Warm CPU acoustic verifiers

README.md

@@ -335,10 +335,11 @@ Median-F0 只在 multi-chunk lattice 真的出現兩條以上可選路徑、必
 時，才對 retained DP candidates 延遲計算一次；單一路徑、single/K32、semantic reject、
 joined 與 final waveform 不再支付無用的 pYIN latency。真正進入多路徑排序時，pYIN 演算法與
 transition score 定義沒有變更。
-Pinned ECAPA release speaker encoder 會在 Space 啟動階段於 CPU 完成載入，並以固定、非語音、
-不保存的合成波形各跑一次 local-batch 與 release-third dry-run；因此第一次公開合成不再承擔
-SpeechBrain import、權重初始化、encoder 建構及第一次 CPU kernel/page-fault 的冷啟動成本。
-正式請求仍重新 encode 真實 waveform；warmup 不寫入 request cache，也不改 speaker gate 數值。
+Pinned ECAPA、SQUIM 與 pYIN runtime 會在 Space 啟動階段於 CPU 完成載入，並以固定、非語音、
+不保存的合成波形執行 local-batch、release-third 與 acoustic dry-run；因此第一次公開合成
+不再承擔 SpeechBrain import、權重初始化、encoder 建構及第一次 CPU kernel/page-fault/JIT
+的冷啟動成本。正式請求仍對真實 waveform 重跑所有量測；warmup 不寫入 request cache，
+也不改 speaker、SQUIM、F0 transition 或 echo gate 數值。
 
 Generation work 以實際產生的 chunks 與其 speech units 累計：初始 trajectory 花費
 `chunk count` 與全文 speech units；每次 refill 再加 1 chunk 與該段 units。任一 request 最多

app.py

@@ -564,10 +564,10 @@ def _get_ecapa_encoder():
     return _ECAPA_ENCODER
 
 
-def _preload_release_speaker_runtime() -> None:
-    """Move deterministic CPU speaker load and first forwards out of requests."""
+def _preload_release_cpu_runtime() -> None:
+    """Move deterministic CPU verifier loads and first forwards out of requests."""
 
-    print("[BlueMagpie] preloading release speaker verifier on CPU ...")
+    print("[BlueMagpie] preloading release CPU verifiers ...")
     encoder = _get_ecapa_encoder()
     if encoder is None:
         raise RuntimeError("release speaker verifier preload failed")
@@ -575,7 +575,8 @@ def _preload_release_speaker_runtime() -> None:
     timeline = np.arange(sample_count, dtype=np.float64) / float(SR)
     warmup_waveform = np.asarray(
         0.08 * np.sin(2.0 * np.pi * 180.0 * timeline)
-        + 0.03 * np.sin(2.0 * np.pi * 263.0 * timeline),
+        + 0.04 * np.sin(2.0 * np.pi * 360.0 * timeline)
+        + 0.02 * np.sin(2.0 * np.pi * 540.0 * timeline),
         dtype=np.float32,
     )
     anchor = _speaker_anchor_array(DEFAULT_CENTROID)
@@ -601,7 +602,30 @@ def _preload_release_speaker_runtime() -> None:
         for evidence in warmup_evidence
     ):
         raise RuntimeError("release speaker verifier warmup failed")
-    print("[BlueMagpie] release speaker verifier ready")
+    squim_evidence = squim_objective_evidence_from_audio(
+        warmup_waveform,
+        SR,
+    )
+    if not all(
+        np.isfinite(value)
+        for value in (
+            squim_evidence.stoi,
+            squim_evidence.pesq,
+            squim_evidence.si_sdr,
+        )
+    ):
+        raise RuntimeError("release SQUIM verifier warmup failed")
+    median_f0_hz = active_audio_median_f0_hz(
+        warmup_waveform,
+        SR,
+    )
+    if (
+        median_f0_hz is None
+        or not np.isfinite(median_f0_hz)
+        or median_f0_hz < 1.0
+    ):
+        raise RuntimeError("release F0 transition runtime warmup failed")
+    print("[BlueMagpie] release CPU verifiers ready")
 
 
 def _apply_speed(
@@ -982,7 +1006,7 @@ def _speaker_anchor_array(centroid: torch.Tensor) -> np.ndarray:
     return np.asarray(anchor / norm, dtype=np.float32)
 
 
-_preload_release_speaker_runtime()
+_preload_release_cpu_runtime()
 
 
 def _generate_trajectory(

tests/test_inference_cleanup.py

@@ -633,7 +633,7 @@ def test_space_defers_transition_f0_to_actual_multi_path_ranking():
     assert "_apply_speed(\n            audio,\n            final_speed" in generate
 
 
-def test_space_preloads_release_speaker_runtime_before_serving_requests():
+def test_space_preloads_release_cpu_runtime_before_serving_requests():
     source = (ROOT / "app.py").read_text(encoding="utf-8")
     tree = ast.parse(source)
     functions = {
@@ -647,15 +647,17 @@ def test_space_preloads_release_speaker_runtime_before_serving_requests():
         if isinstance(node, ast.Expr)
         and isinstance(node.value, ast.Call)
         and isinstance(node.value.func, ast.Name)
-        and node.value.func.id == "_preload_release_speaker_runtime"
+        and node.value.func.id == "_preload_release_cpu_runtime"
     ]
 
     assert len(preload_calls) == 1
     assert preload_calls[0].lineno > functions["_speaker_anchor_array"].end_lineno
     preload_source = ast.get_source_segment(
         source,
-        functions["_preload_release_speaker_runtime"],
+        functions["_preload_release_cpu_runtime"],
     )
     assert "speaker_evidence_from_audio(" in preload_source
     assert "release_speaker_evidence_from_audio(" in preload_source
+    assert "squim_objective_evidence_from_audio(" in preload_source
+    assert "active_audio_median_f0_hz(" in preload_source
     assert "np.arange(sample_count" in preload_source