llm_inference_engine_benchmark

This repository aims to evaluate various open-source inference frameworks, analyzing their strengths and weaknesses.

Thanks for the res: pandada8/llm-inference-benchmark: LLM 推理服务性能测试 (github.com) https://github.com/ninehills/llm-inference-benchmark

Information

• Hardware: Nvidia H800
• LLM model: llama2
• llm inference engine: llama.cpp, vllm, fastllm
• 指标:
  • Thoughtput
  • none-first Token Latency
  • First Token Latency

Run

llama.cpp

• In the llama.cpp directory, run the llama.cpp server.

~/code/llama.cpp$ CUDA_VISIBLE_DEVICES=0 ./server --n-gpu-layers 999 --host 127.0.0.1 --port 8082 -m models/llama-2-7b/llama-2-7b-7B-F16.gguf

--n-gpu-layers the number of layer load to GPU. Set 999 making all layers load to GPU.

--host Set the hostname or ip address to listen. Default 127.0.0.1.

--port Set the port to listen. Default: 8080

-m The model path.

more parameters and detail can be learned from: llama.cpp and llama.cpp-server and other literature.

• In the llm_inference_engine_benchmark directory, run the test code for llama.cpp.

python benchmark.py --model llama --backend llama.cpp --endpoint http://127.0.0.1:8082

## draw the result
python draw.py

--endpoint The endpoint for the server.

--model llama meanings the family of llama.

vllm

• run the vllm serve

python -m vllm.entrypoints.openai.api_server  --model model_executor/models/Llama-2-7b-hf/ --host 127.0.0.1 --port 8082

• run the test code for vllm.

python benchmark.py --model llama --backend vllm --endpoint http://127.0.0.1:8082

fastllm

• run the fastllm serve

~/code/fastllm/build$ ./webui -p ../mode_zoo/llama2-7b.flm --port 8082

• run the test code for fastllm.

python benchmark.py --model llama --backend fastllm --endpoint http://127.0.0.1:8082

Result

Thoughput

Averge Token Latency

First Token Latency