Multi-threaded libraries (e.g., numpy and PyTorch on CPUs come to mind) are well supported. In scenarios where many processes are each running heavily multi-threaded computations, it can help to pin specific processes to specific cores (e.g., using tools like psutil) to avoid contention.
The scenario where a Ray task forks is probably not very well supported. You can certainly start a subprocess from within a Ray task, but I think forking could easily cause issues.
You can definitely use Ray + Jax, but you probably need to avoid forking a process within a Ray worker.
Our intention from the start was for Ray to be general purpose. And the core Ray APIs are quite general (basically just scheduling a Python function somewhere in a cluster or instantiating a Python class as a process somewhere in the cluster).
We had AI use cases in mind from the start, since we were grad students in AI. But the generality has really been important since AI workloads encompass a huge variety of computational patterns (allreduce style communication patterns on GPUs for training, embarrassingly parallel data processing workloads on spot instances, and so on).
Yeah, mmap, I think this is the relevant line [1].
Fun fact, very early on, we used to create one mmapped file per serialized object, but that very quickly broke down.
Then we switched to mmapping one large file at the start and storing all of the serialized objects in that file. But then as objects get allocated and deallocated, you need to manage the memory inside of that mmapped file, and we just repurposed a malloc implementation to handle that.
Your right that the serialization / deserialization overhead can quickly exceed the compute time. To avoid this you have to get a lot of small things right. And given our focus on ML workloads, this is particularly important when sharing large numerical arrays between processes (especially processes running on the same node).
One of the key things is to make sure the serialized object is stored in a data format where the serialized object does not need to be "transformed" in order to access it. For example, a numpy array can be created in O(1) time from a serialized blob by initializing a Python object with the right shape and dtype and a pointer to the right offset in the serialized blob. We also use projects like Apache Arrow that put a lot of care into this.
Example in more detail:
Imagine the object you are passing from process A to process B is a 1GB numpy array of floats. In the serialization step, process A produces a serialized blob of bytes that is basically just the 1GB numpy array plus a little bit of metadata. Process A writes that serialized blob into shared memory. This step of "writing into shared memory" still involves O(N) work, where N is the size of the array (though you can have multiple threads do the memcpy in parallel and be limited just by memory bandwidth).
In the deserialization step, process B accesses the same shared memory blob (process A and B are on the same machine). It reads a tiny bit of metadata to figure out the type of the serialized object and shape and so on. Then it constructs a numpy array with the correct shape and type and with a pointer to the actual data in shared memory at the right offset. Therefore it doesn't need to touch all of the bytes of data, it just does O(1) work instead of O(N).
That's the basic idea. You can imagine generalizing this beyond numpy arrays, but it's most effective for objects that include large numerical data (e.g., objects that include numpy arrays).
There are a bunch of little details to get right, e.g., serializing directly into shared memory instead of creating a serialized copy in process A and then copying it into shared memory. Doing the write into shared memory in parallel with a bunch of threads. Getting the deserialization right. You also have to make sure that the starting addresses of the numpy arrays are 64-byte aligned (if memory serves) so that you don't accidentally trigger a copy later on.
1. This is truly impressive work from AWS. Patrick Ames began speaking about this a couple years ago, though at this point the blog post is probably the best reference. https://www.youtube.com/watch?v=h7svj_oAY14
2. This is not a "typical" Ray use case. I'm not aware of any other exabyte scale data processing workloads. Our bread and butter is ML workloads: training, inference, and unstructured data processing.
3. We have a data processing library called Ray Data for ingesting and processing data, often done in conjunction with training and inference. However, I believe in this particular use case, the heavy lifting is largely done with Ray's core APIs (tasks & actors), which are lower level and more flexible, which makes sense for highly custom use cases. Most Ray users use the Ray libraries (train, data, serve), but power users often use the Ray core APIs.
4. Since people often ask about data processing with Ray and Spark, Spark use cases tend to be more geared toward structured data and CPU processing. If you are joining a bunch of tables together or running SQL queries, Spark is going to be way better. If you're working with unstructured data (images, text, video, audio, etc), need mixed CPU & GPU compute, are doing deep learning and running inference, etc, then Ray is going to be much better.
Thanks for the feedback, we'll improve the landing page!
The models (and current prices) right now are
- Llama-2-7B ($0.25 / million tokens)
- Llama-2-13B ($0.50 / million tokens)
- Llama-2-70B ($1 / million tokens)
- Code Llama ($1 / million tokens)
I think for fine-tuned GPT-3.5 to be competitive with GPT-4 on your use cases (assistance with Angular), you'd have to fine-tune on enough data that it really resembles pre-training more than fine-tuning. And it wouldn't be worth the hassle unless you're building a product around it.
That said, many valuable LLM products / features are more narrow in scope and can see a huge lift from fine-tuning. We've run a bunch of experiments on this (e.g., SQL query generation is a good example), where fine-tuning even the 7B Llama-2 model outperforms GPT-4 (surprisingly) [1]. That's a very different type of problem from teaching software engineering of course.
I think of fine-tuning as an avenue to significantly reduce LLM inference costs, so I think this is an exciting development. You're right if you compare GPT-3.5-turbo to fine-tuned GPT-3.5-turbo, but if it's anything like fine-tuning the Llama-2 models, you'll be able to achieve GPT-4 level performance for a wide range of practical use cases (SQL query generation is an example), but probably not for math or coding (at least not without fine-tuning on a significant amount of data).
In fact, we've seen GPT-4 level performance from even the 7B Llama-2 model after fine-tuning. [1]
> Llama 2 might by some measures be close to GPT 3.5, but it’s nowhere near GPT 4
I think you're right about this, and benchmarks we've run at Anyscale support this conclusion [1].
The caveat there (which I think will be a big boon for open models) is that techniques like fine-tuning makes a HUGE difference and can bridge the quality gap between Llama-2 and GPT-4 for many (but not all) problems.
If you want to try out the Llama-2 models (7B, 13B, 70B), you can get started very easily with Anyscale Endpoints (~2 min). https://app.endpoints.anyscale.com/
Data processing workloads are quite common on Ray, especially with unstructured data.
Also, I work on Ray, which is the underlying framework used here, but all the work in the post was done by the Amazon team.