Akide Liu

PhD Candidate @ Monash University

[Akide Liu @Akide_Liu]

@thsottiaux ChatGPT Pro account deactivated on Aug 25. Personal research account since late 2022. Years of notes, code, drafts, and files are locked inside, and I cannot export them.
Official appeal is filed. This looks like a false positive from local code-debugging request bursts, not abus

[Akide Liu @Akide_Liu]

ChatGPT Pro account deactivated on Aug 25. Personal research account since late 2022. Years of notes, code, drafts, and files are locked inside, and I cannot export them.
Please have a human review it. Happy to verify identity privately. @OpenAI @thsottiaux

[Zeyu Zhang @SteveZeyuZhang]

📢 Excited to share!
@Akide_Liu will be presenting our Spotlight paper, “FPSAttention: Training-Aware FP8 and Sparsity Co-Design for Fast Video Diffusion,” at @NeurIPSConf San Diego!

📍 Exhibit Hall C, D, E — #5011
⏰ Fri, Dec 5, 2025·4:30 PM - 7:30 PM PST

[Akide Liu @Akide_Liu]

@NamanyayG It's a good point, wish to join the trial program.

[Akide Liu @Akide_Liu]

@supremeZhuang @thu_hanyz @SteveZeyuZhang

[Akide Liu @Akide_Liu]

🚀 Our latest work on accelerating video diffusion models, FPSAttention, is out now!
🎬 FPSAttention: Training-Aware FP8 and Sparsity Co-Design for Fast Video Diffusion

[Akide Liu @Akide_Liu]

📄 Paper: https://t.co/TmfDLn1izr
🌐 Project page: https://t.co/BhxB2s2ioW
📬 Feel free to reach out: [email protected]

You’re welcome to share and follow—let’s push the boundaries of efficient video generation together!

[Akide Liu @Akide_Liu]

💥 Trained on Wan2.1’s 1.3B and 14B models and evaluated on VBench, FPSAttention delivers:

⚡️ 7.09× speedup on attention operations
⚡️ 4.96× end-to-end speedup for 720p video generation — all nearly lossless compromising generation quality

[Akide Liu @Akide_Liu]

1️⃣ Unified 3D tile granularity supporting both quantization and sparsity
2️⃣ A denoising-step-aware adaptive strategy
3️⃣ Native hardware-friendly kernel combining FlashAttention and Hopper architecture optimizations

[Akide Liu @Akide_Liu]

We propose a training-aware co-design of FP8 quantization and structured sparsity, tailored for optimizing the 3D bi-directional attention in video generation. The method introduces three key innovations

[Akide Liu @Akide_Liu]

I just found a really neat tool to quickly smash out a new database schema. Highly recommended! https://t.co/OnnWI1VuoC

[Akide Liu @Akide_Liu]

Ray-Ban glasses special offer, only $ 24,99 only today! https://t.co/iJsX72Gjt1