Blog
-
GPU Architecture and Optimization for Large Language Models
An overview of GPU architecture and practical optimization techniques for training and serving large language models.
-
A Simple Summary of Modern Long Video Understanding and Memory System
A simple summary of long video understanding and memory systems.
-
Implementing AffectGPT-R1: A Reasoning Model for Multimodal Emotion Recognition
The cold start stage uses large scale, coarse grained descriptive emotion data so that the model first acquires basic multimodal emotion understanding and output format alignment. It also provides a stable initialization for later reinforce
-
An Brief Introduction to GRPO
A practical introduction to Group Relative Policy Optimization (GRPO), covering why it removes the critic from PPO-style LLM training, how group-relative advantages are computed, and how to implement the core idea with TRL and PyTorch.
-
A Brief Survey on Training-Free Long Video Generation
This is a research report about training-free long video generation, including four main sections: 1) A general review of current development of training-free long video generation from different aspects. 2) A summary of base models and benchmarks. 3) Discussion, limitation and potential works. 4) Conclusion