關於 LLM 的數字

Hacker News Daily 上看到的文章,講 LLM 的各種數字 (大多都是費用):「Numbers every LLM developer should know (github.com/ray-project)」,原文在「Numbers every LLM Developer should know」這邊。

其中第一條就蠻重要的,如果你是用 API 依照 token 收費的話,叫 API 長話短說會省不少錢 XD

40-90: Amount saved by appending “Be Concise” to your prompt

第二條是給個感覺,換算 word 與 token,不過這邊講的應該是英文的:

1.3:1 -- Average tokens per word

後面也有蠻多數字的,都是讓你有個感覺。都讀過後就可以把 cheatsheet 留下來:

llama.cpp 開始支援 GPU 了

前陣子因為重灌桌機,所以在重建許多環境... 其中一個就是 llama.cpp,連到專案頁面上時意外發現這兩個新的 feature:

OpenBLAS support
cuBLAS and CLBlast support

這代表可以用 GPU 加速了,所以就照著說明試著編一個版本測試。

編好後就跑了 7B 的 model,看起來快不少,然後改跑 13B 的 model,也可以把完整 40 個 layer 都丟進 3060 (12GB 版本) 的 GPU 上:

./main -m models/13B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 -ngl 40

從 log 可以看到 40 layers 到都 GPU 上面,吃了 7.5GB 左右:

llama.cpp: loading model from models/13B/ggml-model-q4_0.bin
llama_model_load_internal: format     = ggjt v2 (latest)
llama_model_load_internal: n_vocab    = 32000
llama_model_load_internal: n_ctx      = 512
llama_model_load_internal: n_embd     = 5120
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 40
llama_model_load_internal: n_layer    = 40
llama_model_load_internal: n_rot      = 128
llama_model_load_internal: ftype      = 2 (mostly Q4_0)
llama_model_load_internal: n_ff       = 13824
llama_model_load_internal: n_parts    = 1
llama_model_load_internal: model size = 13B
llama_model_load_internal: ggml ctx size =  90.75 KB
llama_model_load_internal: mem required  = 9807.48 MB (+ 1608.00 MB per state)
llama_model_load_internal: [cublas] offloading 40 layers to GPU
llama_model_load_internal: [cublas] total VRAM used: 7562 MB
llama_init_from_file: kv self size  =  400.00 MB

30B 的 model 我也試著丟上去跑,但只能丟 28 layers 上去 (全部是 60 layers),再多 GPU 的記憶體就撐不住了。

但能用 GPU 算是一個很大的進展,現在這版只快了一半的時間,不知道後面還有沒有 tune 的空間...

透過 WebGPU 跑的 Web LLM

Simon Willison 這邊看到的玩法,透過 WebGPU 在瀏覽器上面直接跑 LLM 的 demo:「Web LLM runs the vicuna-7b Large Language Model entirely in your browser, and it’s very impressive」,專案在「Web LLM」這邊,可以直接玩。

不過要注意一下瀏覽器的支援度,如果是 Chrome 的話需要 113+,但目前 stable 還是 112;而 Firefox 的話我試過在 about:config 裡面用 dom.webgpu.enabled 打開 WebGPU 支援,但重開瀏覽器後還是跑不動?(也有可能是 Linux 環境的關係)

Update:應該是 Linux 環境的關係,我在 Linux 下用 dev channel (114) 也不行。

話說有 WebGPU 後是不是開始要擋 GPU 挖礦了...

目前可商用的 LLM

Ask Hacker News Weekly 上看到的討論,有人問了目前可商用的 LLM 有哪些:「Ask HN: Open source LLM for commercial use?」。

有人提到 GoogleFlan 應該是目前最能打的?在 Hugging Face 上可以下載到:

I've seen this question asked repeatedly in many LLaMa threads, currently the best models that are truly open are the released models from the Flan family by Google, which includes Flan-T5[0] and Flan-UL2[1]. According to its paper, Flan-UL2 performs slightly better than Flan-T5-XXL.

然後差不多是 GPT-3 的等級,離 GPT-3.5 或是演伸出來的 ChatGPT 都還有段距離。但如果針對特定情境下 tune 的話應該還是能用的:

These models perform slightly better than GPT-3 under some tasks[2], but they're still far from achieving the results from GPT-3.5 and GPT-4. This becomes evident when you try to use them in the real world; they're not "good enough" for general use cases, unlike ChatGPT models. However, if you can restrict your use case to one particular domain, you can achieve pretty good results by further fine-tuning these models.

另外一則回覆有提到一些其他的 model:

The ones I saw mentioned so far were Flan, Cerebras, GPT-J, and RWKV.

Not yet mentioned:

* Pythia https://github.com/EleutherAI/pythia

* GLM-130B https://github.com/THUDM/GLM-130B - see also ChatGLM-6B https://github.com/THUDM/ChatGLM-6B

* GPT-NeoX-20B https://huggingface.co/EleutherAI/gpt-neox-20b

* GeoV-9B https://github.com/geov-ai/geov

* BLOOM https://huggingface.co/bigscience/bloom and BLOOMZ https://huggingface.co/bigscience/bloomz

看起來如果有需要用的話是可以從這裡面挖看看...

GPT 的進程 (或是 LLM 的進程)

前幾天不知道在哪邊看到「Five years of GPT progress」這篇,裡面整理了這五年 GPT/LLM 的進程,算是回顧性質的文章,裡面當然有提到技術改善的地方 (像是參數大小,類神經網路層的架構差異),另外裡面都有原始論文或是資料的連結,然後作者也有描述一些當時的背景,對於要釐清歷史脈絡也蠻有幫助的。

GPTGPT-2GPT-3 這三個 OpenAI 的作品開始講,然後提到 GPT-3 帶出來的新紀元。

接著提到的是各家都開始進來參與的年代,Jurassic-1 (AI21 Labs)、Megatron-Turing NLG (Nvidia)、Gopher (DeepMind)、Chinchilla (DeepMind)、PaLM (Google AI)。

然後是 LLaMa (Facebook),第一個有參數夠大,而且效能夠好的 model,被放出來讓大家玩的 LLM。

最後又回到 OpenAI 的 GPT-4

這樣整理讀起來清晰不少,但要注意裡面的發展不是線性關係,彼此之間互相影響交錯在跑 (因為中間還是有很多其他的論文互相影響)。

llama.cpp 的載入速度加速

Hacker News 上看到「Llama.cpp 30B runs with only 6GB of RAM now (github.com/ggerganov)」這個消息,原 pull request 在「Make loading weights 10-100x faster #613」這邊。

這個 PR 的作者 Justine Tunney 在 PR 上有提到他改變 model 檔案格式,以便改用 mmap(),大幅降低了需要預先讀取的時間 (因為變成 lazy-loading style),而且這也讓系統可以利用 cache page,避免了 double buffering 的問題:

This was accomplished by changing the file format so we can mmap() weights directly into memory without having to read() or copy them thereby ensuring the kernel can make its file cache pages directly accessible to our inference processes; and secondly, that the file cache pages are much less likely to get evicted (which would force loads to hit disk) because they're no longer competing with memory pages that were needlessly created by gigabytes of standard i/o.

這讓我想到在資料庫領域中,PostgreSQL 也會用 mmap() 操作,有點類似的概念。

另外 Justine Tunney 在這邊的 comment 有提到一個意外觀察到的現象,他發現實際在計算的時候用到的 model 內容意外的少:他用一個簡單的 prompt 測試,發現 20GB 的 30B model 檔案在他的 Intel 機器上實際只用到了 1.6GB 左右:

If I run 30B on my Intel machine:

[...]

As we can see, 400k page faults happen, which means only 1.6 gigabytes ((411522 * 4096) / (1024 * 1024)) of the 20 gigabyte weights file actually needed to be used.

這點他還在懷疑是不是他的修改有 bug,但目前他覺得不太像,也看不太出來:

Now, since my change is so new, it's possible my theory is wrong and this is just a bug. I don't actually understand the inner workings of LLaMA 30B well enough to know why it's sparse. Maybe we made some kind of rare mistake where llama.cpp is somehow evaluating 30B as though it were the 7B model. Anything's possible, however I don't think it's likely. I was pretty careful in writing this change, to compare the deterministic output of the LLaMA model, before and after the Git commit occurred. I haven't however actually found the time to reconcile the output of LLaMA C++ with something like PyTorch. It'd be great if someone could help with that, and possibly help us know why, from more a data science (rather than systems engineering perspective) why 30B is sparse.

如果不是 bug 的話,這其實冒出了一個很有趣的訊號,表示這些 model 是有可能再瘦身的?

Alpaca.cpp 有 13B 與 30B 的 model 可以玩了

在「Stanford Alpaca 與 Alpaca.cpp」這邊有提到 Alpaca 有放出 7B 的 model 可以用,後來 community 也照著同樣的方式訓練出對應的 13B 與 30B 的 model 了,在 Alpaca.cpp 專案的 README.md 有寫怎麼取得。

看起來比較大的限制是記憶體,13B 與 30B 的記憶體建議是 >10GB RAM 與 >32GB RAM,不知道剛好 32GB RAM 跑不跑得動,先拿 32GB RAM 的機器跑看看,如果真的不行的話家裡桌機是 64GB RAM,應該是不會有大問題...

不過 13B 的 model 檔案位置是放在 IPFS 上,透過這幾個 gateway 下載起來都有點慢,有需要的人用 BitTorrent 可能會快一點?

Update:測了一下 13B 的版本:

> Explain FAANG.
FAANG refers to five of America's largest technology companies - Facebook, Apple Inc., Amazon, Netflix and Google (Alphabet). These are some of the most valuable brands in today’s marketplace with a combined value that exceeded $3 trillion as at 2018.

以及 30B 的版本:

> Explain FAANG.
FAANG stands for Facebook, Amazon, Apple, Netflix and Google - five of the most powerful technology companies in today's world. These tech giants have become increasingly influential over recent years due to their dominance in various markets such as social media platforms (Facebook), e-commerce websites (Amazon) or streaming services (Netflix).

Fake GitHub Star 的生意

昨天在 Hacker News 首頁上看到「Tracking the Fake GitHub Star Black Market (dagster.io)」這篇,原文在「Tracking the Fake GitHub Star Black Market with Dagster, dbt and BigQuery」這邊。

作者群想要偵測 GitHub 上面 fake star 的行為,所以就跑去找黑市買,然後找到了兩家,Baddhi Shop (1000 個 $64) 與 GitHub24 (每個 €0.85,大約是 $0.91),價錢差異很大,「品質」差異也很大:貴的 star 在一個月後還是存在,而便宜的看起來有一些有被 GitHub 偵測到而清除掉:

A month later, all 100 GitHub24 stars still stood, but only three-quarters of the fake Baddhi Shop stars remained. We suspect the rest were purged by GitHub’s integrity teams.

接下來就是想要系統化分析,切入點是 GH Archive 這個服務,可以直接下載 GitHub 全站上的 public evnets 資料:

GH Archive is a project to record the public GitHub timeline, archive it, and make it easily accessible for further analysis.

想要偵測兩種不同的 fake account,第一種是 obvious fake account,定義成這樣:

  • Created in 2022 or later
  • Followers <=1
  • Following <= 1
  • Public gists == 0
  • Public repos <=4
  • Email, hireable, bio, blog, and twitter username are empty
  • Star date == account creation date == account updated date

從定義就可以看出來完全就是灌水帳號,開出來就沒在動的。從 screenshot 可以看出這種帳號長的都一樣:

另外一種則是透過演算法去分析,這邊拿 unsupervised clustering 類的演算法分析出來的結果,可以看到抓到很多:

最近 NN 類的 machine learning 演算法太多,看到這些傳統的 machine learning 演算法還是覺得頗新鮮的...

Stanford Alpaca 與 Alpaca.cpp

先前提到的「玩最近 Facebook Research (Meta) 放出來的 LLaMA」,在 2/24 公佈後,過沒多久 Stanford CRFM 團隊就丟出 fine-tuned 過的版本:「Alpaca: A Strong, Replicable Instruction-Following Model」,對應的討論在「Alpaca: A strong open-source instruction-following model (stanford.edu)」。

這個 tuning 是基於 LLaMA-7B:

We are releasing our findings about an instruction-following language model, dubbed Alpaca, which is fine-tuned from Meta’s LLaMA 7B model. We train the Alpaca model on 52K instruction-following demonstrations generated in the style of self-instruct using text-davinci-003. On the self-instruct evaluation set, Alpaca shows many behaviors similar to OpenAI’s text-davinci-003, but is also surprisingly small and easy/cheap to reproduce.

後來就再被包成一個專案 (改自 llama.cpp),讓大家可以很方便的跑:「Alpaca.cpp」。

剛好可以比較,用 llama.cpp 跑的話是:

$ ./main -m ./models/7B/ggml-model-q4_0.bin -t 12 -n 128 -p 'Explain FAANG. '

Explain FAANG. 🙂
The acronym “FAAN” is used for the five tech giants (Facebook, Apple, Amazon and Netflix) that are
 taking over our lives:
15 November 2018 / by Amaraksha S [end of text]

而用 alpaca.cpp 跑的話回答品質好不少:

> Explain FAANG.
FAANG (or FANNG) refers to five technology companies, Facebook Inc., Apple Inc., Amazon.com Inc., Netflix Inc and Alphabet's Google division that have become some of the most influential in recent years due to their rapid growth rates as well as their market capitalization values: 
Facebook (FB), $596 billion; Apple (AAPL) ,$1,084.72 billion; Amazon (AMZN), $394.5 billion ; Netflix( NFLX ), $154.8 billion and Google's parent company Alphabet Inc., ($GOOG,$GOOGL).
These companies have become some of the most influential in recent years due to their rapid growth rates as well as their market capitalization values:  Facebook (FB),$307.15 billion; Apple(AAPL) , $924.86 billion ; Amazon (AMZN ), $783.

可以看到 instruction-following 後的差異,很像是 GPT-3.5 API 與 ChatGPT API 的輸出結果差異。

而這只是 LLaMA-7B 的成果,好像可以期待有人對 LLaMA-65B 做些事情?

玩最近 Facebook Research (Meta) 放出來的 LLaMA

很多地方應該都有提到 Facebook Research (Meta) 放出來的 LLaMA 了,對應的論文是「LLaMA: Open and Efficient Foundation Language Models」這篇,但這邊論文提到的 open 並不是一般常見的 open 定義,而只是常見的行銷詞彙而已,實際上只是 free for charging with constraints。

另外要注意 LLaMA 是個 LLM 而已,跟 ChatGPT 不算是同樣性質的東西,能對比應該是 GPT-3 (或是 GPT-3.5)。

主要是 ChatGPT 多了 SLRL 的步驟,而產出來的東西更接近商業化產品要的結果。

LLaMA 的特點在於效能不錯,可以用 LLaMA-13B 打贏 GPT-3 (175B),另外這次訓練出來最大的 LLaMA-65B 則可以站上第一梯隊 (與 DeepMindChinchilla-70BGoogle ResearchPaLM-540B):

LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla-70B and PaLM-540B.

但跟以前差異最大的是,這次 Facebook Research 決定把訓練完後的 model 檔案放出來,所以就有了後續很多的進展:

We release all our models to the research community.

首先一開始 Facebook Research 要求使用者填表單才提供下載 (2/24 的時候),但三月初的時候 GitHub 上有人直接把 BitTorrentmagnet 連結附上去,送了一個 pull request:「Save bandwidth by using a torrent to distribute more efficiently #73」,所以你就有「方法」可以取得 model 檔案,但還是可以注意一下使用限制:

To maintain integrity and prevent misuse, we are releasing our model under a noncommercial license focused on research use cases. Access to the model will be granted on a case-by-case basis to academic researchers; those affiliated with organizations in government, civil society, and academia; and industry research laboratories around the world. People interested in applying for access can find the link to the application in our research paper.

除了可以透過 BitTorrent 下載外,comment 裡面也有 IPFS 的連結可以下載。

有需要下載這包檔案的人要注意檔案很大,大約 240GB,其中 65B model 佔了 128GB 左右。

被放出來以後就開始有很多人在上面包起來用,其中目前比較完整的應該是「Running LLaMA 7B and 13B on a 64GB M2 MacBook Pro with llama.cpp」這篇提到的方法,裡面提到的 ggerganov/llama.cpp 專案同時支援了 WindowsmacOSLinux,而且可以用 CPU 跑,速度也不慢。

試著用 llama.cpp 跑,65B 的 model 在家裡桌機有 64GB RAM 的情況下是可以應硬扛的,跑降到 4bits 的 model 大約吃 41GB RAM 左右。

比較驚訝的是 efficient 這個部份,拿 CPU 跑 65B 版本是跑得動的。

隨便丟個「文字接龍」進去 LLaMA-65B 讓他接,可以看到接出來的東西還可以:

main: prompt: 'The main difference between javascript and java are '
main: number of tokens in prompt = 10
[...]

The main difference between javascript and java are 1)Java is a compiled language while JavaScript
is not. So that the performance of Javascript code can be lower than Java Code, but they run on different platforms: browser vs server or local machine;
2)'javac' to compile .java file to bytecode(for JVM), and 'jar'(to make executable jar files for client systems).
3)Java uses more memory (RAM) that javascript. So the Javascript code is generally small than Java
codes.(A web application written in JavaScript may run faster on a PC with 64 MB of RAM compared to one running using applets.) While an advantage offered by J

如果是 LLaMA-7B 的話會快很多,但回答就有點微妙了:

main: prompt: 'The main difference between javascript and java are '
main: number of tokens in prompt = 10
[...]

The main difference between javascript and java are 1) Javascript is interpreted, while Java compiles to native machine code. (Compiled languages have a performance advantage over Interpreted ones
.)
2) The language standardization process for JavaScript has not been completed at the time of this writing which leaves it open to being hacked or changed by those who care enough about your website to do so... In Java, everything is set in stone. (Or at least as much code that can be shipped on a CD-ROM disk)
Sorry if my response was not clear - but you're right! I think that "2" above should really have said: "(

訓練所花的資源的部份,可以從論文裡面看到,如果是 2048 張 A100 的話大約要跑五個月 (照這個語氣,實際上大概不是這個數字):

Finally, we estimate that we used 2048 A100-80GB for a period of approximately 5 months to develop our models.

另外也有列出 GPU hours 可以參考: