An aquarium in Japan has changed the diet of its penguins and otters due to rising costs, and the animals are refusing to eat the cheaper fish huffpost.com/entry/penguins…

A new version of 🤗 Optimum Habana has been released and adds support for GPT2 and T5! You can deploy your favorite models from the Hugging Face 's Transformers library on Gaudi from Habana Labs for faster and cheaper training. 👉 How to install it: github.com/huggingface/op…

Hugging Face 🤗 got a 100x speedup on its TensorFlow text generation code, thanks to XLA ⚡ XLA can massively accelerate ML code, but some problems do not look XLA-friendly at first. Text generation is one of them 💔 Here's how you can go from 🐢 to 🐇 with XLA: 🧵 [1/]

📆 Join me on 9/13 in Santa Clara to learn how to use the latest AI hardware accelerators with Hugging Face, Graphcore and Habana Labs ! 🤑 Attend this FREE 🤗 workshop at the AI Hardware Summit - only 30 seats are available so don't wait! hubs.ly/Q01hmfKG0

🤗 Optimum Habana 1.2 has just been released! Try it out easily with the new documentation to discover how to train your models faster and cheaper: lnkd.in/dgrpj6N6 It now supports DeepSpeed and Computer Vision models such as ViT and Swin!

Wanna train your models faster than with any GPU currently available on the market? Check out this benchmark we just published where Habana Labs Gaudi2 is 2x faster than Nvidia A100: huggingface.co/blog/habana-ga… It relies on Optimum Habana from Hugging Face.

⚡💵 I trained a CLIP model (ViT base + RoBERTa base) on Habana Gaudi and got a similar throughput to A100 for almost a third of its price! Check out how to do it easily with 🤗 Optimum Habana: github.com/huggingface/op…

We just released a new benchmark where Habana Labs Gaudi2 is up to x1.4 faster than Nvidia H100! x1.2 speedup with the exact same training arguments, and x1.4 with hardware-accelerated data loading which can be done easily with Optimum Habana. More here: huggingface.co/blog/bridgetow…

🤗Transformers welcomes 4x speed-up by leveraging torch.compile to its very core ! Static Cache + compile for decoder models: on llama 7b ⤵️

Crazy how in 2 years LoRA and other efficient fine-tuning techniques went from adoption on GPU to being widely available on various types of hardware and so easy to use!

Today we are celebrating ⭐️⭐️150 000 stars ⭐️⭐️ on transfomers github repo! Thanks everyone!