Update app.py

app.py

@@ -97,41 +97,42 @@ async def tts(ws: WebSocket):
     await ws.accept()
     try:
         req = json.loads(await ws.receive_text())
-        text  = req.get("text","")
-        voice = req.get("voice","Jakob")
+        ids, attn = build_inputs(req.get("text", ""), req.get("voice", "Jakob"))
 
-        ids, attn = build_prompt(text, voice)
-        past      = None
-        buf       = []
+        past  = None        # Cache
+        buf   = []
 
         while True:
-            out = model.generate(
+            gen = model.generate(
                 input_ids=ids if past is None else None,
                 attention_mask=attn if past is None else None,
                 past_key_values=past,
                 max_new_tokens=CHUNK_TOKENS,
                 logits_processor=[MASKER],
-                do_sample=True, temperature=0.7, top_p=0.95,
-                use_cache=True,
+                do_sample=True, top_p=0.95, temperature=0.7,
                 return_dict_in_generate=True,
+                use_cache=True,
+                return_legacy_cache=True,       # ⚡ wichtig
             )
-            pkv = out.past_key_values
-            if isinstance(pkv, Cache):
-                pkv = pkv.to_legacy()
-            past_kvs = pkv
-            newtok = out.sequences[0,-out.num_generated_tokens:].tolist()
-
-            for t in newtok:
-                if t==EOS_TOKEN:
-                    raise StopIteration
-                if t==NEW_BLOCK_TOKEN:
-                    buf.clear(); continue
-                buf.append(t-AUDIO_BASE)
-                if len(buf)==7:
-                    await ws.send_bytes(decode_snac(buf))
+
+            # ⚡ Legacy‑Cache weitergeben
+            past = gen.past_key_values
+
+            # die tatsächlich erzeugten neuen Tokens
+            new_tok = gen.sequences[0, -gen.num_generated_tokens :].tolist()
+
+            for t in new_tok:
+                if t == EOS_TOKEN:
+                    raise StopAsyncIteration
+                if t == NEW_BLOCK_TOKEN:
+                    buf.clear()
+                    continue
+                buf.append(t - AUDIO_BASE)
+                if len(buf) == 7:
+                    await ws.send_bytes(decode_block(buf))
                     buf.clear()
 
-            # ab jetzt nur noch mit Cache weiter‑generieren
+            # ab jetzt nur Cache, keine neuen IDs mehr nötig
             ids, attn = None, None
 
     except (StopIteration, WebSocketDisconnect):