Inside one Qwen-style decoder block Diagram

Learned projections create queries, keys and values. RoPE supplies position information to queries and keys; the causal mask blocks future positions.
Attention mixes permitted value vectors, then an output projection returns to the residual width. Grouped-query attention shares KV heads across groups of query heads.
A checkpoint may also normalize queries and keys. Read the actual model implementation rather than assuming every decoder is identical.

Save the incoming block vector. Add the attention output to that saved vector.
Save that sum. Add the MLP output to it. Side wires are omitted from this main-branch schematic, not from the actual computation.

The MLP transforms features within each token position; attention is the part that mixes information between positions.
RMSNorm regulates vector scale. Neither normalization nor attention is a separate training phase.

Repeat blocks for the checkpoint's layer count, then apply final normalization and the vocabulary head.
This is a Qwen-style teaching schematic, not the exact architecture of every transformer or a claim about Jev internals.