prefill vs decode, TTFT, ITL, throughput, goodput
Used in: Day 1 · Step 1 Day 1 · Step 3 (rewatch)
Videos
Every video, in the order the course plays them. Each one plays inside its step; here you can find one again.
Day 1 · Guardrails and first calls
Section titled “Day 1 · Guardrails and first calls”Day 2 · Local stack on the Mac
Section titled “Day 2 · Local stack on the Mac”
what an inference server does, containers, key flags
Used in: Day 2 · Step 1 Day 9 · Step 3 (rewatch)
Day 3 · Benchmark like a field engineer
Section titled “Day 3 · Benchmark like a field engineer”
continuous batching, chunked prefill, speculation
Used in: Day 3 · Step 1 Day 9 · Step 2 (rewatch)
roofline, decode ceiling = bandwidth ÷ bytes
Used in: Day 3 · Step 1 Day 4 · Step 1 (rewatch)
Day 4 · Memory levers
Section titled “Day 4 · Memory levers”
Day 5 · Agent workloads
Section titled “Day 5 · Agent workloads”
Day 6 · Structured output and batch
Section titled “Day 6 · Structured output and batch”
serving layers, capacity modes, maturity curve
Used in: Day 6 · Step 2 Day 10 · Step 1 (rewatch)
Day 7 · Training day, managed
Section titled “Day 7 · Training day, managed”
4-bit base + adapter on your own box
Used in: Day 7 · Step 2 Day 8 · Step 1 (optional)
Day 9 · Scale-out
Section titled “Day 9 · Scale-out”
Day 11 · Training methods in miniature
Section titled “Day 11 · Training methods in miniature”
training loop, 16 bytes/param, ZeRO/FSDP, when full FT pays
Used in: Day 11 · Step 1
Parameter-Efficient Fine-Tuning
LoRA maths, QLoRA, DoRA, adapters, IA³, multi-LoRA
Used in: Day 11 · Step 1
Day 12 · Preference tuning
Section titled “Day 12 · Preference tuning”
Direct Preference Optimization
the DPO loss, a worked step, pairs from product data
Used in: Day 12 · Step 2