README.md 15.9 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
4
5
6
</div>

--------------------------------------------------------------------------------

Ying Sheng's avatar
Ying Sheng committed
7
| [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/) | [**Paper**](https://arxiv.org/abs/2312.07104) |
Lianmin Zheng's avatar
Lianmin Zheng committed
8

Ying Sheng's avatar
Ying Sheng committed
9
10
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
11

12
The core features include:
Ying Sheng's avatar
Ying Sheng committed
13
- **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
14
- **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
15

Ying Sheng's avatar
Ying Sheng committed
16
## News
Ying Sheng's avatar
Ying Sheng committed
17
18
19
- [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
20

Ying Sheng's avatar
Ying Sheng committed
21
22
23
<details>
<summary>More</summary>

Ying Sheng's avatar
Ying Sheng committed
24
- [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
25
26
27
28
- [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
29
30
31
## Contents
- [Install](#install)
- [Backend: SGLang Runtime (SRT)](#backend-sglang-runtime-srt)
Ying Sheng's avatar
Ying Sheng committed
32
- [Frontend: Structured Generation Language (SGLang)](#frontend-structured-generation-language-sglang)
Lianmin Zheng's avatar
Lianmin Zheng committed
33
34
35
36
37
38
- [Benchmark And Performance](#benchmark-and-performance)
- [Roadmap](#roadmap)
- [Citation And Acknowledgment](#citation-and-acknowledgment)

## Install

Lianmin Zheng's avatar
Lianmin Zheng committed
39
40
### Method 1: With pip
```
41
pip install --upgrade pip
Lianmin Zheng's avatar
Lianmin Zheng committed
42
pip install "sglang[all]"
Lianmin Zheng's avatar
Lianmin Zheng committed
43

Lianmin Zheng's avatar
Lianmin Zheng committed
44
45
46
# Install FlashInfer CUDA kernels
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
```
47

Lianmin Zheng's avatar
Lianmin Zheng committed
48
### Method 2: From source
Lianmin Zheng's avatar
Lianmin Zheng committed
49
```
50
git clone https://github.com/sgl-project/sglang.git
Lianmin Zheng's avatar
Lianmin Zheng committed
51
52
cd sglang

Lianmin Zheng's avatar
Lianmin Zheng committed
53
pip install --upgrade pip
Lianmin Zheng's avatar
Lianmin Zheng committed
54
55
pip install -e "python[all]"

Lianmin Zheng's avatar
Lianmin Zheng committed
56
57
58
# Install FlashInfer CUDA kernels
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
```
59

Lianmin Zheng's avatar
Lianmin Zheng committed
60
### Method 3: Using docker
Ying Sheng's avatar
Ying Sheng committed
61
The docker images are available on Docker Hub as [lmsysorg/sglang](https://hub.docker.com/r/lmsysorg/sglang/tags), built from [Dockerfile](docker).
Ying Sheng's avatar
Ying Sheng committed
62
Repalce `<secret>` below with your huggingface hub [token](https://huggingface.co/docs/hub/en/security-tokens).
Ying Sheng's avatar
Ying Sheng committed
63

Liangsheng Yin's avatar
Liangsheng Yin committed
64
65
66
67
68
69
70
```bash
docker run --gpus all \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
Ying Sheng's avatar
Ying Sheng committed
71
    python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --host 0.0.0.0 --port 30000
Liangsheng Yin's avatar
Liangsheng Yin committed
72
73
```

Lianmin Zheng's avatar
Lianmin Zheng committed
74
### Common Notes
Lianmin Zheng's avatar
Lianmin Zheng committed
75
76
- If you cannot install FlashInfer, check out its [installation](https://docs.flashinfer.ai/installation.html#) page. If you still cannot install it, you can use the slower Triton kernels by adding `--disable-flashinfer` when launching the server.
- 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
77

Ying Sheng's avatar
Ying Sheng committed
78
79
80
## Backend: SGLang Runtime (SRT)
The SGLang Runtime (SRT) is an efficient serving engine.

Ying Sheng's avatar
Ying Sheng committed
81
### Quick Start
Ying Sheng's avatar
Ying Sheng committed
82
83
84
85
86
87
88
89
90
91
92
93
94
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
128
129
130
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 format [here](docs/sampling_params.md).

### 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
131
It supports streaming, vision, and most features of the Chat/Completions/Models endpoints specified by the [OpenAI API Reference](https://platform.openai.com/docs/api-reference/).
Ying Sheng's avatar
Ying Sheng committed
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146

### Additional Server Arguments
- Add `--tp 2` to enable tensor parallelism. If it indicates `peer access is not supported between these two devices`, add `--enable-p2p-check` option.
```
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --tp 2
```
- Add `--dp 2` to enable data parallelism. It can also be used together with tp. Data parallelism is better for throughput if there is enough memory.
```
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000 --dp 2 --tp 2
```
- If you see out-of-memory errors during serving, please 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 --port 30000 --mem-fraction-static 0.7
```
- See [hyperparameter_tuning.md](docs/hyperparameter_tuning.md) on tuning hyperparameters for better performance.
147
- 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
148
149
```
# Node 0
150
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
151
152

# Node 1
153
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
154
155
```
- If the model does not have a template in the Hugging Face tokenizer, you can specify a [custom chat template](docs/custom_chat_template.md).
156
- To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
157
- To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
Ying Sheng's avatar
Ying Sheng committed
158
159
160

### Supported Models

161
- Llama / Llama 2 / Llama 3 / Llama 3.1
Ying Sheng's avatar
Ying Sheng committed
162
163
164
165
- Mistral / Mixtral
- Gemma / Gemma 2
- Qwen / Qwen 2 / Qwen 2 MoE
- LLaVA 1.5 / 1.6
Ying Sheng's avatar
Ying Sheng committed
166
167
168
  - `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
169
- LLaVA-NeXT-Video
Ying Sheng's avatar
Ying Sheng committed
170
  - see [examples/usage/llava_video](examples/usage/llava_video)
Ying Sheng's avatar
Ying Sheng committed
171
172
173
174
175
176
177
178
- Yi-VL
  - see [srt_example_yi_vl.py](examples/quick_start/srt_example_yi_vl.py).
- StableLM
- Command-R
- DBRX
- Grok
- ChatGLM
- InternLM 2
zhyncs's avatar
zhyncs committed
179
- Mistral NeMo
Ying Sheng's avatar
Ying Sheng committed
180
181
182

Instructions for supporting a new model are [here](https://github.com/sgl-project/sglang/blob/main/docs/model_support.md).

Ying Sheng's avatar
Ying Sheng committed
183
184
185
186
187
188
189
190
191
192
193
### Benchmark Performance

- Benchmark a single static batch. Run the following command without launching a server. The arguments are the same as those for `launch_server.py`.
  ```
  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
  ```

Ying Sheng's avatar
Ying Sheng committed
194
195
196
197
## Frontend: Structured Generation Language (SGLang)
The frontend language can be used with local models or API models.

### Quick Start
Lianmin Zheng's avatar
Lianmin Zheng committed
198
199
The example below shows how to use sglang to answer a mulit-turn question.

Ying Sheng's avatar
Ying Sheng committed
200
#### Using Local Models
201
First, launch a server with
Lianmin Zheng's avatar
Lianmin Zheng committed
202
```
Ying Sheng's avatar
Ying Sheng committed
203
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000
Lianmin Zheng's avatar
Lianmin Zheng committed
204
205
```

206
207
Then, connect to the server and answer a multi-turn question.

Lianmin Zheng's avatar
Lianmin Zheng committed
208
```python
209
from sglang import function, system, user, assistant, gen, set_default_backend, RuntimeEndpoint
Lianmin Zheng's avatar
Lianmin Zheng committed
210
211
212
213
214
215
216
217
218

@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))

219
set_default_backend(RuntimeEndpoint("http://localhost:30000"))
Lianmin Zheng's avatar
Lianmin Zheng committed
220
221
222
223
224
225
226
227

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"])
228
229

print(state["answer_1"])
Lianmin Zheng's avatar
Lianmin Zheng committed
230
231
```

Ying Sheng's avatar
Ying Sheng committed
232
#### Using OpenAI Models
233
Set the OpenAI API Key
Lianmin Zheng's avatar
Lianmin Zheng committed
234
```
235
export OPENAI_API_KEY=sk-******
Lianmin Zheng's avatar
Lianmin Zheng committed
236
237
```

238
Then, answer a multi-turn question.
Lianmin Zheng's avatar
Lianmin Zheng committed
239
```python
240
from sglang import function, system, user, assistant, gen, set_default_backend, OpenAI
Lianmin Zheng's avatar
Lianmin Zheng committed
241
242
243
244
245
246
247
248
249

@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))

250
set_default_backend(OpenAI("gpt-3.5-turbo"))
Lianmin Zheng's avatar
Lianmin Zheng committed
251
252
253
254
255
256
257
258

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"])
259
260

print(state["answer_1"])
Lianmin Zheng's avatar
Lianmin Zheng committed
261
262
```

Ying Sheng's avatar
Ying Sheng committed
263
#### More Examples
Lianmin Zheng's avatar
Lianmin Zheng committed
264

265
Anthropic and VertexAI (Gemini) models are also supported.
Lianmin Zheng's avatar
Lianmin Zheng committed
266
267
You can find more examples at [examples/quick_start](examples/quick_start).

Ying Sheng's avatar
Ying Sheng committed
268
### Language Feature
Lianmin Zheng's avatar
Lianmin Zheng committed
269
270
271
272
273
To begin with, import sglang.
```python
import sglang as sgl
```

Lianmin Zheng's avatar
Lianmin Zheng committed
274
`sglang` provides some simple primitives such as `gen`, `select`, `fork`, `image`.
Lianmin Zheng's avatar
Lianmin Zheng committed
275
276
You can implement your prompt flow in a function decorated by `sgl.function`.
You can then invoke the function with `run` or `run_batch`.
277
The system will manage the state, chat template, parallelism and batching for you.
Lianmin Zheng's avatar
Lianmin Zheng committed
278

279
280
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
281
#### Control Flow
Lianmin Zheng's avatar
Lianmin Zheng committed
282
283
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
284
285
```python
@sgl.function
286
287
288
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
289
290
291

    if s["tool"] == "calculator":
        s += "The math expression is" + sgl.gen("expression")
292
293
    elif s["tool"] == "search engine":
        s += "The key word to search is" + sgl.gen("word")
Lianmin Zheng's avatar
Lianmin Zheng committed
294
```
Lianmin Zheng's avatar
Lianmin Zheng committed
295

Ying Sheng's avatar
Ying Sheng committed
296
#### Parallelism
Lianmin Zheng's avatar
Lianmin Zheng committed
297
298
299
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
300
301
302
303
304
305
306
307
```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
308
    forks = s.fork(2)
Lianmin Zheng's avatar
Lianmin Zheng committed
309
310
311
312
313
314
315
316
    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
317

Ying Sheng's avatar
Ying Sheng committed
318
#### Multi Modality
Lianmin Zheng's avatar
Lianmin Zheng committed
319
320
Use `sgl.image` to pass an image as input.

Lianmin Zheng's avatar
Lianmin Zheng committed
321
322
```python
@sgl.function
Lianmin Zheng's avatar
Lianmin Zheng committed
323
def image_qa(s, image_file, question):
Lianmin Zheng's avatar
Lianmin Zheng committed
324
    s += sgl.user(sgl.image(image_file) + question)
Lianmin Zheng's avatar
Lianmin Zheng committed
325
    s += sgl.assistant(sgl.gen("answer", max_tokens=256)
Lianmin Zheng's avatar
Lianmin Zheng committed
326
327
```

328
329
See also [srt_example_llava.py](examples/quick_start/srt_example_llava.py).

Ying Sheng's avatar
Ying Sheng committed
330
#### Constrained Decoding
331
332
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
333

Lianmin Zheng's avatar
Lianmin Zheng committed
334
```python
Lianmin Zheng's avatar
Lianmin Zheng committed
335
@sgl.function
Lianmin Zheng's avatar
Lianmin Zheng committed
336
337
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
338
    s += "A: " + sgl.gen(
Lianmin Zheng's avatar
Lianmin Zheng committed
339
340
341
342
343
        "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
344

Ying Sheng's avatar
Ying Sheng committed
345
#### JSON Decoding
Lianmin Zheng's avatar
Lianmin Zheng committed
346
Use `regex` to specify a JSON schema with a regular expression.
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367

```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
368
    s += name + " is a character in Harry Potter. Please fill in the following information about this character.\n"
369
370
371
    s += sgl.gen("json_output", max_tokens=256, regex=character_regex)
```

Lianmin Zheng's avatar
Lianmin Zheng committed
372
See also [json_decode.py](examples/usage/json_decode.py) for an additional example on specifying formats with Pydantic models.
373

Ying Sheng's avatar
Ying Sheng committed
374
#### Batching
Lianmin Zheng's avatar
Lianmin Zheng committed
375
376
Use `run_batch` to run a batch of requests with continuous batching.

Lianmin Zheng's avatar
Lianmin Zheng committed
377
378
379
380
381
382
383
384
385
386
387
388
```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
389
    progress_bar=True
Lianmin Zheng's avatar
Lianmin Zheng committed
390
391
)
```
Lianmin Zheng's avatar
Lianmin Zheng committed
392

Ying Sheng's avatar
Ying Sheng committed
393
#### Streaming
Lianmin Zheng's avatar
Lianmin Zheng committed
394
395
Add `stream=True` to enable streaming.

Lianmin Zheng's avatar
Lianmin Zheng committed
396
397
398
399
400
401
```python
@sgl.function
def text_qa(s, question):
    s += "Q: " + question + "\n"
    s += "A:" + sgl.gen("answer", stop="\n")

402
state = text_qa.run(
Lianmin Zheng's avatar
Lianmin Zheng committed
403
    question="What is the capital of France?",
Lianmin Zheng's avatar
Lianmin Zheng committed
404
405
406
    temperature=0.1,
    stream=True
)
Lianmin Zheng's avatar
Lianmin Zheng committed
407

Lianmin Zheng's avatar
Lianmin Zheng committed
408
409
410
for out in state.text_iter():
    print(out, end="", flush=True)
```
Lianmin Zheng's avatar
Lianmin Zheng committed
411

Ying Sheng's avatar
Ying Sheng committed
412
#### Tips and Implementation Details
413
414
- 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
415

Lianmin Zheng's avatar
Lianmin Zheng committed
416

Ying Sheng's avatar
Ying Sheng committed
417
418
419
## 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
420

Ying Sheng's avatar
Ying Sheng committed
421
Learn more at this [blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/).
Lianmin Zheng's avatar
Lianmin Zheng committed
422

Lianmin Zheng's avatar
Lianmin Zheng committed
423
## Roadmap
Ying Sheng's avatar
Ying Sheng committed
424
[Development Roadmap (2024 Q3)](https://github.com/sgl-project/sglang/issues/634)
Lianmin Zheng's avatar
Lianmin Zheng committed
425
426

## Citation And Acknowledgment
Ying Sheng's avatar
Ying Sheng committed
427
428
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).