🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Hardware-Efficient Attention for Fast Decoding Princeton optimizes decoding by maximizing arithmetic intensity (FLOPs/byte) for better memory–compute efficiency: - GTA (Grouped-Tied Attention) Ties key/value states + partial RoPE → 2× arithmetic intensity vs. GQA, ½ KV cache,

Hardware-Efficient Attention for Fast Decoding "We first propose Grouped-Tied Attention (GTA), a simple variant that combines and reuses key and value states, reducing memory transfers without compromising model quality. We then introduce Grouped Latent Attention (GLA), a

We've been thinking about what the "ideal" architecture should look like in the era where inference is driving AI progress. GTA & GLA are steps in this direction: attention variants tailored for inference: high arithmetic intensity (make GPUs go brr even during decoding), easy to

Great project where we rethink attention for inference: Grouped-Tied Attn (GTA) ties the KV, and Grouped Latent Attn (GLA) shards low-rank latents across GPUs. Results: high arithmetic intensity, high quality models and parallel-friendly. Kudos to the team !!

🦆🚀QuACK🦆🚀: new SOL mem-bound kernel library without a single line of CUDA C++ all straight in Python thanks to CuTe-DSL. On H100 with 3TB/s, it performs 33%-50% faster than highly optimized libraries like PyTorch's torch.compile and Liger. 🤯 With Ted Zadouri and Tri Dao

Getting mem-bound kernels to speed-of-light isn't a dark art, it's just about getting the a couple of details right. We wrote a tutorial on how to do this, with code you can directly use. Thanks to the new CuTe-DSL, we can hit speed-of-light without a single line of CUDA C++.

🦆QuACK: blazing fast cute-DSL GPU kernels with 3TB/s goodness! Optimizing your kernels as much as possible is important... unless you are okay with leaving throughput on the table. check out this work from vlaw, Ted Zadouri and Tri Dao

CuTe DSL feels almost unreal: minimal Python code hits peak memory throughput on H100, as we show in QuACK. Can't wait for the addition of kernels optimized for Blackwell in QuACK 🦆

Congrats to Parastoo Abtahi, Tri Dao and Alex Lombardi on being named 2025 Google Research Scholars. 🎉 The @googleresearch scholars program funds world-class research conducted by early-career professors. bit.ly/4kvpvFx