This is the whole chapter in one sitting. Larry opened his terminal and went down the software stack — CUDA and kernels, fusion and torch.compile, the file formats and inference-first runtimes, then the serving engines (vLLM, SGLang, TensorRT-LLM), Dynamo at rack scale, and the profiler that tells you what to fix. Twelve questions decide whether it stuck.
Score 80% or better to pass. Every option explains itself, so a wrong pick still teaches — and you can take another pass whenever you like.
01 · The exam
Twelve questions across the chapter.
The CUDA host/device model and the load-compute-store kernel shape, fusion and torch.compile, safetensors and the ONNX/TensorRT trade-off, vLLM's three layers and PagedAttention, SGLang's RadixAttention, TensorRT-LLM's compile-for-throughput bargain, Dynamo's rack-scale role, and measure-before-you-optimise.
Check your understanding0 / 12
Question 1
What is vLLM, in one line?
Question 2
A request flows through vLLM's three layers in what order?
Question 3
What problem does PagedAttention solve for the KV cache?
Question 4
What's the standard production hygiene when deploying vLLM?
Question 5
When does SGLang win over vLLM?
Question 6
RadixAttention speeds up serving by:
Question 7
What trade-off defines TensorRT-LLM?
Question 8
Where does NVIDIA Dynamo sit in the stack?
Question 9
Dynamo's router chooses where to send a request based on:
Question 10
What's the guiding principle of the benchmarking-and-profiling lesson?
Question 11
Why should you run Nsight Compute (ncu) on a canary rather than live production?
Question 12
A team is choosing a serving stack. Which sequence matches the chapter's advice?