- Tweets 110
- Following 288
- Followers 2,128
- Likes 279
Personal update: I am excited to share that I will join @GoogleDeepMind next week after defending my PhD thesis @MITEECS earlier last month. I will be working on generative models that simulate the physical world. Looking forward to the new journey ahead in 2025!
71
49
2,185
126,321
🚀 We're thrilled to introduce HART, an efficient AR model that generates stunning 1024x1024 images! 🎨✨
HART delivers:
⚡️ 4.5-7.7x higher throughput
🔋 6.9-13.4x less compute
🔥 top-notch FID & CLIP scores, rivaling diffusion models in quality!
Code: tinyurl.com/nkvpnhyk
8
25
138
24,482
Excited to share my #MLSys 2024 best paper 🏆 presentation on AWQ. AWQ democratizes edge LLM deployment 💻 and has been downloaded over 1 million times on Huggingface 🙌! piped.video/dcINVsqxQgQ?si=WEuH…
1
10
38
15,030
AWQ website: hanlab.mit.edu/projects/awq
Paper: arxiv.org/abs/2306.00978
Code: github.com/mit-han-lab/llm-a…
Joint work with @jilin_14, @jmtang42, @Shang_mit, Wei-Ming Chen, Wei-Chen Wang, @Guangxuan_Xiao, Xingyu Dang, Prof. @gan_chuang, Prof. @songhan_mit.
2
12
1,308
📄 Paper: arxiv.org/abs/2410.10812
🌐 Project: hanlab.mit.edu/projects/hart
🖥 Demo: hart.mit.edu
💻 Code: github.com/mit-han-lab/hart
1
1
11
1,167
🔥Welcome to try out QServe! TRT-LLM efficiency⚡️ + PyTorch flexibility 😄, your LLM serving turn-key solution 🔑
🔥🎉Thrilled to introduce QServe, our latest breakthrough in efficient LLM serving with W4-A8-KV4 quantization.
🚀⚡1.2-3.5x higher throughput over TensorRT-LLM.
💵 Matches TensorRT-LLM’s A100 throughput with 3x cheaper L40S GPUs.
👐 Code: github.com/mit-han-lab/qserv… (1/4)
4
1,081
✨ How it works: We decompose continuous latents into two parts:
🔹 Discrete tokens for the big picture, modeled by a scalable-resolution AR transformer
🔸 Residual tokens for image details, handled by a lightweight diffusion module (37M parameters, 8 sampling steps)
1
6
1,407
What an achievement! Congrats to the team!
Our latest update to our Gemini 2.0 Flash Thinking model (available here: goo.gle/4jsCqZC) scores 73.3% on AIME (math) & 74.2% on GPQA Diamond (science) benchmarks. Thanks for all your feedback, this represents super fast progress from our first release just this past Dec! Latest version also includes code execution, a 1M token content window & a reduced likelihood of thought-answer contradictions. We’ve been pioneering these types of planning systems for over a decade, starting with programs like AlphaGo, and it is exciting to see the powerful combination of these ideas with the most capable foundation models.
4
709
Thank you! I think my research at GDM will be a continuation of recent publications of MIT HAN Lab~
1
1
253
Thank you Ryan! Boston is indeed a great place and I hope that I will be able to come back occasionally
1
29