flashinfer is a kernel library for LLM serving. It can be used in SGLang runtime to accelerate attention computation.
You can install flashinfer via pip as follows for CUDA 12.1.
pip install flashinfer -i https://flashinfer.ai/whl/cu121/You can look for other CUDA versions in https://github.com/flashinfer-ai/flashinfer?tab=readme-ov-file#installation. If there is no desire version for your environment, please build it from source (the compilation takes a long time).
Add --model-mode flashinfer argument to enable flashinfer when launching a server.
Example:
python -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000 --model-mode flashinfer