24 lines
610 B
Markdown
24 lines
610 B
Markdown
## Flashinfer Mode
|
|
|
|
[flashinfer](https://github.com/flashinfer-ai/flashinfer) is a kernel library for LLM serving.
|
|
It can be used in SGLang runtime to accelerate attention computation.
|
|
|
|
### Install flashinfer
|
|
|
|
Note: The compilation can take a very long time.
|
|
|
|
```bash
|
|
git submodule update --init --recursive
|
|
pip install 3rdparty/flashinfer/python
|
|
```
|
|
|
|
### Run a Server With Flashinfer Mode
|
|
|
|
Add `--model-mode flashinfer` argument to enable flashinfer when launching a server.
|
|
|
|
Example:
|
|
|
|
```bash
|
|
python -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000 --model-mode flashinfer
|
|
```
|