| """In-process A/B of the served call: blocking vs non-blocking trace replay (TtHamer.nonblocking_replay), |
| interleaved in blocks so host-load drift hits both equally. usage: ab_replay.py [N]""" |
| import os, sys, time, json, statistics |
| from pathlib import Path |
| _REPO = Path(__file__).resolve().parents[1] |
| sys.path.insert(0, str(_REPO)) |
| os.environ.setdefault("TT_FUSED_HASH", "0") |
| import torch |
| from hamer import Hamer, build_from_reference, load_hamer_checkpoint |
| from hamer.preprocess import preprocess |
| N = int(sys.argv[1]) if len(sys.argv) > 1 else 400 |
| hf = Path.home() / ".cache/huggingface/hub/models--changh95--hamer-weights/snapshots/e8217b46d54b4e8afaaf0dd6efb6b62341c97d32" |
| ref = Hamer().eval(); load_hamer_checkpoint(ref, hf / "hamer.ckpt", hf / "mano_mean_params.npz") |
| media = _REPO.parent / "media" |
| x, _ = preprocess(media / "sample1.jpg", bbox_xyxy=(20, 170, 295, 260), is_right=True) |
| tt = build_from_reference(ref, device_id=0, strict=True) |
| res = {True: [], False: []} |
| with torch.no_grad(): |
| tt(x); tt(x) |
| o = {m: None for m in res} |
| for blk in range(N // 20): |
| for mode in (True, False) if blk % 2 == 0 else (False, True): |
| tt.nonblocking_replay = mode |
| for _ in range(10): |
| t0 = time.perf_counter(); out = tt(x); res[mode].append((time.perf_counter() - t0) * 1e3) |
| o[mode] = out |
| print("identical outputs:", torch.equal(o[True], o[False])) |
| for m, v in res.items(): |
| print(f"AB nonblocking={m}: median {statistics.median(v):.3f} ms min {min(v):.3f} p90 {sorted(v)[int(0.9 * len(v))]:.3f} n={len(v)}") |
| tt.close() |
|
|
