chore: OpenVINO MetricGAN GPU spike (do not merge)

Arc B580 lost the kill test: batch-14 mask p50 103 ms vs Torch CPU 70 ms enhance / OpenVINO CPU 5 ms. Production stays on known-good CPU workers (tag known-good-cpu-2026-08-28). Hop stays 200 ms.
2026-08-28 07:18:54 +00:00
parent 736eccb932
commit 42443420be
6 changed files with 307 additions and 0 deletions
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""CPU baseline: thread-capped MetricGAN enhance_batch (project .venv)."""
+from __future__ import annotations
+
+import argparse
+import os
+import statistics
+import time
+
+os.environ.setdefault("OMP_NUM_THREADS", "1")
+os.environ.setdefault("MKL_NUM_THREADS", "1")
+os.environ.setdefault("OPENBLAS_NUM_THREADS", "1")
+
+import torch
+from speechbrain.inference.enhancement import SpectralMaskEnhancement
+
+SAVEDIR = "/root/.cache/speechbrain-enhancement"
+
+
+def pct(xs, p):
+    xs = sorted(xs)
+    i = int(round((p / 100.0) * (len(xs) - 1)))
+    return xs[max(0, min(i, len(xs) - 1))]
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--batch", type=int, default=14)
+    ap.add_argument("--samples", type=int, default=16000)
+    ap.add_argument("--iters", type=int, default=30)
+    ap.add_argument("--warmup", type=int, default=5)
+    args = ap.parse_args()
+
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    m = SpectralMaskEnhancement.from_hparams(
+        source="speechbrain/metricgan-plus-voicebank",
+        savedir=SAVEDIR,
+        run_opts={"device": "cpu"},
+    )
+    wav = torch.randn(args.batch, args.samples)
+    lengths = torch.ones(args.batch)
+    with torch.no_grad():
+        for _ in range(args.warmup):
+            m.enhance_batch(wav, lengths=lengths)
+        times = []
+        for _ in range(args.iters):
+            t0 = time.perf_counter()
+            m.enhance_batch(wav, lengths=lengths)
+            times.append((time.perf_counter() - t0) * 1000.0)
+        feats = m.compute_features(wav)
+        net = m.mods.enhance_model
+        for _ in range(args.warmup):
+            net(feats, lengths=lengths)
+        mask_times = []
+        for _ in range(args.iters):
+            t0 = time.perf_counter()
+            net(feats, lengths=lengths)
+            mask_times.append((time.perf_counter() - t0) * 1000.0)
+
+    def report(name, xs):
+        per = [x / args.batch for x in xs]
+        print(
+            f"{name} batch={args.batch} n={len(xs)} "
+            f"p50={statistics.median(xs):.2f}ms p95={pct(xs,95):.2f}ms "
+            f"p99={pct(xs,99):.2f}ms "
+            f"per_cam_p50={statistics.median(per):.2f}ms "
+            f"per_cam_p95={pct(per,95):.2f}ms"
+        )
+
+    report("cpu_enhance_batch", times)
+    report("cpu_mask_only", mask_times)
+
+
+if __name__ == "__main__":
+    main()