Read with a question in mind
These explanations connect a visual model to an engineering mechanism. Each linked lesson tells you exactly what to look for in the reference.
How does an error reach early layers?
Follow local sensitivities, not a search for one guilty neuron.
What does attention actually mix?
Follow the query, key, and value for one token.
Why does decode wait on memory?
Count how often weights and historical KV are read, then ask what a larger batch reuses.
Book chapter · How To Scale Your Model · Transformer inference ↗
Which optimization matches the symptom?
Use the scheduler, cache, graph, and parallelism guidance as experiments rather than universal settings.
- Chunked prefill: stop one reader blocking every writer →
- Prefix caching: reuse the shared first pages →
- Fusion and CUDA graphs: less dispatch, fewer round trips →
Documentation · vLLM · Optimization and tuning ↗
When should a model choose its next action?
Compare deterministic workflows with agent loops and identify the extra failure paths.
Engineering article · Anthropic · Building effective agents ↗