Upload app.py with huggingface_hub

app.py

@@ -180,6 +180,7 @@ def stream(session_id: str) -> str:
     cur_reset = cur_seed = 0
     force_reenc = False
     txt_cache, aud_cache = {}, {}
+    CHUNK = 5                                   # frames per yield (smaller = finer delivery, lower buffer floor)
     cur_bank_ver = 0
     while time.time() - t0 < 55.0:
         c = read_slot(session_id)
@@ -212,7 +213,7 @@ def stream(session_id: str) -> str:
                 print("[bank] error:", repr(e), flush=True)
         toks = []
         gen_t = time.time()
-        for _ in range(10):                        # per-frame slot read -> steering applies ~instantly
+        for _ in range(CHUNK):                        # per-frame slot read -> steering applies ~instantly
             c = read_slot(session_id) or c
             prompts = c.get("prompts") or ["instrumental music"]
             weights = c.get("weights") or [1.0] * len(prompts)
@@ -279,7 +280,7 @@ 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
+        frame_ms = (time.time() - gen_t) * 1000.0 / CHUNK     # real per-frame inference time
         if audio.shape[1] > 0:
             b64 = base64.b64encode(_float_to_int16(audio[0].float().cpu().numpy()).astype("<i2").tobytes()).decode("ascii")
             yield f"{frame_ms:.1f}|{b64}"