2025 Technical Notes（3）
Technical notes during 2025 (3).
Summary: FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
Summary for paper ‘FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness’
Summary: DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving
Summary for paper ‘DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving’
Summary: Fast Inference from Transformers via Speculative Decoding
Summary for paper ‘Fast Inference from Transformers via Speculative Decoding’
Summary: MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation
Summary for paper ‘MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation’
Summary: Infinite Retrieval: Attention Enhanced LLMs in Long-Context Processing
Summary for paper ‘Infinite Retrieval: Attention Enhanced LLMs in Long-Context Processing’
Summary: Efficient Memory Management for Large Language Model Serving with PagedAttention
Summary for paper ‘Efficient Memory Management for Large Language Model Serving with PagedAttention’
Summary: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Summary for paper ‘Incentivizing Reasoning Capability in LLMs via Reinforcement Learning’
TVM: 2D Depth Conv GPU Optimization
This blog demonstrates optimization techniques for 2D depth Convolution in GPU using TVM, including block and thread organization, memory hierarchy exploitation and dimension fuse, etc.
TVM: GEMM GPU Optimization
This blog demonstrates optimization techniques for GEMM in GPU using TVM, including thread organization and memory hierarchy exploitation.