Update app.py

app.py

@@ -122,15 +122,15 @@ async def tts(ws: WebSocket):
             # Update buffer_pos based on current buffer length before generation
             masker.buffer_pos = len(buf)
 
-            # --- Mini‑Generate (Cache Re-enabled) -------------------------------------------
+            # --- Mini‑Generate (Cache Disabled for Debugging) -------------------------------------------
             gen = model.generate(
-                input_ids       = ids if past is None else torch.tensor([[last_tok]], device=device), # Reverted
-                attention_mask  = attn if past is None else None, # Reverted
-                past_key_values = past, # Re-enabled cache
+                input_ids      = ids, # Always use full sequence
+                attention_mask = attn, # Always use full attention mask
+                # past_key_values= past, # Disabled cache
                 max_new_tokens = CHUNK_TOKENS,
                 logits_processor=[masker],
                 do_sample=True, temperature=0.7, top_p=0.95,
-                use_cache=True, # Re-enabled cache
+                use_cache=False, # Disabled cache
                 return_dict_in_generate=True,
                 return_legacy_cache=True
             )
@@ -142,10 +142,10 @@ async def tts(ws: WebSocket):
                 break
             offset_len += len(new)
 
-            # ----- Update past and last_tok (Cache Re-enabled) ---------
-            # ids = torch.tensor([seq], device=device) # Removed
-            # attn = torch.ones_like(ids) # Removed
-            past = gen.past_key_values # Re-enabled cache access
+            # ----- Update ids and attn with the full sequence (Cache Disabled) ---------
+            ids = torch.tensor([seq], device=device) # Re-added
+            attn = torch.ones_like(ids) # Re-added
+            # past = gen.past_key_values # Disabled cache access
             last_tok = new[-1]
 
             print("new tokens:", new[:25], flush=True)