code/scripts
| script | purpose | device |
|---|---|---|
fetch_samples.sh |
downloads the public sample data that is not redistributable (sha256-checked) | no |
bench.py |
stage breakdown of warm forwards (host pre-processing with decode / squash split, staging copy, H2D, one blocking replay, D2H with copy / host-conversion split, host post-processing, e2e p50 / p99, back-to-back replays), AICLK / power sampling, per-input check against the stored fp32 CPU reference; --variant, --dispatch, --cqs, --input / --golden, --json (the card / OPT_BASELINE numbers) |
yes, via bin/devrun |
profile_ops.py |
one eager forward (with m:<block> / c:<layer> signposts) + one traced replay between the eager / trace signposts under the device profiler (python -m tracy -r -p -v --op-support-count 6000 ...: a frame is more than the default 1000 programs) |
yes |
profile_summary.py |
CPU-only summary of a profile_ops.py ops CSV: programs per frame, kernel sum vs span, op-to-op gaps, tables per category / block / op code / layer, top ops, the per-layer cost split (compute, conv data movement, activation split, fp32 sum / bias / cast) (the OPT_BASELINE.md profile tables) |
no |
visualizer_report.py |
ttnn-visualizer memory + graph report (db.sqlite, config.json) of one eager forward; never 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 |
renders media/*.png (BEV / overlay views) from the shipped sample |
yes |
ref_golden.py |
goldens of the fp32 CPU reference (tt_sceneseg.reference): small ones in tests/goldens/, the smoke references samples/highway_normal_1.<variant>.reference.json, large ones (module taps, public frames, comma10k with the training ROI crop) in the workspace's research/sceneseg/goldens/ (research venv, CPU) |
no |
make_weights_upload.py |
stages the re-hosted weights for changh95/sceneseg-p150 weights/ (PLAN.md D10): the deployed ONNX byte for byte, the .pth as bit-identical safetensors, LICENSE / NOTICE, provenance.json (Drive ids, sha256), in the git-ignored weights_upload/ with UPLOAD.md for the orchestrator (research venv, CPU) |
no |
Everything here ships in code/ on the Hub and inside the image (source.extra_code lists scripts).