@jiqizhixin

–

14 March 2025 10h31

GitHub: https://
github.com/thu-pacman/chi
tu
…
By the way, Chitu (赤兔) is the name of a great horse in ancient China.

Development Team Optimizes GeMM and MoE for FP8 Processing

By

–

14 March 2025 10h31

A member of the development team told us: "We've optimized a series of key operators, such as GeMM and MoE, at the instruction level to achieve native processing capabilities for FP8 data."

Chitu Matches vLLM Performance on H20 DeepSeek-R1

By

–

14 March 2025 10h31

On an H20 (96GB) cluster, Chitu performs comparably to vLLM when deploying DeepSeek-R1-671B.

DeepSeek-R1-671B Outperforms vLLM on A800 Cluster

By

–

14 March 2025 10h31

And it outperforms vLLM when deploying DeepSeek-R1-671B on an A800 (40GB) cluster.

Chitu: High-Performance Open-Source LLM Inference Framework

By

–

14 March 2025 10h31

Yet another open-source project from China: Chitu! A high-performance inference framework for LLMs, designed for efficiency, flexibility, and availability. Chitu supports various mainstream models, including DeepSeek, the LLaMA series, Mixtral, and more.

Meta researchers replace normalization with Dynamic Tanh in Transformers

By

–

14 March 2025 8h30

Imagine a Transformer model without normalization. This is exactly what's proposed in a new paper from Meta, NYU, MIT, and Princeton. The authors found that normalization layers can be replaced with something called Dynamic Tanh (DyT). It looks like this: DyT(x)=γ ∗ tanh(αx)+β.

Chain of Thought Enhances AI Image Generation Step by Step

By

–

08 February 2025 8h03

Can We Generate Images with CoT? Let’s Verify and Reinforce Image Generation Step by Step https://
arxiv.org/pdf/2501.13926

8 February 2025

ByteDance Introduces OmniHuman-1 for Conditioned Human Animation

By

–

07 February 2025 4h45

🔥🔥🔥ByteDance introduces OmniHuman-1: Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Modelshttps://t.co/nGWC8HyoLE https://t.co/CSzIVEFT8i pic.twitter.com/BWwMelPbfq
— 机器之心 JIQIZHIXIN (@jiqizhixin) 7 février 2025

ByteDance introduces OmniHuman-1: Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Models https://
omnihuman-lab.github.io https://
arxiv.org/abs/2502.01061

7 February 2025

LIMO: Less Data Improves Reasoning Performance

By

–

07 February 2025 4h42

LIMO: Less is More for Reasoning With only 817 examples, it outperforms previous models, proving less data can enhance reasoning. https://
arxiv.org/pdf/2502.03387 https://
github.com/GAIR-NLP/LIMO https://
huggingface.co/datasets/GAIR/
LIMO
…