek / DeepSeek 2
Ying Sheng's avatar
Ying Sheng committed
227
- LLaVA 1.5 / 1.6
Ying Sheng's avatar
Ying Sheng committed
228
229
230
  - `python -m sglang.launch_server --model-path liuhaotian/llava-v1.5-7b --tokenizer-path llava-hf/llava-1.5-7b-hf --chat-template vicuna_v1.1 --port 30000`
  - `python -m sglang.launch_server --model-path liuhaotian/llava-v1.6-vicuna-7b --tokenizer-path llava-hf/llava-1.5-7b-hf --chat-template vicuna_v1.1 --port 30000`
  - `python -m sglang.launch_server --model-path liuhaotian/llava-v1.6-34b --tokenizer-path liuhaotian/llava-v1.6-34b-tokenizer --port 30000`
Ying Sheng's avatar
Ying Sheng committed
231
- LLaVA-NeXT-Video
Ying Sheng's avatar
Ying Sheng committed
232
  - see [examples/usage/llava_video](examples/usage/llava_video)
Ying Sheng's avatar
Ying Sheng committed
233
234
235
236
237
238
239
240
241
- Yi-VL
  - see [srt_example_yi_vl.py](examples/quick_start/srt_example_yi_vl.py).
- StableLM
- Command-R
- DBRX
- Grok
- ChatGLM
- InternLM 2

242
Instructions for supporting a new model are [here](https://github.com/sgl-project/sglang/blob/main/docs/en/model_support.md).
Ying Sheng's avatar
Ying Sheng committed
243

Lianmin Zheng's avatar
Lianmin Zheng committed
244
245
246
247
248
249
250
251
252
253
254
#### Use Models From ModelScope
To use model from [ModelScope](https://www.modelscope.cn), setting environment variable SGLANG_USE_MODELSCOPE.
```
export SGLANG_USE_MODELSCOPE=true
```
Launch [Qwen2-7B-Instruct](https://www.modelscope.cn/models/qwen/qwen2-7b-instruct) Server
```
SGLANG_USE_MODELSCOPE=true python -m sglang.launch_server --model-path qwen/Qwen2-7B-Instruct --port 30000
```    

#### Run Llama 3.1 405B
Ying Sheng's avatar
Ying Sheng committed
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269

```bash
## 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
# replace the `172.16.4.52:20000` with your own first node ip address and port, disable CUDA Graph temporarily

# on the first node
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 --mem-frac 0.75

# on the second
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 --mem-frac 0.75
```

Ying Sheng's avatar
Ying Sheng committed
270
271
### Benchmark Performance

Ying Sheng's avatar
Ying Sheng committed
272
- 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, consider using `sglang.bench_serving`.
Ying Sheng's avatar
Ying Sheng committed
273
  ```
274
  python -m sglang.bench_latency --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch 32 --input-len 256 --output-len 32
Ying Sheng's avatar
Ying Sheng committed
275
276
277
  ```
- Benchmark online serving. Launch a server first and run the following command.
  ```
278
  python3 -m sglang.bench_serving --backend sglang --num-prompt 10
Ying Sheng's avatar
Ying Sheng committed
279
280
  ```

Ying Sheng's avatar
Ying Sheng committed
281
## Frontend: Structured Generation Language (SGLang)
282
The frontend language can be used with local models or API models. It is an alternative to the OpenAI API. You may found it easier to use for complex prompting workflow.
Ying Sheng's avatar
Ying Sheng committed
283
284

### Quick Start
Lianmin Zheng's avatar
Lianmin Zheng committed
285
286
The example below shows how to use sglang to answer a mulit-turn question.

Ying Sheng's avatar
Ying Sheng committed
287
#### Using Local Models
288
First, launch a server with
Lianmin Zheng's avatar
Lianmin Zheng committed
289
```
Ying Sheng's avatar
Ying Sheng committed
290
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000
Lianmin Zheng's avatar
Lianmin Zheng committed
291
292
```

293
294
Then, connect to the server and answer a multi-turn question.

Lianmin Zheng's avatar
Lianmin Zheng committed
295
```python
296
from sglang import function, system, user, assistant, gen, set_default_backend, RuntimeEndpoint
Lianmin Zheng's avatar
Lianmin Zheng committed
297
298
299
300
301
302
303
304
305

@function
def multi_turn_question(s, question_1, question_2):
    s += system("You are a helpful assistant.")
    s += user(question_1)
    s += assistant(gen("answer_1", max_tokens=256))
    s += user(question_2)
    s += assistant(gen("answer_2", max_tokens=256))

306
set_default_backend(RuntimeEndpoint("http://localhost:30000"))
Lianmin Zheng's avatar
Lianmin Zheng committed
307
308
309
310
311
312
313
314

state = multi_turn_question.run(
    question_1="What is the capital of the United States?",
    question_2="List two local attractions.",
)

for m in state.messages():
    print(m["role"], ":", m["content"])
315
316

print(state["answer_1"])
Lianmin Zheng's avatar
Lianmin Zheng committed
317
318
```

Ying Sheng's avatar
Ying Sheng committed
319
#### Using OpenAI Models
320
Set the OpenAI API Key
Lianmin Zheng's avatar
Lianmin Zheng committed
321
```
322
export OPENAI_API_KEY=sk-******
Lianmin Zheng's avatar
Lianmin Zheng committed
323
324
```

325
Then, answer a multi-turn question.
Lianmin Zheng's avatar
Lianmin Zheng committed
326
```python
327
from sglang import function, system, user, assistant, gen, set_default_backend, OpenAI
Lianmin Zheng's avatar
Lianmin Zheng committed
328
329
330
331
332
333
334
335
336

@function
def multi_turn_question(s, question_1, question_2):
    s += system("You are a helpful assistant.")
    s += user(question_1)
    s += assistant(gen("answer_1", max_tokens=256))
    s += user(question_2)
    s += assistant(gen("answer_2", max_tokens=256))

337
set_default_backend(OpenAI("gpt-3.5-turbo"))
Lianmin Zheng's avatar
Lianmin Zheng committed
338
339
340
341
342
343
344
345

state = multi_turn_question.run(
    question_1="What is the capital of the United States?",
    question_2="List two local attractions.",
)

for m in state.messages():
    print(m["role"], ":", m["content"])
346
347

print(state["answer_1"])
Lianmin Zheng's avatar
Lianmin Zheng committed
348
349
```

Ying Sheng's avatar
Ying Sheng committed
350
#### More Examples
Lianmin Zheng's avatar
Lianmin Zheng committed
351

352
Anthropic and VertexAI (Gemini) models are also supported.
Lianmin Zheng's avatar
Lianmin Zheng committed
353
354
You can find more examples at [examples/quick_start](examples/quick_start).

Ying Sheng's avatar
Ying Sheng committed
355
### Language Feature
Lianmin Zheng's avatar
Lianmin Zheng committed
356
357
358
359
360
To begin with, import sglang.
```python
import sglang as sgl
```

Lianmin Zheng's avatar
Lianmin Zheng committed
361
`sglang` provides some simple primitives such as `gen`, `select`, `fork`, `image`.
Lianmin Zheng's avatar
Lianmin Zheng committed
362
363
You can implement your prompt flow in a function decorated by `sgl.function`.
You can then invoke the function with `run` or `run_batch`.
364
The system will manage the state, chat template, parallelism and batching for you.
Lianmin Zheng's avatar
Lianmin Zheng committed
365

366
367
The complete code for the examples below can be found at [readme_examples.py](examples/usage/readme_examples.py)

Ying Sheng's avatar
Ying Sheng committed
368
#### Control Flow
Lianmin Zheng's avatar
Lianmin Zheng committed
369
370
You can use any Python code within the function body, including control flow, nested function calls, and external libraries.

Lianmin Zheng's avatar
Lianmin Zheng committed
371
372
```python
@sgl.function
373
374
375
def tool_use(s, question):
    s += "To answer this question: " + question + ". "
    s += "I need to use a " + sgl.gen("tool", choices=["calculator", "search engine"]) + ". "
Lianmin Zheng's avatar
Lianmin Zheng committed
376
377
378

    if s["tool"] == "calculator":
        s += "The math expression is" + sgl.gen("expression")
379
380
    elif s["tool"] == "search engine":
        s += "The key word to search is" + sgl.gen("word")
Lianmin Zheng's avatar
Lianmin Zheng committed
381
```
Lianmin Zheng's avatar
Lianmin Zheng committed
382

Ying Sheng's avatar
Ying Sheng committed
383
#### Parallelism
Lianmin Zheng's avatar
Lianmin Zheng committed
384
385
386
Use `fork` to launch parallel prompts.
Because `sgl.gen` is non-blocking, the for loop below issues two generation calls in parallel.

Lianmin Zheng's avatar
Lianmin Zheng committed
387
388
389
390
391
392
393
394
```python
@sgl.function
def tip_suggestion(s):
    s += (
        "Here are two tips for staying healthy: "
        "1. Balanced Diet. 2. Regular Exercise.\n\n"
    )

Lianmin Zheng's avatar
Lianmin Zheng committed
395
    forks = s.fork(2)
Lianmin Zheng's avatar
Lianmin Zheng committed
396
397
398
399
400
401
402
403
    for i, f in enumerate(forks):
        f += f"Now, expand tip {i+1} into a paragraph:\n"
        f += sgl.gen(f"detailed_tip", max_tokens=256, stop="\n\n")

    s += "Tip 1:" + forks[0]["detailed_tip"] + "\n"
    s += "Tip 2:" + forks[1]["detailed_tip"] + "\n"
    s += "In summary" + sgl.gen("summary")
```
Lianmin Zheng's avatar
Lianmin Zheng committed
404

Ying Sheng's avatar
Ying Sheng committed
405
#### Multi Modality
Lianmin Zheng's avatar
Lianmin Zheng committed
406
407
Use `sgl.image` to pass an image as input.

Lianmin Zheng's avatar
Lianmin Zheng committed
408
409
```python
@sgl.function
Lianmin Zheng's avatar
Lianmin Zheng committed
410
def image_qa(s, image_file, question):
Lianmin Zheng's avatar
Lianmin Zheng committed
411
    s += sgl.user(sgl.image(image_file) + question)
Lianmin Zheng's avatar
Lianmin Zheng committed
412
    s += sgl.assistant(sgl.gen("answer", max_tokens=256)
Lianmin Zheng's avatar
Lianmin Zheng committed
413
414
```

415
416
See also [srt_example_llava.py](examples/quick_start/srt_example_llava.py).

Ying Sheng's avatar
Ying Sheng committed
417
#### Constrained Decoding
418
419
Use `regex` to specify a regular expression as a decoding constraint.
This is only supported for local models.
Lianmin Zheng's avatar
Lianmin Zheng committed
420

Lianmin Zheng's avatar
Lianmin Zheng committed
421
```python
Lianmin Zheng's avatar
Lianmin Zheng committed
422
@sgl.function
Lianmin Zheng's avatar
Lianmin Zheng committed
423
424
def regular_expression_gen(s):
    s += "Q: What is the IP address of the Google DNS servers?\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
425
    s += "A: " + sgl.gen(
Lianmin Zheng's avatar
Lianmin Zheng committed
426
427
428
429
430
        "answer",
        temperature=0,
        regex=r"((25[0-5]|2[0-4]\d|[01]?\d\d?).){3}(25[0-5]|2[0-4]\d|[01]?\d\d?)",
    )
```
Lianmin Zheng's avatar
Lianmin Zheng committed
431

Ying Sheng's avatar
Ying Sheng committed
432
#### JSON Decoding
Lianmin Zheng's avatar
Lianmin Zheng committed
433
Use `regex` to specify a JSON schema with a regular expression.
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454

```python
character_regex = (
    r"""\{\n"""
    + r"""    "name": "[\w\d\s]{1,16}",\n"""
    + r"""    "house": "(Gryffindor|Slytherin|Ravenclaw|Hufflepuff)",\n"""
    + r"""    "blood status": "(Pure-blood|Half-blood|Muggle-born)",\n"""
    + r"""    "occupation": "(student|teacher|auror|ministry of magic|death eater|order of the phoenix)",\n"""
    + r"""    "wand": \{\n"""
    + r"""        "wood": "[\w\d\s]{1,16}",\n"""
    + r"""        "core": "[\w\d\s]{1,16}",\n"""
    + r"""        "length": [0-9]{1,2}\.[0-9]{0,2}\n"""
    + r"""    \},\n"""
    + r"""    "alive": "(Alive|Deceased)",\n"""
    + r"""    "patronus": "[\w\d\s]{1,16}",\n"""
    + r"""    "bogart": "[\w\d\s]{1,16}"\n"""
    + r"""\}"""
)

@sgl.function
def character_gen(s, name):
Lianmin Zheng's avatar
Lianmin Zheng committed
455
    s += name + " is a character in Harry Potter. Please fill in the following information about this character.\n"
456
457
458
    s += sgl.gen("json_output", max_tokens=256, regex=character_regex)
```

Lianmin Zheng's avatar
Lianmin Zheng committed
459
See also [json_decode.py](examples/usage/json_decode.py) for an additional example on specifying formats with Pydantic models.
460

Ying Sheng's avatar
Ying Sheng committed
461
#### Batching
Lianmin Zheng's avatar
Lianmin Zheng committed
462
463
Use `run_batch` to run a batch of requests with continuous batching.

Lianmin Zheng's avatar
Lianmin Zheng committed
464
465
466
467
468
469
470
471
472
473
474
475
```python
@sgl.function
def text_qa(s, question):
    s += "Q: " + question + "\n"
    s += "A:" + sgl.gen("answer", stop="\n")

states = text_qa.run_batch(
    [
        {"question": "What is the capital of the United Kingdom?"},
        {"question": "What is the capital of France?"},
        {"question": "What is the capital of Japan?"},
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
476
    progress_bar=True
Lianmin Zheng's avatar
Lianmin Zheng committed
477
478
)
```
Lianmin Zheng's avatar
Lianmin Zheng committed
479

Ying Sheng's avatar
Ying Sheng committed
480
#### Streaming
Lianmin Zheng's avatar
Lianmin Zheng committed
481
482
Add `stream=True` to enable streaming.

Lianmin Zheng's avatar
Lianmin Zheng committed
483
484
485
486
487
488
```python
@sgl.function
def text_qa(s, question):
    s += "Q: " + question + "\n"
    s += "A:" + sgl.gen("answer", stop="\n")

489
state = text_qa.run(
Lianmin Zheng's avatar
Lianmin Zheng committed
490
    question="What is the capital of France?",
Lianmin Zheng's avatar
Lianmin Zheng committed
491
492
493
    temperature=0.1,
    stream=True
)
Lianmin Zheng's avatar
Lianmin Zheng committed
494

Lianmin Zheng's avatar
Lianmin Zheng committed
495
496
497
for out in state.text_iter():
    print(out, end="", flush=True)
```
Lianmin Zheng's avatar
Lianmin Zheng committed
498

499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
#### Roles

Use `sgl.system``sgl.user` and `sgl.assistant` to set roles when using Chat models. You can also define more complex role prompts using begin and end tokens.

```python
@sgl.function
def chat_example(s):
    s += sgl.system("You are a helpful assistant.")
    # Same as: s += s.system("You are a helpful assistant.")

    with s.user():
        s += "Question: What is the capital of France?"

    s += sgl.assistant_begin()
    s += "Answer: " + sgl.gen(max_tokens=100, stop="\n")
    s += sgl.assistant_end()
```

Ying Sheng's avatar
Ying Sheng committed
517
#### Tips and Implementation Details
518
519
- The `choices` argument in `sgl.gen` is implemented by computing the [token-length normalized log probabilities](https://blog.eleuther.ai/multiple-choice-normalization/) of all choices and selecting the one with the highest probability.
- The `regex` argument in `sgl.gen` is implemented through autoregressive decoding with logit bias masking, according to the constraints set by the regex. It is compatible with `temperature=0` and `temperature != 0`.
Lianmin Zheng's avatar
Lianmin Zheng committed
520

Lianmin Zheng's avatar
Lianmin Zheng committed
521

Ying Sheng's avatar
Ying Sheng committed
522
523
524
## Benchmark And Performance
![8b_throughput](https://lmsys.org/images/blog/sglang_llama3/8b_throughput.svg)
![70b_fp8_throughput](https://lmsys.org/images/blog/sglang_llama3/70b_fp8_throughput.svg)
Lianmin Zheng's avatar
Lianmin Zheng committed
525

Ying Sheng's avatar
Ying Sheng committed
526
Learn more at this [blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/).
Lianmin Zheng's avatar
Lianmin Zheng committed
527

Lianmin Zheng's avatar
Lianmin Zheng committed
528
## Roadmap
Ying Sheng's avatar
Ying Sheng committed
529
[Development Roadmap (2024 Q3)](https://github.com/sgl-project/sglang/issues/634)
Lianmin Zheng's avatar
Lianmin Zheng committed
530
531

## Citation And Acknowledgment
Ying Sheng's avatar
Ying Sheng committed
532
533
Please cite our paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.
We also learned from the design and reused code from the following projects: [Guidance](https://github.com/guidance-ai/guidance), [vLLM](https://github.com/vllm-project/vllm), [LightLLM](https://github.com/ModelTC/lightllm), [FlashInfer](https://github.com/flashinfer-ai/flashinfer), [Outlines](https://github.com/outlines-dev/outlines), and [LMQL](https://github.com/eth-sri/lmql).