Build it: one codebase, model by model. Each stop shows only what changed.
The 2017 encoder-decoder, post-norm.
Decoder-only, weight tying, real weight names.
A trainer later stops can reuse.
Pre-norm, RMSNorm, SwiGLU, RoPE, GQA, and a KV cache.
A sliding window, then a mixture of experts.
MLA, DeepSeekMoE, multi-token prediction.
Linear attention and a hybrid stack.
Two encoders and a contrastive loss.
A prefix-LM mask over image and text tokens.
Walkthroughs merged from Deep Cuts were inspired by Umar Jamil.