EDGE FDE
Advanced PremiumAI Full-Stack Operator: Train, Convert, Ship On-Device
Build progressively realistic replicas of the whole stack until you can reason about tokens, throughput, and hardware without hand-waving.
Explain transformer internals, train small LMs, run SFT and preference tuning, serve with vLLM, write basic Triton, and hold earned opinions on clusters and the semiconductor supply chain.
You do not need to pretrain an 800B model to understand the shape of the problem. You do need to build progressively realistic replicas of the whole stack.
What you'll learn
- Easy: the stack has a shape. Linear maps, a decoder, a train loop, a serve loop, a cluster, and a hardware chain. Name each layer before you rent a GPU.
- Hard: train 10M to 100M, fine-tune, preference-tune, serve with vLLM, write a Triton kernel, and estimate KV bytes and checkpoint size.
- God: earned opinions. You can walk from a projection identity to PagedAttention to HBM and power without inventing a statistic.
Weekly rhythm
- Monday: read and prove the claim for the week.
- Tuesday and Wednesday: code the replica.
- Thursday: finish the assignment.
- Friday: write the opinion checkpoint in your own words.
- Weekend: optional stretch. Do not collect tabs.
Core projects
- Linear algebra and optimiser lab (weeks 1 to 2).
- TinyGPT from scratch (week 3).
- Pretrain a 10M to 100M LM (week 5).
- DDP / FSDP / ZeRO memory simulator (week 6).
- SFT plus typed synthetic data (weeks 4 and 7).
- DPO on a small preference set (week 8).
- vLLM serve, Triton kernel, and the capstone write-up (weeks 9 to 12).
Environment setup
Use Python 3.11 or newer and uv. Weeks 1 to 8 run on a laptop CPU (Mac or Linux). Add packages in phases so later adds keep the CPU torch wheels. Use --index, not --index-url. That flag is not saved, and a later uv add can swap torch to a CUDA wheel. --index writes the CPU index into pyproject.toml.
vLLM, Triton, and CUDA wheels are for later GPU weeks (9 and 10). They may need a CUDA host. Do not install them in this CPU project. Do not pretend they install on CPU-only machines.
curl -LsSf https://astral.sh/uv/install.sh | sh
mkdir -p ai-full-stack-operator && cd ai-full-stack-operator
uv init --python 3.11
uv venv
uv add numpy
uv add torch --index pytorch-cpu=https://download.pytorch.org/whl/cpu
uv add pydantic
uv run python -c "import numpy, torch, pydantic; print(numpy.__version__, torch.__version__, pydantic.__version__); print('cuda:', torch.cuda.is_available())"Week 1 stays in NumPy first. Week 2 uses torch. Week 4 uses pydantic. Always run with uv run.
Suggested repo structure
ai-full-stack-operator/
week01_linalg/
week02_optim/
week03_tinygpt/
week04_data/
week05_pretrain/
week06_distributed/
week07_sft/
week08_preference/
week09_vllm/
week10_triton/
week11_cluster/
week12_capstone/
shared/
train_step.py
synthetic.pyLessons
- Proof-based linear algebra for deep learning
- Non-convex optimisation
- Build a transformer from scratch (TinyGPT)
- Scaling laws, data, and synthetic data
- Pretrain a small language model
- Distributed training simulation
- Supervised fine-tuning
- Preference tuning, PPO, DPO, and RLHF intuition
- vLLM internals and production serving
- Triton and hand-rolled kernels
- Cluster orchestration
- Semiconductor supply chain and capstone
Key concepts
- Replicas beat tourism. You do not need an 800B pretrain to see the shape.
- A decoder is linear maps, a causal mask, and residuals you can draw.
- Train, SFT, and DPO are the same next-token machine with different data contracts.
- Serving is KV cache, batching, and tensor parallel. Systems are failure modes. Hardware is HBM, packaging, network, and power.
Takeaways
- You can explain transformer internals without hand-waving.
- You can train and fine-tune a small LM and talk about instability with logs.
- You can serve with vLLM, write a basic Triton kernel, and size a cache.
- You leave with earned opinions, not rented slides.
Agent briefs
- Course SKILL.md
- 01-proof-based-linear-algebra.llms.md
- 02-non-convex-optimisation.llms.md
- 03-tinygpt-from-scratch.llms.md
- 04-scaling-laws-and-synthetic-data.llms.md
- 05-pretrain-a-small-lm.llms.md
- 06-distributed-training-simulation.llms.md
- 07-supervised-fine-tuning.llms.md
- 08-preference-tuning-and-rlhf.llms.md
- 09-vllm-internals-and-serving.llms.md
- 10-triton-and-hand-rolled-kernels.llms.md
- 11-cluster-orchestration.llms.md
- 12-supply-chain-and-capstone.llms.md
Related courses
Model architectures in plain English · LLM LoRA for your writing style · Turn feedback into a better local model · The agent harness (and how to improve it)