Part 0: Foundations

The model side of the tracer and, from Pass 2, the autograd engine every later model trains with. Pass 1 has one chapter: a byte-level bigram fitted by counting, whose logits use NumPy matmul and whose weights you write as a safetensors checkpoint the Rust engine serves. Pass 2 retrains the same bigram with your autograd (L0.5 takes over bigram.py) behind the same checkpoint contract.

Course passes: 1 (L0.0, gate MS-P1), 2 (L0.1 to L0.6, gate MS-P2).

Before you start: Python and NumPy and matrix multiplication in Python.

#	Module	Chapter	Kind	Pass
1	L0.0	Byte bigram: counts, logits, safetensors v0	build	1
2	L0.1	Tensor, broadcasting backward, and no_grad	build	2
3	L0.2	The op library and gradcheck_all	build	2
4	L0.3	Losses with a fused backward	build	2
5	L0.4	Module system and basic layers	build	2
6	L0.5	Training loop and the autograd bigram	build	2
7	L0.6	Safetensors for every dtype, atomic checkpoints, and the token stream	build	2