Kit README (offline start)

01 Watch18 videosabout 57 minutes 02 ReadLab bookplus the field guide 03 Run19 lessons00 to 18, with code 04 SourceVideo sourceRemotion + Kokoro

Hands-on material for learning AI engineering and field engineering, in one folder. It has three parts that work together, plus the source for the videos:

fireworks-field-engineer-kit/
├── 01-videos/        18 narrated explainers, numbered 01–18 (watch)
├── 02-lab-book/      the lab book web page + the earlier field guide (read)
├── 03-labs/          the code: lessons 00–18, one folder per lesson (run)
└── 04-video-source/  Remotion + Kokoro project that made the videos (optional)

The loop for every topic: watch the video (3–4 min) → read the matching lab card in the lab book → cd 03-labs/lessons/NN-…, read its README, run the scripts. Your numbers land in 03-labs/results/, and lesson 15 turns them into a customer sizing memo.

---

Get started in 5 steps (about 15 minutes, nothing to download)

1. Open the lab book. Double-click 02-lab-book/field-engineer-lab-book.html. Keep it on the left of your screen.
2. Watch 01-videos/01-inference-101.mp4.
3. Set up the code in a terminal on the right:

   cd fireworks-field-engineer-kit/03-labs
   python3 -m venv .venv && source .venv/bin/activate
   pip install -r requirements.txt && pip install -e .
   cp .env.example .env

4. Check it all works offline, using a built-in fake model server:

   make mock &          # fake LLM server on :9000
   make smoke           # runs every lesson end to end, ~2 min, should end "✓ smoke test passed"

5. Start lesson 00: bash lessons/00-setup/setup_mac.sh installs Ollama, llama.cpp and MLX. Then follow the table below.

Fireworks costs money only in the lessons marked $. Run make fw-check after each one to confirm nothing is still billing.

---

The course, in order

Watch the video(s), then do the lesson. Lesson folders are in 03-labs/lessons/.

Part 1: Serving and inference (days 1–10)

day	watch (01-videos/)	then do (03-labs/lessons/)	lab card	cost
1	01-inference-101.mp4 · 09-serving-stack.mp4 (first half)	00-setup → 01-latency-harness	L1 · F1	~$0.02
2	09-serving-stack.mp4	02-three-local-servers	L2	free
3	10-benchmarking.mp4 · 05-batching.mp4 · 03-gpu-bandwidth.mp4	03-concurrency-sweep	L3	free
4	04-quantization.mp4 · 02-kv-cache.mp4	04-quantization → 05-kv-cache	L4 · L5	free
5	11-prefix-caching.mp4	06-prefix-caching	L6 · F2	~$0.03
6	08-platform.mp4	07-structured-output → 08-batch-api	F3 · F4	~$0.03
7	06-training.mp4 · 12-qlora.mp4	09-eval-harness → 10-lora-fireworks	F6 · F5	$ ~1 GPU-h
8	12-qlora.mp4	11-lora-local-mlx	L8	free
9	07-offloading.mp4 · 13-scale-out.mp4	12-moe-vs-dense → 13-speculative-decoding → 14-spark-track	L7 · F7 · L9	$ ~½ GPU-h
10	08-platform.mp4 (again)	15-capstone	Capstone	free

Part 2: How models are trained (days 11–12)

day	watch (01-videos/)	then do (03-labs/lessons/)	lab card	cost
11	14-full-finetuning.mp4 · 15-sft.mp4 · 16-peft.mp4	16-training-toy (part 1) → 17-full-vs-peft-mlx	T1 · T2	free
12	17-rlhf.mp4 · 18-dpo.mp4	16-training-toy (part 2) → 18-preference-dpo	T1 · T3	free · $ optional

---

The 18 videos

Numbers match the “Video N of 18” on each title card and the “Watch:” line in each lesson README.

#	file	min	what it explains	run next
01	01-inference-101.mp4	3:02	prefill vs decode, TTFT, ITL, throughput, goodput	lesson 01
02	02-kv-cache.mp4	3:02	KV cache maths, PagedAttention, FP8 KV	lesson 05
03	03-gpu-bandwidth.mp4	3:10	roofline, decode ceiling = bandwidth ÷ bytes	lessons 03–04
04	04-quantization.mp4	2:43	FP16 → FP8 → 4-bit, NVFP4, KV quantization	lesson 04
05	05-batching.mp4	2:58	continuous batching, chunked prefill, speculation	lessons 03, 13
06	06-training.mp4	3:14	overview: SFT, LoRA, DPO, RFT and how to choose	lessons 10–11
07	07-offloading.mp4	3:07	sizing, offloading, MoE, tensor vs pipeline parallel	lesson 12
08	08-platform.mp4	2:51	serving layers, capacity modes, maturity curve	lesson 15
09	09-serving-stack.mp4	2:56	what an inference server does, containers, key flags	lessons 02, 14
10	10-benchmarking.mp4	2:22	traffic profiles, concurrency sweeps, percentiles	lesson 03
11	11-prefix-caching.mp4	2:37	shared prefixes, RadixAttention, prompt caching	lesson 06
12	12-qlora.mp4	2:49	4-bit base + adapter on your own box	lessons 11, 10
13	13-scale-out.mp4	2:41	speculative decoding, acceptance rate, two boxes	lessons 13–14
14	14-full-finetuning.mp4	4:17	training loop, 16 bytes/param, ZeRO/FSDP, when full FT pays	lesson 17
15	15-sft.mp4	4:02	data format, loss masking, data volume, loss curves	lesson 17
16	16-peft.mp4	3:50	LoRA maths, QLoRA, DoRA, adapters, IA³, multi-LoRA	lesson 17
17	17-rlhf.mp4	4:10	reward model, PPO, KL leash, reward hacking, GRPO/RFT	lesson 16
18	18-dpo.mp4	3:50	the DPO loss, a worked step, pairs from product data	lesson 18

Total ≈ 57 minutes. The subtitles are burned in.

The 19 lessons (03-labs/lessons/)

#	folder	what you do	runs on
00	00-setup	install Ollama, llama.cpp, MLX; Fireworks guardrails	Mac
01	01-latency-harness	measure TTFT and ITL; the harness every later lesson reuses	Mac → FW
02	02-three-local-servers	one model served by Ollama, llama.cpp and MLX	Mac
03	03-concurrency-sweep	the throughput vs TTFT curve that sizes a deployment	Mac
04	04-quantization	Q8 vs Q4 vs Q3: speed, ceiling maths, quality check	Mac
05	05-kv-cache	KV cache maths, then break and fix long context	Mac
06	06-prefix-caching	shared prefixes: the ~20× TTFT win, and the anti-pattern	Mac → FW
07	07-structured-output	JSON validity: prompt vs json_object vs json_schema	Mac → FW
08	08-batch-api	evals at half price with the Batch API	FW
09	09-eval-harness	quality + p95 + $/1k tasks in one table, LLM judge	Mac → FW
10	10-lora-fireworks	a real LoRA fine-tune: train → deploy → eval → delete	FW $
11	11-lora-local-mlx	the same LoRA on your Mac (QLoRA)	Mac
12	12-moe-vs-dense	infer MoE active parameters from decode speed	Mac
13	13-speculative-decoding	draft models, acceptance rate, when it hurts	Mac → FW $
14	14-spark-track	vLLM and SGLang on the DGX Spark, same harness	Spark
15	15-capstone	build the customer sizing memo from your results	anywhere
16	16-training-toy	full vs LoRA, SFT → reward model → RLHF vs DPO, in numpy	anywhere
17	17-full-vs-peft-mlx	full vs LoRA vs DoRA on a real model; data checker	Mac
18	18-preference-dpo	DPO on your Mac (and optionally Fireworks)	Mac → FW

Every lesson folder has a README.md with the same parts: what and why, the code to read first, commands to run, expected output, self-check questions and a short explanation of the result.

What’s in 02-lab-book

• field-engineer-lab-book.html is the main companion page. It covers the role, what Fireworks sells, spend control, all lab cards (F1–F7, L1–L9, T1–T3, each naming its lesson folder), calculators, your handbook mapped to practice, the 12-day plan and talk tracks. It works offline; a hosted copy is also in your Claude artifacts.
• inference-field-guide.html is the earlier visual field guide, with interactive explainers and the frameworks chapter.

What’s in 04-video-source (optional)

This is the Remotion project that rendered the videos. Use it to change a script, the voice or the pacing. On your Mac:

cd 04-video-source && npm install && npx remotion browser ensure
pip install mlx-audio misaki soundfile numpy
python scripts/tts_mlx.py                 # regenerate narration with Kokoro on MLX
npm run render                            # → out/<id>.mp4

If something doesn’t work

symptom	fix
make smoke says connection refused	the mock isn’t running: make mock & first
a lesson says Connection refused	that server isn’t running; make check shows what is up
a Fireworks model id is not found	ids change; copy a current one from the model library and pass --model
page shows odd symbols	open the .html in a browser, not a text editor
unsure whether you’re still paying	make fw-check lists anything still billing