How to reproduce the benchmark results of SGLang
Prerequisite
Install the latest SGLang
git clone https://github.com/sgl-project/sglang.git
cd sglang
pip install --upgrade pip
pip install -e "python[all]"
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
Set up HF_TOKEN
# Change the token to a real and usable one, with access permissions for the Llama 3 models.
export HF_TOKEN=hf_token
Launch the server
# Meta-Llama-3.1-8B-Instruct
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --enable-torch-compile --disable-radix-cache
# Meta-Llama-3.1-70B-Instruct
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-70B-Instruct --disable-radix-cache --tp 8
# Meta-Llama-3-70B-Instruct-FP8
python -m sglang.launch_server --model-path neuralmagic/Meta-Llama-3-70B-Instruct-FP8 --disable-radix-cache --tp 8
Benchmark
Offline benchmark
# Random dataset, Input [512, 1024], Output [512, 1024], num prompts 3k
python3 -m sglang.bench_serving --backend sglang --dataset-name random --num-prompts 3000 --random-input 1024 --random-output 1024 --random-range-ratio 0.5 --output-file sglang_offline_benchmark.jsonl
# Random dataset, Input [2048, 4096], Output [512, 1024], num prompts 3k
python3 -m sglang.bench_serving --backend sglang --dataset-name random --num-prompts 3000 --random-input 4096 --random-output 1024 --random-range-ratio 0.5 --output-file sglang_offline_benchmark.jsonl
# Random dataset, Input [512, 1024], Output [256, 512], num prompts 3k
python3 -m sglang.bench_serving --backend sglang --dataset-name random --num-prompts 3000 --random-input 1024 --random-output 512 --random-range-ratio 0.5 --output-file sglang_offline_benchmark.jsonl
# Random dataset, Input [2048, 4096], Output [256, 512], num prompts 3k
python3 -m sglang.bench_serving --backend sglang --dataset-name random --num-prompts 3000 --random-input 4096 --random-output 512 --random-range-ratio 0.5 --output-file sglang_offline_benchmark.jsonl
# ShareGPT dataset, num prompts 3k
python3 -m sglang.bench_serving --backend sglang --num-prompts 3000 --output-file sglang_offline_benchmark.jsonl
# get output token throughput
cat sglang_offline_benchmark.jsonl | cut -d':' -f12 | cut -d',' -f1
Online benchmark
# Random dataset, Input [512, 4096], Output [128, 1024], request rate 1, num prompts 300
python3 -m sglang.bench_serving --backend sglang --dataset-name random --random-input 4096 --random-output 1024 --random-range-ratio 0.125 --num-prompts 300 --request-rate 1 --output-file sglang_online_benchmark.jsonl
# Random dataset, Input [512, 4096], Output [128, 1024], request rate 2, num prompts 600
python3 -m sglang.bench_serving --backend sglang --dataset-name random --random-input 4096 --random-output 1024 --random-range-ratio 0.125 --num-prompts 600 --request-rate 2 --output-file sglang_online_benchmark.jsonl
# Random dataset, Input [512, 4096], Output [128, 1024], request rate 4, num prompts 1200
python3 -m sglang.bench_serving --backend sglang --dataset-name random --random-input 4096 --random-output 1024 --random-range-ratio 0.125 --num-prompts 1200 --request-rate 4 --output-file sglang_online_benchmark.jsonl
# Random dataset, Input [512, 4096], Output [128, 1024], request rate 8, num prompts 2400
python3 -m sglang.bench_serving --backend sglang --dataset-name random --random-input 4096 --random-output 1024 --random-range-ratio 0.125 --num-prompts 2400 --request-rate 8 --output-file sglang_online_benchmark.jsonl
# Random dataset, Input [512, 4096], Output [128, 1024], request rate 16, num prompts 3200
python3 -m sglang.bench_serving --backend sglang --dataset-name random --random-input 4096 --random-output 1024 --random-range-ratio 0.125 --num-prompts 3200 --request-rate 16 --output-file sglang_online_benchmark.jsonl
# get median e2e latency
cat sglang_online_benchmark.jsonl | cut -d':' -f9 | cut -d',' -f1
Other
We tried using vLLM 0.5.3.post1, but it often crashes under high loads, so we are using the older version, vLLM 0.5.2.
Preparation for TensorRT LLM can refer to https://github.com/sgl-project/tensorrt-demo. Specifically, we used a batch size of 512, a max input length of 8192, and a max number of tokens of 8192. The instance count for preprocessing and postprocessing in Triton Server is 16.