Update app.py

app.py

@@ -100,8 +100,7 @@ async def tts(ws: WebSocket):
         voice = req.get("voice", "Jakob")
 
         ids, attn  = build_prompt(text, voice)
-        # Initialized StaticCache
-        past = StaticCache(config=model.config, max_batch_size=1, max_cache_len=ids.size(1) + CHUNK_TOKENS, device=device, dtype=model.dtype)
+        past       = None # Reverted past initialization
         offset_len = ids.size(1)          # wie viele Tokens existieren schon
         last_tok   = None # Initialized last_tok
         buf         = []
@@ -109,17 +108,18 @@ async def tts(ws: WebSocket):
         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) -------------------------------------------
+            # --- Mini‑Generate (StaticCache via cache_implementation) -------------------------------------------
             gen = model.generate(
-                input_ids       = ids if past.get_seq_length() == 0 else torch.tensor([[last_tok]], device=device), # Use cache seq length
-                attention_mask  = attn if past.get_seq_length() == 0 else None, # Use cache seq length
-                past_key_values = past, # Use StaticCache instance
+                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, # Pass past (will be None initially, then the cache object)
                 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
+                return_legacy_cache=True,
+                cache_implementation="static" # Enabled StaticCache via implementation
             )
             print(f"DEBUG: After generate - type of gen.past_key_values: {type(gen.past_key_values)}", flush=True) # Added logging
 
@@ -133,7 +133,7 @@ async def tts(ws: WebSocket):
             # ----- 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
+            past = 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
             last_tok = new[-1]