Upload app.py with huggingface_hub

app.py

@@ -95,6 +95,7 @@ async def set_collider(request: Request):
         "reset": int(body.get("reset", 0)),
         "seed": int(body.get("seed", 0)),
         "bank_op": body.get("bank_op", prev.get("bank_op")),
+        "ts": time.time(),
     })
     return {"ok": True}
 
@@ -197,6 +198,7 @@ def stream(session_id: str) -> str:
                 prev_active = set(active)
                 cur_sig, last_enc = sig, time.time()
                 force_reenc = onsetmode and bool(onsets)
+                _lt0 = time.time()
                 # ---- style tokens on GPU: embed (text+audio), weight-blend, tokenize ----
                 emb = torch.zeros(style_model.embedding_dim, device=dev, dtype=torch.float32)
                 tot = float(sum(w for w in weights if w > 0)) or 1.0
@@ -223,7 +225,8 @@ def stream(session_id: str) -> str:
                         if e is None:
                             e = style_model.embed(lbl).float(); txt_cache[lbl] = e
                     emb = emb + (w / tot) * e
-                tokens = style_model.tokenize(emb).tolist()   # onset(2)->continuation(1) on the next frame
+                tokens = style_model.tokenize(emb).tolist()
+                _lt1 = time.time()
                 nvec = []                                  # per-pitch: -1 masked / 0 off / 1 cont / 2 onset / 3 on
                 for pitch in range(128):
                     if pitch in active:
@@ -239,6 +242,7 @@ def stream(session_id: str) -> str:
                 cond = model._conditioning((list(tokens) + [-1] * model.num_musiccoca)[:model.num_musiccoca],
                                            nvec, drm, cfgs)
                 source = model.model.encode(cond).to(dt)
+                print(f"[lat] pickup={(_lt0-float(c.get('ts',_lt0)))*1000:.0f}ms style+tok={(_lt1-_lt0)*1000:.0f}ms encode={(time.time()-_lt1)*1000:.0f}ms", flush=True)
             sampler = make_sampler(c.get("temperature", 1.1), c.get("top_k", 50), gen)
             toks.append(model.model.decoder.step_f(dstate, source, sampler=sampler,
                         temporal_step=model._temporal_step, depth_step=model._depth_step))