"""C simultaneous streaming requests to a running zonos3-server; per stream: time to first audio and whether playback (starting at the first chunk) would ever run dry. Stdlib only. python3 stream_bench.py --port 1919 --conc 1,4,8,16 --tag vulkan-b8 """ import argparse, array, base64, http.client, json, os, statistics as st, threading, time, wave ap = argparse.ArgumentParser() ap.add_argument("--port", type=int, default=1919) ap.add_argument("--conc", default="1,2,4,8") ap.add_argument("--rounds", type=int, default=2) ap.add_argument("--tag", required=True) ap.add_argument("--prompts", default="~/omnispace-b70/bench/prompts.jsonl") ap.add_argument("--min-len", type=int, default=60) ap.add_argument("--save", action="store_true", help="write each stream to bench//wav and a timings.json for score.py") a = ap.parse_args() BPS = 44100 * 4 # float32 mono PCM BUFFERS = (0.0, 0.1, 0.2) emb = base64.b64encode(open(os.path.expanduser( "~/omnispace-b70/models/voices/Savannah_en-American_F_1.embedding.npy"), "rb").read()).decode() texts = [json.loads(l)["text"] for l in open(os.path.expanduser(a.prompts))] texts = [t for t in texts if len(t) > a.min_len] def one(i, seed, out): body = json.dumps({"text": texts[i % len(texts)], "stream": True, "seed": seed, "speaker_embedding_base64": emb, "speaker_mode": "embedding"}) c = http.client.HTTPConnection("127.0.0.1", a.port, timeout=600) t0 = time.perf_counter() c.request("POST", "/v2/generate", body, {"Content-Type": "application/json"}) r = c.getresponse() if r.status != 200: out[i] = {"error": f"{r.status} {r.read()[:200]!r}"} return got, first, arrivals, pcm = 0, None, [], bytearray() # (arrival time, audio seconds received before this chunk) while True: chunk = r.read1(65536) if not chunk: break now = time.perf_counter() if first is None: first = now arrivals.append((now, got / BPS)) got += len(chunk) if a.save: pcm += chunk end = time.perf_counter() res = {"ttfa_ms": round(1000 * (first - t0), 1), "audio_s": round(got / BPS, 2), "wall_s": round(end - t0, 2)} for buf in BUFFERS: # playback starts buf seconds after the first chunk; a stall is the player running dry short = [(now - first - buf) - before for now, before in arrivals[1:]] res[f"stall_ms_b{int(buf * 1000)}"] = round(1000 * max([0.0] + short), 1) if a.save: d = os.path.expanduser(f"~/omnispace-b70/bench/{a.tag}/wav"); os.makedirs(d, exist_ok=True) x = array.array("f"); x.frombytes(bytes(pcm[: len(pcm) // 4 * 4])) res["wav"] = f"{d}/c{len(out)}_s{seed}.wav" with wave.open(res["wav"], "wb") as w: w.setnchannels(1); w.setsampwidth(2); w.setframerate(44100) w.writeframes(array.array("h", (max(-32768, min(32767, int(v * 32767))) for v in x)).tobytes()) res["text"] = texts[i % len(texts)]; res["id"] = f"c{len(out)}_s{seed}" out[i] = res rows, SAVED = [], [] for C in [int(x) for x in a.conc.split(",")]: for rnd in range(a.rounds): out = [None] * C th = [threading.Thread(target=one, args=(i, 1000 * rnd + i, out)) for i in range(C)] t = time.perf_counter() [x.start() for x in th] [x.join() for x in th] wall = time.perf_counter() - t ok = [o for o in out if o and "error" not in o] res = {"tag": a.tag, "conc": C, "round": rnd, "ok": len(ok), "errors": [o["error"] for o in out if o and "error" in o][:2], "ttfa_ms_median": st.median(o["ttfa_ms"] for o in ok) if ok else None, "ttfa_ms_max": max(o["ttfa_ms"] for o in ok) if ok else None, **{f"stalled_b{int(b * 1000)}": sum(o[f"stall_ms_b{int(b * 1000)}"] > 0 for o in ok) for b in BUFFERS}, **{f"worst_ms_b{int(b * 1000)}": max((o[f"stall_ms_b{int(b * 1000)}"] for o in ok), default=None) for b in BUFFERS}, "audio_s_per_wall_s": round(sum(o["audio_s"] for o in ok) / wall, 1)} rows.append(res) SAVED.extend(o for o in ok if "wav" in o) print(json.dumps(res), flush=True) os.makedirs(os.path.expanduser("~/omnispace-b70/bench/stream"), exist_ok=True) if a.save: json.dump(SAVED, open(os.path.expanduser(f"~/omnispace-b70/bench/{a.tag}/timings.json"), "w"), indent=1) json.dump(rows, open(os.path.expanduser(f"~/omnispace-b70/bench/stream/{a.tag}.json"), "w"), indent=1)