LLM推理系统问答
Explain why FlashAttention is faster than standard attention. What is the key bottleneck it addresses?
原题What is kernel fusion and why does it improve inference performance?
原题How does the online softmax trick enable tiled attention computation?
原题What improvements did FlashAttention-3 introduce for the H100, and how does it exploit H100-specific features?
原题When would you use Triton instead of writing raw CUDA? What are the trade-offs?
原题Explain continuous batching. Why is it superior to static batching for LLM serving?
原题What is “prefill piracy” and how does chunked prefill address it?
原题Describe the architecture of vLLM. What are its key innovations and where are its limitations?
原题How does disaggregated prefill/decode serving work? What are the engineering challenges?
原题What metrics would you use to evaluate and compare two LLM serving systems?
原题
暂无题目