@antiAIvo: I've gone through the basics of the Transformer architecture, inference, and training process. Next, I'll start a new l…

I've gone through the basics of the Transformer architecture, inference, and training process. Next, I'll start a new learning track: "Large Model Inference Optimization" This track is a complete causal chain, spanning 6 parts in total: 1. KV Cache (Part 1): Saving on repeated computations 2. KV Cache (Part 2): The memory cost 3. Continuous Batching: Scheduling without waste 4. PagedAttention: How vLLM manages memory 5. RadixAttention: How SGLang reuses prefixes 6. vLLM vs SGLang: A side-by-side comparison to wrap it up Why choose this track? Because it's the intersection of recommendations from multiple AI analyses It's a causal chain: → Saves on repeated computations → Which means facing memory costs → Memory pressure gives rise to scheduling optimizations → After scheduling, you still need to solve how to allocate memory → After allocation, you need to address repeated prefixes across multiple requests → Finally, a comparison of the two technical paths in practice This is also currently one of the most cutting-edge and actively used layers of technology in large model engineering deployment—not just concepts on paper Because of my lack of knowledge in this area I don't yet have a global perspective So I'm temporarily setting this track After gradually learning, things should become clear ahead Of course, it will change midway based on learning progress