Update app.py

app.py

@@ -73,7 +73,7 @@ def build_prompt(text: str, voice: str):
                        prompt_ids,
                        torch.tensor([[128009, 128260]], device=device)], 1)
     attn  = torch.ones_like(ids)
-    return ids, attn
+    return ids, attn # Ensure attention mask is created
 
 def decode_block(block7: list[int]) -> bytes:
     l1,l2,l3=[],[],[]
@@ -130,20 +130,23 @@ async def tts(ws: WebSocket):
                 break
             offset_len += len(new)
 
-            # ----- Update ids and attn for next iteration (Cache Disabled) ---------
-            ids = torch.tensor([seq], device=device)
-            attn = torch.ones_like(ids)
+            # ----- Update past and last_tok (Cache Re-enabled) ---------
+            # ids = torch.tensor([seq], device=device) # Removed full sequence update
+            # attn = torch.ones_like(ids) # Removed full sequence update
+            past = gen.past_key_values # Re-enabled cache update
+            print(f"DEBUG: After cache update - type of past: {type(past)}", flush=True) # Added logging
+            last_tok = new[-1]
 
             print("new tokens:", new[:25], flush=True)
 
             # ----- Token‑Handling ----------------------------------------
             for t in new:
-                if t == EOS_TOKEN:
-                    raise StopIteration
+                if t == EOS_TOKEN: # Re-enabled EOS check
+                    raise StopIteration # Re-enabled EOS check
                 if t == NEW_BLOCK:
                     buf.clear()
                     continue
-                buf.append(t - AUDIO_BASE)
+                buf.append(t - AUDIO_BASE) # Reverted to appending relative token
                 if len(buf) == 7:
                     await ws.send_bytes(decode_block(buf))
                     buf.clear()