
Custom CUDA Kernels for LLMS: From Theory to Production
Expert guide to writing custom CUDA kernels for LLM inference and training: GPU architecture fundamentals, kernel fusion strategies, Flash Attention… From the article "Custom CUDA Kernels for LLMS: From Theory to Production" by Synor, published on Misar.Blog.
In this episode:
0:00 Custom CUDA Kernels Are Presented as the Main Reason LLM Inference
3:32 The GPU Memory Hierarchy Matters Because HBM Is Far Slower
5:47 Kernel Fusion Is the Article’s Single Most Important Optimization
7:53 Flash Attention Improves Long-Context Attention by Tiling Q/K/V
9:39 Tiled Matrix Multiplication Turns the…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.