Files
sglang/benchmark/latency_throughput/README.md
2024-06-27 11:37:49 -07:00

61 lines
1.8 KiB
Markdown

# Benchmark Latency and Throughput
## SGLang
### Launch server
```
python -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000
```
### Benchmark one batch
```
python3 bench_one.py
python3 bench_one.py --batch-size 64
```
### Benchmark online serving with many requests
```
python3 bench_serving.py --backend srt --port 30000 --tokenizer meta-llama/Llama-2-7b-chat-hf --num-prompt 1000 --request-rate 100 --input-len 1024 --output-len 256
```
### Benchmark online serving on the ShareGPT dataset
#### Download data
```
wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
```
#### Run ShareGPT
```
python3 bench_throughput.py --backend srt --port 30000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
```
## Other baselines
### vLLM
```
python3 -m vllm.entrypoints.api_server --model meta-llama/Llama-2-7b-chat-hf --tensor-parallel 1 --disable-log-requests --swap-space 16 --port 21000
```
```
# run synthetic
python3 bench_throughput.py --backend vllm --port 30000 --tokenizer meta-llama/Llama-2-7b-chat-hf --num-prompt 1000 --request-rate 100 --input-len 1024 --output-len 256
```
```
# run ShareGPT
python3 bench_throughput.py --backend vllm --port 21000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
```
### LightLLM
```
python -m lightllm.server.api_server --model_dir ~/model_weights/Llama-2-7b-chat-hf --max_total_token_num 15600 --tokenizer_mode auto --port 22000
```
```
python3 bench_throughput.py --backend lightllm --port 22000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
```