# Phase 121 completion evidence

The final six open locally actionable items were completed on 2026-07-13.

- **121.3.2.5–121.3.2.6 — MACE foundation loaders:** the MACE-MP-0 small and
  medium and MACE-OFF23 small and medium loaders now consume strict portable
  PyTorch state-dict exports, restore every inference tensor, reject mismatched
  architecture/variant/dimensions and reproduce the source model prediction
  exactly after JSON and byte serialization.
- **121.3.6.6 — distributed training:** the Python training runtime supports
  `torchrun` rank discovery, NCCL/Gloo initialization, one-process-per-GPU DDP,
  synchronized BatchNorm, unique-rank distributed sampling, `no_sync` gradient
  accumulation (including a correctly scaled partial final window), mixed
  precision with gradient scaling, globally reduced metrics and FSDP weight-
  matrix/model sharding. CPU inference hosts receive an actionable lazy-
  dependency error rather than failing package import.
- **121.5.1.2 — CHGNet:** the MPtrj 0.3.0 loader restores the graph backbone,
  energy head and charge head, including total-charge configuration.
- **121.5.2.2 — M3GNet:** the MP 2021.2.8 loader accepts Kalika and MatGL
  `model.`-prefixed state-dict names and restores all node, edge, angle and
  energy-head tensors.
- **121.5.3.2 — ALIGNN-FF:** both JARVIS energy-only and energy+force checkpoint
  variants are supported. The force variant exposes energy-conserving Cartesian
  forces; the energy-only variant rejects force requests explicitly.

Official multi-megabyte model weights are intentionally not committed. The
loaders accept JSON state-dict exports so applications can retain the upstream
model license and artifact distribution while obtaining deterministic, validated
local inference.

## Automated verification

- `pnpm exec vitest run --config libs/kalika/ml-potentials/vitest.config.ts --maxWorkers=2`
  — 55 files and 337 tests passed (full library).
- Focused loader suites — 3 files and 16 tests passed.
- `pnpm exec tsc -p libs/kalika/ml-potentials/tsconfig.typecheck.json` — passed.
- `pnpm exec eslint 'libs/kalika/ml-potentials/**/*.ts' --quiet` — passed; seven
  pre-existing lint errors in the phase library were repaired.
- `PYTHONPATH=libs/kalika/ml-potentials/python/src python3 -m pytest libs/kalika/ml-potentials/python/tests -q`
  — 4 passed.
- `cargo test` in `libs/kalika/ml-potentials/rust` — 17 passed.

The configured Nx production build remains blocked before target emission by
pre-existing `TS6307` declaration file-list errors imported from
`@kalika/crystallography`. The phase library's strict standalone typecheck and
all runtime tests pass; no phase-121 TypeScript diagnostic is present.
