# code/scripts

| script | purpose | device |
|---|---|---|
| `make_sample.py` | generates the SHIPPED samples (`samples/synthetic_<variant>.npz`, a deterministic synthetic scan per serve profile; no dataset involved), their calibration presets and their stored fp32 CPU-reference outputs (ONNX Runtime) | no |
| `bench.py` | stage breakdown of a warm forward (pre / H2D / trace / back-to-back / D2H / unsort / post / e2e; AICLK; a check of each output against its stored CPU reference); the card / OPT numbers | yes, via `bin/devrun` |
| `profile_ops.py` | one eager forward with module / layer signposts + one traced replay between signposts under the device profiler (`python -m tracy -r -p -v ...`) | yes |
| `profile_summary.py` | per-module / per-layer / per-op tables from the profiler CSV (`OPT_BASELINE.md`) | no |
| `visualizer_report.py` | ttnn-visualizer memory + graph report of one eager forward (not for timing) | yes |
| `container_smoke.sh` | serve ONE serve profile of the built package (`--profile NAME`; default profile otherwise), run `server/smoke_test.py` against it (asserts ETH dispatch, the 12x10 grid and the profile's pins; compares with the stored CPU reference), keep the evidence (container log, `/info`, the `/predict` output and the result, in `logs/smoke/` of the repo or `--log-dir DIR`), always stop it; one `bin/devrun -t 3600 -k 150` window per profile | yes |
| `make_demo.py` | workspace tool: runs the production API on the shipped samples and the public-dataset frames of `research/frnet/public_data` (`run`, device) and renders the card's `media/` (`render`, CPU) | `run`: yes |
| `make_samples.py` | workspace tool: rebuilds the STAGED PandaSet 019 frame 40 sample (git-ignored `staging_samples_pandaset/`, not shipped) and its CPU references | no |
| `make_goldens.py` | workspace tool: the large module-tap goldens and the host pre-processing hashes of the tests | no |
| `precision_study.py`, `diag_conv_precision.py`, `diag_frame_taps.py` | precision studies of the port: the CPU emulation of the TT numerics, and device diagnostics (per-unit conv precision; per-module device-vs-emulation error on one frame) | the `diag_*` scripts: yes |
| `fetch_samples.sh` | downloads public sample data that may not be redistributed (sha256-checked); no entry is active in this release | no |

Everything here ships in `code/` on the Hub and inside the image (`source.extra_code` lists `scripts`). The workspace
tools need the tt-models research tree (`research/frnet/`) and the raw datasets; they document how the card's
numbers and renders were made but do not run from a Hub clone.