Zhijian Liu (@zhijianliu

Zhijian Liu

214 posts

Zhijian Liu

@zhijianliu_

Assistant Professor @UCSanDiego (running z-lab.ai). Prev: RS @NVIDIA. PhD @MIT. Efficient AI. Views are my own.

California

Joined November 2017

Pinned
Zhijian Liu
@zhijianliu_
Jun 15
🚀 DFlash now runs on SGLang's new default speculative-decoding engine, Spec V2. ⚡️ Hitting >4.3× baseline throughput (1.5× over native MTP) on Qwen 3.5 397B-A17B. Same quality, more speed! ⭐ github.com/z-lab/dflash
17K
Zhijian Liu
@zhijianliu_
Aug 4, 2024
📢 Excited to share that I’ll be joining @UCSanDiego @HDSIUCSD as an Assistant Professor in January 2026! My lab will focus on efficient AI. I'm recruiting PhD students from HDSI/CSE in the Fall 2024 cycle and also looking for RAs/interns! More info see zhijianliu.com.
zhijianliu.com
Zhijian Liu
Zhijian Liu is an Assistant Professor at UC San Diego. His research focuses on efficient machine learning and systems.
92K
Zhijian Liu
@zhijianliu_
Sep 22, 2023
Expanding LLMs' context size requires expensive training. Our new method, LongLoRA, allows for fine-tuning using sparse local attention, but retains dense attention during inference. This enabled us to extend LLaMA2-70B's context length to 32K using a single 8xA100 machine. 🧵
8.1K
Zhijian Liu
@zhijianliu_
Jun 18, 2023
I am at #CVPR2023 this week. Reach out if you are up to for a chat. Looking forward to meeting you all!
9.7K
Zhijian Liu
@zhijianliu_
Mar 31, 2023
Our #CVPR2023 papers are now available online. Check out FlatFormer at arxiv.org/abs/2301.08739 and SparseViT at arxiv.org/abs/2303.17605. More details in this thread. 🧵
Zhijian Liu
@zhijianliu_
Feb 28, 2023
Thrilled to announce that our papers on FlatFormer and SparseViT have been accepted to #CVPR2023! They showcase remarkable efficiency improvements for transformers in 2D and 3D vision. Stay tuned for more updates!
5.5K
Zhijian Liu
@zhijianliu_
Feb 28, 2023
Thrilled to announce that our papers on FlatFormer and SparseViT have been accepted to #CVPR2023! They showcase remarkable efficiency improvements for transformers in 2D and 3D vision. Stay tuned for more updates!
7.6K
Zhijian Liu
@zhijianliu_
May 29, 2022
We are excited to share our latest research, BEVFusion (bevfusion.mit.edu), that achieves SOTA performance on nuScenes for 3D object detection and BEV map segmentation, with half the computation cost. We will soon release our code at github.com/mit-han-lab/be…. Stay tuned!
Zhijian Liu
@zhijianliu_
Aug 4, 2024
Replying to @zhijianliu_
Deeply grateful to my amazing advisor (@songhan_mit) for his unwavering support through my PhD and beyond. Huge thanks to my mentors (@KurtKeutzer, @drmapavone, @ShenlongWang, Philip Harris, and many others) for their guidance. Thanks everyone who has been part of this journey!
6.1K
Zhijian Liu
@zhijianliu_
Jul 4, 2021
Super excited to receive the Qualcomm Innovation Fellowship 2021. Thanks for the great support from my advisor @SongHan_MIT and collaborators! qualcomm.com/research/resea…
Zhijian Liu
@zhijianliu_
Aug 4, 2024
Replying to @zhijianliu_ @songhan_mit and 3 others
Until then, I'll be working at @NVIDIA Research on building efficient foundation models. I'll still be based in Boston, so please say hi if you're around -- I'd love to catch up!
4.9K
Zhijian Liu
@zhijianliu_
Jan 18, 2023
Thrilled to share that BEVFusion was accepted to #ICRA2023. 🥳
Zhijian Liu
@zhijianliu_
May 29, 2022
We are excited to share our latest research, BEVFusion (bevfusion.mit.edu), that achieves SOTA performance on nuScenes for 3D object detection and BEV map segmentation, with half the computation cost. We will soon release our code at github.com/mit-han-lab/be…. Stay tuned!
1.2K
Zhijian Liu
@zhijianliu_
Sep 22, 2023
Replying to @zhijianliu_
🚀 Dive into our code and grab our pre-trained models here: github.com/dvlab-research…. 🤝 A collaborative effort spearheaded by Yukang Chen and teamed with @thesouthfrog, @haotiantang, Xin Lai, @songhan_mit, and Jiaya Jia!
951
Zhijian Liu
@zhijianliu_
Mar 31, 2023
Replying to @zhijianliu_
FlatFormer bridges the 3-4x latency gap between point cloud transformers and sparse convolutional models. The key idea is to trade spatial proximity for better computational regularity. This is a joint effort with Xinyu Yang, @haotiant1998, Shang Yang, and @songhan_mit.
533
Zhijian Liu
@zhijianliu_
Dec 14, 2022
Congratulations and thanks everyone for joining the journey with us!
Song Han
@songhan_mit
Dec 14, 2022
Dear efficientml.ai students, Congratulations on completing the TinyML and Efficient Deep Learning course! I hope that you have found the course to be informative and valuable in learning about the challenges and techniques of deploying neural networks on mobile devices.