Upload app.py with huggingface_hub

app.py

@@ -157,6 +157,18 @@ def stream(session_id: str) -> str:
     while read_slot(session_id) is None and time.time() < deadline:
         time.sleep(0.05)
     gen = torch.Generator(device=dev).manual_seed(0)
+    try:                                          # one-time GPU warm-up so the first audio frame isn't slow
+        style_model.embed("warmup")
+        _wd = mrt_base.model.decoder.init_streaming_f(1, dev, dt)
+        _wc = mrt_base._conditioning([-1] * mrt_base.num_musiccoca, [-1] * 128, [-1],
+                                     [discretize_cfg(1.6, 0.2, 40), discretize_cfg(2.4, 0.2, 40), discretize_cfg(4.0, 1.0, 8)])
+        _ws = mrt_base.model.encode(_wc).to(dt)
+        for _ in range(2):
+            mrt_base.model.decoder.step_f(_wd, _ws, sampler=make_sampler(1.1, 50, gen),
+                                          temporal_step=mrt_base._temporal_step, depth_step=mrt_base._depth_step)
+        print("[warmup] done", flush=True)
+    except Exception as _e:
+        print("[warmup]", repr(_e), flush=True)
     notes, drums = [-1] * 128, [-1]
     cur_name = model = dstate = history = source = cur_tokens = cur_notes = None
     emitted = 0
@@ -283,8 +295,8 @@ def stream(session_id: str) -> str:
             buf_log, buf_t = [], time.time()
         sleep_for = max(0.0, ahead - buf)
         cl = read_client_lead(session_id)                 # client's real Web-Audio queue (ms)
-        if cl is not None and cl - buf * 1000.0 > 100:    # queue above target -> drain it (back-pressure)
-            sleep_for = max(sleep_for, (cl - buf * 1000.0) / 1000.0 * 0.6)
+        if cl is not None and cl - buf * 1000.0 > 200:    # only drain a genuinely-deep queue (target+200ms)
+            sleep_for = max(sleep_for, min((cl - buf * 1000.0) / 1000.0 * 0.3, 0.2))  # gentle + capped
         if sleep_for > 0:
             time.sleep(min(sleep_for, 0.5))