Sound came back out — streamed a byte at a time, then without any text in the middle. Then the pixels: the fused kernels that turn seconds into milliseconds, the one weird cache that skips redundant work, and the two moves that make minute-long video tractable. Prove the second half stuck.
Seven questions. Score 80% or better to finish Chapter 6. Every option explains itself, and you can take another pass any time.
01 · Checkpoint
Generation & scale — speech-out, image gen, video gen.
Streaming TTS and speech-to-speech, image-generation kernel optimization and DeepCache, and the video-generation duo: attention-plus-quant and context parallelism.
Check your understanding0 / 7
Question 1
What makes text-to-speech feel real-time in a voice agent?
Question 2
Comparing a cascade (ASR → LLM → TTS) with a native speech-to-speech model, what most decides whether the conversation feels human?
Question 3
How does optimizing image-generation (diffusion) inference relate to LLM optimization?
Question 4
DeepCache speeds up diffusion by exploiting which observation — and what's the catch when stacking it with other tricks?
Question 5
In a video diffusion transformer (DiT), why does attention become the optimization priority when it barely mattered for images?
Question 6
Context parallelism (e.g. ring attention) does what, and when is it needed?
Question 7
You stack FP8 quantization, feature caching, and few-step distillation on a diffusion pipeline. How should you reason about the combined speedup?