This is a perfect example of doing impactful AI research in academia in the era of LLMs

We have lots of cool algorithms (fine-tuning, LoRA, prompt tuning...) that can adapt a pre-trained model to a downstream task, but what do they MISS?

Surprisingly, we find models trained with these algorithms have trouble with focusing their attention!

A Winning Combination for Large Language Models

TL;DR: Did you find MoE models generalize worse than dense models on downstream tasks? Not any more at the age of instruction tuning!

Surprisingly, we see the “1 + 1 > 2” effect when it comes to MoE + Instruction Tuning. [1/4]

I have been waiting for this plot for a long time -- such a great reference to holistically assess the capabilities and limitations of those 'llama family' open-source models. Most importantly, it provides insights into the potential applications a self-hosted model is capable of

Thanks AK for sharing! Our new work, RPT, can help your robot learn better by masked pre-training on sensorimotor sequences.

🌟 Thrilled to introduce vLLM with Woosuk Kwon!

🚀 vLLM is an open-source LLM inference and serving library that accelerates HuggingFace Transformers by 24x and powers lmsys.org Vicuna and Chatbot Arena.

Github: github.com/vllm-project/v…
Blog: vllm.ai

If you'll be #CVPR2023 on Sunday morning, come check out the LOVEU workshop! We have competitions for video understanding and generative AI video editing.
sites.google.com/view/loveucvpr…

Thanks for reading, now enjoy a video of a horse on mars by the winning team:

Check out breakthroughs in LLM quantization with our new work, which presents a sensitivity-aware non-uniform quantization scheme that outperforms existing methods at 3 & 4 bits, and our 4-bit quantized Vicuna matches FP baseline performance as evaluated by GPT-4! 🔥

SmoothQuant is good for W8A8 LLM quantization, what about low-bit weight-only quantization (e.g., W4A16)? We present Activation-aware Weight Quantization (AWQ) for LLM compression and acceleration: github.com/mit-han-lab/ll… 🧵

Humans pay attention to different objects when performing different tasks. Can vision transformer (ViT) do that as well?

In our recent work, we build a ViT with task-guided attention! 1/n

Visit our website to learn more: sites.google.com/view/absvit

Introducing Vicuna, an open-source chatbot impressing GPT-4!

🚀 Vicuna reaches 90%* quality of ChatGPT/Bard while significantly outperforming other baselines, according to GPT-4's assessment.

Blog: vicuna.lmsys.org
Demo: chat.lmsys.org

Excited to share our latest research GARNET, which improves GNN robustness on large-scale graphs with millions of nodes, and can serve as a plug-and-play module for various graphs and GNN backbones.

Excited to meet with friends and have a Cornell University Artificial Intelligence reunion at my first in-person conference! Horace He

TorchSparse is really a great project to work on — if you are interested in learning more, please come and chat with us at MLSys 2022 :)

My student Kayo Yin needs your help. Her visa has been unnecessarily delayed, which would prevent her from coming to UC Berkeley to start her studies. Despite bringing all required documents, the Department of State refused to process the visa and it could take months to re-process.