Upload app.py with huggingface_hub

app.py

@@ -179,6 +179,7 @@ def stream(session_id: str) -> str:
             except Exception as e:
                 print("[bank] error:", repr(e), flush=True)
         toks = []
+        gen_t = time.time()
         for _ in range(10):                        # per-frame slot read -> steering applies ~instantly
             c = read_slot(session_id) or c
             prompts = c.get("prompts") or ["instrumental music"]
@@ -243,8 +244,10 @@ def stream(session_id: str) -> str:
                         temporal_step=model._temporal_step, depth_step=model._depth_step))
         history = torch.cat([history] + toks, dim=1)
         audio, emitted = model._decode_stream(history, emitted)
+        frame_ms = (time.time() - gen_t) * 1000.0 / 10.0     # real per-frame inference time
         if audio.shape[1] > 0:
-            yield base64.b64encode(_float_to_int16(audio[0].float().cpu().numpy()).astype("<i2").tobytes()).decode("ascii")
+            b64 = base64.b64encode(_float_to_int16(audio[0].float().cpu().numpy()).astype("<i2").tobytes()).decode("ascii")
+            yield f"{frame_ms:.1f}|{b64}"
         buf = (0.3, 0.55, 0.85)[max(0, min(2, int(c.get("buffer", 0))))]
         ahead = (emitted * FRAME_SAMPLES / SR) - (time.time() - t0)
         buf_log.append(ahead)