Update app.py

app.py

@@ -104,22 +104,23 @@ async def tts(ws: WebSocket):
         offset_len = ids.size(1)          # wie viele Tokens existieren schon
         last_tok   = None
         buf         = []
-        masker.buffer_pos = 0 # Removed initialization here
-        # Removed buffer_pos update before generation
 
         while True:
+            print(f"DEBUG: Before generate - past is None: {past is None}", flush=True) # Added logging
+            print(f"DEBUG: Before generate - type of past: {type(past) if past is not None else 'None'}", flush=True) # Added logging
             # --- Mini‑Generate (Cache Re-enabled) -------------------------------------------
             gen = model.generate(
                 input_ids       = ids if past is None else torch.tensor([[last_tok]], device=device), # Re-enabled cache input
                 attention_mask  = attn if past is None else None, # Re-enabled cache attention
                 past_key_values = past, # Re-enabled cache
-                max_new_tokens = CHUNK_TOKENS,
+                max_new_tokens = 1, # Set max_new_tokens to 1 for debugging cache
                 logits_processor=[masker],
                 do_sample=True, temperature=0.7, top_p=0.95,
                 use_cache=True, # Re-enabled cache
                 return_dict_in_generate=True,
                 return_legacy_cache=True
             )
+            print(f"DEBUG: After generate - type of gen.past_key_values: {type(gen.past_key_values)}", flush=True) # Added logging
 
             # ----- neue Tokens heraus schneiden --------------------------
             seq  = gen.sequences[0].tolist()
@@ -132,6 +133,7 @@ async def tts(ws: WebSocket):
             # 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)
@@ -143,16 +145,11 @@ async def tts(ws: WebSocket):
                 if t == NEW_BLOCK:
                     buf.clear()
                     continue
-                # Only append if it's an audio token
-                # Only append if it's an audio token
                 buf.append(t - AUDIO_BASE) # Reverted to appending relative token
-                masker.buffer_pos += 1 # Removed increment here
                 if len(buf) == 7:
                     await ws.send_bytes(decode_block(buf))
                     buf.clear()
                     masker.sent_blocks = 1      # ab jetzt EOS zulässig
-                    masker.buffer_pos = 0 # Removed reset here
-                # Removed else block for skipping non-audio tokens
 
     except (StopIteration, WebSocketDisconnect):
         pass