THE LEARNING PATHS

Build intuition.
Then put it to work.

Follow a sequence through the published library, or rehearse the 25-problem interview review. Work each problem before opening its answer, then use the supporting concepts and experiments to check your reasoning.

Check your starting point

FROM SYMPTOM TO EVIDENCE

Performance & debugging

Decompose latency, challenge benchmark assumptions, and turn a regression into a testable explanation.

12 steps · First concept free · Access shown at each step

Start this path See the steps

Bring with you: Experience running or profiling a model-serving workload.
Work toward: Diagnose latency across serving stages and separate encoder memory from decoder capacity.
A good fit if: You have run or profiled a serving workload and want to turn symptoms into evidence an interviewer can check.
Not in this path: Designing a fleet or writing a kernel from scratch; this path starts from a running system.

Check your understanding 4 tasks to try without the answer

Split a TTFT figure into queueing, prefill and first-token sampling, and name the clock each part is read from.
Redo if You cannot say which part your proposed fix would change.
For a symptom, name the one measurement that separates its two likeliest causes before proposing a fix.
Redo if Your first step is a configuration change.
Report a speedup with the run-to-run spread you measured on the unchanged configuration.
Redo if You compared one run of each.
Recompute a speculative acceptance rate from engine counters and say what resets them.
Redo if Your numerator and denominator come from different scopes.

HOW TO REHEARSE A WORKED ANSWER

Read the question and stop. Write down what you would ask the interviewer, because the missing input is often the point.
Write the chain on paper: inputs with their units, the substitution, the result and one line that checks it is plausible.
Open the answer and find the first place your reasoning diverges. That place matters more than whether the final number matched.
Grade yourself against the Weak, Hire and Strong Hire lines under “How it’s scored”, then fix the step that cost you the grade.
Change one input, such as the GPU, the precision or the context length, and predict the new answer before you recompute it.

A NOTE ON THE CURRICULUM

These are curated study paths through existing material, with access shown at each step. For dedicated lessons, exercises and progress tracking, explore the published courses.

Explore the courses

Build intuition.
Then put it to work.

Performance & debugging

Latency Decomposition: Queue, Prefill, Decode

The Roofline Model for LLM Inference

After a prompt-template deploy, p99 TTFT went from about 100 ms to over a second while GPU utilization did not move. Debug it.

Speculative decoding

You turned on speculative decoding in production. What do you measure to prove it helps, and what is the right baseline to compare against?

Inference Benchmarks That Do Not Lie

Structured output slowed generation. Is the cause grammar compilation, mask construction, or host scheduling?

MoE decode slowed after the traffic mix changed. Are hot experts or communication causing the delay?

Prefill/decode disaggregation increased latency. Is KV transfer, admission, or token delivery on the critical path?

Image requests OOM while the KV pool is mostly free. Separate encoder peaks from decoder memory and choose a fix.

Prefix-Aware and KV-Aware Routing

A router raises prefix-cache hits but worsens tail latency. Replay its decisions and find the missing state.