Megatron-LM
Ongoing research training transformer models at scale (by NVIDIA)
xla
Enabling PyTorch on XLA Devices (e.g. Google TPU) (by pytorch)
Our great sponsors
Megatron-LM | xla | |
---|---|---|
18 | 8 | |
8,561 | 2,291 | |
7.6% | 2.7% | |
9.9 | 9.9 | |
3 days ago | 3 days ago | |
Python | C++ | |
GNU General Public License v3.0 or later | GNU General Public License v3.0 or later |
The number of mentions indicates the total number of mentions that we've tracked plus the number of user suggested alternatives.
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
Megatron-LM
Posts with mentions or reviews of Megatron-LM.
We have used some of these posts to build our list of alternatives
and similar projects. The last one was on 2024-04-23.
-
Apple releases CoreNet, a library for training deep neural networks
https://github.com/NVIDIA/Megatron-LM
This is probably a good baseline to start thinking about LLM training at scale.
- Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping
-
Large Language Models: Compairing Gen2/Gen3 Models (GPT-3, GPT-J, MT5 and More)
This 20B model was trained on the same datasets as its predecessor, aptly named The Pile. Furthermore, the libraries Megatron and DeepSpeed were used to achieve better computing resource utilization, and eventually GPT-NeoX evolved into its own framework for training other LLMs. It was used, for example, as the foundation for Llemma, an open-source model specializing on theorem proving.
- Why async gradient update doesn't get popular in LLM community?
-
[D] Distributes pre-training and fine-tuning
Deepspeed Megatron-LM
-
Why Did Google Brain Exist?
GPU cluster scaling has come a long way. Just checkout the scaling plot here: https://github.com/NVIDIA/Megatron-LM
-
Does Megatron-LM really not communicate during multi-head attention operations?
I found their code that the softmax function conduct all-reduce before they work.
-
I asked ChatGPT to rate the intelligence level of current AI systems out there.
Google's PaLM, Facebook's LLaMA, Nvidia's Megatron, I am missing some surely and Apple sure has something cooking as well but these are the big ones, of course none of them are publicly available, but research papers are reputable. All of the ones mentioned should beat GPT-3 although GPT-3.5 (chatGPT) should be bit better and ability to search (Bing) should level the playing field even further, but Google's PaLM with search functionality should be clearly ahead. This is why people are excited about GPT-4, GPT-3 was way ahead of anyone else when it came out but others were able to catch up since, we'll see if GPT-4 will be another bing jump among LLMs.
-
GPT-4 Will Be 500x Smaller Than People Think - Here Is Why
Found relevant code at https://github.com/nvidia/megatron-lm + all code implementations here
-
People who do “unglamorous” businesses - what do you do and approx how much money do you make?
Involvement in projects utilising stuff like this: https://github.com/NVIDIA/Megatron-LM is my end goal.
xla
Posts with mentions or reviews of xla.
We have used some of these posts to build our list of alternatives
and similar projects. The last one was on 2024-01-26.
-
Who uses Google TPUs for inference in production?
> The PyTorch/XLA Team at Google
Meanwhile you have an issue from 5 years ago with 0 support
https://github.com/pytorch/xla/issues/202
-
Google TPU v5p beats Nvidia H100
PyTorch has had an XLA backend for years. I don't know how performant it is though. https://pytorch.org/xla
-
Why Did Google Brain Exist?
It's curtains for XLA, to be precise. And PyTorch officially supports XLA backend nowadays too ([1]), which kind of makes JAX and PyTorch standing on the same foundation.
1. https://github.com/pytorch/xla
-
Accelerating AI inference?
Pytorch supports other kinds of accelerators (e.g. FPGA, and https://github.com/pytorch/glow), but unless you want to become a ML systems engineer and have money and time to throw away, or a business case to fund it, it is not worth it. In general, both pytorch and tensorflow have hardware abstractions that will compile down to device code. (XLA, https://github.com/pytorch/xla, https://github.com/pytorch/glow). TPUs and GPUs have very different strengths; so getting top performance requires a lot of manual optimizations. Considering the the cost of training LLM, it is time well spent.
-
[D] Colab TPU low performance
While apparently TPUs can theoretically achieve great speedups, getting to the point where they beat a single GPU requires a lot of fiddling around and debugging. A specific setup is required to make it work properly. E.g., here it says that to exploit TPUs you might need a better CPU to keep the TPU busy, than the one in colab. The tutorials I looked at oversimplified the whole matter, the same goes for pytorch-lightning which implies switching to TPU is as easy as changing a single parameter. Furthermore, none of the tutorials I saw (even after specifically searching for that) went into detail about why and how to set up a GCS bucket for data loading.
- How to train large deep learning models as a startup
-
Distributed Training Made Easy with PyTorch-Ignite
XLA on TPUs via pytorch/xla.
-
[P] PyTorch for TensorFlow Users - A Minimal Diff
I don't know of any such trick except for using TensorFlow. In fact, I benchmarked PyTorch XLA vs TensorFlow and found that the former's performance was quite abysmal: PyTorch XLA is very slow on Google Colab. The developers' explanation, as I understood it, was that TF was using features not available to the PyTorch XLA developers and that they therefore could not compete on performance. The situation may be different today, I don't know really.
What are some alternatives?
When comparing Megatron-LM and xla you can also consider the following projects:
DeepSpeed - DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.
NCCL - Optimized primitives for collective multi-GPU communication