Update app.py

app.py

@@ -102,44 +102,44 @@ async def tts(ws: WebSocket):
         ids, attn = build_prompt(text, voice)
         past      = None
         offset_len = ids.size(1)
-        cache_pos  = offset_len        # 0‑basiert
+        past       = None
         last_tok   = None
         buf        = []
 
         while True:
+            next_cache_pos = torch.tensor([offset_len], device=device) if past is not None else None
+
             gen = model.generate(
                 input_ids       = ids if past is None else torch.tensor([[last_tok]], device=device),
                 attention_mask  = attn if past is None else None,
                 past_key_values = past,
-                cache_position  = None if past is None else torch.tensor([cache_pos], device=device),  #  ← **NEU**
+                cache_position  = next_cache_pos,     # **hier nur ab 2. Durchlauf**
                 max_new_tokens  = CHUNK_TOKENS,
                 logits_processor=[masker],
                 do_sample=True, temperature=0.7, top_p=0.95,
                 use_cache=True, return_dict_in_generate=True,
-                return_legacy_cache=False
             )
 
-            new = gen.sequences[0, offset_len:].tolist()
-            if not new:
+            # neu erzeugte Tokens hinter dem bisherigen Ende
+            new_tokens = gen.sequences[0, offset_len:].tolist()
+            if not new_tokens:
                 break
 
-            offset_len += len(new)
-            cache_pos   = offset_len - 1            #  ← **NEU**
-            past        = gen.past_key_values
-            last_tok    = new[-1]
+            offset_len += len(new_tokens)            # Cache ist jetzt größer
+            past        = gen.past_key_values        # Cache zurück für nächste Runde
+            last_tok    = new_tokens[-1]
 
-            # ----- Token‑Handling ----------------------------------------
-            for t in new:
-                if t == EOS_TOKEN: # Re-enabled EOS check
-                    raise StopIteration # Re-enabled EOS check
+            for t in new_tokens:
+                if t == EOS_TOKEN:
+                    raise StopIteration
                 if t == NEW_BLOCK:
                     buf.clear()
                     continue
-                buf.append(t - AUDIO_BASE) # Reverted to appending relative token
+                buf.append(t - AUDIO_BASE)
                 if len(buf) == 7:
                     await ws.send_bytes(decode_block(buf))
                     buf.clear()
-                    masker.sent_blocks = 1      # ab jetzt EOS zulässig
+                    masker.sent_blocks = 1            # ab jetzt darf EOS
 
     except (StopIteration, WebSocketDisconnect):
         pass