Llama-2-7B-Chat
本文档将介绍如何在高通平台上通过 Qualcomm® Genie 使用 NPU 硬件加速推理 Llama-2-7B-Chat 模型
模型细节
| 模型 | 量化方式 | 上下文长度 |
|---|---|---|
| Llama-2-7B-Chat | W4A16 | 4096 |
支持设备
提示
请参考 SoC 架构对照表 查寻当前设备 SoC 的 DSP 架构
-
此示例支持 v73 DSP 架构的高通平台 SoC
dsp_arch v73 -
运行设备
设备 SoC dsp_arch Fogwise® AIRbox Q900 QCS9075 v73
下载 qcom-qairt 依赖
- QCS6490
- QCS9075
Device
sudo apt install qcom-qnn-sdk-v68 qcom-genie-sdk-v68
Device
sudo apt install qcom-qnn-sdk-v73 qcom-genie-sdk-v73
导入环境变量
Device
export ADSP_LIBRARY_PATH=/usr/lib/aarch64-linux-gnu
下载模型
提示
请在 python 虚拟环境中安装 modelscope python 包,虚拟环境使用请参考 Python 虚拟环境使用
Device
pip3 install modelscope
modelscope download --model radxa/Llama-2-7B-Chat-w4a16-4096-v73 --local_dir ./Llama-2-7B-Chat-w4a16-4096-v73
推理模型
Device
cd Llama-2-7B-Chat-w4a16-4096-v73
构建 prompt
prompt 支持以文件形式或者参数形式传入
- prompt
- prompt_file
<s>[INST] <<SYS>>You are a helpful AI Assistant.<</SYS>>[/INST]</s><s>[INST]What is France's capital?[/INST]
Device
vim chat.txt
<s>[INST] <<SYS>>You are a helpful AI Assistant.<</SYS>>[/INST]</s><s>[INST]Introduce yourself.[/INST]
执行推理
- prompt
- prompt_file
Device
genie-t2t-run -c Llama2-7B-Chat-htp.json -p "<s>[INST] <<SYS>>You are a helpful AI Assistant.<</SYS>>[/INST]</s><s>[INST]Introduce yourself.[/INST]"
Device
genie-t2t-run -c Llama2-7B-Chat-htp.json --prompt_file chat.txt
(.venv) rock@radxa-airbox-q900:/mnt/ssd/qualcomm/Llama-2-7B-Chat$ genie-t2t-run -c Llama2-7B-Chat-htp.json -p "<s>[INST] <<SYS>>You are a helpful AI Assistant.<</SYS>>[/INST]</s><s>[INST]Introduce yourself.[/INST]" --profile profile2.txt
Using libGenie.so version 1.14.0
/prj/qct/webtech_scratch20/mlg_user_admin/qaisw_source_repo/rel/qairt-2.42.0/release/snpe_src/avante-tools/prebuilt/dsp/hexagon-sdk-5.5.5/ipc/fastrpc/rpcmem/src/rpcmem_android.c:38:dummy call to rpcmem_init, rpcmem APIs will be used from libxdsprpc
[INFO] "Using create From Binary"
[INFO] "Allocated total size = 361500672 across 8 buffers"
[PROMPT]: <s>[INST] <<SYS>>You are a helpful AI Assistant.<</SYS>>[/INST]</s><s>[INST]Introduce yourself.[/INST]
[BEGIN]: Hello! My name is LLaMA, I'm a large language model trained by a team of researcher at Meta AI. My primary function is to assist and communicate with users in a helpful and informative manner. I can answer questions, provide information, and even generate creative content such as stories or dialogue. Is there something specific you would like to know or discuss? I'm here to help![END]
/prj/qct/webtech_scratch20/mlg_user_admin/qaisw_source_repo/rel/qairt-2.42.0/release/snpe_src/avante-tools/prebuilt/dsp/hexagon-sdk-5.5.5/ipc/fastrpc/rpcmem/src/rpcmem_android.c:42:dummy call to rpcmem_deinit, rpcmem APIs will be used from libxdsprpc
性能参考
可以使用 --profile 选项开启性能分析功能
genie-t2t-run -c Llama2-7B-Chat-htp.json --prompt_file chat.txt --profile profile.txt
| Fogwise® AIRbox Q900 | |
|---|---|
| GenieDialog_create | 3,305,162 us |
| num-prompt-tokens | 38 |
| prompt-processing-rate | 20.969196319580078 toks/sec |
| time-to-first-token | 1,812,196 us |
| num-generated-tokens | 88 |
| token-generation-rate | 11.676922798156738, toks/sec |
| token-generation-time | 7,536,277 us |
| GenieDialog_free | 217,879 us |
指标含义解析
| 指标 | 含义解释 |
|---|---|
| GenieDialog_create | 初始化一个会话对象的时间。包括模型加载、上下文准备、内存分配等。 |
| num-prompt-tokens | 本次输入给模型的 prompt(提示词)的 token 数量,也就是模型接收到的文本拆分成的最小单元数量。 |
| prompt-processing-rate | 模型处理输入 prompt 的速度,单位为 token 每秒(toks/sec),表示模型分析 prompt 并准备生成输出的效率。 |
| time-to-first-token | 从开始处理到生成第一个输出 token 所花的时间,反映模型响应的延迟。 |
| num-generated-tokens | 模型在本次生成中实际输出的 token 数量,也就是模型生成的文本长度(以 token 为单位)。 |
| token-generation-rate | 模型生成 token 的速度,单位为 token 每秒(toks/sec),反映生成效率。 |
| token-generation-time | 模型生成所有输出 token 总共花费的时间,单位通常为微秒(us)。 |
| GenieDialog_free | 释放会话对象的时间,包括释放内存和清理资源。 |
Genie 官方文档
如果想深入了解 Qualcomm® Genie 的使用方法与详细 API 请参考