| """Round 14: why the server's model call (serve_gate model p50) is slower than bench's back-to-back served call. |
| Times TtHamer.__call__ (fresh u8 crop each call) after different host-side gaps: none (bench), a busy-wait gap, a |
| sleeping gap, and the server's own work in between (decode + native preprocess of a media frame). Run with the serve env |
| (eth dispatch, 12x10).""" |
| import json, os, statistics, sys, time, io |
| from pathlib import Path |
| _REPO = Path(__file__).resolve().parents[1] |
| sys.path.insert(0, str(_REPO)) |
| os.environ["HAMER_DISPATCH"] = "eth" |
| import torch |
| from PIL import Image |
| from hamer import Hamer, build_from_reference, load_hamer_checkpoint |
| from hamer.preprocess import preprocess_pil |
|
|
| 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") |
| tt = build_from_reference(ref, device_id=0, strict=True) |
| media = _REPO.parent / "media"; boxes = json.loads((media / "bboxes.json").read_text()) |
| reqs = [((media / n).read_bytes(), e["bbox"], bool(e.get("is_right", 1))) for n, es in sorted(boxes.items()) for e in es] |
| xs = [preprocess_pil(Image.open(io.BytesIO(r[0])), r[1], is_right=r[2], raw=True)[0] for r in reqs] |
| with torch.inference_mode(): |
| tt(xs[0]); tt(xs[1]); assert tt.trace_active |
| N = int(os.environ.get("N", "200")) |
|
|
| def run(mode, gap_ms=0.0): |
| ts = [] |
| with torch.inference_mode(): |
| for i in range(N): |
| if mode == "busy": |
| e = time.perf_counter() + gap_ms / 1e3 |
| while time.perf_counter() < e: pass |
| elif mode == "sleep": |
| time.sleep(gap_ms / 1e3) |
| elif mode == "work": |
| r = reqs[i % len(reqs)] |
| im = Image.open(io.BytesIO(r[0])); im.load() |
| x = preprocess_pil(im, r[1], is_right=r[2], raw=True)[0] |
| x = xs[i % len(xs)] if mode != "work" else x |
| t0 = time.perf_counter(); tt(x); ts.append((time.perf_counter() - t0) * 1e3) |
| ts = ts[10:] |
| print(f"[gap] {mode:5s} gap {gap_ms:4.1f} ms: call median {statistics.median(ts):.3f} min {min(ts):.3f} p90 {sorted(ts)[int(.9*len(ts))]:.3f}", flush=True) |
|
|
| for rep in range(2): |
| run("none"); run("busy", 1.0); run("busy", 2.0); run("busy", 5.0); run("sleep", 2.0); run("work") |
| tt.close() |
|
|
