Correctness:

- Copy kernels should use high-performance ldst instructions where bank constraints allow.
- Must ask models to review for weird hacks thoroughly and remove them.
- Also duplicate code.
- Should automate finding tuning parameters at some point.
- Refactor code across modules e.g. effective_memory_elements is in ipu-exchange for some reason.
- Adjust placements to minimize delays in exchange/pairing (has not worked ever).
- Keep cost models updated.
- Does Add work?
- Why so much standard memory reliance? Scaling issue.
- Something something "unmaterializable plan".
- Fix attention (works, slow).
- Paired transfer usability (has not worked ever).
- Is exchange sufficiently dense? Cutover delays?
- Factorize cost models.
- Remove redundant formats.
- Move memory classes somewhere.
- Exchange density?
- Put EXECUTABLE_MEMORY_LIMIT and such in the right place.
- Code sharing with deferred things.
- prepare_attention_blocks is also wrong.
- finish_gemm_plan etc is suspicious
- plans_for_operation maybe duplicates something
- ParallelGridProxy?!
- IPU21_PLANNED_DATA_BYTES factor
- objective duplication
- `let column_grain` is really long. Lots of long chains like this → hard to read.
- Generally wrong abstractions. Trace through code?
- Merge topology and target
- Fix copy descriptors

Completeness:

- LayerNorm.
- MAP head.
- Other ops? Subtract, but that's pretty trivial.
- Inplace, fusion, or something.
- Positional encoding.
- Input projection (probably this is just batched MLP).
- FP8 support!
- Run full model.