sync app + active models

restoflow/app.py

@@ -433,7 +433,9 @@ def _run_upload(filepath, start_s=0.0, end_s=0.0):
     except RuntimeError as e:                      # separation unavailable on this deployment
         return _empty_upload(f"⚠️ {e}")
 
-    # ---- pass 1: process OVERLAPPING 3s windows (hop 2.5s) independently ----
+    # ---- separate the WHOLE clip ONCE (the CPU bottleneck) — demucs/audio-sep split internally — then
+    #      process OVERLAPPING 3s windows by SLICING the pre-separated stems (no per-window re-separation) ----
+    full = separate_chunk(sep, a)
     windows = []; summ = {}
     for i in range(0, max(1, total), hop):
         seg = a[:, i:i + chunk]; orig = seg.shape[1]
@@ -441,7 +443,12 @@ def _run_upload(filepath, start_s=0.0, end_s=0.0):
             break
         if orig < chunk:
             seg = np.pad(seg, ((0, 0), (0, chunk - orig)))   # pad silence then crop output back
-        stems = separate_chunk(sep, seg)
+        stems = {}
+        for s, fs in full.items():
+            w = fs[:, i:i + orig]
+            if w.shape[1] < chunk:
+                w = np.pad(w, ((0, 0), (0, chunk - w.shape[1])))
+            stems[s] = w
         present_stems = [s for s, au in stems.items()
                          if (lambda rp: rp[0] > PRESENT_RMS or rp[1] > PRESENT_PEAK)(rms_peak_db(au.mean(0)))]
         enc = encode_batch([stems[s] for s in present_stems] + [seg])     # one batched encode (stems + mix)
@@ -542,7 +549,10 @@ def get_separator():
         ) from e
 
 def separate_chunk(sep, seg_2S):
-    """seg_2S [2,S] @ SR -> {stem: [2,S] float32}, via the same htdemucs v4 weights the cache used."""
+    """[2,S] @ SR -> {stem:[2,S]} via the same htdemucs v4 weights the cache used. CPU-tuned: shifts=0
+    (no test-time augmentation — the big needless cost), light overlap, no_grad (demucs does in-place ops
+    that throw under inference_mode). Called ONCE on the whole clip (demucs splits internally) — NOT per
+    overlapping window — so live separation stays ~1 forward instead of N."""
     if sep["kind"] == "audio_separator":
         tmp = "/tmp/app_sep_in.wav"; sf.write(tmp, seg_2S.T, SR)
         files = sep["sep"].separate(tmp)
@@ -554,12 +564,11 @@ def separate_chunk(sep, seg_2S):
                     au, _ = sf.read(os.path.join("/tmp/app_sep", f) if not os.path.isabs(f) else f, dtype="float32")
                     out[s] = (au.T if au.ndim == 2 else np.repeat(au[None], 2, 0))
         return out
-    # demucs package: per-mix normalize (the standard htdemucs recipe), shifts=1 overlap=0.25 to match.
     wav = torch.as_tensor(seg_2S, dtype=torch.float32, device=DEVICE)        # [2,S]
     ref = wav.mean(0); mu, sd = ref.mean(), ref.std().clamp_min(1e-8)
-    with torch.inference_mode():
+    with torch.no_grad():
         src = sep["apply"](sep["model"], ((wav - mu) / sd)[None],
-                           shifts=1, overlap=0.25, split=True, device=DEVICE)[0]   # [4,2,S]
+                           shifts=0, overlap=0.1, split=True, device=DEVICE)[0]   # [4,2,S]
     src = (src * sd + mu).cpu().numpy()
     return {s: src[i] for i, s in enumerate(sep["sources"]) if s in STEMS}