Study routes through real questions
Choose a question you cannot yet answer, then follow the connected lessons. Return to the question after reading the references.
Why is the first token slow?
Separate queueing, preprocessing, prefill, and transfers. Diagnose the stage before changing batch size.
Why are token streams jerky?
Look for long prefills, too much admitted work, network stalls, or gateway buffering.
Why did memory run out?
Rebuild the ledger with real context lengths and concurrency. Distinguish allocated space, active cache, fragmentation, and workspace.
Why did more GPUs barely help?
Inspect the split, per-token collectives, uneven experts, and CPU or network bottlenecks.
How do I reduce the bill?
Try doing less work, sharing repeated work, and choosing the right service class. Keep quality and latency constraints fixed.