RetroSearch Browse

Home - News ( United States | United Kingdom | Italy | Germany ) - Football scores

Showing content from https://github.com/4paradigm/llm-perf-bench below:

4paradigm/llm-perf-bench: This is a benchmarking tool to help evaluate the performance of a given LLM and hardware platform.

llm-perf-benchmark 是一个针对大语言模型多用户推理服务的性能测试工具。

编写接口函数query_model()

使用项目提供的模板model_query.py.template，并针对所测试的大语言模型进行修改

cp model_query.py.template model_query.py
vim model_query.py

执行python3 llm_benchmark.py启动测试

python3 llm_benchmark.py --name sagegpt_test --desc "Test SageGPT on 2x4PD AccXPU with TGI 1.0.3" --work-dir ./ --test-set ./resources/testset.txt

详细参数列表请参照：

usage: llm_benchmark.py [-h] [--name TEST_NAME] [--desc TEST_DESCRIPTION]
                        [--work-dir WORK_DIR] [--test-set TEST_SET_PATH]
                        [--test-duration TEST_DURATION]
                        [--test-runs NUM_USERS [NUM_USERS ...]]

options:
  -h, --help            show this help message and exit
  --name TEST_NAME      identifier of the benchmark
  --desc TEST_DESCRIPTION
                        description of this benchmark run
  --work-dir WORK_DIR   working directory
  --test-set TEST_SET_PATH
                        path of the testset for llm test
  --test-duration TEST_DURATION
                        length of each test run in seconds
  --test-runs NUM_USERS [NUM_USERS ...]
                        list of concurrent users for each test run

运行工具后将记录以下关键指标：

[并发]：当前测试的并发用户数。

[问答/错误]：对于已经发送的请求，记录当前累计收到的答复数以及累计收到的错误答复或者答复超时的数量。

[平均延迟]：记录当前系统的平均响应延迟。

[出字均速]：记录系统生成文本的平均速度，即每秒钟生成的字符数。

[总吞吐]：统计系统的总体性能，表示每秒钟能够生成的总字符数。
```
23:16:06 [并发] 1 [问答/错误] 3/0 [平均延迟] 0.1秒 [出字均速] 每秒56.4字 [总吞吐] 59.5 字/秒
```
对应输出的测试结果在$work_dir/output目录下

RetroSearch is an open source project built by @garambo | Open a GitHub Issue

Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo

HTML: 3.2 | Encoding: UTF-8 | Version: 0.7.4