cuda 4
- FreeToken: 284B on a Gaming Desktop, 753B on One Workstation GPU - and the Ratio That Makes It Work
- NVIDIA CUDA Tile Architecture: High-Level Tile-Based GPU Programming
- FlexAttention and FlashAttention-3: Unifying Custom Attention in CuTe DSL
- Building an End-to-End LLM from Scratch: Analyzing Karpathy's LLM101n Curriculum