Input tokens become token embeddings plus positional information, then flow through a stack of decoder blocks, and finally output logits used for next-token prediction. Each block contains causal self-attention, a feed-forward network, residual connections, and layer normalization — there's no separate encoder, unlike translation-style architectures.

The full path is: input tokens → token embeddings + positional information → stack of decoder transformer blocks → output logits → next-token prediction. Unlike encoder-decoder models used for translation (like the original Transformer or T5), GPT-style models drop the encoder entirely and use only a stack of decoder blocks with causal attention, which naturally models P(next token | previous tokens) — exactly what generation needs. This architecture is what powers chat, completion, summarization, coding, reasoning, and tool-calling models, since all of those tasks reduce to "generate the next token given everything so far."