Kudos to the HuggingFace Accelerate team for making it trivial to switch ZeRO backends.

I had a homegrown Accelerate-based training loop setup to work with Deepspeed ZeRO.

I wanted to try FSDP - and I just needed to change the Accelerate config to use FSDP instead of Deepspeed

Announcing 4-bit Mixtral 8x7B on 🤗Transformers!

Run the new Mistal MoE with minimal performance degradation on your local computer (24Go) 🔥

Stay tuned as more quants are coming soon using AWQ. We are also looking into sparsification with Tim Dettmers

huggingface.co/TheBloke/Mixtr…

Today Eric Ries (creator of Lean Startup & LTSE) & I (fast.ai /Kaggle) are launching a new kind of R&D lab: answer.ai. We're backed by $10m of funding from Decibel.

For-profit R&D labs are rare today, but have an amazing history... 🧵

v4.36 is out, and Mixtral is in 🫡

github.com/huggingface/tr…

Quadratic attention has been indispensable for information-dense modalities such as language... until now.

Announcing Mamba: a new SSM arch. that has linear-time scaling, ultra long context, and most importantly--outperforms Transformers everywhere we've tried.

With Tri Dao 1/

This is the first pass on the new chapter for ML Engineering:

The AI Battlefield Engineering - What You Need To Know

github.com/stas00/ml-engi…

This a WIP and your feedback for improvement is always welcome.

Today is a new achievement for Hugging Face Accelerate, we've passed 50 million downloads on pypi! In honor of this I wanted to highlight some history about the framework, where we've gone to, and how the community is relying on it more and more:

1/6

I have started working on a new guide, called

The Art of Debugging

which is a brain dump based on almost 3 decades of developing software. So far I have the initial draft of the first chapter:

Fast Debugging Methodology

github.com/stas00/the-art…

A lot more to come...

I

Excited to announce a new Hugging Face space to help with one of machine learning's biggest questions:

How much space does {X} model take in vRAM? And most importantly: when using `device_map='auto'`

huggingface.co/spaces/hf-acce…

Yesterday was my last day at Hugging Face. The past three years have been exhilarating and I am very proud of what the team has accomplished during that time!

Taking a bit of a break with opensource full time (though I will still contribute to Transformers and Accelerate)

Super excited to welcome our new investors Salesforce Ventures, Google, Amazon, NVIDIA, AMD, Intel, Qualcomm Ventures, IBM & Sound Ventures who all participated in Hugging Face’s $235M series D at a $4.5B valuation to celebrate the crossing of 1,000,000 models, datasets and apps

LLMs just got faster and lighter with 🤗 Transformers x AutoGPTQ !

You can now load your models from Hugging Face with GPTQ quantization. Enjoy faster inference speed and lower memory usage than existing supported quantization schemes 🚀

Blogpost: huggingface.co/blog/gptq-inte…

Introducing IDEFICS, the first open state-of-the-art visual language model at the 80B scale!

The model accepts arbitrary sequences of images and texts and produces text. A bit like a multimodal ChatGPT!

Blogpost: huggingface.co/blog/idefics
Playground:
huggingface.co/spaces/Hugging…

Shout out to HF 🤗Acellerate for making it so easy to split LLMs over multiple GPUs with device_map=”auto”

It just works, all the time

Very underrated esp since most people have GPUs < 24GB vRAM

Wow! Hugging Face accelerate's growth is truly taking off!

A brief timeline history to gain 1M monthly downloads:

1MM -> 2MM: 3 months
2MM -> 3MM: 3 months
3MM -> 4MM: 2 months
4MM -> 5MM: 2 weeks 🤯

How long until we hit 6MM? 🤔

Huge personal update! After 1 year at 🤗 through sweat, tears and tokenizers I am now a core maintainer of transformers! 🎉
Huge thanks to Sylvain Gugger , amy, Lysandre and Nicolas Patry for their precious guidance along the way. It's only the beginning! 💪

TRL 🤗 Hugging Face

Excited to announce that we're doubling down on our efforts to democratize RLHF and reinforcement learning with TRL, new addition to the Hugging Face family, developed and led by team member Leandro von Werra 🎉🎉

Train your first RLHF model 👉github.com/huggingface/trl

Hugging Face is now part of the PyTorch Foundation as a premier member 🤝

We have been collaborating with the PyTorch team for the past four years and are committed to supporting the project.

We share an objective: to lower the barrier of entry to ML.

pytorch.org/blog/hugging-f…

Excited to share our latest creation: the Object Detection Leaderboard on Hugging Face! 🤗 📣

Check it out: huggingface.co/spaces/rafaelp…

Compare performances 📊 of various models on different datasets and metrics. Got a model? Bring it over and see how it fares! 🤗

Today is a very exciting day for Hugging Face accelerate: in just 2 months we went from 20 million to 30 million downloads! It's been amazing seeing the growth and use throughout the @pytorch ecosystem, and we're excited for the times to come!