引言
LLaMA全称是Large Language Model Meta AI,是由Meta AI研究人员发布的一个预训练语言模型。与最近爆火的ChatGPT相比,LLaMA架构更小,但训练过程和单GPU推理速度更快,成本更低。今天在刷推特的时候无意中看到了这样一条消息,@ggerganov在GitHub上发布了llama.cpp,使用了4-bit量化将模型尽可能缩小,并能在多种移动设备上运行。这我就不淡定了,正好手里有个闲置的Raspberry Pi 4B 4GB版本,赶紧搞起。
Added a new section to my post highlighting the advances we've already seen in the last two days: LLaMA runs on a 4GB RaspberryPi and a Pixel 6 phone now! https://t.co/pWOv6PP85b
— Simon Willison (@simonw) March 13, 2023
配置树莓派
如果你已经配置好了树莓派,且gcc版本为10以上的就可以跳过这一部分了
以下是我配置树莓派的步骤
首先通过Raspberry Pi Imager刷入系统至SD卡 - 系统版本为Ubuntu Server 22.10 (64 bit) - 注意这里一定要使用64 bit的系统,不然后续可能无法编译
配置系统
直接用命令行升级软件包并安装相关依赖,这里给出参考命令
sudo su
apt-get update
apt-get upgrade
# 一些会用到的工具
apt-get install gcc g++ python3 python3-pip
# 安装python依赖
python3 -m pip install torch numpy sentencepiece
llama.cpp
首先可以先下载LLaMa模型,7B的模型大概是28GB左右,网速不好的同学可以提前开始下载,以免到时候还需要苦等。可以在网上搜索到泄漏的下载磁力链接,用你喜欢的任意P2P / Torrent软件下载即可。
接下来就可以进入llama.cpp按照里面的教程进行操作了。虽然看似比较简单,但里面还是有一些小坑的地方的。我在这里简单阐述下我遇到的问题以及我是怎么解决的。
构建二进制
# build this repo
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make
首先需要克隆仓库到本地,并进行make。我在这一步发现两个问题
- Rasbian Buster自带的gcc版本较老(gcc 8)且无法升级,因此才需要在前面重装系统并安装最新版本的。目前我使用的是gcc 12。
make
的时候提示不支持的选项。
makefile
里面处理一些平台特定的flag的时候是读取系统平台并储存到UNAME_S
,UNAME_P
和UNAME_M
里面,之后通过这三个参数判断编译器相关的选项。如果这里通过不了的话可以考虑看看makefile
里相近平台的参数并指定。make UNAME_P=armv7 UNAME_M=armv7
就可以指定为armv7
的编译选项。
- make的时候提示內联错误
‘always_inline’
`‘vdotq_s32’`
这个问题比较坑爹。我一开始以为是编译器的问题,换了好几种选项都不行。后来发现是作者在03.14针对Apple Sillicon做的一个优化 Use vdotq_s32 to improve performance #67 - Merged
。目前暂不清楚为什么树莓派无法处理这个,已经向作者反馈了。既然无法针对性的做处理,我尝试了直接revert对应的commit。之后就没有编译错误可以继续了。在我报出这个问题之后两小时,另一位贡献者@Ronsor
已经提出了一个PR去修复这个问题了 Don't use vdotq_s32 if it's not available #139
,点赞。
预处理模型
# obtain the original LLaMA model weights and place them in ./models
ls ./models
65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
# install Python dependencies
python3 -m pip install torch numpy sentencepiece
# convert the 7B model to ggml FP16 format
python3 convert-pth-to-ggml.py models/7B/ 1
# quantize the model to 4-bits
./quantize.sh 7B
我们将前面下载下来的模型放到llama.cpp/models
文件夹,主要包含7B
模型文件夹和tokenizer.model
分词器模型。然后使用convert-pth-to-ggml.py
进行预处理转换成FP16
精度,最后使用./quantize.sh
脚本进行4 bit
量化以进一步缩小。
这一步主要遇到的这么两个问题
- 使用
scp从Mac
上传文件到pi
上,稍微配置了一会儿,主要是等待时间较长。 - 在
pi
上运行convert-pth-to-ggml.py
这一步的时候,消耗内存太大OOM进程直接被系统kill掉了。
简单看了下convert-pth-to-ggml.py
,似乎都是做一些浮点精度的转换,最后生成的也是通用的模型格式ggml
。于是我决定尝试先用Mac做ggml转换,然后拷贝到Pi上作进一步的处理ggml-model-f16.bin
。实操发现这样是可行的,Pi也可以成功的运行quantize.sh
量化脚本。
使用llama.cpp
到这里我们的安装就已经结束了,紧张又兴奋的使用时间开始了。先来跑一个简单的Hello World。
./main -m ./models/7B/ggml-model-q4_0.bin -p "Hello world!" -t 8 -n 512
main: seed = 167882008main: seed = 167882008main: seed = 1678820083
llama_model_load: loading model from './models/7B/ggml-model-q4_0.bin' - please wait ...
llama_model_load: n_vocab = 32000
llama_model_load: n_ctx = 512
llama_model_load: n_embd = 4096
llama_model_load: n_mult = 256
llama_model_load: n_head = 32
llama_model_load: n_layer = 32
llama_model_load: n_rot = 128
llama_model_load: f16 = 2
llama_model_load: n_ff = 11008
llama_model_load: n_parts = 1
llama_model_load: ggml ctx size = 4529.34 MB
llama_model_load: memory_size = 512.00 MB, n_mem = 16384
llama_model_load: loading model part 1/1 from './models/7B/ggml-model-q4_0.bin'
llama_model_load: .................................... done
llama_model_load: model size = 4017.27 MB / num tensors = 291
system_info: n_threads = 4 / 4 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | VSX = 0 |
main: prompt: 'Hello world'
main: number of tokens in prompt = 3
1 -> ''
10994 -> 'Hello'
3186 -> ' world'
sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000, repeat_last_n = 64, repeat_penalty = 1.300000
Hello world! | Welcome to the new site.
We've been working hard on a few updates around here and it feels really great to finally share this
main: mem per token = 14368644 bytes
main: load time = 113492.53 ms
main: sample time = 987.31 ms
main: predict time = 3881120.50 ms / 121285.02 ms per token
main: total time = 4215260.00 ms
我们可以看到,生成了这句30词的回复,Pi用了1小时10分钟,生成速度约为121 s/Token。有可能是我编译出了问题,这个后续再修复一下吧。当然由于时间太慢了,我后面的几个用例都直接在Mac上运行了。
再跑一个官方例子
./main -m ./models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -t 8 -n 512
Building a website can be done in 10 simple steps:
Start with the end goal. Think about what your audience wants to achieve before they ever visit you site – and once there, how do these users interact on it? This is an important step as it will help formulate all of those great ideas into something that makes sense for their target marketplace!
2) Get Inspired & Develop a Mood Board: It's always good to have some kinda inspiration board. If you’re thinking about starting from scratch, use Pinterest or other sources (like Behance and Dribbble), as they will give your ideas a direction when it comes time for the visual design phase!
3) Strategize Your Content: Here's where things start to get tricky. You need some kinda content strategy in place here – do you have all of that information ready at hand? Now is also a good time as any to think about wireframes and how users will interact with your site once its built!
4) Do It With A Purpose: Are there particular goals for the user’s journey on this website, or simply trying to build an online presence? If so you may want to consider using marketing automation software like Hubspot that allows marketers and salespeople to work together better. This will help them organize your content into a more logical hierarchy – which is one of those things we love doing!
5) Visual Design: It's time for the fun part, where you think about colors (orange!) layout & typography, etcetera - and then start to design it using Illustrator or Photoshop. At this point in your project its important that all parties involved have a clear understanding of what needs to be done!
6) User Interface Design: This is when the UI designer comes into play – they should know how things interact with one another, understand user goals and create an experience so seamless it feels like second nature. The better this step goes down, you will end up having a much more engaging website at hand!
7) Coding Starts: After all of that planning is done here’s where the coding process begins – usually with HTML & CSS first and then followed by Javascript to make sure your site works flawlessly across multiple browsers.
8 ) Test, Test Again And Then Test Some More…: This phase can go on for a while (we know this from experience
main: mem per token = 14434244 bytes
main: load time = 4570.61 ms
main: sample time = 448.29 ms
main: predict time = 138987.47 ms / 271.99 ms per token
main: total time = 146495.30 ms
在Mac上运行的速度还是非常快的(废话),生成512个词的回复用了2分26秒,平均0.27 s/Token
。
此外,仓库里还包含了一个对于android设备的编译选项和实际运行的视频,有兴趣的读者可以自己研究下。也有人尝试在三星Galaxy S22 Ultra上运行这个模型,速度竟然可以达到1.2s/Token
。
小结
目前整个项目还是处于一个比较初级的形态,在RPi上生成速度相当慢,几乎不可用。此外,作者也在README中提到了一些限制
- 目前很难判断量化是否影响了模型生成的质量,需要一些更严谨的基准测试。
- 目前模型并没有使用MacOS提供的Accelerate框架,因为对于整个解码器中大部分张量形状来说,使用ARM_NEON的内部实现和使用Accelerate框架并没有什么性能差别。
除开性能之外,项目的指引还需要更加完善一些才能够更好的帮到大家。
但无论如何,在训练和运行语言模型逐渐要求海量算力的今天,能有这样一个模型让个人设备也能跑起来一个甚至还挺好用的模型,还是挺让人感动的。不知道是巧合还是什么,第一个跑通的老兄在Raspberry Pi 4 4GB #58里面展示的例句是:“The first man on the moon was Neil Armstrong..." 总要有人尝试,去发现,去做第一个登上月球的人,不是吗?
参考文献
如果你想进一步了解相关内容的话,可以阅读下面的文章
作者:Mestrace
文章来源:github
推荐阅读
- 编译入门那些事儿(6):LLVM Metadata System
- 聊聊 DiT 和 GenTron
- 阿里提出Mamba in Mamba | 比现有SOTA提速10倍,GPU使用减少73.4%
- 编译入门那些事儿(5):Jump Table 优化介绍
更多嵌入式AI干货请关注嵌入式AI专栏。欢迎添加极术小姐姐微信(id:aijishu20)加入技术交流群,请备注研究方向。