Update app.py

app.py

@@ -99,49 +99,34 @@ async def tts(ws: WebSocket):
         text  = req.get("text", "")
         voice = req.get("voice", "Jakob")
 
-        ids, attn  = build_prompt(text, voice)
-        past       = None # Reverted past initialization
-        offset_len = ids.size(1)          # wie viele Tokens existieren schon
-        last_tok   = None # Initialized last_tok
-        buf         = []
-        past_key_values = DynamicCache()
+        ids, attn = build_prompt(text, voice)
+        past      = None
+        offset_len = ids.size(1)
+        cache_pos  = offset_len - 1        # 0‑basiert
+        last_tok   = None
+        buf        = []
+
         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 (StaticCache via cache_implementation) -------------------------------------------
             gen = model.generate(
-                input_ids       = ids if past is None else torch.tensor([[last_tok]], device=device), # Use past is None check
-                attention_mask  = attn if past is None else None, # Use past is None check
-                past_key_values = past_key_values, # Pass past (will be None initially, then the cache object)
-                max_new_tokens = 1, # Set max_new_tokens to 1 for debugging cache
+                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**
+                max_new_tokens  = CHUNK_TOKENS,
                 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,
-                #cache_implementation="static" # Enabled StaticCache via implementation
+                use_cache=True, return_dict_in_generate=True,
+                return_legacy_cache=False
             )
-            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()
-            new  = seq[offset_len:]
-            if not new:                         # nichts -> fertig
+            new = gen.sequences[0, offset_len:].tolist()
+            if not new:
                 break
-            offset_len += len(new)
-
-            # ----- 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
-            #pkv = gen.past_key_values # Update past with the cache object returned by generate
-            print(f"DEBUG: After cache update - type of past: {type(past)}", flush=True) # Added logging
-            #if isinstance(pkv, StaticCache): pkv = pkv.to_legacy()
-            past = gen.past_key_values
-            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)
+            offset_len += len(new)
+            cache_pos   = offset_len - 1            #  ← **NEU**
+            past        = gen.past_key_values
+            last_tok    = new[-1]
 
             # ----- Token‑Handling ----------------------------------------
             for t in new: