ThirdAI Uses Ray for Parallel Training of Billion-Parameter NN on Commodity CPUs(anyscale.com)
anyscale.com
ThirdAI Uses Ray for Parallel Training of Billion-Parameter NN on Commodity CPUs
https://www.anyscale.com/blog/how-thirdai-uses-ray-for-parallel-training-of-billion-parameter-neural-networks-on-commodity-cpus
16 comments
Their methods rely on heuristics to sparsify NNs. They compare their sparse and dense methods to PyTorch and Tensorflow and get worse perf on dense inferences but better on sparse. It would be useful to know if PyTorch runs faster or slower than their methods using their sparsified matrix techniques.
> It would be useful to know if PyTorch runs faster or slower than their methods using their sparsified matrix techniques.
Probably not an easy feat. Sparse support in pytorch is minimal.
Probably not an easy feat. Sparse support in pytorch is minimal.
Very cool tech! I remember seeing the SLIDE paper and then being shocked at the lack of high profile follow-up work.
Is there any plan to open source BOLT? This would be extremely valuable to the community and in reducing Nvidia's chokehold.
Is there any plan to open source BOLT? This would be extremely valuable to the community and in reducing Nvidia's chokehold.
Thanks for reading! This is a great question. While we currently don't have immediate plans to open source BOLT, we are still in the very early stages of the company and may consider it down the road. Although BOLT is currently closed-source, we do offer free trial licenses https://www.thirdai.com/try-bolt/
Yes but do they pay Anyscale anything for this use of Ray?
IIUC Anyscale is managed Ray, and also works on Ray full-time. However Ray is open source and you can run it yourself. It's fairly easy to get started with, too.
This is all open source Ray. None of the AWS benchmarks were run on Anyscale platform. You can install and run OSS Ray on any of the popular clouds, including your laptop.
All the scripts for you to run in your OSS Ray cluster are here: https://github.com/ThirdAILabs/Public-Benchmarks/blob/main/c...
All the scripts for you to run in your OSS Ray cluster are here: https://github.com/ThirdAILabs/Public-Benchmarks/blob/main/c...
Founder of Anyscale is the Berkeley student who helped start Ray and is usually happy to give free advice. Anyscale is to Ray as Databricks is to Spark.
you can set up a self-hosted Ray infra to achieve this.
Seems to require Tensorflow. That makes it basically irrelevant?
Hi, I am one of the co-authors of the blog from ThirdAI. Our software has no dependency on TensorFlow or any other deep learning framework. You can see the libraries we import for training and try it out for yourself in our public github repo: https://github.com/ThirdAILabs/Public-Benchmarks/blob/main/c...
Ray does not require Tensorflow, tho the cited workflow might use tf.
Here’s a random example of pytorch on ray: https://www.ray.io/ray-sgd
Here’s a random example of pytorch on ray: https://www.ray.io/ray-sgd
PyTorch too?[1]
[1] https://docs.ray.io/en/latest/train/api/doc/ray.train.torch....
[1] https://docs.ray.io/en/latest/train/api/doc/ray.train.torch....
Ray provides integrations for PyTorch and a variety of other frameworks, but doesn't depend on them (unless you're using those integrations of course). The core library is just general purpose distributed compute.
by skipping it you can play a little slower but probably a lot cheaper