Song Han (@songhan_mit) 's Twitter Profile
Song Han

@songhan_mit

Efficient AI computing

ID: 1102729037192089600

linkhttps://songhan.mit.edu calendar_today05-03-2019 00:34:24

219 Tweet

8,8K Followers

158 Following

Song Han (@songhan_mit) 's Twitter Profile Photo

Hybrid AutoRegressive Transformer is on MIT homepage today: news.mit.edu/2025/ai-tool-g… We thank NVIDIA for donating us the DGX server!

Hybrid AutoRegressive Transformer is on MIT homepage today: 

news.mit.edu/2025/ai-tool-g… 

We thank <a href="/nvidia/">NVIDIA</a> for donating us the DGX server!
Song Han (@songhan_mit) 's Twitter Profile Photo

Explore CoT-VLA where VILA-U, a unified visual language model for both understanding and generation, generates visual chain of thought: imagine the next frame before you move.

NVIDIA AI Developer (@nvidiaaidev) 's Twitter Profile Photo

NVIDIA and Massachusetts Institute of Technology (MIT) researchers unveil HART, a hybrid AI tool that generates high-quality images 9x faster. Perfect for training self-driving cars, robots, and more. Read the full story ➡️ nvda.ws/3G6Zy0y

NVIDIA and <a href="/MIT/">Massachusetts Institute of Technology (MIT)</a> researchers unveil HART, a hybrid AI tool that generates high-quality images 9x faster. Perfect for training self-driving cars, robots, and more.

Read the full story ➡️ nvda.ws/3G6Zy0y
Zhuoyang Zhang (@zhuoyang_zhang) 's Twitter Profile Photo

Welcome to our ICLR2025 VILA-U poster session: 24 Apr 10:00-12:30 at Hall 3+Hall 2B #210. In the past several months, VILA-U has gained remarkable attention, attracting lots of researchers to explore native unified multimodal models. Proud to be one of the first several works in

Welcome to our ICLR2025 VILA-U poster session: 24 Apr 10:00-12:30 at Hall 3+Hall 2B #210. In the past several months, VILA-U has gained remarkable attention, attracting lots of researchers to explore native unified multimodal models. Proud to be one of the first several works in
Casper Hansen (@casper_hansen_) 's Twitter Profile Photo

2.1k stars, 2+ million downloads, and 7000+ models on Huggingface later, and I am officially ready to retire my long-time project AutoAWQ ⚡️ Proud to say that AutoAWQ has been adopted by the vLLM and will now be maintained by 55+ contributors 🥳

2.1k stars, 2+ million downloads, and 7000+ models on Huggingface later, and I am officially ready to retire my long-time project AutoAWQ ⚡️

Proud to say that AutoAWQ has been adopted by the <a href="/vllm_project/">vLLM</a> and will now be maintained by 55+ contributors 🥳
Muyang Li (@lmxyy1999) 's Twitter Profile Photo

🚀 How to run 12B FLUX.1 on your local laptop with 2-3× speedup? Come check out our #SVDQuant (#ICLR2025 Spotlight) poster session! 🎉 🗓️ When: Friday, Apr 25, 10–12:30 (Singapore time) 📍 Where: Hall 3 + Hall 2B, Poster 169 📌 Poster: tinyurl.com/poster-svdquant 🎮 Demo:

🚀 How to run 12B FLUX.1 on your local laptop with 2-3× speedup? Come check out our #SVDQuant (#ICLR2025 Spotlight) poster session! 🎉 
🗓️ When: Friday, Apr 25, 10–12:30 (Singapore time)
📍 Where: Hall 3 + Hall 2B, Poster 169
📌 Poster: tinyurl.com/poster-svdquant
🎮 Demo:
Haocheng Xi (@haochengxiucb) 's Twitter Profile Photo

🚀 COAT: Memory Efficient FP8 Training @ICLR 2025 📍 Hall 3 + Hall 2B Poster #566 🗓 Sat, Apr 26 | 3:00–5:30 PM Singapore Time ✅ 1.54x Memory Efficiency, 1.43x Speedup, near lossless performance! ✅ Check our poster about FP8 Training by Compressing Optimizer states and

🚀 COAT: Memory Efficient FP8 Training @ICLR 2025 
📍 Hall 3 + Hall 2B Poster #566 
🗓 Sat, Apr 26 | 3:00–5:30 PM Singapore Time

✅ 1.54x Memory Efficiency, 1.43x Speedup, near lossless performance!

✅ Check our poster about FP8 Training by Compressing Optimizer states and
Enze Xie (@xieenze_jr) 's Twitter Profile Photo

🚀The code for Fast-dLLM is now open-source! 💥 Fast-dLLM achieves a 27.6× end-to-end speedup on 1024-token sequences with less than 2% accuracy drop. Check out the code here: github.com/NVlabs/Fast-dL…

Yi Wu (@jxwuyi) 's Twitter Profile Photo

We release fully async RL system AReaL-boba² for LLM & SOTA code RL w. Qwen3-14B! Qwen #opensource 🚀system&algorithm co-design → 2.77x faster ✅ 69.1 on LiveCodeBench 🔥 multi-turn RL ready 🔗 Project: github.com/inclusionAI/AR… 📄 Paper: arxiv.org/pdf/2505.24298 1/3👇

We release fully async RL system AReaL-boba² for LLM &amp; SOTA code RL w. Qwen3-14B! <a href="/Alibaba_Qwen/">Qwen</a> #opensource 
🚀system&amp;algorithm co-design → 2.77x faster 
✅ 69.1 on LiveCodeBench 
🔥 multi-turn RL ready
🔗 Project: github.com/inclusionAI/AR…
📄 Paper: arxiv.org/pdf/2505.24298
1/3👇
Zhuoyang Zhang (@zhuoyang_zhang) 's Twitter Profile Photo

Excited to attend CVPR! I will be at the poster sessions for our two papers: COT-VLA and NVILA. Welcome to chat about Efficient and Unified VLMs and VLAs. See you in Nashville!

Excited to attend CVPR! I will be at the poster sessions for our two papers: COT-VLA and NVILA. Welcome to chat about Efficient and Unified VLMs and VLAs. See you in Nashville!
Hongxu (Danny) Yin (@yin_hongxu) 's Twitter Profile Photo

Happy to attend #CVPR2025 at Nashville. We will present new techniques on improving foundation models including VILA and RADIO series towards agentic and physical AI. Together by team NVIDIA and our amazing academic collaborators. Greater news - all are now open sourced!