ptv3-p150 / code /scripts /README.md
changh95's picture
tt-model push ptv3-p150 (container)
5662e7f verified
|
Raw History Blame Contribute Delete
2.36 kB

code/scripts

script purpose device
fetch_samples.sh template for downloading non-redistributable sample data (sha256-checked); no entry yet: the T4, PandaSet and nuScenes validation frames live only in the tt-models workspace (research/ptv3/) no
bench.py stage breakdown of warm forwards (pre split / tables / H2D / trace A / trace B / D2H / post / e2e, p50 / p99; b2b replays; AICLK; output checked against the stored CPU reference); --dispatch / --cqs / --variant for the D14 matrix (the card / OPT numbers) yes, via bin/devrun
profile_ops.py per trace one eager forward (module / layer signposts) + one traced replay between signposts under the device profiler (python -m tracy -r -p -v --op-support-count 6000 ...) yes
profile_summary.py the OPT_BASELINE tables from the profile's ops CSV (per trace: modules, layers, op codes, top ops, op-to-op gaps) no
visualizer_report.py ttnn-visualizer memory + graph report (db.sqlite) 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 run: the served model (variant seg_det) on point clouds, per-point outputs + boxes saved per input (device); compare / bev / gif: bird's-eye views with tt_ptv3.viz (p150 vs the CPU reference, the class differences, a sequence as GIF; host only) -- the BEV images of media/ run only
make_goldens.py / make_samples.py CPU-reference goldens for the tests / the samples (the shipped test_pcd.npz from autoware_universe test.pcd; the PandaSet frame into the git-ignored staging_samples_pandaset/) and their stored reference outputs per variant (research venv) no
bringup.py bring-up driver of the port (crop, watcher, eager + 2-call trace, taps vs the reference) yes

Everything here ships in code/ on the Hub and inside the image (source.extra_code lists scripts).