Context
AGENTS.md step 5 says to use the same tiny causal model as the Phase 3 sharding workload before scaling. Phase 3 Sections 2–4 sketch mesh kinds, GSPMD propagation, and collective coercions. The compiler has none of those.
A first @mesh / @sharded implementation would invent
surface tokens and typed-IR fields that Section 2 has not
locked against the bootstrap .kina grammar.
Column-parallel lm_head is already expressible:
slice the vocab axis, two matmuls, concat the
shards. That is the Section 4 all-gather written as
existing ops.
Decision
Specify the first Phase 3 workload as the existing
composition in
05-storylm0-workload.md:
2-way column-parallel lm_head on storylm0_forward
widths. No new operator. Do not add @mesh or @sharded
in this slice. Do not scale to the ~100M exit demo.
Consequences
storylm0_tp_vocab.kinaand a.sonumeric check land without a new compiler operator.- Visitors can see vocab-parallel logits that match the unsharded forward; they still cannot parse a mesh or run the 4-device simulator.
- Head-parallel attention is Section 6. Projection shards are Section 7. Mesh syntax, GSPMD, and collective lowering remain follow-on work.