在 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?