Update app.py

app.py

@@ -107,12 +107,12 @@ async def tts(ws: WebSocket):
         buf        = []
 
         while True:
-            next_cache_pos = torch.tensor([offset_len], device=device) if past is not None else None
+            next_cache_pos = torch.tensor([past.get_seq_length()], device=device)
             gen = model.generate(
                 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 next_cache_pos,
+                cache_position  = next_cache_pos,
                 max_new_tokens  = CHUNK_TOKENS,
                 logits_processor=[masker],
                 do_sample=True, temperature=0.7, top_p=0.95,