Llama.cpp is now 100x faster

pierotofy · April 1, 2023, 3:38am

Make loading weights 10-100x faster

ggerganov:master ← jart:loader

opened 11:51PM - 29 Mar 23 UTC

This is a breaking change that's going to give us three benefits: 1. Your inf…erence commands should load 100x faster 2. You may be able to safely load models 2x larger 3. You can run many concurrent inference processes 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. The new file format supports single-file models like LLaMA 7b, and it also supports multi-file models like LLaMA 13B. Our Python tool now merges the foo.1, foo.2, etc. files back into a single file so that the C++ code which maps it doesn't need to reshape data every time. That's made llama.cpp so much simpler. Much of its load code has now been deleted. Furthermore, this change ensures that tensors are aligned properly on a 32-byte boundary. That opens the door to seeing if we can get additional performance gains on some microprocessors, by using ops that require memory alignment. Lastly note that both POSIX and the Windows platform are supported The issue this PR solves is https://github.com/ggerganov/llama.cpp/issues/91 This PR was written in collaboration with @slaren. This PR is also rebased on PR #586 **so please do not squash merge**! Use either merge or rebase.

argosopentech · April 4, 2023, 1:56pm

I think “10-100x faster” is probably an exaggeration. My understanding is that the mmap system call maps a file into memory. This gives the OS more flexibility and control over loading the model weights into memory but I don’t think it’s going to be 10x faster. It probably is faster though especially for a single inference run.

It’s very exciting to see all of this open source experimentation being done!

pierotofy · April 4, 2023, 3:00pm

Agreed, the way they named the PR is a bit misleading.